Base_JSON/prefixO/json/O12/O12-4003.json

{
    "paper_id": "O12-4003",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T08:02:45.858921Z"
    },
    "title": "Enhancement of Feature Engineering for Conditional Random Field Learning in Chinese Word Segmentation Using Unlabeled Data",
    "authors": [
        {
            "first": "Mike",
            "middle": [],
            "last": "Tian-Jian",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Jiang",
            "middle": [],
            "last": "\uff0a+",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "National Tsing Hua University",
                "location": {
                    "settlement": "Hsinchu",
                    "country": "Taiwan"
                }
            },
            "email": ""
        },
        {
            "first": "Cheng-Wei",
            "middle": [],
            "last": "Shih",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "National Tsing Hua University",
                "location": {
                    "settlement": "Hsinchu",
                    "country": "Taiwan"
                }
            },
            "email": ""
        },
        {
            "first": "Ting-Hao",
            "middle": [],
            "last": "Yang",
            "suffix": "",
            "affiliation": {},
            "email": "tinghaoyang@iis.sinica.edu.tw"
        },
        {
            "first": "Chan-Hung",
            "middle": [],
            "last": "Kuo",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Richard",
            "middle": [],
            "last": "Tzong",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Han",
            "middle": [],
            "last": "Tsai",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Yuan Ze University",
                "location": {
                    "settlement": "Taoyuan",
                    "country": "Taiwan"
                }
            },
            "email": "thtsai@saturn.yzu.edu.tw"
        },
        {
            "first": "Wen-Lian",
            "middle": [],
            "last": "Hsu",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "National Tsing Hua University",
                "location": {
                    "settlement": "Hsinchu",
                    "country": "Taiwan"
                }
            },
            "email": "hsu@iis.sinica.edu.tw"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "This work proposes a unified view of several features based on frequent strings extracted from unlabeled data that improve the conditional random fields (CRF) model for Chinese word segmentation (CWS). These features include character-based n-gram (CNG), accessor variety based string (AVS) and its variation of left-right co-existed feature (LRAVS), term-contributed frequency (TCF), and term-contributed boundary (TCB) with a specific manner of boundary overlapping. For the experiments, the baseline is the 6-tag, a state-of-the-art labeling scheme of CRF-based CWS, and the data set is acquired from the 2005 CWS Bakeoff of Special Interest Group on Chinese Language Processing (SIGHAN) of the Association for Computational Linguistics (ACL) and SIGHAN CWS Bakeoff 2010. The experimental results show that all of these features improve the performance of the baseline system in terms of recall, precision, and their harmonic average as F 1 measure score, on both accuracy (F) and out-of-vocabulary recognition (F OOV). In particular, this work presents compound features involving LRAVS/AVS and TCF/TCB that are competitive with other types of features for CRF-based CWS in terms of F and F OOV , respectively.",
    "pdf_parse": {
        "paper_id": "O12-4003",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "This work proposes a unified view of several features based on frequent strings extracted from unlabeled data that improve the conditional random fields (CRF) model for Chinese word segmentation (CWS). These features include character-based n-gram (CNG), accessor variety based string (AVS) and its variation of left-right co-existed feature (LRAVS), term-contributed frequency (TCF), and term-contributed boundary (TCB) with a specific manner of boundary overlapping. For the experiments, the baseline is the 6-tag, a state-of-the-art labeling scheme of CRF-based CWS, and the data set is acquired from the 2005 CWS Bakeoff of Special Interest Group on Chinese Language Processing (SIGHAN) of the Association for Computational Linguistics (ACL) and SIGHAN CWS Bakeoff 2010. The experimental results show that all of these features improve the performance of the baseline system in terms of recall, precision, and their harmonic average as F 1 measure score, on both accuracy (F) and out-of-vocabulary recognition (F OOV). In particular, this work presents compound features involving LRAVS/AVS and TCF/TCB that are competitive with other types of features for CRF-based CWS in terms of F and F OOV , respectively.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Many intelligent text processing tasks, such as information retrieval, text-to-speech, and machine translation assume the ready availability of a tokenization into words, which is relatively straightforward in languages with word delimiters (e.g., space) but is a little difficult for Asian languages, such as Chinese and Japanese.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction Background",
                "sec_num": "1."
            },
            {
                "text": "Chinese word segmentation (CWS) has been an active area of research in computational linguistics for two decades. SIGHAN, the Special Interest Group for Chinese Language Processing of the Association for Computational Linguistics, has conducted five word segmentation bakeoffs (Emerson, 2005; Jin & Chen, 2007; Levow, 2006; Sproat & Emerson, 2003; Zhao & Liu, 2010) . After years of intensive research, CWS has achieved high accuracy, but the issue of out-of-vocabulary (OOV) word recognition remains.",
                "cite_spans": [
                    {
                        "start": 277,
                        "end": 292,
                        "text": "(Emerson, 2005;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 293,
                        "end": 310,
                        "text": "Jin & Chen, 2007;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 311,
                        "end": 323,
                        "text": "Levow, 2006;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 324,
                        "end": 347,
                        "text": "Sproat & Emerson, 2003;",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 348,
                        "end": 365,
                        "text": "Zhao & Liu, 2010)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction Background",
                "sec_num": "1."
            },
            {
                "text": "Traditional approaches for CWS adopt a dictionary and rules to segment unlabeled texts, such as the work of Ma and Chen (2003) . In recent years, there has been a potent trend of using statistical machine learning models, especially the conditional random fields (CRF) (Lafferty et al., 2001) , which displays moderate performance for the sequential labeling problem and achieves competitive results with character-position based methods (Zhao et al., 2010) .",
                "cite_spans": [
                    {
                        "start": 108,
                        "end": 126,
                        "text": "Ma and Chen (2003)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 269,
                        "end": 292,
                        "text": "(Lafferty et al., 2001)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 438,
                        "end": 457,
                        "text": "(Zhao et al., 2010)",
                        "ref_id": "BIBREF37"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The State of the Art of CWS",
                "sec_num": null
            },
            {
                "text": "In this work, unsupervised feature selection for CWS is based on frequent strings that are extracted automatically from unlabeled corpora. For convenience, these features are referred to as unsupervised features in the rest of this paper. Unsupervised features are suitable for closed training evaluation where external resources or extra information is not allowed, especially for cross-domain tasks, such as SIGHAN CWS bakeoff 2010 (Zhao & Liu, 2010) . Without proper knowledge, the closed training evaluation of word segmentation can be difficult with OOV words, where frequent strings collected from the test data may help. For incorporating unsupervised features into character-position based CRF for CWS, Zhao and Kit (2007) tried strings based on accessor variety (AV), which was developed by Feng et al. (2004) , and based on co-occurrence strings (COS). Jiang et al. (2010) applied a feature similar to COS, called term-contributed boundary (TCB).",
                "cite_spans": [
                    {
                        "start": 434,
                        "end": 452,
                        "text": "(Zhao & Liu, 2010)",
                        "ref_id": null
                    },
                    {
                        "start": 711,
                        "end": 730,
                        "text": "Zhao and Kit (2007)",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 800,
                        "end": 818,
                        "text": "Feng et al. (2004)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 863,
                        "end": 882,
                        "text": "Jiang et al. (2010)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unsupervised Feature Selection for CWS",
                "sec_num": null
            },
            {
                "text": "According to Zhao and Kit (2007) , AV-based string (AVS) is one of the most effective unsupervised features for CWS by character-position based CRF. One motivation here is to seek deeper understanding of AVS's success. This work suspects that, since AVS is designed to keep overlapping substrings via the outer structure of a string while COS/TCB is usually selected via the inner structure of a string with its longest-first (i.e., non-overlapping) nature before integration into CRF, combining overlapping and outer information with",
                "cite_spans": [
                    {
                        "start": 13,
                        "end": 32,
                        "text": "Zhao and Kit (2007)",
                        "ref_id": "BIBREF38"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unsupervised Feature Selection for CWS",
                "sec_num": null
            },
            {
                "text": "Field Learning in Chinese Word Segmentation Using Unlabeled Data non-overlapping and inner information may enhance CRF-based CWS. Hence, a series of experiments is conducted to examine this hypothesis.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Enhancement of Feature Engineering for Conditional Random 47",
                "sec_num": null
            },
            {
                "text": "The remainder of the article is organized as follows. Section 2 briefly introduces CRF. Common unsupervised features based on the concept of frequent strings are explained in Section 3. Section 4 discusses related works. Section 5 describes the design of the labeling scheme and feature templates, along with a framework that is able to encode those overlapping features in a unified way. Details about the experiment are reported in Section 6. Finally, the conclusion is presented in Section 7.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Enhancement of Feature Engineering for Conditional Random 47",
                "sec_num": null
            },
            {
                "text": "Conditional random fields (CRF) are undirected graphical models trained to maximize a conditional probability of random variables X and Y, and the concept is well established for the sequential labeling problem (Lafferty et al., 2001) . Given an input sequence (or observation sequence) 1 ... T X x x = and a label sequence 1 ... T Y y y = , a conditional probability of linear-chain CRF with parameters 1 ... n \u03bb \u03bb \u039b = can be defined as:",
                "cite_spans": [
                    {
                        "start": 211,
                        "end": 234,
                        "text": "(Lafferty et al., 2001)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conditional Random Fields",
                "sec_num": "2."
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "1 1 X 1 ( | ) exp ( , , , ) T k k t t t k P Y X f y y X t Z \u03bb \u03bb \u2212 = \u239b \u239e = \u239c \u239f \u239d \u23a0 \u2211 \u2211 .",
                        "eq_num": "(1)"
                    }
                ],
                "section": "Conditional Random Fields",
                "sec_num": "2."
            },
            {
                "text": "where Z X is the normalization constant that makes probability of all label sequences sum to one; 1 ( , , , )",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conditional Random Fields",
                "sec_num": "2."
            },
            {
                "text": "k t t f y y X t \u2212",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conditional Random Fields",
                "sec_num": "2."
            },
            {
                "text": "is a feature function which is often binary valued, but can be real valued; and k \u03bb is a learned weight associated with feature k f .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conditional Random Fields",
                "sec_num": "2."
            },
            {
                "text": "The feature functions can measure any aspect of state transition , and the entire observation sequence X is centered at the current position t.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conditional Random Fields",
                "sec_num": "2."
            },
            {
                "text": "Given the model defined in (1), the most probable labeling sequence for an input sequence X is as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conditional Random Fields",
                "sec_num": "2."
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "* argmax ( | ) Y y P Y X \u039b = .",
                        "eq_num": "(2)"
                    }
                ],
                "section": "Conditional Random Fields",
                "sec_num": "2."
            },
            {
                "text": "Equation (2) can be efficiently calculated by dynamic programming using the Viterbi algorithm. More details about the concepts of CRF and learning parameters could be found in Wallach (2004) . For sequential labeling tasks, like CWS, a linear-chain CRF is currently one of the most popular choices.",
                "cite_spans": [
                    {
                        "start": 176,
                        "end": 190,
                        "text": "Wallach (2004)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conditional Random Fields",
                "sec_num": "2."
            },
            {
                "text": "The word boundary and the word frequency are the standard notions of frequency in corpus-based natural language processing. Word-based n-gram is an intuitive and effective solution of language modeling. For languages without explicit word boundaries, such as Chinese, character-based n-gram (CNG) is usually insufficient. For example, consider some sample texts in Chinese:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Character-based N-gram",
                "sec_num": "3.1"
            },
            {
                "text": "\"\u81ea\u7136\u79d1\u5b78\u7684\u91cd\u8981\u6027\" (the importance of natural science), and \"\u81ea\u7136\u79d1\u5b78\u7684\u7814\u7a76\u662f\u552f\u4e00\u7684\u9014\u5f91\" (natural science research is the only way), where many character-based n-grams can be extracted, but some of them are out of context, such as \"\u7136\u79d1\" (so; discipline) and \"\u5b78\u7684\" (study; of), even when they are relatively frequent. For the purpose of interpreting overlapping behavior of frequent strings, however, character-based n-grams could still be useful for baseline analysis and implementation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Character-based N-gram",
                "sec_num": "3.1"
            },
            {
                "text": "The lack of correct information about the actual boundary and frequency of a multi-character/word expression's occurrence has been researched in different languages. The distortion of phrase boundaries and frequencies was first observed in the Vodis Corpus, where the word-based bigram \"RAIL ENQUIRIES\" and word-based trigram \"BRITISH RAIL ENQUIRIES\" were estimated and reported by O'Boyle (1993) and Ha et al. (2005) . Both of them occur 73 times, which is a large number for such a small corpus. \"ENQUIRIES\" follows \"RAIL\" with a very high probability when \"BRITISH\" precedes it. When \"RAIL\" is preceded by words other than \"BRITISH,\" however, \"ENQUIRIES\" does not occur, but words like \"TICKET\" or \"JOURNEY\" may. Thus, the bigram \"RAIL ENQUIRIES\" gives a misleading probability that \"RAIL\" is followed by \"ENQUIRIES\" irrespective of what precedes it.",
                "cite_spans": [
                    {
                        "start": 390,
                        "end": 396,
                        "text": "(1993)",
                        "ref_id": null
                    },
                    {
                        "start": 401,
                        "end": 417,
                        "text": "Ha et al. (2005)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reduced N-gram",
                "sec_num": "3.2"
            },
            {
                "text": "A common solution to this problem is that, if some n-grams consist of others, then the frequencies of the shorter ones have to be discounted with the frequencies of the longer ones. For Chinese, Lin & Yu (2011) reported a similar problem and its corresponding solution in the sense of reduced n-gram of Chinese characters. By excluding n-grams with their numbers of appearance that fully depend on other superstrings, \"\u7136\u79d1\" and \"\u5b78\u7684\" from the sample texts in the previous sub-section are no longer candidates of the string. Zhao and Kit (2007) described the same concept briefly as co-occurrence string (COS). Sung et al. (2008) invented a specific data structure for suffix array algorithm to calculate exact boundaries of phrase-alike string and their frequencies called term-contributed boundaries (TCB) and term-contributed frequencies (TCF), respectively, to analogize similarities and differences",
                "cite_spans": [
                    {
                        "start": 522,
                        "end": 541,
                        "text": "Zhao and Kit (2007)",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 608,
                        "end": 626,
                        "text": "Sung et al. (2008)",
                        "ref_id": "BIBREF30"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reduced N-gram",
                "sec_num": "3.2"
            },
            {
                "text": "Field Learning in Chinese Word Segmentation Using Unlabeled Data with the term frequencies. Since this work uses the program of TCB and TCF (namely YASA, yet another suffix array) for experiments, the family of reduced n-gram will be referred as TCB hereafter for convenience. Feng et al. (2004) proposed accessor variety (AV) to measure the likelihood a substring is a Chinese word. Another measurement, called boundary entropy or branching entropy (BE), exists in some works (Chang & Su, 1997; Cohen et al., 2007; Huang & Powers, 2003; Tanaka-Ishii, 2005; Tung & Lee, 1994) . The basic idea behind those measurements is closely related to one particular perspective of n-gram and information theory, cross-entropy or perplexity. According to Zhao and Kit (2007) , AV and BE both assume that the border of a potential Chinese word is located where the uncertainty of successive character increases. They believe that AV and BE are the discrete and continuous version, respectively, of a fundamental work of Harris (1970) , and they decided to adopt AVS as an unsupervised feature for CRF-based CWS. This work follows their choice in hope of producing a comparable study. AV of a string s is defined as:",
                "cite_spans": [
                    {
                        "start": 277,
                        "end": 295,
                        "text": "Feng et al. (2004)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 477,
                        "end": 495,
                        "text": "(Chang & Su, 1997;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 496,
                        "end": 515,
                        "text": "Cohen et al., 2007;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 516,
                        "end": 537,
                        "text": "Huang & Powers, 2003;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 538,
                        "end": 557,
                        "text": "Tanaka-Ishii, 2005;",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 558,
                        "end": 575,
                        "text": "Tung & Lee, 1994)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 744,
                        "end": 763,
                        "text": "Zhao and Kit (2007)",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 1008,
                        "end": 1021,
                        "text": "Harris (1970)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Enhancement of Feature Engineering for Conditional Random 49",
                "sec_num": null
            },
            {
                "text": "( ) min{ ( ), ( )} av av AV s L s R s = .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Uncertainty of Succeeding Character",
                "sec_num": "3.3"
            },
            {
                "text": "(3)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Uncertainty of Succeeding Character",
                "sec_num": "3.3"
            },
            {
                "text": "In (3), L av (s) and R av (s) are defined as the number of distinct preceding and succeeding characters, respectively, except, when the adjacent character is absent because of a sentence boundary, the pseudo-character of sentence beginning or sentence ending will be accumulated. Feng et al. (2004) also developed more heuristic rules to remove strings that contain known words or adhesive characters. For the strict meaning of unsupervised feature and for the sake of simplicity, these additional rules are dropped in this study.",
                "cite_spans": [
                    {
                        "start": 280,
                        "end": 298,
                        "text": "Feng et al. (2004)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Uncertainty of Succeeding Character",
                "sec_num": "3.3"
            },
            {
                "text": "Since a recent work of Sun and Xu (2011) used both L av (s) and R av (s) as features of CRF, this work will apply a similar approach, which is denoted as LRAVS, to make a thorough comparison.",
                "cite_spans": [
                    {
                        "start": 23,
                        "end": 40,
                        "text": "Sun and Xu (2011)",
                        "ref_id": "BIBREF28"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Uncertainty of Succeeding Character",
                "sec_num": "3.3"
            },
            {
                "text": "Besides previous works of TCB and TCF extraction (Sung et al., 2008) , Chinese frequent strings (Lin & Yu, 2001) , and reduced n-gram (Ha et al., 2005) , which have already been mentioned, the article about a linear algorithm for frequency of substring with reduction (L\u00fc & Zhang, 2005 ) also falls into this category. Most of these projects focused on the computational complexity of algorithms. Broader algorithms for frequent string extraction are suffix array (Manber & Myers, 1993) and PAT-tree (Chien, 1997) .",
                "cite_spans": [
                    {
                        "start": 49,
                        "end": 68,
                        "text": "(Sung et al., 2008)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 96,
                        "end": 112,
                        "text": "(Lin & Yu, 2001)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 134,
                        "end": 151,
                        "text": "(Ha et al., 2005)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 268,
                        "end": 285,
                        "text": "(L\u00fc & Zhang, 2005",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 464,
                        "end": 486,
                        "text": "(Manber & Myers, 1993)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 500,
                        "end": 513,
                        "text": "(Chien, 1997)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Frequent String Extraction Algorithm",
                "sec_num": "4.1"
            },
            {
                "text": "Zhao and Kit have explored several unsupervised strategies with their unified goodness measurement of logarithm ranking (Zhao & Kit, 2007) , including frequency of substring with reduction (L\u00fc & Zhang, 2005) , description length gain (Kit & Wilks, 1999) , accessor variety (Feng et al., 2004) , and boundary/branching entropy (Chang & Su, 1997; Cohen et al., 2007; Huang & Powers, 2003; Tanaka-Ishii, 2005; Tung & Lee, 1994) . Unlike the technique described in this paper for incorporating unsupervised features into supervised CRF learning, those methods usually filter out word-alike candidates using their own scoring mechanism directly as unsupervised word segmentation.",
                "cite_spans": [
                    {
                        "start": 120,
                        "end": 138,
                        "text": "(Zhao & Kit, 2007)",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 189,
                        "end": 207,
                        "text": "(L\u00fc & Zhang, 2005)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 234,
                        "end": 253,
                        "text": "(Kit & Wilks, 1999)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 273,
                        "end": 292,
                        "text": "(Feng et al., 2004)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 326,
                        "end": 344,
                        "text": "(Chang & Su, 1997;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 345,
                        "end": 364,
                        "text": "Cohen et al., 2007;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 365,
                        "end": 386,
                        "text": "Huang & Powers, 2003;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 387,
                        "end": 406,
                        "text": "Tanaka-Ishii, 2005;",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 407,
                        "end": 424,
                        "text": "Tung & Lee, 1994)",
                        "ref_id": "BIBREF33"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unsupervised Word Segmentation Method",
                "sec_num": "4.2"
            },
            {
                "text": "Subword based tagging of Zhang et al. (2006) utilizes confidence measurement. Other overlapping ambiguity resolution approaches are Na\u00efve Bayesian classifiers (Li et al., 2003) , mutual information, difference of t-test (Sun et al., 1997) , and sorted table look-up (Qiao et al., 2008) . These works concentrate on overlapping of words according to some (supervised) standard, rather than overlapping of substrings from unsupervised selection.",
                "cite_spans": [
                    {
                        "start": 25,
                        "end": 44,
                        "text": "Zhang et al. (2006)",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 159,
                        "end": 176,
                        "text": "(Li et al., 2003)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 220,
                        "end": 238,
                        "text": "(Sun et al., 1997)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 266,
                        "end": 285,
                        "text": "(Qiao et al., 2008)",
                        "ref_id": "BIBREF23"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overlapping Ambiguity Resolution",
                "sec_num": "4.3"
            },
            {
                "text": "In this study, the CRF label set for CWS prediction adopts the 6-tag approach of Zhao et al. (2010) , which achieves very competitive performance and is one of the most fine-grained character position based labeling schemes. According to Zhao et al. (2010) , since less than 1% of Chinese words are longer than five characters in most corpora from SIGHAN CWS bakeoffs 2003 , 2006 , the coverage of a 6-tag approach should be sufficient. This configuration of CRF without additional unsupervised features is also the control group of the experiment. Table 1 provides a sample of labeled training data. For the sample text \"\u53cd\u800c (contrarily) / \u6703 (make) / \u6b32\u901f\u5247\uf967\u9054 (more haste, less speed)\" (on the contrary, haste makes waste), the tag B stands for the beginning character of a word, while C and D represent the second character and the third character of a word, respectively. The ending character of a word is tagged as E. Once a word consists of more than four characters, the tag for all of the middle characters between D and E is I. Finally, the tag S is reserved specifically for single-character words.",
                "cite_spans": [
                    {
                        "start": 81,
                        "end": 99,
                        "text": "Zhao et al. (2010)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 238,
                        "end": 256,
                        "text": "Zhao et al. (2010)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 348,
                        "end": 372,
                        "text": "SIGHAN CWS bakeoffs 2003",
                        "ref_id": null
                    },
                    {
                        "start": 373,
                        "end": 379,
                        "text": ", 2006",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 549,
                        "end": 556,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Character Position Based Labels",
                "sec_num": "5.1"
            },
            {
                "text": "\u53cd B \u800c E \u6703 S \u6b32 B \u901f C \u5247 D \uf967 I \u9054 E Enhancement",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Character Label",
                "sec_num": null
            },
            {
                "text": "Feature instances are generated from templates based on the work of Ratnaparkhi (1996) . Table 2 explains their abilities. C -1 , C 0 , and C 1 stand for the input tokens individually bound to the prediction label at the current position. For example, in Table 1 , if the current position is at the label I, features generated by C -1 , C 0 , and C 1 are \"\u5247,\" \"\uf967,\" and \"\u9054,\" respectively. Meanwhile, for window size 2, C -1 C 0 , C 0 C 1 , and C -1 C 1 expands features of the label I to \"\u5247\uf967,\" \"\uf967\u9054,\" and \"\u5247\u9054,\" respectively. One may argue that the feature template should expand to five tokens to cover the whole range of the 6-tag approach; however, according to Zhao et al. (2010) , the context window size in three tokens is effective to catch parameters of the 6-tag approach for most strings that do not exceed five characters. Our pilot test for this case also showed that context window size in two tokens would be sufficient without a significant decrease in performance (Jiang et al., 2010) .",
                "cite_spans": [
                    {
                        "start": 68,
                        "end": 86,
                        "text": "Ratnaparkhi (1996)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 662,
                        "end": 680,
                        "text": "Zhao et al. (2010)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 977,
                        "end": 997,
                        "text": "(Jiang et al., 2010)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 89,
                        "end": 96,
                        "text": "Table 2",
                        "ref_id": "TABREF1"
                    },
                    {
                        "start": 255,
                        "end": 262,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Feature Templates",
                "sec_num": "5.2"
            },
            {
                "text": "Unsupervised features that will be introduced in the next subsection are generated by the same template, except the binding target moves column by column, as listed in tables of the next subsection. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Feature Templates",
                "sec_num": "5.2"
            },
            {
                "text": "To our knowledge, TCF, which is designed to fulfill a symmetrical comparison between the properties of inner pattern (CNG, TCF, or COS/TCB) vs. outer pattern (AVS) and between overlapping string (CNG, AVS, or TCF) vs. maximally matched string (COS/TCB), has not been evaluated in any previous work. In short, while the original version of COS/TCB selects the maximally matched string (i.e., non-overlapping string) as the feature (Feng et al., 2004; Jiang et al., 2010; Zhao & Kit, 2007) , TCF collects features of reduced n-gram from every character position with additional rank of likelihood converted from term-contributed frequency, as its name implies. To compare different types of overlapping strings as unsupervised features systematically, this work extends the previous work of Zhao and Kit (2007) into a unified representation of features. The representation accommodates both character position of a string and the string's likelihood ranked in the logarithm. Formally, the ranking function for a string s with a score x counted by CNG, AVS, or TCF is defined as:",
                "cite_spans": [
                    {
                        "start": 430,
                        "end": 449,
                        "text": "(Feng et al., 2004;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 450,
                        "end": 469,
                        "text": "Jiang et al., 2010;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 470,
                        "end": 487,
                        "text": "Zhao & Kit, 2007)",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 789,
                        "end": 808,
                        "text": "Zhao and Kit (2007)",
                        "ref_id": "BIBREF38"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unified Feature Representation of CNG/AVS/TCF/TCB",
                "sec_num": "5.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "1 ( ) , 2 2 r r f s r if x + = \u2264 < .",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Unified Feature Representation of CNG/AVS/TCF/TCB",
                "sec_num": "5.3"
            },
            {
                "text": "The logarithm ranking mechanism in (4) is inspired by Zipf's law with the intention to alleviate the potential data sparseness problem of infrequent strings. The rank r and the corresponding character positions of a string then are concatenated as feature tokens. To give the reader a clearer picture about what feature tokens look like, a sample representation, which is denoted in regex as \"[0-9]+[B|C|D|I|E|S]\" for rank and character position, of CNG, AVS, or TCF is demonstrated and explained by Figure 1 and Table 3 . For example, judging by strings with two characters, one of the strings \"\u53cd\u800c\" gets rank r = 3; therefore, the column of two-character feature tokens has \"\u53cd\" denoted as 3B and \"\u800c\" denoted as 3E. If another two-character string \"\u800c\u6703\" competes with \"\u53cd\u800c\" at the position of \"\u800c\" with a lower rank r = 0, then 3E is selected for feature representation of the token at a certain position.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 500,
                        "end": 508,
                        "text": "Figure 1",
                        "ref_id": "FIGREF1"
                    },
                    {
                        "start": 513,
                        "end": 520,
                        "text": "Table 3",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Unified Feature Representation of CNG/AVS/TCF/TCB",
                "sec_num": "5.3"
            },
            {
                "text": "Note that, when the string \"\u5247\uf967\" conflicts with the string \"\uf967\u9054\" at the position of \"\uf967\" with the same rank r = 3, the corresponding character position with rank of the leftmost string, which is 3E in this case, is applied arbitrarily.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unified Feature Representation of CNG/AVS/TCF/TCB",
                "sec_num": "5.3"
            },
            {
                "text": "Although those are indeed common situations of overlapping strings, this work simply implements the above rules by Zhao and Kit (2007) for the sake of compatibility. In fact, pilot tests have been done with a more complicated representation, like 3E-0B for \"\u800c\" and 3E-3B for \"\uf967,\" to keep the overlapping information within each column, but the test result shows no significant differences in terms of accuracy and OOV recognition. Since the statistics of the pilot tests could be redundant, they are omitted in this paper.",
                "cite_spans": [
                    {
                        "start": 115,
                        "end": 134,
                        "text": "Zhao and Kit (2007)",
                        "ref_id": "BIBREF38"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unified Feature Representation of CNG/AVS/TCF/TCB",
                "sec_num": "5.3"
            },
            {
                "text": "To make an informative comparison, this work also applies the original version of non-overlapping COS/TCB features that is without ranks and is selected by the forward maximum matching algorithm (Feng et al., 2004; Jiang et al., 2010; Zhao & Kit, 2007) . Table  4 illustrates a sample representation of features in this case. Notably, there are several features encoded as -1 individually to represent that the desired string is unseen. For the non-overlapping siblings of the reduced n-grams family, such as COS/TCB, either the string is always occupied by other superstrings or it simply does not appear more than once.",
                "cite_spans": [
                    {
                        "start": 195,
                        "end": 214,
                        "text": "(Feng et al., 2004;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 215,
                        "end": 234,
                        "text": "Jiang et al., 2010;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 235,
                        "end": 252,
                        "text": "Zhao & Kit, 2007)",
                        "ref_id": "BIBREF38"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 255,
                        "end": 263,
                        "text": "Table  4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Unified Feature Representation of CNG/AVS/TCF/TCB",
                "sec_num": "5.3"
            },
            {
                "text": "Non-overlapping COS/TCB strings.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Table 4. Sample of the unified feature representation for",
                "sec_num": null
            },
            {
                "text": "\u53cd B B \u800c C E \u6703 E S \u6b32 -1 B \u901f -1 C \u5247 -1 D \uf967 -1 I \u9054 -1 E",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Input Original COS/TCB Feature Label",
                "sec_num": null
            },
            {
                "text": "The length of a string is limited to five characters for the sake of efficiency and consistency with the 6-tag approach.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Input Original COS/TCB Feature Label",
                "sec_num": null
            },
            {
                "text": "CRF++ 0.54 (http://crfpp.sourceforge.net/) employs L-BFGS optimization and the tunable hyper-parameter (CRF++ training function argument \"-c\"), i.e., the Gaussian prior, set to 100 throughout the whole experiment.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "6."
            },
            {
                "text": "The corpora used for the experiment are from the SIGHAN CWS bakeoff 2005 (Emerson, 2005) and SIGHAN CWS bakeoff 2010 (Zhao & Liu, 2010) . SIGHAN 2005 comes with four different standards, including Academia Sinica (AS), City University of Hong Kong (CityU), Microsoft Research (MSR), and Peking University (PKU). SIGHAN 2010 provides a Traditional Chinese corpus and a Simplified Chinese corpus. Each corpus has training/test sets of four domains, including literature, computers, medicine, and finance, that are denoted as domains A, B, C, and D, respectively. For comparison, statistics on most corpora of SIGHAN 2003 SIGHAN , 2006 SIGHAN , and 2008 that have been obtained are listed in the appendix.",
                "cite_spans": [
                    {
                        "start": 73,
                        "end": 88,
                        "text": "(Emerson, 2005)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 117,
                        "end": 135,
                        "text": "(Zhao & Liu, 2010)",
                        "ref_id": null
                    },
                    {
                        "start": 607,
                        "end": 618,
                        "text": "SIGHAN 2003",
                        "ref_id": null
                    },
                    {
                        "start": 619,
                        "end": 632,
                        "text": "SIGHAN , 2006",
                        "ref_id": null
                    },
                    {
                        "start": 633,
                        "end": 650,
                        "text": "SIGHAN , and 2008",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data Set",
                "sec_num": "6.1"
            },
            {
                "text": "Unsupervised features are collected according to pairs of corresponding training/test corpora. CNG and AVS are arranged with the help from SRILM (Stolcke, 2002) . TCB strings and their ranks converted from TCF are calculated by YASA (Sung et al., 2008) . To distinguish the ranked and overlapping features of TCB/TCF from those of the original version of non-overlapping COS/TCB-based features, the former are denoted as TCF to indicate the score source of frequency for ranking, and the abbreviation of the later remains as TCB.",
                "cite_spans": [
                    {
                        "start": 145,
                        "end": 160,
                        "text": "(Stolcke, 2002)",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 233,
                        "end": 252,
                        "text": "(Sung et al., 2008)",
                        "ref_id": "BIBREF30"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unsupervised Feature Selection",
                "sec_num": "6.2"
            },
            {
                "text": "The evaluation metrics of CWS task are adopted from SIGHAN bakeoffs, including test precision (P), test recall (R), and their harmonic average F 1 measure score (F), as (5), (6), and 7, respectively. For performance of OOV, formulae that are similar to P/R/F are employed. To estimate the differences of performance between configurations of CWS experiments, this work uses the confidence level, which has been applied since SIGHAN CWS bakeoff 2003 (Sproat & Emerson, 2003) . The confidence level assumes that the recall (or precision) X of accuracy (or OOV recognition) represents the probability that a word (or OOV word) will be identified from N words in total and that a binomial distribution is appropriate for the experiment. Confidence levels of P, R, P OOV , and R OOV appear in Tables 5-10 under the columns C P , C R , C Poov , and C Roov , respectively, and they are calculated at the 95% confidence interval with the formula \u00b12 \u221a ([X(1-X)] / N). Two configurations of CWS experiments then are considered to be statistically different at a 95% confidence level if one of their C P , C R , ",
                "cite_spans": [
                    {
                        "start": 449,
                        "end": 473,
                        "text": "(Sproat & Emerson, 2003)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 550,
                        "end": 570,
                        "text": "(or OOV recognition)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Metrics",
                "sec_num": "6.3"
            },
            {
                "text": "the number of words that are correctly segmented 100% the number of words in the gold standard R = \u00d7 .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Metrics",
                "sec_num": "6.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "2 P R F P R \u00d7 \u00d7 = + .",
                        "eq_num": "(6)"
                    }
                ],
                "section": "Evaluation Metrics",
                "sec_num": "6.3"
            },
            {
                "text": "The most significant type of error is unintentionally segmented alphanumeric sequences, such as English words or factoids in Arabic numerals. Rather than developing another set of feature templates for non-Chinese characters that may violate the rules of closed training evaluation, post-processing, which is mentioned in the official report of SIGHAN CWS bakeoff 2005 (Emerson, 2005) , has been applied to remove spaces between non-Chinese characters in the gold standard data of the AS corpus manually, since there are no urgent expectations of correct segmentation on non-Chinese text. In SIGHAN 2005 and 2006, however, some participants used character types, such as digits, date/time specific Chinese characters, English letters, punctuation, and others (Chinese characters) as extra features, which triggered a debate of closed training criteria (Zhao et al., 2010) . Consequently, SIGHAN 2010 decided to allow four types of characters, distinguished as Chinese characters, English letters, digits, and punctuation. This work provides preliminary tests on non-Chinese patterns extracted from SIGHAN 2010 unlabeled training corpora A and B, extra features of character types (in character based trigram, T -1 T 0 T 1 , where T can be E, D, P, or C for alphabets, digits, punctuations, or Chinese characters, respectively), and their combinations to verify the performance impact of these special treatments, as shown in Table 5 -Table 8 . On the one hand, the statistics indicate that the character types perform well and stably on most of the corpora. On the other hand, the features, such as AVS and TCF, may still need help from non-Chinese patterns of unlabeled training corpora A and B. As a matter of fact, our other preliminary test suggests that SIGHAN 2010 test corpora contain a lot of OOV and inconsistent segments from non-Chinese text (for example, inconsistency of usage on full-width or half-width non-Chinese characters, some English words and factoids being segmented but some of them not, etc.), which only can be memorized from the non-Chinese patterns. Consequently, the experimental results of SIGHAN 2010 corpora involve non-Chinese treatment based on the combination of the extra character type features and the non-Chinese patterns, but the experimental results of SIGHAN 2005 corpora do not. This empirical decision implies that CWS benchmarking corpus should be prepared more carefully to avoid unpredictable side effects from non-Chinese text. Note that the treatment does not use unlabeled training corpora A and B separately. Further discussions are mainly based on this treatment, hopefully without loss of generality and of interest for comparative studies. Numbers in bold face and italic style indicate the best and the second best results of a certain evaluation metric, respectively, except for the topline and the best record from each year of SIGHAN bakeoffs. Configurations with the same values of confidence level on P or R are underlined, but only records that have the same confidence level on both P and R should be considered as statistically insignificant, and this phenomenon did not occur in our experiment results.",
                "cite_spans": [
                    {
                        "start": 369,
                        "end": 384,
                        "text": "(Emerson, 2005)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 852,
                        "end": 871,
                        "text": "(Zhao et al., 2010)",
                        "ref_id": "BIBREF37"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 1425,
                        "end": 1441,
                        "text": "Table 5 -Table 8",
                        "ref_id": "TABREF4"
                    }
                ],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.4"
            },
            {
                "text": "Unlike the previous work, which showed a relatively clearer trend of feature selection (Jiang et al., 2011) , CWS performance may vary between different CWS standards and domains in this study. Considering either the best or second best records in terms of F, feature combinations consisting of LRAVS or AVS usually outperform, except on MSR of SIGHAN 2005 corpora. Nevertheless, in terms of F OOV , feature combinations consisting of TCF or TCB consistently increase in performance on every corpus. Similar situations also can be recognized from the experiments on some of the SIGHAN 2003 SIGHAN , 2006 SIGHAN , and 2008  please refer to the appendix for details. This complicated phenomenon indicates that, since CWS studies usually struggle with incremental and small improvements, different CWS standards and/or domains can make comparative research difficult and cause experimental results of related works to be incompatible. For equipping supervised CWS with unsupervised feature selection from unlabeled data, the experimental results of this work suggests that using LRAVS+TCF with more careful non-Chinese text treatments and CRF parameter tuning (e.g., more cross-validations to find a specific hyper-parameter of Gaussian prior) would be a very good choice. Nevertheless, it is still worth noting that the best performance of this work in terms of F is found on the best official records on traditional Chinese domain B (Computer) of SIGHAN 2010 corpora and all of the SIGHAN 2005 corpora except the PKU corpus. This is especially true when this work does not apply any special treatment of character type and non-Chinese text that many other related works do on SIGHAN 2005 corpora. Note that \"Our Baseline/Topline\" in the following tables indicates where official baseline/topline suffered from official release script for maximum matching malfunctions on data in UTF-8 encoding and/or some uncertain incompatibilities between obtained corpora and official ones that caused inconsistent statistics during experiment reproductions. It has been observed that using any of the unsupervised features could create short patterns for the CRF learner, which might break more English words than using the 6-tag approach alone. AVS, TCF, and TCB, however, resolve more overlapping ambiguities of Chinese words than the 6-tag approach and CNG. Interestingly, even for the unsupervised feature without rank or overlapping information, TCB/TCF successfully recognizes \"\u4f9d\u9760 / \u5355\u4f4d / \u7684 / \u7ebd\u5e26 / \u6765 / \u7ef4\u6301,\" while the 6-tag approach sees this phrase incorrectly as \"\u4f9d \u9760 / \u5355\u4f4d / \u7684 / \u7ebd / \u5e26\u6765 / \u7ef4\u6301.\" TCB/TCF also saves more factoids, such as \"\u4e00\u4e8c\u4e5d\uff0e \u4e5d / \u5de6\u53f3\" (129.9 / around) from scattered tokens, such as \"\u4e00\u4e8c\u4e5d / \uff0e / \u4e5d / \u5de6\u53f3\" (129 / point / 9 / around).",
                "cite_spans": [
                    {
                        "start": 87,
                        "end": 107,
                        "text": "(Jiang et al., 2011)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 578,
                        "end": 589,
                        "text": "SIGHAN 2003",
                        "ref_id": null
                    },
                    {
                        "start": 590,
                        "end": 603,
                        "text": "SIGHAN , 2006",
                        "ref_id": null
                    },
                    {
                        "start": 604,
                        "end": 621,
                        "text": "SIGHAN , and 2008",
                        "ref_id": null
                    },
                    {
                        "start": 622,
                        "end": 622,
                        "text": "",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.4"
            },
            {
                "text": "The above observations suggest that the quality of a string as a word-like candidate should be an important factor for the unsupervised feature injected CRF learner. Relatively speaking, CNG probably brings in too much noise. Feature combinations of LRAVS and TCF usually improve F and F OOV , respectively. Improvements are significant in terms of C R , C P , C Roov , and C Poov ,, which confirms the hypothesis mentioned at the end of Section 1.3 that, combining information from the outer pattern of a substring (i.e., LRAVS) with information from the inner pattern of a substring (i.e., TCF) into a compound of unsupervised feature could help improving CWS performance of supervised labeling scheme of CRF. Nevertheless, since AVS or TCB sometimes gain better results, fine-tuning of feature engineering according to different corpora and segmentation standards is necessary.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.4"
            },
            {
                "text": "This work provides a unified view of CRF-based CWS integrated with unsupervised features via frequent string, and it reasons that, since LRAVS comes with inner structure and TCF comes with outer structure of overlapping string, utilizing their compound features could be more useful than applying one of them solely. The thorough experimental results show that the compound features of LRAVS and TCF usually obtain competitive performance in terms of F and F OOV , respectively. Sometimes, AVS and TCB may contribute more, but generally combining the outer pattern of a substring (i.e., LRAVS or AVS) with the inner pattern of a substring (i.e., TCF or TCB) into a compound of unsupervised features could help improve CWS performance of a supervised labeling scheme of CRF. Recommended future investigation is unknown word extraction and named entity recognition using AVS (Li et al., 2010) and TCF/TCB (Chang & Lee, 2003; Zhang et al., 2010) as features for more complicated CRF (Sun & Nan, 2010 ",
                "cite_spans": [
                    {
                        "start": 873,
                        "end": 890,
                        "text": "(Li et al., 2010)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 903,
                        "end": 922,
                        "text": "(Chang & Lee, 2003;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 923,
                        "end": 942,
                        "text": "Zhang et al., 2010)",
                        "ref_id": "BIBREF35"
                    },
                    {
                        "start": 980,
                        "end": 996,
                        "text": "(Sun & Nan, 2010",
                        "ref_id": "BIBREF29"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "7."
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "An Unsupervised Iterative Method for Chinese New Lexicon Extraction",
                "authors": [
                    {
                        "first": "J.-S",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "K.-Y",
                        "middle": [],
                        "last": "Su",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Proc. Computational Linguistics and Chinese Language Processing",
                "volume": "2",
                "issue": "",
                "pages": "97--148",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chang, J.-S., & Su, K.-Y. (1997). An Unsupervised Iterative Method for Chinese New Lexicon Extraction. in Proc. Computational Linguistics and Chinese Language Processing, 2(2), 97-148.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Automatic Chinese unknown word extraction using small-corpus-based method",
                "authors": [
                    {
                        "first": "T.-H",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "C.-H",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. International Conference on Natural Language Processing and Knowledge Engineering",
                "volume": "",
                "issue": "",
                "pages": "459--464",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chang, T.-H., & Lee, C.-H. (2003). Automatic Chinese unknown word extraction using small-corpus-based method. in Proc. International Conference on Natural Language Processing and Knowledge Engineering, 459-464.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "PAT-tree-based Keyword Extraction for Chinese Information Retrieval",
                "authors": [
                    {
                        "first": "L.-F",
                        "middle": [],
                        "last": "Chien",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Proc. 20th Annnual International ACM SIGIR Conference on Research and Development in Information Retrieval",
                "volume": "",
                "issue": "",
                "pages": "50--58",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chien, L.-F. (1997). PAT-tree-based Keyword Extraction for Chinese Information Retrieval. in Proc. 20th Annnual International ACM SIGIR Conference on Research and Development in Information Retrieval, 50-58.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Voting Experts: An Unsupervised Algorithm for Segmenting Sequences",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Cohen",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Adams",
                        "suffix": ""
                    },
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Heeringa",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Intelligent Data Analysis",
                "volume": "11",
                "issue": "6",
                "pages": "607--625",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Cohen, P., Adams, N., & Heeringa, B. (2007). Voting Experts: An Unsupervised Algorithm for Segmenting Sequences. Intelligent Data Analysis, 11(6), 607-625.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "The Second International Chinese Word Segmentation Bakeoff",
                "authors": [
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Emerson",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proc. 4th SIGHAN Workshop on Chinese Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Emerson, T. (2005). The Second International Chinese Word Segmentation Bakeoff. in Proc. 4th SIGHAN Workshop on Chinese Language Processing.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Accessor Variety Criteria for Chinese Word Extraction",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Feng",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Deng",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Zheng",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Computational Linguistics",
                "volume": "30",
                "issue": "1",
                "pages": "75--93",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Feng, H., Chen, K., Deng, X., & Zheng, W. (2004). Accessor Variety Criteria for Chinese Word Extraction. Computational Linguistics, 30(1), 75-93.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Reduced N-Grams for Chinese Evaluation",
                "authors": [
                    {
                        "first": "L",
                        "middle": [
                            "Q"
                        ],
                        "last": "Ha",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Seymour",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Hanna",
                        "suffix": ""
                    },
                    {
                        "first": "F",
                        "middle": [
                            "J"
                        ],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Computational Linguistics and Chinese Language Processing",
                "volume": "10",
                "issue": "",
                "pages": "19--34",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ha, L. Q., Seymour, R., Hanna, P., & Smith, F. J. (2005). Reduced N-Grams for Chinese Evaluation. Computational Linguistics and Chinese Language Processing, 10(1), 19-34.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Morpheme Boundaries within Words. Paper presented at the Structural and Transformational Linguistics",
                "authors": [
                    {
                        "first": "Z",
                        "middle": [
                            "S"
                        ],
                        "last": "Harris",
                        "suffix": ""
                    }
                ],
                "year": 1970,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Harris, Z. S. (1970). Morpheme Boundaries within Words. Paper presented at the Structural and Transformational Linguistics.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Chinese Word Segmentation based on contextual entropy",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "H"
                        ],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Powers",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. 17th Asian Pacific Conference on Language, Information and Computation",
                "volume": "",
                "issue": "",
                "pages": "152--158",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Huang, J. H., & Powers, D. (2003). Chinese Word Segmentation based on contextual entropy. in Proc. 17th Asian Pacific Conference on Language, Information and Computation, 152-158.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Enhancement of Unsupervised Feature Selection for Conditional Random Fields Learning in Chinese Word Segmentation",
                "authors": [
                    {
                        "first": "T.-J",
                        "middle": [],
                        "last": "Jiang",
                        "suffix": ""
                    },
                    {
                        "first": "W.-L",
                        "middle": [],
                        "last": "Hsu",
                        "suffix": ""
                    },
                    {
                        "first": "C.-H",
                        "middle": [],
                        "last": "Kuo",
                        "suffix": ""
                    },
                    {
                        "first": "T.-H",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proc. 7th IEEE International Conference on Natural Language Processing and Knowledge Engineering",
                "volume": "",
                "issue": "",
                "pages": "382--389",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jiang, T.-J., Hsu, W.-L., Kuo, C.-H., & Yang, T.-H. (2011). Enhancement of Unsupervised Feature Selection for Conditional Random Fields Learning in Chinese Word Segmentation. in Proc. 7th IEEE International Conference on Natural Language Processing and Knowledge Engineering, 382-389.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Term Contributed Boundary Tagging by Conditional Random Fields for SIGHAN 2010 Chinese Word Segmentation Bakeoff",
                "authors": [
                    {
                        "first": "T.-J",
                        "middle": [],
                        "last": "Jiang",
                        "suffix": ""
                    },
                    {
                        "first": "S.-H",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "C.-L",
                        "middle": [],
                        "last": "Sung",
                        "suffix": ""
                    },
                    {
                        "first": "W.-L",
                        "middle": [],
                        "last": "Hsu",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proc. 1st CIPS-SIGHAN Joint Conf. on Chinese Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jiang, T.-J., Liu, S.-H., Sung, C.-L., & Hsu, W.-L. (2010). Term Contributed Boundary Tagging by Conditional Random Fields for SIGHAN 2010 Chinese Word Segmentation Bakeoff. in Proc. 1st CIPS-SIGHAN Joint Conf. on Chinese Language Processing, Beijing, China.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "The Fourth International Chinese Language Processing Bakeoff : Chinese Word Segmentation, Named Entity Recognition and Chinese POS Tagging",
                "authors": [
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Jin",
                        "suffix": ""
                    },
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proc. 6th SIGHAN Workshop on Chinese Language Processing",
                "volume": "",
                "issue": "",
                "pages": "69--81",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jin, G., & Chen, X. (2007). The Fourth International Chinese Language Processing Bakeoff : Chinese Word Segmentation, Named Entity Recognition and Chinese POS Tagging. in Proc. 6th SIGHAN Workshop on Chinese Language Processing, 69-81.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Unsupervised learning of word boundary with description length gain",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Kit",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Wilks",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Proc. CoNLL-99",
                "volume": "",
                "issue": "",
                "pages": "1--6",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kit, C., & Wilks, Y. (1999). Unsupervised learning of word boundary with description length gain. in Proc. CoNLL-99, 1-6.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Statistical Substring Reduction in Linear Time",
                "authors": [
                    {
                        "first": "X",
                        "middle": [],
                        "last": "L\u00fc",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proc. 1st Internal Joint Conference on Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "L\u00fc, X., & Zhang, L. (2005). Statistical Substring Reduction in Linear Time. in Proc. 1st Internal Joint Conference on Natural Language Processing.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Conditional Random Fields Enhancement of Feature Engineering for Conditional Random 73",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Lafferty",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Mccallum",
                        "suffix": ""
                    },
                    {
                        "first": "F",
                        "middle": [
                            "C N"
                        ],
                        "last": "Pereira",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Lafferty, J., McCallum, A., & Pereira, F. C. N. (2001). Conditional Random Fields Enhancement of Feature Engineering for Conditional Random 73",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Field Learning in Chinese Word Segmentation Using Unlabeled Data Probabilistic Models for Segmenting and Labeling Sequence Data",
                "authors": [],
                "year": null,
                "venue": "Proc. ICML",
                "volume": "",
                "issue": "",
                "pages": "282--289",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Field Learning in Chinese Word Segmentation Using Unlabeled Data Probabilistic Models for Segmenting and Labeling Sequence Data. in Proc. ICML. 282-289.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "The Third International Chinese Language Processing Bakeoff Word Segmentation and Named Entity Recognition",
                "authors": [
                    {
                        "first": "G.-A",
                        "middle": [],
                        "last": "Levow",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proc. 5th SIGHAN Workshop on Chinese Language Processing",
                "volume": "",
                "issue": "",
                "pages": "108--117",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Levow, G.-A. (2006). The Third International Chinese Language Processing Bakeoff Word Segmentation and Named Entity Recognition. in Proc. 5th SIGHAN Workshop on Chinese Language Processing, 108-117.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "A Hybrid Model Combining CRF with Boundary Templates for Chinese Person Name Recognition",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Z",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Z",
                        "middle": [],
                        "last": "Ding",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "International Journal Advanced Intelligent",
                "volume": "2",
                "issue": "1",
                "pages": "73--80",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Li, L., Li, Z., Ding, Z., & Huang, D. (2010). A Hybrid Model Combining CRF with Boundary Templates for Chinese Person Name Recognition. International Journal Advanced Intelligent, 2(1), 73-80.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Unsupervised Training for Overlapping Ambiguity Resolution in Chinese Word Segmentation",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. 2nd SIGHAN Workshop on Chinese Language Processing",
                "volume": "17",
                "issue": "",
                "pages": "1--7",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Li, M., Gao, J., Huang, C., & Li, J. (2003). Unsupervised Training for Overlapping Ambiguity Resolution in Chinese Word Segmentation. in Proc. 2nd SIGHAN Workshop on Chinese Language Processing, 17, 1-7.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Extracting Chinese Frequent Strings without a Dictionary from a Chinese Corpus and its Applications",
                "authors": [
                    {
                        "first": "Y.-J",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    },
                    {
                        "first": "M.-S",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "J. Information Science and Engineering",
                "volume": "17",
                "issue": "",
                "pages": "805--824",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Lin, Y.-J., & Yu, M.-S. (2001). Extracting Chinese Frequent Strings without a Dictionary from a Chinese Corpus and its Applications. J. Information Science and Engineering, 17, 805-824.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Introduction to CKIP Chinese Word Segmentation System for the First International Chinese Word Segmentation Bakeoff",
                "authors": [
                    {
                        "first": "W.-Y",
                        "middle": [],
                        "last": "Ma",
                        "suffix": ""
                    },
                    {
                        "first": "K.-J",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. 2nd SIGHAN Workshop on Chinese Language Processing",
                "volume": "17",
                "issue": "",
                "pages": "168--171",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ma, W.-Y., & Chen, K.-J. (2003). Introduction to CKIP Chinese Word Segmentation System for the First International Chinese Word Segmentation Bakeoff. in Proc. 2nd SIGHAN Workshop on Chinese Language Processing, 17, 168-171.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Suffix arrays: a new method for on-line string searches",
                "authors": [
                    {
                        "first": "U",
                        "middle": [],
                        "last": "Manber",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Myers",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "SIAM J. Computing",
                "volume": "22",
                "issue": "5",
                "pages": "935--948",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Manber, U., & Myers, G. (1993). Suffix arrays: a new method for on-line string searches. SIAM J. Computing, 22(5), 935-948.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "A Study of an N-Gram Language Model for Speech Recognition",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "O'boyle",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "O'Boyle, P. (1993). A Study of an N-Gram Language Model for Speech Recognition. (Ph.D.), Queen's University Belfast.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Statistical Properties of Overlapping Ambiguities in Chinese Word Segmentation and a Strategy for Their Disambiguation",
                "authors": [
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Qiao",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Menzel",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proc. Text, Speech and Dialogue",
                "volume": "",
                "issue": "",
                "pages": "177--186",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Qiao, W., Sun, M., & Menzel, W. (2008). Statistical Properties of Overlapping Ambiguities in Chinese Word Segmentation and a Strategy for Their Disambiguation. in Proc. Text, Speech and Dialogue, 177-186.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "A Maximum Entropy Model for Part-of-Speech Tagging",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Ratnaparkhi",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "Proc. Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "133--142",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ratnaparkhi, A. (1996). A Maximum Entropy Model for Part-of-Speech Tagging. in Proc. Empirical Methods in Natural Language Processing, 133-142.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "The First International Chinese Word Segmentation Bakeoff",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Sproat",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Emerson",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. 2nd SIGHAN Workshop on Chinese Language Processing",
                "volume": "17",
                "issue": "",
                "pages": "133--143",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sproat, R., & Emerson, T. (2003). The First International Chinese Word Segmentation Bakeoff. in Proc. 2nd SIGHAN Workshop on Chinese Language Processing, 17, 133-143.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "SRILM -An Extensible Language Modeling Toolkit",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Stolcke",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proc. Spoken Language Processing",
                "volume": "",
                "issue": "",
                "pages": "901--904",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Stolcke, A. (2002). SRILM -An Extensible Language Modeling Toolkit. in Proc. Spoken Language Processing, 901-904.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Using Character Bigram for Ambiguity Resolution In Chinese Word Segmentation (In Chinese)",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "N"
                        ],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "F",
                        "middle": [],
                        "last": "Lu",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [
                            "Y"
                        ],
                        "last": "Shen",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Computer Research and Development",
                "volume": "34",
                "issue": "5",
                "pages": "332--339",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sun, M., Huang, C. N., Lu, F., & Shen, D. Y. (1997). Using Character Bigram for Ambiguity Resolution In Chinese Word Segmentation (In Chinese). Computer Research and Development, 34(5), 332-339.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Enhancing Chinese Word Segmentation Using Unlabeled Data",
                "authors": [
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proc. Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "970--979",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sun, W., & Xu, J. (2011). Enhancing Chinese Word Segmentation Using Unlabeled Data. in Proc. Empirical Methods in Natural Language Processing, 970-979.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Chinese base phrases chunking based on latent semi-CRF model",
                "authors": [
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Nan",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proc. International Conference on Natural Language Processing and Knowledge Engineering",
                "volume": "",
                "issue": "",
                "pages": "1--7",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sun, X., & Nan, X. (2010). Chinese base phrases chunking based on latent semi-CRF model. in Proc. International Conference on Natural Language Processing and Knowledge Engineering, 1-7.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Compute the Term Contributed Frequency",
                "authors": [
                    {
                        "first": "C.-L",
                        "middle": [],
                        "last": "Sung",
                        "suffix": ""
                    },
                    {
                        "first": "H.-C",
                        "middle": [],
                        "last": "Yen",
                        "suffix": ""
                    },
                    {
                        "first": "W.-L",
                        "middle": [],
                        "last": "Hsu",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sung, C.-L., Yen, H.-C., & Hsu, W.-L. (2008). Compute the Term Contributed Frequency. in Mike Tian-Jian Jiang et al.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Proc. 8th Int. Conference Intelligent System Design and Application",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "2",
                "issue": "",
                "pages": "325--328",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Proc. 8th Int. Conference Intelligent System Design and Application, 2, 325-328.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Entropy as an Indicator of Context Boundaries: An Experiment Using a Web Search Engine",
                "authors": [
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Tanaka-Ishii",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proc. Internal Joint Conference on Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "93--105",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tanaka-Ishii, K. (2005). Entropy as an Indicator of Context Boundaries: An Experiment Using a Web Search Engine. in Proc. Internal Joint Conference on Natural Language Processing, 93-105.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "Identification of Unkown Words from Corpus",
                "authors": [
                    {
                        "first": "C.-H",
                        "middle": [],
                        "last": "Tung",
                        "suffix": ""
                    },
                    {
                        "first": "H.-J",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 1994,
                "venue": "Computational Proc. Chinese and Oriental Languages",
                "volume": "8",
                "issue": "",
                "pages": "131--145",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tung, C.-H., & Lee, H.-J. (1994). Identification of Unkown Words from Corpus. Computational Proc. Chinese and Oriental Languages, 8, 131-145.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Conditional Random Fields An Introduction",
                "authors": [
                    {
                        "first": "H",
                        "middle": [
                            "M"
                        ],
                        "last": "Wallach",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Wallach, H. M. (2004). Conditional Random Fields An Introduction. (MS-CIS-04-21).",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "A pragmatic model for new Chinese word extraction",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Shi",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proc. International Conference on Natural Language Processing and Knowledge Engineering",
                "volume": "",
                "issue": "",
                "pages": "1--8",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhang, H., Huang, H., Zhu, C., & Shi, S. (2010). A pragmatic model for new Chinese word extraction. in Proc. International Conference on Natural Language Processing and Knowledge Engineering, 1-8.",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "Subword-based Tagging for Confidence-dependent Chinese Word Segmentation",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Kikui",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Sumita",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proc. COLING/ACL",
                "volume": "",
                "issue": "",
                "pages": "961--968",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhang, R., Kikui, G., & Sumita, E. (2006). Subword-based Tagging for Confidence-dependent Chinese Word Segmentation. in Proc. COLING/ACL, 961-968.",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "title": "A Unified Character-Based Tagging Framework for Chinese Word Segmentation",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    },
                    {
                        "first": "C.-N",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "B.-L",
                        "middle": [],
                        "last": "Lu",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "ACM Trans. on Asian Language Information Processing",
                "volume": "9",
                "issue": "2",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhao, H., Huang, C.-N., Li, M., & Lu, B.-L. (2010). A Unified Character-Based Tagging Framework for Chinese Word Segmentation. ACM Trans. on Asian Language Information Processing, 9(2).",
                "links": null
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "Incorporating Global Information into Supervised Learning for Chinese Word Segmentation",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Kit",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proc. 10th PACLIC",
                "volume": "",
                "issue": "",
                "pages": "66--74",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhao, H., & Kit, C. (2007). Incorporating Global Information into Supervised Learning for Chinese Word Segmentation. in Proc. 10th PACLIC, 66-74.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF1": {
                "text": "Example of overlapping strings with ranks.",
                "num": null,
                "type_str": "figure",
                "uris": null
            },
            "TABREF1": {
                "text": "",
                "content": "<table><tr><td>C -1 , C 0 , C 1</td><td>Previous, current, or next token</td></tr><tr><td>C -1 C 0</td><td>Previous and current tokens</td></tr><tr><td>C 0 C 1</td><td>Current and next tokens</td></tr><tr><td>C -1 C 1</td><td>Previous and next tokens</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF2": {
                "text": "",
                "content": "<table><tr><td>Input</td><td>1 char</td><td>2 char</td><td colspan=\"2\">Unsupervised Feature 3 char 4 char</td><td>5 char</td><td>Label</td></tr><tr><td>\u53cd</td><td>5S</td><td>3B</td><td>4B</td><td>0B</td><td>0B</td><td>B</td></tr><tr><td>\u800c</td><td>6S</td><td>3E</td><td>4C</td><td>0C</td><td>0C</td><td>E</td></tr><tr><td>\u6703</td><td>6S</td><td>0E</td><td>4E</td><td>0D</td><td>0D</td><td>S</td></tr><tr><td>\u6b32</td><td>4S</td><td>0E</td><td>0E</td><td>0E</td><td>0I</td><td>B</td></tr><tr><td>\u901f</td><td>4S</td><td>0E</td><td>0E</td><td>0E</td><td>0E</td><td>C</td></tr><tr><td>\u5247</td><td>6S</td><td>3B</td><td>0E</td><td>0E</td><td>0E</td><td>D</td></tr><tr><td>\uf967</td><td>7S</td><td>3E</td><td>0E</td><td>0E</td><td>0E</td><td>I</td></tr><tr><td>\u9054</td><td>5S</td><td>3E</td><td>0E</td><td>0E</td><td>0E</td><td>E</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF3": {
                "text": "Enhancement of Feature Engineering for Conditional Random 55 Field Learning in Chinese Word Segmentation Using Unlabeled Data C Poov , or C Roov is different.",
                "content": "<table><tr><td>P =</td><td>the number of words thatare correctly segmented 100% \u00d7 the number of words that are segmented</td><td>.</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF4": {
                "text": "",
                "content": "<table><tr><td colspan=\"2\">Domain Feature</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td/><td>Original 6-tag</td><td/><td>92.16 \u00b10.002869</td><td/><td>91.63 \u00b10.002956</td><td>91.89</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td/><td>92.32 \u00b10.002842</td><td/><td>91.27 \u00b10.003013</td><td>91.79</td></tr><tr><td>A</td><td>+(Character Type)</td><td/><td>92.70 \u00b10.002777</td><td/><td>92.33 \u00b10.002840</td><td>92.51</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td/><td>92.71 \u00b10.002775</td><td/><td>92.33 \u00b10.002841</td><td>92.52</td></tr><tr><td/><td>Original 6-tag</td><td/><td>77.44 \u00b10.004558</td><td/><td>86.72 \u00b10.003701</td><td>81.82</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td/><td>89.85 \u00b10.003294</td><td/><td>83.62 \u00b10.004036</td><td>86.62</td></tr><tr><td>B</td><td/><td/><td/><td/><td/></tr><tr><td/><td>+(Character Type)</td><td/><td>91.68 \u00b10.003013</td><td/><td>93.58 \u00b10.002673</td><td>92.62</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td/><td>92.93 \u00b10.002795</td><td/><td>91.19 \u00b10.003091</td><td>92.05</td></tr><tr><td/><td>Original 6-tag</td><td/><td>89.61 \u00b10.003466</td><td/><td>90.64 \u00b10.003309</td><td>90.12</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td/><td>90.87 \u00b10.003272</td><td/><td>89.77 \u00b10.003443</td><td>90.32</td></tr><tr><td>C</td><td>+(Character Type)</td><td/><td>91.11 \u00b10.003233</td><td/><td>92.02 \u00b10.003078</td><td>91.56</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td/><td>91.54 \u00b10.003161</td><td/><td>91.29 \u00b10.003203</td><td>91.42</td></tr><tr><td/><td>Original 6-tag</td><td/><td>89.82 \u00b10.003367</td><td/><td>91.24 \u00b10.003148</td><td>90.52</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td/><td>93.48 \u00b10.002749</td><td/><td>91.06 \u00b10.003176</td><td>92.25</td></tr><tr><td>D</td><td>+(Character Type)</td><td/><td>92.35 \u00b10.002960</td><td/><td>93.99 \u00b10.002646</td><td>93.16</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF5": {
                "text": "",
                "content": "<table><tr><td colspan=\"2\">Domain Feature</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td/><td>Original 6-tag</td><td colspan=\"2\">55.52 \u00b10.019647</td><td colspan=\"2\">52.00 \u00b10.019752</td><td>53.71</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td colspan=\"2\">53.71 \u00b10.019714</td><td colspan=\"2\">52.34 \u00b10.019746</td><td>53.01</td></tr><tr><td>A</td><td>+(Character Type)</td><td colspan=\"2\">62.42 \u00b10.019149</td><td colspan=\"2\">58.86 \u00b10.019455</td><td>60.59</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td colspan=\"2\">61.77 \u00b10.019212</td><td colspan=\"2\">59.24 \u00b10.019427</td><td>60.48</td></tr><tr><td/><td>Original 6-tag</td><td colspan=\"2\">36.06 \u00b10.014105</td><td colspan=\"2\">20.49 \u00b10.011855</td><td>26.13</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td colspan=\"2\">41.38 \u00b10.014467</td><td colspan=\"2\">52.17 \u00b10.014673</td><td>46.16</td></tr><tr><td>B</td><td>+(Character Type)</td><td colspan=\"2\">76.27 \u00b10.012496</td><td colspan=\"2\">71.40 \u00b10.013274</td><td>73.76</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td colspan=\"2\">67.49 \u00b10.013759</td><td colspan=\"2\">76.28 \u00b10.012495</td><td>71.62</td></tr><tr><td/><td>Original 6-tag</td><td colspan=\"2\">59.69 \u00b10.016736</td><td colspan=\"2\">49.40 \u00b10.017059</td><td>54.06</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td colspan=\"2\">58.80 \u00b10.016793</td><td colspan=\"2\">54.76 \u00b10.016982</td><td>56.71</td></tr><tr><td>C</td><td>+(Character Type)</td><td colspan=\"2\">68.14 \u00b10.015898</td><td colspan=\"2\">59.69 \u00b10.016736</td><td>63.64</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td colspan=\"2\">66.03 \u00b10.016159</td><td colspan=\"2\">60.54 \u00b10.016677</td><td>63.17</td></tr><tr><td/><td>Original 6-tag</td><td colspan=\"2\">48.79 \u00b10.018869</td><td colspan=\"2\">35.90 \u00b10.018109</td><td>41.36</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td colspan=\"2\">53.98 \u00b10.018815</td><td colspan=\"2\">55.56 \u00b10.018757</td><td>54.76</td></tr><tr><td>D</td><td>+(Character Type)</td><td colspan=\"2\">68.81 \u00b10.017487</td><td colspan=\"2\">57.73 \u00b10.018648</td><td>62.79</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td colspan=\"2\">68.64 \u00b10.017514</td><td colspan=\"2\">66.30 \u00b10.017844</td><td>67.45</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF6": {
                "text": "",
                "content": "<table><tr><td colspan=\"2\">Domain Feature</td><td>R OOV C Roov</td><td>P OOV C Poov</td><td>F OOV</td></tr><tr><td/><td>Original 6-tag</td><td>72.50 \u00b10.015297</td><td>57.20 \u00b10.016951</td><td>63.95</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td>71.62 \u00b10.015446</td><td>57.04 \u00b10.016959</td><td>63.50</td></tr><tr><td>A</td><td>+(Character Type)</td><td>75.45 \u00b10.014745</td><td>67.72 \u00b10.016017</td><td>71.38</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td>75.60 \u00b10.014715</td><td>68.44 \u00b10.015923</td><td>71.84</td></tr><tr><td/><td>Original 6-tag</td><td>76.46 \u00b10.014455</td><td>71.38 \u00b10.015399</td><td>73.83</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td>68.49 \u00b10.015828</td><td>65.20 \u00b10.016229</td><td>66.80</td></tr><tr><td>B</td><td/><td/><td/><td/></tr><tr><td/><td>+(Character Type)</td><td>80.44 \u00b10.013514</td><td>81.81 \u00b10.013143</td><td>81.12</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td>74.07 \u00b10.014931</td><td>76.40 \u00b10.014466</td><td>75.22</td></tr><tr><td/><td>Original 6-tag</td><td>73.48 \u00b10.015336</td><td>58.33 \u00b10.017128</td><td>65.03</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td>69.69 \u00b10.015968</td><td>56.31 \u00b10.017232</td><td>62.29</td></tr><tr><td>C</td><td/><td/><td/><td/></tr><tr><td/><td>+(Character Type)</td><td>76.91 \u00b10.014641</td><td>68.87 \u00b10.016087</td><td>72.67</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td>75.97 \u00b10.014843</td><td>68.18 \u00b10.016181</td><td>71.87</td></tr><tr><td/><td>Original 6-tag</td><td>78.54 \u00b10.013963</td><td>66.01 \u00b10.016110</td><td>71.73</td></tr><tr><td/><td>+(Non-Chinese Pattern)</td><td>75.53 \u00b10.014622</td><td>63.69 \u00b10.016355</td><td>69.11</td></tr><tr><td>D</td><td/><td/><td/><td/></tr><tr><td/><td>+(Character Type)</td><td>81.58 \u00b10.013184</td><td>76.99 \u00b10.014315</td><td>79.22</td></tr><tr><td/><td>+(Non-Chinese Pattern, Character Type)</td><td>80.64 \u00b10.013438</td><td>76.22 \u00b10.014481</td><td>78.37</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF7": {
                "text": "Enhancement of Feature Engineering for Conditional Random 59 Field Learning in Chinese Word Segmentation Using Unlabeled Data",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>94.50</td><td>\u00b10.001308</td><td>95.74</td><td>\u00b10.001159</td><td>95.12</td></tr><tr><td>CNG</td><td>95.12</td><td>\u00b10.001236</td><td>95.53</td><td>\u00b10.001186</td><td>95.32</td></tr><tr><td>AVS</td><td>95.14</td><td>\u00b10.001234</td><td>95.86</td><td>\u00b10.001143</td><td>95.50</td></tr><tr><td>TCB</td><td>94.48</td><td>\u00b10.001311</td><td>95.73</td><td>\u00b10.001160</td><td>95.10</td></tr><tr><td>TCF</td><td>94.86</td><td>\u00b10.001267</td><td>95.92</td><td>\u00b10.001135</td><td>95.39</td></tr><tr><td>AVS+TCB</td><td>95.21</td><td>\u00b10.001226</td><td>95.96</td><td>\u00b10.001130</td><td>95.58</td></tr><tr><td>AVS+TCF</td><td>95.27</td><td>\u00b10.001218</td><td>96.02</td><td>\u00b10.001121</td><td>95.65</td></tr><tr><td>LRAVS</td><td>94.88</td><td>\u00b10.001265</td><td>95.91</td><td>\u00b10.001136</td><td>95.39</td></tr><tr><td>LRAVS+TCB</td><td>95.03</td><td>\u00b10.001247</td><td>96.02</td><td>\u00b10.001122</td><td>95.52</td></tr><tr><td>LRAVS+TCF</td><td>95.00</td><td>\u00b10.001251</td><td>96.01</td><td>\u00b10.001124</td><td>95.50</td></tr><tr><td>2005 Best</td><td>95.10</td><td>\u00b10.001230</td><td>95.20</td><td>\u00b10.001220</td><td>95.20</td></tr><tr><td>2005 Baseline</td><td>85.70</td><td>\u00b10.002000</td><td>90.90</td><td>\u00b10.001643</td><td>88.20</td></tr><tr><td>Our Baseline</td><td>86.40</td><td>\u00b10.001967</td><td>91.15</td><td>\u00b10.001629</td><td>88.71</td></tr><tr><td>2005 Topline</td><td>98.50</td><td>\u00b10.000694</td><td>97.90</td><td>\u00b10.000819</td><td>98.20</td></tr><tr><td>Our Topline</td><td>98.64</td><td>\u00b10.000665</td><td>97.97</td><td>\u00b10.000809</td><td>98.30</td></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>66.09</td><td>\u00b10.012356</td><td>61.85</td><td>\u00b10.012678</td><td>63.90</td></tr><tr><td>CNG</td><td>67.39</td><td>\u00b10.012235</td><td>66.81</td><td>\u00b10.01229</td><td>67.10</td></tr><tr><td>AVS</td><td>68.93</td><td>\u00b10.012078</td><td>70.73</td><td>\u00b10.011875</td><td>69.82</td></tr><tr><td>TCB</td><td>66.16</td><td>\u00b10.012349</td><td>64.02</td><td>\u00b10.012668</td><td>64.02</td></tr><tr><td>TCF</td><td>70.27</td><td>\u00b10.011929</td><td>63.89</td><td>\u00b10.012536</td><td>66.93</td></tr><tr><td>AVS+TCB</td><td>69.31</td><td>\u00b10.012037</td><td>71.49</td><td>\u00b10.011783</td><td>70.38</td></tr><tr><td>AVS+TCF</td><td>69.59</td><td>\u00b10.012006</td><td>70.94</td><td>\u00b10.011850</td><td>70.26</td></tr><tr><td>LRAVS</td><td>66.31</td><td>\u00b10.012336</td><td>67.07</td><td>\u00b10.012266</td><td>66.69</td></tr><tr><td>LRAVS+TCB</td><td>67.33</td><td>\u00b10.012241</td><td>67.91</td><td>\u00b10.012184</td><td>67.62</td></tr><tr><td>LRAVS+TCF</td><td>69.82</td><td>\u00b10.011981</td><td>66.15</td><td>\u00b10.012350</td><td>67.94</td></tr><tr><td>2005 Best</td><td>69.60</td><td>\u00b10.012005</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2005 Baseline</td><td>0.40</td><td>\u00b10.001647</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>1.41</td><td>\u00b10.003080</td><td>3.08</td><td>\u00b10.004512</td><td>1.94</td></tr><tr><td>2005 Topline</td><td>99.60</td><td>\u00b10.001647</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.59</td><td>\u00b10.001677</td><td>95.48</td><td>\u00b10.005420</td><td>97.49</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF8": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>94.82</td><td>\u00b10.002207</td><td>94.64</td><td>\u00b10.002245</td><td>94.73</td></tr><tr><td>CNG</td><td>95.55</td><td>\u00b10.002055</td><td>94.39</td><td>\u00b10.002292</td><td>94.97</td></tr><tr><td>AVS</td><td>95.27</td><td>\u00b10.002115</td><td>94.93</td><td>\u00b10.002185</td><td>95.10</td></tr><tr><td>TCB</td><td>95.21</td><td>\u00b10.002129</td><td>94.93</td><td>\u00b10.002186</td><td>95.07</td></tr><tr><td>TCF</td><td>95.30</td><td>\u00b10.002107</td><td>94.96</td><td>\u00b10.002180</td><td>95.13</td></tr><tr><td>AVS+TCB</td><td>95.34</td><td>\u00b10.002100</td><td>95.13</td><td>\u00b10.002145</td><td>95.23</td></tr><tr><td>AVS+TCF</td><td>95.39</td><td>\u00b10.002088</td><td>95.15</td><td>\u00b10.002140</td><td>95.27</td></tr><tr><td>LRAVS</td><td>95.35</td><td>\u00b10.002099</td><td>95.08</td><td>\u00b10.002155</td><td>95.21</td></tr><tr><td>LRAVS+TCB</td><td>95.45</td><td>\u00b10.002077</td><td>95.21</td><td>\u00b10.002127</td><td>95.33</td></tr><tr><td>LRAVS+TCF</td><td>95.41</td><td>\u00b10.002085</td><td>95.20</td><td>\u00b10.002130</td><td>95.30</td></tr><tr><td>2005 Best</td><td>94.60</td><td>\u00b10.002230</td><td>94.10</td><td>\u00b10.002330</td><td>94.30</td></tr><tr><td>2005 Baseline</td><td>79.00</td><td>\u00b10.004026</td><td>88.20</td><td>\u00b10.003189</td><td>83.30</td></tr><tr><td>Our Baseline</td><td>83.84</td><td>\u00b10.003667</td><td>90.81</td><td>\u00b10.002877</td><td>87.19</td></tr><tr><td>2005 Topline</td><td>99.10</td><td>\u00b10.000934</td><td>98.80</td><td>\u00b10.001076</td><td>98.20</td></tr><tr><td>Our Topline</td><td>99.24</td><td>\u00b10.000867</td><td>98.90</td><td>\u00b10.001040</td><td>99.07</td></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>69.15</td><td>\u00b10.016141</td><td>65.54</td><td>\u00b10.016609</td><td>67.30</td></tr><tr><td>CNG</td><td>69.68</td><td>\u00b10.016063</td><td>69.41</td><td>\u00b10.016104</td><td>69.55</td></tr><tr><td>AVS</td><td>70.48</td><td>\u00b10.015942</td><td>71.90</td><td>\u00b10.015709</td><td>71.18</td></tr><tr><td>TCB</td><td>71.83</td><td>\u00b10.015721</td><td>70.12</td><td>\u00b10.016236</td><td>70.12</td></tr><tr><td>TCF</td><td>72.39</td><td>\u00b10.015624</td><td>68.76</td><td>\u00b10.016198</td><td>70.53</td></tr><tr><td>AVS+TCB</td><td>71.14</td><td>\u00b10.015836</td><td>72.70</td><td>\u00b10.01557</td><td>71.91</td></tr><tr><td>AVS+TCF</td><td>70.97</td><td>\u00b10.015863</td><td>72.77</td><td>\u00b10.015556</td><td>71.86</td></tr><tr><td>LRAVS</td><td>69.78</td><td>\u00b10.016048</td><td>72.09</td><td>\u00b10.015676</td><td>70.92</td></tr><tr><td>LRAVS+TCB</td><td>70.57</td><td>\u00b10.015926</td><td>73.06</td><td>\u00b10.015505</td><td>71.80</td></tr><tr><td>LRAVS+TCF</td><td>71.17</td><td>\u00b10.015831</td><td>73.22</td><td>\u00b10.015475</td><td>72.18</td></tr><tr><td>2005 Best</td><td>69.80</td><td>\u00b10.016046</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2005 Baseline</td><td>0.00</td><td>\u00b10.000000</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>16.22</td><td>\u00b10.012882</td><td>33.91</td><td>\u00b10.016544</td><td>21.94</td></tr><tr><td>2005 Topline</td><td>99.70</td><td>\u00b10.001911</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.74</td><td>\u00b10.001794</td><td>98.82</td><td>\u00b10.003771</td><td>99.28</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF9": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>93.73</td><td>\u00b10.001512</td><td>92.70</td><td>\u00b10.001623</td><td>93.21</td></tr><tr><td>CNG</td><td>94.36</td><td>\u00b10.001438</td><td>93.57</td><td>\u00b10.001530</td><td>93.96</td></tr><tr><td>AVS</td><td>94.21</td><td>\u00b10.001457</td><td>93.24</td><td>\u00b10.001566</td><td>93.72</td></tr><tr><td>TCB</td><td>93.97</td><td>\u00b10.001485</td><td>92.76</td><td>\u00b10.001616</td><td>93.36</td></tr><tr><td>TCF</td><td>93.94</td><td>\u00b10.001488</td><td>92.81</td><td>\u00b10.001611</td><td>93.37</td></tr><tr><td>AVS+TCB</td><td>94.33</td><td>\u00b10.001443</td><td>93.31</td><td>\u00b10.001559</td><td>93.81</td></tr><tr><td>AVS+TCF</td><td>94.25</td><td>\u00b10.001451</td><td>93.44</td><td>\u00b10.001544</td><td>93.85</td></tr><tr><td>LRAVS</td><td>94.34</td><td>\u00b10.001441</td><td>93.48</td><td>\u00b10.001540</td><td>93.91</td></tr><tr><td>LRAVS+TCB</td><td>94.32</td><td>\u00b10.001443</td><td>93.44</td><td>\u00b10.001544</td><td>93.88</td></tr><tr><td>LRAVS+TCF</td><td>93.91</td><td>\u00b10.001492</td><td>92.20</td><td>\u00b10.001672</td><td>93.05</td></tr><tr><td>2005 Best</td><td>94.60</td><td>\u00b10.001400</td><td>95.30</td><td>\u00b10.001310</td><td>95.00</td></tr><tr><td>2005 Baseline</td><td>83.60</td><td>\u00b10.002292</td><td>90.40</td><td>\u00b10.001824</td><td>86.90</td></tr><tr><td>Our Baseline</td><td>84.29</td><td>\u00b10.002269</td><td>90.68</td><td>\u00b10.001813</td><td>87.37</td></tr><tr><td>2005 Topline</td><td>98.80</td><td>\u00b10.000674</td><td>98.50</td><td>\u00b10.000752</td><td>98.70</td></tr><tr><td>Our Topline</td><td>98.96</td><td>\u00b10.000634</td><td>98.62</td><td>\u00b10.000726</td><td>98.79</td></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>57.48</td><td>\u00b10.012083</td><td>48.04</td><td>\u00b10.012211</td><td>52.33</td></tr><tr><td>CNG</td><td>65.58</td><td>\u00b10.011612</td><td>57.87</td><td>\u00b10.012068</td><td>61.48</td></tr><tr><td>AVS</td><td>62.69</td><td>\u00b10.011821</td><td>55.60</td><td>\u00b10.012144</td><td>58.93</td></tr><tr><td>TCB</td><td>60.07</td><td>\u00b10.011970</td><td>54.87</td><td>\u00b10.012220</td><td>54.87</td></tr><tr><td>TCF</td><td>60.39</td><td>\u00b10.011954</td><td>50.41</td><td>\u00b10.012220</td><td>54.95</td></tr><tr><td>AVS+TCB</td><td>64.02</td><td>\u00b10.011730</td><td>56.97</td><td>\u00b10.012101</td><td>60.29</td></tr><tr><td>AVS+TCF</td><td>63.80</td><td>\u00b10.011746</td><td>56.06</td><td>\u00b10.012130</td><td>59.68</td></tr><tr><td>LRAVS</td><td>65.02</td><td>\u00b10.011656</td><td>57.31</td><td>\u00b10.012089</td><td>60.92</td></tr><tr><td>LRAVS+TCB</td><td>65.42</td><td>\u00b10.011625</td><td>57.60</td><td>\u00b10.012079</td><td>61.26</td></tr><tr><td>LRAVS+TCF</td><td>60.42</td><td>\u00b10.011952</td><td>48.92</td><td>\u00b10.012218</td><td>54.07</td></tr><tr><td>2005 Best</td><td>63.60</td><td>\u00b10.011760</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2005 Baseline</td><td>5.90</td><td>\u00b10.005759</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>6.86</td><td>\u00b10.006178</td><td>6.10</td><td>\u00b10.005850</td><td>6.46</td></tr><tr><td>2005 Topline</td><td>99.40</td><td>\u00b10.001888</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.37</td><td>\u00b10.001938</td><td>97.72</td><td>\u00b10.003645</td><td>98.54</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF10": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>90.95</td><td>\u00b10.003129</td><td>92.46</td><td>\u00b10.002880</td><td>91.70</td></tr><tr><td>CNG</td><td>91.45</td><td>\u00b10.003050</td><td>92.36</td><td>\u00b10.002898</td><td>91.90</td></tr><tr><td>AVS</td><td>91.25</td><td>\u00b10.003081</td><td>92.72</td><td>\u00b10.002833</td><td>91.98</td></tr><tr><td>TCB</td><td>91.21</td><td>\u00b10.003087</td><td>92.53</td><td>\u00b10.002867</td><td>91.87</td></tr><tr><td>TCF</td><td>90.86</td><td>\u00b10.003143</td><td>92.62</td><td>\u00b10.002852</td><td>91.73</td></tr><tr><td>AVS+TCB</td><td>91.60</td><td>\u00b10.003026</td><td>92.67</td><td>\u00b10.002842</td><td>92.13</td></tr><tr><td>AVS+TCF</td><td>90.81</td><td>\u00b10.003151</td><td>92.16</td><td>\u00b10.002932</td><td>91.48</td></tr><tr><td>LRAVS</td><td>91.71</td><td>\u00b10.003007</td><td>92.61</td><td>\u00b10.002854</td><td>92.16</td></tr><tr><td>LRAVS+TCB</td><td>91.97</td><td>\u00b10.002963</td><td>92.76</td><td>\u00b10.002826</td><td>92.37</td></tr><tr><td>LRAVS+TCF</td><td>91.28</td><td>\u00b10.003077</td><td>92.60</td><td>\u00b10.002856</td><td>91.93</td></tr><tr><td>2010 Best</td><td>95.00</td><td>\u00b10.002320</td><td>95.30</td><td>\u00b10.002250</td><td>95.10</td></tr><tr><td>2010 Baseline</td><td>63.20</td><td>\u00b10.005132</td><td>85.60</td><td>\u00b10.003736</td><td>72.70</td></tr><tr><td>Our Baseline</td><td>63.26</td><td>\u00b10.005258</td><td>85.68</td><td>\u00b10.003820</td><td>72.78</td></tr><tr><td>2010 Topline</td><td>99.30</td><td>\u00b10.000887</td><td>99.10</td><td>\u00b10.001005</td><td>99.20</td></tr><tr><td>Our Topline</td><td>99.25</td><td>\u00b10.000940</td><td>99.06</td><td>\u00b10.001052</td><td>99.16</td></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>70.62</td><td>\u00b10.013380</td><td>67.66</td><td>\u00b10.013740</td><td>69.11</td></tr><tr><td>CNG</td><td>70.38</td><td>\u00b10.013412</td><td>65.17</td><td>\u00b10.013994</td><td>67.67</td></tr><tr><td>AVS</td><td>69.85</td><td>\u00b10.013479</td><td>66.16</td><td>\u00b10.013898</td><td>67.96</td></tr><tr><td>TCB</td><td>71.23</td><td>\u00b10.013297</td><td>69.66</td><td>\u00b10.013684</td><td>69.66</td></tr><tr><td>TCF</td><td>72.01</td><td>\u00b10.013187</td><td>66.02</td><td>\u00b10.013913</td><td>68.89</td></tr><tr><td>AVS+TCB</td><td>70.25</td><td>\u00b10.013429</td><td>67.22</td><td>\u00b10.013788</td><td>68.70</td></tr><tr><td>AVS+TCF</td><td>69.63</td><td>\u00b10.013507</td><td>63.73</td><td>\u00b10.014123</td><td>66.55</td></tr><tr><td>LRAVS</td><td>71.25</td><td>\u00b10.013294</td><td>68.25</td><td>\u00b10.013673</td><td>69.72</td></tr><tr><td>LRAVS+TCB</td><td>71.81</td><td>\u00b10.013216</td><td>69.47</td><td>\u00b10.013528</td><td>70.62</td></tr><tr><td>LRAVS+TCF</td><td>70.92</td><td>\u00b10.013340</td><td>66.13</td><td>\u00b10.013902</td><td>68.44</td></tr><tr><td>2010 Best</td><td>82.70</td><td>\u00b10.011111</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2010 Baseline</td><td>16.30</td><td>\u00b10.010850</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>16.65</td><td>\u00b10.010944</td><td>6.39</td><td>\u00b10.007185</td><td>9.24</td></tr><tr><td>2010 Topline</td><td>99.00</td><td>\u00b10.002923</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.00</td><td>\u00b10.002930</td><td>98.08</td><td>\u00b10.004028</td><td>98.54</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF11": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>93.01</td><td>\u00b10.002838</td><td>93.74</td><td>\u00b10.002697</td><td>93.38</td></tr><tr><td>CNG</td><td>94.40</td><td>\u00b10.002561</td><td>93.66</td><td>\u00b10.002714</td><td>94.02</td></tr><tr><td>AVS</td><td>93.54</td><td>\u00b10.002736</td><td>94.30</td><td>\u00b10.002581</td><td>93.92</td></tr><tr><td>TCB</td><td>93.35</td><td>\u00b10.002774</td><td>94.14</td><td>\u00b10.002614</td><td>93.74</td></tr><tr><td>TCF</td><td>93.10</td><td>\u00b10.002822</td><td>93.88</td><td>\u00b10.002669</td><td>93.49</td></tr><tr><td>AVS+TCB</td><td>94.56</td><td>\u00b10.002526</td><td>94.49</td><td>\u00b10.002540</td><td>94.53</td></tr><tr><td>AVS+TCF</td><td>94.05</td><td>\u00b10.002633</td><td>94.10</td><td>\u00b10.002624</td><td>94.08</td></tr><tr><td>LRAVS</td><td>94.30</td><td>\u00b10.002582</td><td>94.13</td><td>\u00b10.002616</td><td>94.21</td></tr><tr><td>LRAVS+TCB</td><td>94.36</td><td>\u00b10.002568</td><td>94.16</td><td>\u00b10.002611</td><td>94.26</td></tr><tr><td>LRAVS+TCF</td><td>94.36</td><td>\u00b10.002569</td><td>94.19</td><td>\u00b10.002604</td><td>94.28</td></tr><tr><td>2010 Best</td><td>96.00</td><td>\u00b10.002160</td><td>95.90</td><td>\u00b10.002180</td><td>95.90</td></tr><tr><td>2010 Baseline</td><td>80.30</td><td>\u00b10.004377</td><td>91.40</td><td>\u00b10.003085</td><td>85.50</td></tr><tr><td>Our Baseline</td><td>80.26</td><td>\u00b10.004431</td><td>91.41</td><td>\u00b10.003119</td><td>85.48</td></tr><tr><td>2010 Topline</td><td>99.50</td><td>\u00b10.000776</td><td>99.40</td><td>\u00b10.000850</td><td>99.40</td></tr><tr><td>Our Topline</td><td>99.56</td><td>\u00b10.000734</td><td>99.47</td><td>\u00b10.000810</td><td>99.52</td></tr><tr><td colspan=\"6\">Table 24. Non-Chinese treatment performance comparison of OOV on SIGHAN 2010</td></tr><tr><td colspan=\"4\">simplified Chinese domain D (Finance) corpus.</td><td/><td/></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>67.60</td><td>\u00b10.017666</td><td>61.28</td><td>\u00b10.018388</td><td>64.28</td></tr><tr><td>CNG</td><td>73.53</td><td>\u00b10.016655</td><td>67.77</td><td>\u00b10.017642</td><td>70.53</td></tr><tr><td>AVS</td><td>71.10</td><td>\u00b10.017111</td><td>64.17</td><td>\u00b10.018101</td><td>67.46</td></tr><tr><td>TCB</td><td>70.58</td><td>\u00b10.017201</td><td>66.44</td><td>\u00b10.018250</td><td>66.44</td></tr><tr><td>TCF</td><td>70.13</td><td>\u00b10.017277</td><td>61.19</td><td>\u00b10.018396</td><td>65.35</td></tr><tr><td>AVS+TCB</td><td>73.80</td><td>\u00b10.016598</td><td>70.79</td><td>\u00b10.017166</td><td>72.26</td></tr><tr><td>AVS+TCF</td><td>70.76</td><td>\u00b10.017172</td><td>67.73</td><td>\u00b10.017648</td><td>69.21</td></tr><tr><td>LRAVS</td><td>71.66</td><td>\u00b10.017012</td><td>68.54</td><td>\u00b10.017528</td><td>70.07</td></tr><tr><td>LRAVS+TCB</td><td>72.63</td><td>\u00b10.016831</td><td>69.82</td><td>\u00b10.017328</td><td>71.20</td></tr><tr><td>LRAVS+TCF</td><td>72.38</td><td>\u00b10.016878</td><td>69.40</td><td>\u00b10.017396</td><td>70.86</td></tr><tr><td>2010 Best</td><td>82.70</td><td>\u00b10.014279</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2010 Baseline</td><td>23.30</td><td>\u00b10.015958</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>23.32</td><td>\u00b10.015963</td><td>14.15</td><td>\u00b10.013157</td><td>17.61</td></tr><tr><td>2010 Topline</td><td>99.50</td><td>\u00b10.002663</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.72</td><td>\u00b10.001985</td><td>99.34</td><td>\u00b10.003047</td><td>99.53</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF12": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>95.15</td><td>\u00b10.002122</td><td>93.20</td><td>\u00b10.002487</td><td>94.17</td></tr><tr><td>CNG</td><td>95.60</td><td>\u00b10.002027</td><td>93.16</td><td>\u00b10.002494</td><td>94.36</td></tr><tr><td>AVS</td><td>95.67</td><td>\u00b10.002012</td><td>93.83</td><td>\u00b10.002378</td><td>94.74</td></tr><tr><td>TCB</td><td>95.21</td><td>\u00b10.002111</td><td>93.25</td><td>\u00b10.002480</td><td>94.22</td></tr><tr><td>TCF</td><td>95.28</td><td>\u00b10.002095</td><td>93.42</td><td>\u00b10.002450</td><td>94.34</td></tr><tr><td>AVS+TCB</td><td>95.62</td><td>\u00b10.002023</td><td>93.72</td><td>\u00b10.002398</td><td>94.66</td></tr><tr><td>AVS+TCF</td><td>95.74</td><td>\u00b10.001996</td><td>93.83</td><td>\u00b10.002378</td><td>94.77</td></tr><tr><td>LRAVS</td><td>95.57</td><td>\u00b10.002034</td><td>93.79</td><td>\u00b10.002384</td><td>94.67</td></tr><tr><td>LRAVS+TCB</td><td>95.63</td><td>\u00b10.002020</td><td>93.85</td><td>\u00b10.002373</td><td>94.73</td></tr><tr><td>LRAVS+TCF</td><td>95.55</td><td>\u00b10.002038</td><td>93.81</td><td>\u00b10.002381</td><td>94.67</td></tr><tr><td>2010 Best</td><td>95.70</td><td>\u00b10.001950</td><td>94.80</td><td>\u00b10.002130</td><td>95.20</td></tr><tr><td>2010 Baseline</td><td>70.10</td><td>\u00b10.004390</td><td>87.30</td><td>\u00b10.003193</td><td>77.80</td></tr><tr><td>Our Baseline</td><td>70.15</td><td>\u00b10.004522</td><td>87.33</td><td>\u00b10.003286</td><td>77.80</td></tr><tr><td>2010 Topline</td><td>99.10</td><td>\u00b10.000906</td><td>98.80</td><td>\u00b10.001044</td><td>99.00</td></tr><tr><td>Our Topline</td><td>99.38</td><td>\u00b10.000778</td><td>98.85</td><td>\u00b10.001055</td><td>99.11</td></tr><tr><td colspan=\"6\">Table 28. Non-Chinese-Pattern performance comparison of OOV on SIGHAN 2010</td></tr><tr><td colspan=\"4\">traditional Chinese domain B (Computer) corpus.</td><td/><td/></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>58.79</td><td>\u00b10.016769</td><td>68.17</td><td>\u00b10.015871</td><td>63.14</td></tr><tr><td>CNG</td><td>61.77</td><td>\u00b10.016556</td><td>70.16</td><td>\u00b10.015589</td><td>65.70</td></tr><tr><td>AVS</td><td>60.59</td><td>\u00b10.016649</td><td>72.29</td><td>\u00b10.015248</td><td>65.93</td></tr><tr><td>TCB</td><td>59.09</td><td>\u00b10.016751</td><td>68.81</td><td>\u00b10.015784</td><td>63.58</td></tr><tr><td>TCF</td><td>59.34</td><td>\u00b10.016735</td><td>69.21</td><td>\u00b10.015727</td><td>63.89</td></tr><tr><td>AVS+TCB</td><td>60.89</td><td>\u00b10.016626</td><td>72.24</td><td>\u00b10.015257</td><td>66.08</td></tr><tr><td>AVS+TCF</td><td>61.35</td><td>\u00b10.01659</td><td>72.90</td><td>\u00b10.015143</td><td>66.63</td></tr><tr><td>LRAVS</td><td>61.67</td><td>\u00b10.016564</td><td>72.84</td><td>\u00b10.015155</td><td>66.79</td></tr><tr><td>LRAVS+TCB</td><td>61.82</td><td>\u00b10.016552</td><td>73.07</td><td>\u00b10.015113</td><td>66.98</td></tr><tr><td>LRAVS+TCF</td><td>61.55</td><td>\u00b10.016574</td><td>72.94</td><td>\u00b10.015135</td><td>66.76</td></tr><tr><td>2010 Best</td><td>66.60</td><td>\u00b10.016069</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2010 Baseline</td><td>1.00</td><td>\u00b10.003390</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>1.03</td><td>\u00b10.003445</td><td>0.55</td><td>\u00b10.002515</td><td>0.72</td></tr><tr><td>2010 Topline</td><td>99.60</td><td>\u00b10.002150</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.34</td><td>\u00b10.002765</td><td>99.41</td><td>\u00b10.002609</td><td>99.37</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF13": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>95.52</td><td>\u00b10.001925</td><td>95.46</td><td>\u00b10.001937</td><td>95.49</td></tr><tr><td>CNG</td><td>96.13</td><td>\u00b10.001794</td><td>95.04</td><td>\u00b10.002020</td><td>95.58</td></tr><tr><td>AVS</td><td>95.99</td><td>\u00b10.001825</td><td>95.79</td><td>\u00b10.001868</td><td>95.89</td></tr><tr><td>TCB</td><td>95.55</td><td>\u00b10.001918</td><td>95.51</td><td>\u00b10.001927</td><td>95.53</td></tr><tr><td>TCF</td><td>95.61</td><td>\u00b10.001907</td><td>95.57</td><td>\u00b10.001915</td><td>95.59</td></tr><tr><td>AVS+TCB</td><td>95.93</td><td>\u00b10.001839</td><td>95.77</td><td>\u00b10.001874</td><td>95.85</td></tr><tr><td>AVS+TCF</td><td>95.99</td><td>\u00b10.001825</td><td>95.88</td><td>\u00b10.001850</td><td>95.93</td></tr><tr><td>LRAVS</td><td>96.02</td><td>\u00b10.001820</td><td>95.73</td><td>\u00b10.001881</td><td>95.87</td></tr><tr><td>LRAVS+TCB</td><td>96.04</td><td>\u00b10.001814</td><td>95.82</td><td>\u00b10.001862</td><td>95.93</td></tr><tr><td>LRAVS+TCF</td><td>95.94</td><td>\u00b10.001836</td><td>95.71</td><td>\u00b10.001885</td><td>95.83</td></tr><tr><td>2010 Best</td><td>96.20</td><td>\u00b10.001760</td><td>96.40</td><td>\u00b10.001720</td><td>96.30</td></tr><tr><td>2010 Baseline</td><td>82.60</td><td>\u00b10.003492</td><td>88.80</td><td>\u00b10.002905</td><td>85.50</td></tr><tr><td>Our Baseline</td><td>82.56</td><td>\u00b10.003531</td><td>88.77</td><td>\u00b10.002937</td><td>85.55</td></tr><tr><td>2010 Topline</td><td>98.60</td><td>\u00b10.001082</td><td>98.10</td><td>\u00b10.001258</td><td>98.40</td></tr><tr><td>Our Topline</td><td>98.63</td><td>\u00b10.001081</td><td>98.10</td><td>\u00b10.00127</td><td>98.36</td></tr><tr><td colspan=\"6\">Table 32. Non-Chinese treatment performance comparison of OOV on SIGHAN 2010</td></tr><tr><td colspan=\"4\">traditional Chinese domain D (Finance) corpus.</td><td/><td/></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>80.45</td><td>\u00b10.013488</td><td>76.61</td><td>\u00b10.014398</td><td>78.48</td></tr><tr><td>CNG</td><td>82.96</td><td>\u00b10.012787</td><td>78.16</td><td>\u00b10.014053</td><td>80.49</td></tr><tr><td>AVS</td><td>81.33</td><td>\u00b10.013253</td><td>81.28</td><td>\u00b10.013267</td><td>81.30</td></tr><tr><td>TCB</td><td>80.99</td><td>\u00b10.013346</td><td>77.44</td><td>\u00b10.014216</td><td>79.17</td></tr><tr><td>TCF</td><td>80.92</td><td>\u00b10.013363</td><td>77.26</td><td>\u00b10.014255</td><td>79.05</td></tr><tr><td>AVS+TCB</td><td>80.99</td><td>\u00b10.013346</td><td>81.55</td><td>\u00b10.013193</td><td>81.27</td></tr><tr><td>AVS+TCF</td><td>80.99</td><td>\u00b10.013346</td><td>81.96</td><td>\u00b10.013077</td><td>81.47</td></tr><tr><td>LRAVS</td><td>82.62</td><td>\u00b10.012889</td><td>82.10</td><td>\u00b10.013038</td><td>82.36</td></tr><tr><td>LRAVS+TCB</td><td>82.18</td><td>\u00b10.013016</td><td>82.44</td><td>\u00b10.012942</td><td>82.31</td></tr><tr><td>LRAVS+TCF</td><td>81.86</td><td>\u00b10.013105</td><td>82.04</td><td>\u00b10.013054</td><td>81.95</td></tr><tr><td>2010 Best</td><td>81.20</td><td>\u00b10.013288</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2010 Baseline</td><td>0.60</td><td>\u00b10.002627</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>0.60</td><td>\u00b10.002618</td><td>2.28</td><td>\u00b10.005078</td><td>0.95</td></tr><tr><td>2010 Topline</td><td>99.70</td><td>\u00b10.001860</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.69</td><td>\u00b10.001902</td><td>98.54</td><td>\u00b10.004076</td><td>99.11</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF15": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>94.77</td><td>\u00b10.002381</td><td>94.79</td><td>\u00b10.002377</td><td>94.78</td></tr><tr><td>CNG</td><td>95.24</td><td>\u00b10.002278</td><td>95.48</td><td>\u00b10.002222</td><td>95.36</td></tr><tr><td>AVS</td><td>95.13</td><td>\u00b10.002302</td><td>95.20</td><td>\u00b10.002286</td><td>95.17</td></tr><tr><td>TCB</td><td>94.84</td><td>\u00b10.002367</td><td>94.87</td><td>\u00b10.002360</td><td>94.85</td></tr><tr><td>TCF</td><td>94.78</td><td>\u00b10.002380</td><td>94.77</td><td>\u00b10.002382</td><td>94.77</td></tr><tr><td>AVS+TCB</td><td>95.18</td><td>\u00b10.002291</td><td>95.24</td><td>\u00b10.002278</td><td>95.21</td></tr><tr><td>AVS+TCF</td><td>95.08</td><td>\u00b10.002313</td><td>95.19</td><td>\u00b10.002288</td><td>95.14</td></tr><tr><td>LRAVS</td><td>95.00</td><td>\u00b10.002332</td><td>95.21</td><td>\u00b10.002284</td><td>95.10</td></tr><tr><td>LRAVS+TCB</td><td>95.18</td><td>\u00b10.002292</td><td>95.33</td><td>\u00b10.002256</td><td>95.26</td></tr><tr><td>LRAVS+TCF</td><td>95.00</td><td>\u00b10.002330</td><td>95.27</td><td>\u00b10.002271</td><td>95.14</td></tr><tr><td>2003 Best</td><td>93.40</td><td>\u00b10.002700</td><td>94.70</td><td>\u00b10.002400</td><td>94.00</td></tr><tr><td>2003 Baseline</td><td>83.00</td><td>\u00b10.004018</td><td>90.80</td><td>\u00b10.003092</td><td>86.70</td></tr><tr><td>Our Baseline</td><td>82.97</td><td>\u00b10.004021</td><td>90.77</td><td>\u00b10.003097</td><td>86.69</td></tr><tr><td>2003 Topline</td><td>99.10</td><td>\u00b10.001010</td><td>98.60</td><td>\u00b10.001257</td><td>98.90</td></tr><tr><td>Our Topline</td><td>99.10</td><td>\u00b10.001009</td><td>98.62</td><td>\u00b10.001249</td><td>98.86</td></tr><tr><td colspan=\"5\">Table 36. Performance comparison of OOV on SIGHAN 2003 CityU corpus.</td><td/></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>75.80</td><td>\u00b10.017149</td><td>66.07</td><td>\u00b10.018969</td><td>70.60</td></tr><tr><td>CNG</td><td>77.25</td><td>\u00b10.016796</td><td>73.25</td><td>\u00b10.017735</td><td>75.20</td></tr><tr><td>AVS</td><td>75.16</td><td>\u00b10.017311</td><td>71.79</td><td>\u00b10.018030</td><td>73.44</td></tr><tr><td>TCB</td><td>76.20</td><td>\u00b10.017061</td><td>66.63</td><td>\u00b10.018891</td><td>71.10</td></tr><tr><td>TCF</td><td>76.28</td><td>\u00b10.017041</td><td>66.38</td><td>\u00b10.018927</td><td>70.99</td></tr><tr><td>AVS+TCB</td><td>75.44</td><td>\u00b10.017245</td><td>72.06</td><td>\u00b10.017977</td><td>73.71</td></tr><tr><td>AVS+TCF</td><td>74.88</td><td>\u00b10.017376</td><td>71.66</td><td>\u00b10.018055</td><td>73.23</td></tr><tr><td>LRAVS</td><td>74.12</td><td>\u00b10.017548</td><td>72.01</td><td>\u00b10.017987</td><td>73.05</td></tr><tr><td>LRAVS+TCB</td><td>74.88</td><td>\u00b10.017376</td><td>72.92</td><td>\u00b10.017804</td><td>73.89</td></tr><tr><td>LRAVS+TCF</td><td>74.32</td><td>\u00b10.017503</td><td>72.23</td><td>\u00b10.017943</td><td>73.26</td></tr><tr><td>2003 Best</td><td>62.50</td><td>\u00b10.019396</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2003 Baseline</td><td>3.70</td><td>\u00b10.007563</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>3.69</td><td>\u00b10.007555</td><td>5.20</td><td>\u00b10.008896</td><td>4.32</td></tr><tr><td>2003 Topline</td><td>99.60</td><td>\u00b10.002529</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.60</td><td>\u00b10.002533</td><td>98.65</td><td>\u00b10.004626</td><td>99.12</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF16": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>87.30</td><td>\u00b10.003334</td><td>86.83</td><td>\u00b10.003385</td><td>87.06</td></tr><tr><td>CNG</td><td>89.61</td><td>\u00b10.003054</td><td>88.66</td><td>\u00b10.003175</td><td>89.13</td></tr><tr><td>AVS</td><td>89.38</td><td>\u00b10.003085</td><td>88.06</td><td>\u00b10.003246</td><td>88.71</td></tr><tr><td>TCB</td><td>87.46</td><td>\u00b10.003315</td><td>86.86</td><td>\u00b10.003382</td><td>87.16</td></tr><tr><td>TCF</td><td>87.18</td><td>\u00b10.003347</td><td>86.45</td><td>\u00b10.003426</td><td>86.81</td></tr><tr><td>AVS+TCB</td><td>89.31</td><td>\u00b10.003092</td><td>88.08</td><td>\u00b10.003244</td><td>88.69</td></tr><tr><td>AVS+TCF</td><td>89.39</td><td>\u00b10.003082</td><td>88.17</td><td>\u00b10.003233</td><td>88.78</td></tr><tr><td>LRAVS</td><td>89.30</td><td>\u00b10.003094</td><td>88.21</td><td>\u00b10.003228</td><td>88.75</td></tr><tr><td>LRAVS+TCB</td><td>89.37</td><td>\u00b10.003086</td><td>88.09</td><td>\u00b10.003243</td><td>88.72</td></tr><tr><td>LRAVS+TCF</td><td>89.31</td><td>\u00b10.003093</td><td>88.07</td><td>\u00b10.003244</td><td>88.68</td></tr><tr><td>2003 Best</td><td>87.50</td><td>\u00b10.003300</td><td>86.60</td><td>\u00b10.003200</td><td>88.10</td></tr><tr><td>2003 Baseline</td><td>66.30</td><td>\u00b10.004731</td><td>80.00</td><td>\u00b10.004004</td><td>72.50</td></tr><tr><td>Our Baseline</td><td>66.33</td><td>\u00b10.004730</td><td>80.01</td><td>\u00b10.004003</td><td>72.53</td></tr><tr><td>2003 Topline</td><td>98.80</td><td>\u00b10.001090</td><td>98.20</td><td>\u00b10.001331</td><td>98.50</td></tr><tr><td>Our Topline</td><td>98.84</td><td>\u00b10.001072</td><td>98.19</td><td>\u00b10.001333</td><td>98.52</td></tr><tr><td colspan=\"5\">Table 40. Performance comparison of OOV on SIGHAN 2003 CTB corpus.</td><td/></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>69.85</td><td>\u00b10.010805</td><td>62.24</td><td>\u00b10.011415</td><td>65.83</td></tr><tr><td>CNG</td><td>71.79</td><td>\u00b10.010596</td><td>71.31</td><td>\u00b10.010650</td><td>71.55</td></tr><tr><td>AVS</td><td>70.59</td><td>\u00b10.010728</td><td>69.61</td><td>\u00b10.010830</td><td>70.09</td></tr><tr><td>TCB</td><td>70.23</td><td>\u00b10.010766</td><td>62.51</td><td>\u00b10.011398</td><td>66.14</td></tr><tr><td>TCF</td><td>69.49</td><td>\u00b10.010841</td><td>61.91</td><td>\u00b10.011434</td><td>65.48</td></tr><tr><td>AVS+TCB</td><td>70.73</td><td>\u00b10.010714</td><td>70.05</td><td>\u00b10.010785</td><td>70.39</td></tr><tr><td>AVS+TCF</td><td>70.95</td><td>\u00b10.010690</td><td>69.80</td><td>\u00b10.010811</td><td>70.37</td></tr><tr><td>LRAVS</td><td>70.35</td><td>\u00b10.010753</td><td>69.98</td><td>\u00b10.010793</td><td>70.16</td></tr><tr><td>LRAVS+TCB</td><td>70.58</td><td>\u00b10.010730</td><td>70.49</td><td>\u00b10.010739</td><td>70.53</td></tr><tr><td>LRAVS+TCF</td><td>70.24</td><td>\u00b10.010765</td><td>70.05</td><td>\u00b10.010785</td><td>70.15</td></tr><tr><td>2003 Best</td><td>70.50</td><td>\u00b10.010738</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2003 Baseline</td><td>6.20</td><td>\u00b10.005678</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>6.24</td><td>\u00b10.005694</td><td>8.36</td><td>\u00b10.006516</td><td>7.14</td></tr><tr><td>2003 Topline</td><td>99.00</td><td>\u00b10.002343</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.02</td><td>\u00b10.002324</td><td>97.46</td><td>\u00b10.003703</td><td>98.23</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF17": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>96.92</td><td>\u00b10.000736</td><td>96.88</td><td>\u00b10.000741</td><td>96.90</td></tr><tr><td>CNG</td><td>97.26</td><td>\u00b10.000696</td><td>97.21</td><td>\u00b10.000701</td><td>97.23</td></tr><tr><td>AVS</td><td>97.31</td><td>\u00b10.000690</td><td>97.34</td><td>\u00b10.000686</td><td>97.32</td></tr><tr><td>TCB</td><td>96.95</td><td>\u00b10.000733</td><td>96.89</td><td>\u00b10.000740</td><td>96.92</td></tr><tr><td>TCF</td><td>96.96</td><td>\u00b10.000732</td><td>96.90</td><td>\u00b10.000739</td><td>96.93</td></tr><tr><td>AVS+TCB</td><td>97.32</td><td>\u00b10.000689</td><td>97.32</td><td>\u00b10.000689</td><td>97.32</td></tr><tr><td>AVS+TCF</td><td>97.35</td><td>\u00b10.000685</td><td>97.32</td><td>\u00b10.000688</td><td>97.33</td></tr><tr><td>LRAVS</td><td>97.35</td><td>\u00b10.000684</td><td>97.32</td><td>\u00b10.000688</td><td>97.34</td></tr><tr><td>LRAVS+TCB</td><td>97.34</td><td>\u00b10.000686</td><td>97.33</td><td>\u00b10.000687</td><td>97.34</td></tr><tr><td>LRAVS+TCF</td><td>97.23</td><td>\u00b10.000700</td><td>97.26</td><td>\u00b10.000696</td><td>97.24</td></tr><tr><td>2006 Best</td><td>97.20</td><td>\u00b10.000703</td><td>97.30</td><td>\u00b10.000691</td><td>97.20</td></tr><tr><td>2006 Baseline</td><td>88.20</td><td>\u00b10.002134</td><td>93.00</td><td>\u00b10.001687</td><td>90.60</td></tr><tr><td>Our Baseline</td><td>88.22</td><td>\u00b10.001374</td><td>93.06</td><td>\u00b10.001083</td><td>90.57</td></tr><tr><td>2006 Topline</td><td>98.50</td><td>\u00b10.000804</td><td>98.20</td><td>\u00b10.000879</td><td>98.40</td></tr><tr><td>Our Topline</td><td>98.55</td><td>\u00b10.00051</td><td>98.19</td><td>\u00b10.000568</td><td>98.37</td></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>78.35</td><td>\u00b10.008738</td><td>69.60</td><td>\u00b10.009759</td><td>73.72</td></tr><tr><td>CNG</td><td>79.66</td><td>\u00b10.008540</td><td>76.97</td><td>\u00b10.008932</td><td>78.29</td></tr><tr><td>AVS</td><td>79.27</td><td>\u00b10.008600</td><td>78.08</td><td>\u00b10.008777</td><td>78.67</td></tr><tr><td>TCB</td><td>78.55</td><td>\u00b10.008708</td><td>69.97</td><td>\u00b10.009725</td><td>74.01</td></tr><tr><td>TCF</td><td>78.94</td><td>\u00b10.008651</td><td>69.94</td><td>\u00b10.009728</td><td>74.17</td></tr><tr><td>AVS+TCB</td><td>79.31</td><td>\u00b10.008595</td><td>77.93</td><td>\u00b10.008798</td><td>78.61</td></tr><tr><td>AVS+TCF</td><td>79.70</td><td>\u00b10.008533</td><td>78.30</td><td>\u00b10.008745</td><td>78.99</td></tr><tr><td>LRAVS</td><td>79.84</td><td>\u00b10.008512</td><td>78.32</td><td>\u00b10.008742</td><td>79.07</td></tr><tr><td>LRAVS+TCB</td><td>79.82</td><td>\u00b10.008514</td><td>78.57</td><td>\u00b10.008706</td><td>79.19</td></tr><tr><td>LRAVS+TCF</td><td>79.48</td><td>\u00b10.008568</td><td>77.93</td><td>\u00b10.008798</td><td>78.70</td></tr><tr><td>2006 Best</td><td>78.70</td><td>\u00b10.008686</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2006 Baseline</td><td>0.90</td><td>\u00b10.002004</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>0.95</td><td>\u00b10.002053</td><td>2.47</td><td>\u00b10.003293</td><td>1.37</td></tr><tr><td>2006 Topline</td><td>99.30</td><td>\u00b10.001769</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.31</td><td>\u00b10.001752</td><td>95.22</td><td>\u00b10.004526</td><td>97.22</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF18": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>96.44</td><td>\u00b10.001169</td><td>95.71</td><td>\u00b10.001279</td><td>96.08</td></tr><tr><td>CNG</td><td>96.19</td><td>\u00b10.001208</td><td>95.58</td><td>\u00b10.001298</td><td>95.88</td></tr><tr><td>AVS</td><td>96.30</td><td>\u00b10.001191</td><td>95.84</td><td>\u00b10.001260</td><td>96.07</td></tr><tr><td>TCB</td><td>96.40</td><td>\u00b10.001177</td><td>95.74</td><td>\u00b10.001275</td><td>96.07</td></tr><tr><td>TCF</td><td>96.35</td><td>\u00b10.001183</td><td>95.69</td><td>\u00b10.001283</td><td>96.02</td></tr><tr><td>AVS+TC</td><td>96.38</td><td>\u00b10.001180</td><td>95.87</td><td>\u00b10.001256</td><td>96.12</td></tr><tr><td>AVS+TCF</td><td>96.40</td><td>\u00b10.001177</td><td>95.73</td><td>\u00b10.001276</td><td>96.06</td></tr><tr><td>LRAVS</td><td>96.22</td><td>\u00b10.001203</td><td>95.85</td><td>\u00b10.001259</td><td>96.04</td></tr><tr><td>LRAVS+TCB</td><td>96.24</td><td>\u00b10.001200</td><td>95.88</td><td>\u00b10.001255</td><td>96.06</td></tr><tr><td>LRAVS+TC</td><td>96.16</td><td>\u00b10.001213</td><td>95.85</td><td>\u00b10.001259</td><td>96.01</td></tr><tr><td>2006 Best</td><td>96.10</td><td>\u00b10.001222</td><td>96.40</td><td>\u00b10.001176</td><td>96.30</td></tr><tr><td>2006 Baseline</td><td>90.00</td><td>\u00b10.001984</td><td>94.90</td><td>\u00b10.001455</td><td>92.40</td></tr><tr><td>Our Baseline</td><td>90.03</td><td>\u00b10.001891</td><td>94.94</td><td>\u00b10.001384</td><td>92.42</td></tr><tr><td>2006 Topline</td><td>99.30</td><td>\u00b10.000551</td><td>99.10</td><td>\u00b10.000625</td><td>99.20</td></tr><tr><td>Our Topline</td><td>99.28</td><td>\u00b10.000534</td><td>99.08</td><td>\u00b10.000603</td><td>99.18</td></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>66.57</td><td>\u00b10.016171</td><td>55.62</td><td>\u00b10.017031</td><td>60.60</td></tr><tr><td>CNG</td><td>61.60</td><td>\u00b10.016672</td><td>58.23</td><td>\u00b10.016906</td><td>59.87</td></tr><tr><td>AVS</td><td>64.60</td><td>\u00b10.016393</td><td>60.83</td><td>\u00b10.016733</td><td>62.66</td></tr><tr><td>TCB</td><td>66.86</td><td>\u00b10.016136</td><td>55.95</td><td>\u00b10.017018</td><td>60.92</td></tr><tr><td>TCF</td><td>66.42</td><td>\u00b10.016189</td><td>54.67</td><td>\u00b10.017065</td><td>59.97</td></tr><tr><td>AVS+TCB</td><td>64.72</td><td>\u00b10.016380</td><td>61.19</td><td>\u00b10.016705</td><td>62.91</td></tr><tr><td>AVS+TCF</td><td>62.78</td><td>\u00b10.016571</td><td>59.86</td><td>\u00b10.016803</td><td>61.28</td></tr><tr><td>LRAVS</td><td>63.92</td><td>\u00b10.016462</td><td>59.94</td><td>\u00b10.016797</td><td>61.87</td></tr><tr><td>LRAVS+TCB</td><td>62.87</td><td>\u00b10.016563</td><td>60.40</td><td>\u00b10.016765</td><td>61.61</td></tr><tr><td>LRAVS+TCF</td><td>62.96</td><td>\u00b10.016554</td><td>59.56</td><td>\u00b10.016824</td><td>61.21</td></tr><tr><td>2006 Best</td><td>61.20</td><td>\u00b10.016704</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>2006 Baseline</td><td>2.20</td><td>\u00b10.005028</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Baseline</td><td>2.17</td><td>\u00b10.004999</td><td>11.13</td><td>\u00b10.010780</td><td>3.64</td></tr><tr><td>2006 Topline</td><td>99.90</td><td>\u00b10.001083</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>Our Topline</td><td>99.85</td><td>\u00b10.001313</td><td>99.24</td><td>\u00b10.002975</td><td>99.55</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF19": {
                "text": "",
                "content": "<table><tr><td>Configuration</td><td>P</td><td>C P</td><td>R</td><td>C R</td><td>F</td></tr><tr><td>6-tag</td><td>95.56</td><td>\u00b10.001682</td><td>95.51</td><td>\u00b10.001691</td><td>95.54</td></tr><tr><td>CNG</td><td>95.54</td><td>\u00b10.001686</td><td>95.53</td><td>\u00b10.001688</td><td>95.54</td></tr><tr><td>AVS</td><td>95.68</td><td>\u00b10.001660</td><td>95.71</td><td>\u00b10.001655</td><td>95.70</td></tr><tr><td>TCB</td><td>95.54</td><td>\u00b10.001687</td><td>95.54</td><td>\u00b10.001687</td><td>95.54</td></tr><tr><td>TCF</td><td>95.52</td><td>\u00b10.001689</td><td>95.54</td><td>\u00b10.001685</td><td>95.53</td></tr><tr><td>AVS+TCB</td><td>95.58</td><td>\u00b10.001680</td><td>95.61</td><td>\u00b10.001674</td><td>95.59</td></tr><tr><td>AVS+TCF</td><td>95.98</td><td>\u00b10.001605</td><td>95.96</td><td>\u00b10.001609</td><td>95.97</td></tr><tr><td>LRAVS</td><td>95.55</td><td>\u00b10.001684</td><td>95.56</td><td>\u00b10.001682</td><td>95.56</td></tr><tr><td>LRAVS+TCB</td><td>95.53</td><td>\u00b10.001687</td><td>95.56</td><td>\u00b10.001683</td><td>95.55</td></tr><tr><td>LRAVS+TCF</td><td>95.69</td><td>\u00b10.001658</td><td>95.72</td><td>\u00b10.001653</td><td>95.71</td></tr><tr><td>2008 Best</td><td>95.96</td><td>\u00b10.001386</td><td>95.83</td><td>\u00b10.001408</td><td>95.89</td></tr><tr><td>2008 Baseline</td><td>84.27</td><td>\u00b10.002563</td><td>88.64</td><td>\u00b10.002234</td><td>86.40</td></tr><tr><td>Our Baseline</td><td>84.05</td><td>\u00b10.002991</td><td>88.86</td><td>\u00b10.002570</td><td>86.39</td></tr><tr><td>2008 Topline</td><td>98.25</td><td>\u00b10.000923</td><td>97.10</td><td>\u00b10.001181</td><td>97.67</td></tr><tr><td>Our Topline</td><td>98.42</td><td>\u00b10.001018</td><td>97.55</td><td>\u00b10.001264</td><td>97.98</td></tr><tr><td>Configuration</td><td>R OOV</td><td>C Roov</td><td>P OOV</td><td>C Poov</td><td>F OOV</td></tr><tr><td>6-tag</td><td>77.63</td><td>\u00b10.014611</td><td>70.56</td><td>\u00b10.01598</td><td>73.92</td></tr><tr><td>CNG</td><td>76.28</td><td>\u00b10.014915</td><td>74.58</td><td>\u00b10.015266</td><td>75.42</td></tr><tr><td>AVS</td><td>77.69</td><td>\u00b10.014597</td><td>75.87</td><td>\u00b10.015001</td><td>76.77</td></tr><tr><td>TCB</td><td>77.69</td><td>\u00b10.014597</td><td>70.71</td><td>\u00b10.015955</td><td>74.04</td></tr><tr><td>TCF</td><td>77.69</td><td>\u00b10.014597</td><td>71.03</td><td>\u00b10.015904</td><td>74.21</td></tr><tr><td>AVS+TCB</td><td>77.20</td><td>\u00b10.014710</td><td>75.14</td><td>\u00b10.015153</td><td>76.16</td></tr><tr><td>AVS+TCF</td><td>78.86</td><td>\u00b10.014316</td><td>77.43</td><td>\u00b10.014657</td><td>78.14</td></tr><tr><td>LRAVS</td><td>77.11</td><td>\u00b10.014731</td><td>75.21</td><td>\u00b10.015139</td><td>76.15</td></tr><tr><td>LRAVS+TCB</td><td>77.04</td><td>\u00b10.014745</td><td>75.19</td><td>\u00b10.015142</td><td>76.11</td></tr><tr><td>LRAVS+TCF</td><td>78.15</td><td>\u00b10.014488</td><td>76.50</td><td>\u00b10.014865</td><td>77.32</td></tr><tr><td>2008 Best</td><td>77.30</td><td>\u00b10.014687</td><td>77.61</td><td>\u00b10.014615</td><td>77.45</td></tr><tr><td>2008 Baseline</td><td>2.83</td><td>\u00b10.005814</td><td>7.69</td><td>\u00b10.009341</td><td>4.14</td></tr><tr><td>Our Baseline</td><td>1.54</td><td>\u00b10.004313</td><td>3.34</td><td>\u00b10.006298</td><td>2.10</td></tr><tr><td>2008 Topline</td><td>99.20</td><td>\u00b10.003123</td><td>97.07</td><td>\u00b10.005913</td><td>98.12</td></tr><tr><td>Our Topline</td><td>99.54</td><td>\u00b10.002375</td><td>97.56</td><td>\u00b10.005409</td><td>98.54</td></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            }
        }
    }
}