{ "paper_id": "O03-1011", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:01:55.020561Z" }, "title": "Chinese Word Auto-Confirmation Agent", "authors": [ { "first": "Jia-Lin", "middle": [], "last": "Tsai", "suffix": "", "affiliation": { "laboratory": "", "institution": "Academia Sinica Nankang", "location": { "settlement": "Taipei", "country": "Taiwan, R.O.C" } }, "email": "tsaijl@iis.sinica.edu.tw" }, { "first": "Cheng-Lung", "middle": [], "last": "Sung", "suffix": "", "affiliation": { "laboratory": "", "institution": "Academia Sinica Nankang", "location": { "settlement": "Taipei", "country": "Taiwan, R.O.C" } }, "email": "clsung@iis.sinica.edu.tw" }, { "first": "Wen-Lian", "middle": [], "last": "Hsu", "suffix": "", "affiliation": { "laboratory": "", "institution": "Academia Sinica Nankang", "location": { "settlement": "Taipei", "country": "Taiwan, R.O.C" } }, "email": "hsu@iis.sinica.edu.tw" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "In various Asian languages, including Chinese, there is no space between words in texts. Thus, most Chinese NLP systems must perform word-segmentation (sentence tokenization). However, successful word-segmentation depends on having a sufficiently large lexicon. On the average, about 3% of the words in text are not contained in a lexicon. Therefore, unknown word identification becomes a bottleneck for Chinese NLP systems. In this paper, we present a Chinese word auto-confirmation (CWAC) agent. CWAC agent uses a hybrid approach that takes advantage of statistical and linguistic approaches. The task of a CWAC agent is to auto-confirm whether an n-gram input (n \u2265 2) is a Chinese word. We design our CWAC agent to satisfy two criteria: (1) a greater than 98% precision rate and a greater than 75% recall rate and (2) domain-independent performance (F-measure). These criteria assure our CWAC agents can work automatically without human intervention. Furthermore, by combining several CWAC agents designed based on different principles, we can construct a multi-CWAC agent through a building-block approach. Three experiments are conducted in this study. The results demonstrate that, for n-gram frequency \u2265 4 in large corpus, our CWAC agent can satisfy the two criteria and achieve 97.82% precision, 77.11% recall, and 86.24% domain-independent F-measure. No existing systems can achieve such a high precision and domain-independent F-measure. The proposed method is our first attempt for constructing a CWAC agent. We will continue develop other CWAC agents and integrating them into a multi-CWAC agent system.", "pdf_parse": { "paper_id": "O03-1011", "_pdf_hash": "", "abstract": [ { "text": "In various Asian languages, including Chinese, there is no space between words in texts. Thus, most Chinese NLP systems must perform word-segmentation (sentence tokenization). However, successful word-segmentation depends on having a sufficiently large lexicon. On the average, about 3% of the words in text are not contained in a lexicon. Therefore, unknown word identification becomes a bottleneck for Chinese NLP systems. In this paper, we present a Chinese word auto-confirmation (CWAC) agent. CWAC agent uses a hybrid approach that takes advantage of statistical and linguistic approaches. The task of a CWAC agent is to auto-confirm whether an n-gram input (n \u2265 2) is a Chinese word. We design our CWAC agent to satisfy two criteria: (1) a greater than 98% precision rate and a greater than 75% recall rate and (2) domain-independent performance (F-measure). These criteria assure our CWAC agents can work automatically without human intervention. Furthermore, by combining several CWAC agents designed based on different principles, we can construct a multi-CWAC agent through a building-block approach. Three experiments are conducted in this study. The results demonstrate that, for n-gram frequency \u2265 4 in large corpus, our CWAC agent can satisfy the two criteria and achieve 97.82% precision, 77.11% recall, and 86.24% domain-independent F-measure. No existing systems can achieve such a high precision and domain-independent F-measure. The proposed method is our first attempt for constructing a CWAC agent. We will continue develop other CWAC agents and integrating them into a multi-CWAC agent system.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "For a human being, efficient word-segmentation (in Chinese) and word sense disambiguation (WSD) arise naturally while a sentence is understood. However, these problems are still difficult for the computer. One reason is that it is hard to create unseen knowledge in the computer from running texts [Dreyfus 1992] . Here, unseen knowledge refers to contextual meaning and unknown lexicon.", "cite_spans": [ { "start": 298, "end": 312, "text": "[Dreyfus 1992]", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Generally, the task of unknown lexicon identification is to identify (1) unknown word (2) unknown word sense, (3) unknown part-of-speech (POS) of a word and 4unknown word pronunciation. Unknown word identification (UWI) is the most essential step in dealing with unknown lexicons. However, UWI is still quite difficult for Chinese NLP. From [Lin et al. 1993 , Lai et al. 2000 , Chen et al. 2002 and Sun et al. 2002 , the difficulty of Chinese UWI is caused by the following problems:", "cite_spans": [ { "start": 341, "end": 357, "text": "[Lin et al. 1993", "ref_id": "BIBREF12" }, { "start": 358, "end": 375, "text": ", Lai et al. 2000", "ref_id": null }, { "start": 376, "end": 394, "text": ", Chen et al. 2002", "ref_id": null }, { "start": 395, "end": 414, "text": "and Sun et al. 2002", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "1. Just as in other Asian languages, Chinese sentences are composed of strings of characters that do not have blank spaces to mark word boundaries.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "2. All Chinese characters can either be a morpheme or a word. Take the Chinese character \u82b1 as an example. It can be either a free morpheme or a word.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "3. Unknown words, which usually are compound words and proper names, are too numerous to list in a machine-readable dictionary (MRD).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "To resolve these issues, statistical, linguistic and hybrid approaches have been developed and investigated. For statistical approaches, researchers use common statistical features, such as maximum entropy [Yu et al. 1998 , Chieu et al. 2002 , association strength [Smadja 1993 , Dunnin 1993 , mutual information [Florian et al. 1999 , Church 2000 , ambiguous matching [Chen & Liu 1992 , Sproat et al. 1996 , and multi-statistical features [Chang et al. 1997] for unknown word detection and extraction. For linguistic approaches, three major types of linguistic rules (knowledge):", "cite_spans": [ { "start": 206, "end": 221, "text": "[Yu et al. 1998", "ref_id": "BIBREF19" }, { "start": 222, "end": 241, "text": ", Chieu et al. 2002", "ref_id": "BIBREF1" }, { "start": 265, "end": 277, "text": "[Smadja 1993", "ref_id": null }, { "start": 278, "end": 291, "text": ", Dunnin 1993", "ref_id": null }, { "start": 313, "end": 333, "text": "[Florian et al. 1999", "ref_id": "BIBREF8" }, { "start": 334, "end": 347, "text": ", Church 2000", "ref_id": "BIBREF4" }, { "start": 369, "end": 385, "text": "[Chen & Liu 1992", "ref_id": null }, { "start": 386, "end": 406, "text": ", Sproat et al. 1996", "ref_id": "BIBREF16" }, { "start": 440, "end": 459, "text": "[Chang et al. 1997]", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "morphology, syntax, and semantics, are used to identify unknown words. Recently, one important trend of UWI follows a hybrid approach so as to take advantage of both merits of statistical and linguistic approaches. Statistical approaches are simple and efficient whereas linguistic approaches are effective in identifying low frequency unknown words [Chang et al. 1997 , Chen et al. 2002 .", "cite_spans": [ { "start": 350, "end": 368, "text": "[Chang et al. 1997", "ref_id": "BIBREF2" }, { "start": 369, "end": 387, "text": ", Chen et al. 2002", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Auto-detection and auto-confirmation are two basic steps in most UWI systems.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Auto-detection is used to detect the possible n-grams candidates from running texts for a better focus, so that in the next auto-confirmation stage, these identification systems need only focus on the set of possible n-grams. In most cases, recall and precision rates are affected by auto-detection and auto-confirmation. Since trade-off would occur between recall and precision, deriving a hybrid approach with precision-recall optimization has become a major challenge [Chang et al. 1997 , Chen et al. 2002 .", "cite_spans": [ { "start": 471, "end": 489, "text": "[Chang et al. 1997", "ref_id": "BIBREF2" }, { "start": 490, "end": 508, "text": ", Chen et al. 2002", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "In this paper, we introduce a Chinese word auto-confirmation (CWAC) agent, which uses a hybrid approach to effectively eliminate human intervention. A CWAC agent is an agent (program) that automatically confirms whether an n-gram input is a Chinese word. We design our CWAC agent to satisfy two criteria: (1) a greater than 98% precision rate and a greater than 75% recall rate and (2) domain-independent performance (F-measure). These criteria assure our CWAC agents can work automatically without human intervention. To our knowledge, no existing system has yet achieved the above criteria.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Furthermore, by combining several CWAC agents designed based on different principles, we can construct a multi-CWAC agent through a building-block approach and service-oriented architecture (such as web services [Graham et al. 2002] ). Figure 1 illustrates one way of a multi-CWAC agent system combining three CWAC agents.", "cite_spans": [ { "start": 212, "end": 232, "text": "[Graham et al. 2002]", "ref_id": null } ], "ref_spans": [ { "start": 236, "end": 246, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "If the number of identified words of a multi-CWAC agent is greater than that of its any single CWAC agent, we believe a multi-CWAC agent could be able to maintain the 98% precision rate and increase its recall rate by merely integrating with more CWAC agents. This article is structured as follows. In Section 2, we will present a method for simulating a CWAC agent. Experimental results and analyses of the CWAC agent will be presented in section 3. Conclusion and future directions will be discussed in Section 4.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "The most frequent 50,000 words were selected from the CKIP lexicon (CKIP [1995] ) to create the system dictionary. From this lexicon, we only use word and POS for our algorithm.", "cite_spans": [ { "start": 67, "end": 79, "text": "(CKIP [1995]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Development of the CWAC agent", "sec_num": "2." }, { "text": "A CWAC agent automatically identifies whether an n-gram input (or, say, n-char string) is a Chinese word. In this paper, an n-gram extractor is developed to extract all n-grams (n \u2265 2 and n-gram frequency \u2265 3) from test sentences as the n-gram input for our CWAC agent (see Figure 2) Figure 3 is the flow chart of the CWAC agent in which the major processes are labeled (1) to (6). The confirmation types, brief descriptions and examples of the CWAC agent, are given in Table 1 . We apply linguistic approach, statistical approach and LFSL (linguistic first, statistical last) approach to develop the CWAC agent. Note in Figure 3 , the processes (5) and (6) are statistical methods, and the remaining four processes are developed from linguistic knowledge. The LFSL approach means a combining process of a linguistic process (such as process 4) and a statistical process (such as process 5).", "cite_spans": [], "ref_spans": [ { "start": 274, "end": 283, "text": "Figure 2)", "ref_id": "FIGREF1" }, { "start": 284, "end": 292, "text": "Figure 3", "ref_id": "FIGREF2" }, { "start": 470, "end": 477, "text": "Table 1", "ref_id": "TABREF1" }, { "start": 621, "end": 629, "text": "Figure 3", "ref_id": "FIGREF2" } ], "eq_spans": [], "section": "Major Processes of the CWAC Agent", "sec_num": "2.1" }, { "text": "The details of these major processes are described below.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Major Processes of the CWAC Agent", "sec_num": "2.1" }, { "text": "Process 1. System dictionary checking: If the n-gram input can be found in the system dictionary, it will be labeled K0 (which means that the n-gram exists in the system dictionary). In Table 1 , the n-gram \u8a08\u7a0b\u8eca is a system word.", "cite_spans": [], "ref_spans": [ { "start": 186, "end": 193, "text": "Table 1", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Major Processes of the CWAC Agent", "sec_num": "2.1" }, { "text": "In this stage, the n-gram input will be segmented by two strategies: left-to-right longest word first (LR-LWF), and right-to-left longest word first (RL-LWF). If LR-LWF and RL-LWF segmentations of the n-gram input are different, the CWAC agent will be triggered to compute the products of all word length for these 1 These n-grams were manually confirmed as is-word in this study 2 These n-grams were manually confirmed as non-word in this study 3 There were no auto-confirming types \"D0\" and \"K0\" in this study segmentations. If both products are equal, the RL-LWF segmentation will be selected. Otherwise, the segmentation with the greatest product will be selected. According to our experiment, the segmentation precision of RL-LWF is, on the average, 1% greater than that of LR-LWF. Take n-gram input \u5c07 \u8ecd\u7528\u7684\u6bdb\u6bef as an example. Its LR-LWF and RL-LWF segmentations are \u5c07\u8ecd/\u7528/\u7684/\u6bdb\u6bef and \u5c07/\u8ecd\u7528/\u7684/\u6bdb\u6bef, respectively. Since both products are equal (2x1x1x2=1x2x1x2), the selected segmentation output for this process is \u5c07/\u8ecd\u7528/\u7684/\u6bdb\u6bef as it is the RL-LFW.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Process 2. Segmentation by system dictionary:", "sec_num": null }, { "text": "The segmentation output from Process 2 is referred to as segmentation2. In this stage, all words in segmentation2 will be compared with the stop word list. There are three types of stop words: begining, middle, and end. The stop word list used in this study is given in Appendix A. These stop words were selected by native Chinese speakers according to those computed beginning, middle, and end single-character words with < 1% of being the beginning, middle, or end words of Hownet [Dong 1999] , respectively. If the first and last words of segmentation2 can be found on the list of begining and end stop words, they will be eliminated from the segmentation2. For those cases in which the word number of seg-mentation2 is greater than 2, middle stop word checking will be triggered. If a middle word in segmentation2 can be found in the middle stop word list, the n-gram input will be split into new strings at any matched stop word. These new strings will be sent to Process 1 as new n-gram input. For example, segmentation2 of the n-gram input \u53ef\u6015\u7684\u8178\u75c5\u6bd2\" is \u53ef\u6015/\u7684/\u8178/\u75c5 \u6bd2. Since there is a middle stop word \"\u7684\" in this segmentation2, the new strings \u53ef\u6015 and \u8178\u75c5\u6bd2 will be sent to Process 1 as new n-gram input.", "cite_spans": [ { "start": 483, "end": 494, "text": "[Dong 1999]", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Process 3. Stop word checking:", "sec_num": null }, { "text": "Once segmentation2 has been processed by Process 3, the result is called segmentation3. If the word number of segmentation3 is 2, POS pattern checking will be triggered. The CWAC agent will first generate all possible POS combinations of the two words using the system dictionary. If the number of generated POS combinations is one and that combination matches one of the POS patterns (", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Process 4. Part-of-Speech (POS) pattern checking:", "sec_num": null }, { "text": "N/V, V/N, N/N, V/V, Adj/N, Adv/N, Adj/V, Adv/V, Adj/Adv, Adv/Adj,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Process 4. Part-of-Speech (POS) pattern checking:", "sec_num": null }, { "text": "Adv/Adv and Adj/Adj) the 2-word string will be tagged as a word and sent to Process 5. This rule-based approach combines syntax knowledge and heuristic observation in order to identify compound words. For example, since the generated POS combination for segmentation3 \u98df\u54c1/\u516c\u53f8 is N/N, \u98df\u54c1\u516c\u53f8 will be sent to Process 5.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Process 4. Part-of-Speech (POS) pattern checking:", "sec_num": null }, { "text": "This stage consists of 4 steps: 1) Thirty randomly selected sentences that include the n-gram input will be extracted from either a large scale or a fixed size corpus. For example, the Chinese sentence \"\u4eba\u4eba\u505a\u74b0\u4fdd\" is a selected sentence that includes the n-gram input \"\u4eba\u4eba\". The details of large scale and fixed size corpus used in this study will be addressed on Subsection 3.2. (Note that the number of selected sentences may be less than thirty and may even be zero due to corpus sparseness.) 2) These selected sentences will be segmented using the system dictionary, and will be segmented by the RL-LWF and LR-LWF technique. 3) For each selected sentence, if the RL-LWF and LR-LWF segmentations are different, the sentence will be regarded as an ambiguous sentence. For example, the Chinese sentence \"\u4eba\u4eba\u505a\u74b0\u4fdd\" is not an ambiguous sentence. 4) Compute the ambiguous ratio of ambiguous sentences to selected sentences. If the ambiguous ratio is less than 50%, the n-gram input will be confirmed as word type K1, K2 or K4 by Process 5 (see Fig. 3) ; other-wise, it will be labeled D1 or D2. According to our observation, the ambiguous ratios of non word n-grams usually are greater than 50%. Process 6. Threshold value checking: In this stage, if the frequency of an n-gram input is greater or equal to 10, it will be labeled as word type K5 by Process 6. According to our experiment, if we directly regard an n-gram input whose frequency is greater than or equal to a certain threshold value as a word, the trade-off frequency of 99% precision rate occurs at the threshold value 7.", "cite_spans": [], "ref_spans": [ { "start": 1034, "end": 1041, "text": "Fig. 3)", "ref_id": "FIGREF2" } ], "eq_spans": [], "section": "Process 5. Segmentation ambiguity checking:", "sec_num": null }, { "text": "The objective of the following experiments is to investigate the performance of the CWAC agent. By this objective, in process1 of the CWAC agent, if an n-gram input is found to be a system word, a temporary system dictionary will be generated. The temporary system is the original system dictionary without this n-gram input. In this case, the n-gram input will be sent to process2 and the temporary system dictionary will be used as system dictionary in both process2 and process5.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiment Results", "sec_num": "3." }, { "text": "Three experiments are conducted in this study. Their results and analysis are given in Sections 3.3, 3.4 and 3.5.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiment Results", "sec_num": "3." }, { "text": "The definition of word is not unique in Chinese [Sciullo et al. 1987 , Sproat et al., 1996 , Huang et al. 1997 , Xia 2000 . As of our knowledge, the Segmentation Standard in China [Liu. et al. 1993] and the Segmentation Standard in Taiwan [CKIP 1996 ] are two of the most famous word-segmentation standards to Chinese. Since the Segmentation Guidelines for the Penn Chinese Treebank (3.0) [Xia 2000] has tried to accommodate the above famous segmentation standards in it, this segmentation was selected as our major guidelines for determining Chinese word. The notion of word in this study includes fixed-phrase words (such as \u6625\u590f\u79cb\u51ac, \u4f60\u4e00\u53e5\u6211\u4e00 \u53e5, \u5967\u6797\u5339\u514b\u904b\u52d5\u6703, etc.) , compounds (such as \u8173\u8e0f\u8eca\u9f8d\u982d, \u592a\u967d\u773c\u93e1, etc.) and simple words (such as \u623f\u5b50, \u8001\u982d\u5152, \u76e4\u5c3c\u897f\u6797, etc.).", "cite_spans": [ { "start": 48, "end": 68, "text": "[Sciullo et al. 1987", "ref_id": "BIBREF14" }, { "start": 69, "end": 90, "text": ", Sproat et al., 1996", "ref_id": "BIBREF16" }, { "start": 91, "end": 110, "text": ", Huang et al. 1997", "ref_id": "BIBREF10" }, { "start": 111, "end": 121, "text": ", Xia 2000", "ref_id": null }, { "start": 180, "end": 198, "text": "[Liu. et al. 1993]", "ref_id": null }, { "start": 239, "end": 249, "text": "[CKIP 1996", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Notion of Word and Evaluation Equations", "sec_num": "3.1" }, { "text": "We use recall, precision and F-measure to evaluate the overall performance of the CWAC agent [Manning et al. 1999] . Precision, recall and F-measure are defined below. Note that the words in following equations (1) and (2) include new words and dictionary words.", "cite_spans": [ { "start": 93, "end": 114, "text": "[Manning et al. 1999]", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Notion of Word and Evaluation Equations", "sec_num": "3.1" }, { "text": "(1) precision = # of correctly identified words / # of identified words", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "recall = # of correctly identified words / # of words", "sec_num": null }, { "text": "(2) F-measure = (2 \u00d7 recall \u00d7 precision) / (recall + precision) (3)", "cite_spans": [ { "start": 43, "end": 63, "text": "(recall + precision)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "recall = # of correctly identified words / # of words", "sec_num": null }, { "text": "In Section 2, we mentioned that the corpus used in process5 of the CWAC agent can be large scale or fixed size. The description of a large scale and a fixed size corpus is given below.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Large Scale Corpus and Fixed Size Corpus", "sec_num": "3.2" }, { "text": "(1) Large scale corpus: In our experiment, texts are collected daily. Texts collected in most Chinese web sites can be used as a large scale corpus. Here, we select OPENFIND [OPENFIND], one of the most popular Chinese search engines, to act as a large scale corpus. If process 5 of the CWAC agent is in large scale corpus mode, it will extract the first thirty matching sentences, including the n-gram input, from the OPENFIND search results. (2) Fixed size corpus: A fixed size corpus is one whose text collection is limited. Here, we use a collection of 14,164,511 Chinese sentences extracted from whole 2002 articles obtained from United Daily News (UDN) web site [UDN] as our fixed size corpus, called 2002 UDN corpus.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Large Scale Corpus and Fixed Size Corpus", "sec_num": "3.2" }, { "text": "The objective of the first experiment is to investigate whether our CWAC agent satisfies criterion 1: the precision rate should be greater than 98% and the recall greater than 75%.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The First Experiment", "sec_num": "3.3" }, { "text": "First, we create a testing corpus, called 2001 UDN corpus, consisting of 4,539,624 Chinese sentences extracted from all 2001 articles on the UDN Web site. The testing corpus includes 10 categories: \u5730\u65b9(local), \u80a1\u5e02(stock), \u79d1\u6280(science), \u65c5\u904a(travel), \u6d88\u8cbb(consuming), \u8ca1\u7d93(financial), \u570b\u969b(world), \u904b\u52d5(sport), \u91ab\u85e5 (health) and \u85dd\u6587(arts). For each category, we randomly select 10,000 sentences to form a test sentence set. We then extract all n-grams from each test sentence set. We then obtain 10 test n-gram sets. All of the extracted n-grams have been manually confirmed as three types: is-word, unsure-word or non-word. In this study, the average percentages of n-grams manually confirmed as is-word, unsure-word, and non-word are 78%, 2% and 20%, respectively. When we compute precision, recall and F-measure, all unsure-word n-grams are excluded. Table 2 shows the results of the CWAC agent in large scale corpus mode. Table 3 shows the results of the CWAC agent in fixed size corpus mode. As shown in Table 2 , the CWAC agent in large scale corpus mode can achieve 96.31% and 97.82% precisions, 77.18% and 77.11% recalls and 85.69% and 86.24% F-measures for n-gram frequencies of \u2265 3 and \u2265 4, respectively. Table 3 shows that the CWAC agent in fixed size corpus mode can achieve 97.32% and 98.11% precisions, 71.44 and 71.61% recalls and 82.39% and 82.79% F-measures.", "cite_spans": [], "ref_spans": [ { "start": 837, "end": 844, "text": "Table 2", "ref_id": "TABREF2" }, { "start": 909, "end": 916, "text": "Table 3", "ref_id": null }, { "start": 992, "end": 999, "text": "Table 2", "ref_id": "TABREF2" }, { "start": 1198, "end": 1205, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "The First Experiment", "sec_num": "3.3" }, { "text": "The hypothesis tests of whether the CWAC agent satisfies criterion 1, H1a and H1b, for this experiment are given below. (One-tailed t-test, reject H 0 if its p-value > 0.05)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The First Experiment", "sec_num": "3.3" }, { "text": "H1a. H 0 : avg. precision \u2264 98%, H 1 : avg. precision > 98% H1b. H 0 : avg. recall \u2264 77%, H 1 : avg. recall > 77%", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The First Experiment", "sec_num": "3.3" }, { "text": "From Tables 2 and 3, we compute the p-values of H1a and H1b for four CWAC modes in Table 4 . Table 4 shows that the CWAC agent passes both hypotheses H1a", "cite_spans": [], "ref_spans": [ { "start": 83, "end": 90, "text": "Table 4", "ref_id": "TABREF3" }, { "start": 93, "end": 100, "text": "Table 4", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "The First Experiment", "sec_num": "3.3" }, { "text": "and H1b in large scale corpus mode with an n-gram frequency of \u2265 4.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The First Experiment", "sec_num": "3.3" }, { "text": "In Chen et al. (2002) , a word that occurs no less than three times in a document is a high frequency word; otherwise, it is a low frequency word. Since a low frequency word in a document could be a high frequency word in our test sentence sets, the results in Tables 2 and 3 can be regarded as an overall evaluation of UWI for low and high frequency words. In Chen et al. (2002) , researchers try to use as much information as possible to identify unknown words in hybrid fashion. Their results have 88%, 84% and 89% precision rates; 67%, 82% and 68% recall rates; 76%, 83%, 78% F-measure rates on low, high, and low/high frequency unknown words, respectively. Table 5 compares some of the famous works on UWI (here, the performance of our CWAC agent was computed solely against \"new words\" exclude words that are already in system dictionary). In Table 5 , the system of [Chen et al. 2002 ] is one of the most famous hybrid approaches on unknown word extraction. Although Lai's system [Lai et al. 2000] achieves the best F-measure 88.45%, but their identifying target (including words and phrases) is different from conventional UWI system. Thus, Lai's result is not included in Table 5 . ", "cite_spans": [ { "start": 3, "end": 21, "text": "Chen et al. (2002)", "ref_id": null }, { "start": 361, "end": 379, "text": "Chen et al. (2002)", "ref_id": null }, { "start": 873, "end": 890, "text": "[Chen et al. 2002", "ref_id": null }, { "start": 987, "end": 1004, "text": "[Lai et al. 2000]", "ref_id": null } ], "ref_spans": [ { "start": 662, "end": 669, "text": "Table 5", "ref_id": "TABREF4" }, { "start": 849, "end": 856, "text": "Table 5", "ref_id": "TABREF4" }, { "start": 1181, "end": 1188, "text": "Table 5", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "The First Experiment", "sec_num": "3.3" }, { "text": "The objective of this experiment is to investigate whether the CWAC agent satisfies criterion 2: the F-measure should be domain-independent.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Second Experiment", "sec_num": "3.4" }, { "text": "The hypothesis test H2 for this experiment is given below. (Two-tailed t-test, reject H 0 if its p-value < 0.05) H2. H 0 : avg. F-measure = \u00b5 0 ; H 1 : avg. F-measure \u2260 \u00b5 0 Table 6 lists the p-values of H2 for four CWAC modes. Table 6 shows that the CWAC agent passes H2 and satisfies criterion 2 in all four CWAC modes. Summing up the results of first and second experiments, we conclude that our method can be used as a CWAC agent in large scale corpus mode when an n-gram frequency is \u2265 4.", "cite_spans": [], "ref_spans": [ { "start": 173, "end": 180, "text": "Table 6", "ref_id": "TABREF5" }, { "start": 227, "end": 234, "text": "Table 6", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Second Experiment", "sec_num": "3.4" }, { "text": "The objective of this experiment is to investigate whether the precision of our CWAC agent is corpus-independent (Q1) and whether its recall is corpus-dependent (Q2). We use large scale and fixed size corpus modes to test Q1 and Q2.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Third Experiment", "sec_num": "3.5" }, { "text": "The hypothesis tests, H3a and H3b, for this experiment are given below. (Two-tailed t-test, reject H 0 if its p-value < 0.05)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Third Experiment", "sec_num": "3.5" }, { "text": "H3a.H 0 : avg. precision of large scale (\u00b51) = avg. precision of fixed size \u00b52H 1 : avg. precision of large scale (\u00b51) \u2260 avg. precision of fixed size (\u00b52)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Third Experiment", "sec_num": "3.5" }, { "text": "H3b.H 0 : avg. recall of large scale (\u00b53) = avg. recall of fixed size (\u00b54) H 1 : avg. recall of large scale (\u00b53) \u2260 avg. recall of fixed size (\u00b54) Table 7 lists the p-values of H3a and H3b for n-gram frequencies of \u2265 3 and \u2265 4. Table 7 shows that H3a is accepted at the 5% significance level. This shows that the precision of the CWAC agent is corpus-independent, since the average precisions of both corpus modes equal at the 5% level. On the other hand, H3b is rejected at the 5% significance level. This shows the recall is corpus-dependent, since the average recalls of both corpus modes are not equal at the 5% level. Tables 8 and 9 were created to sum up the experimental results in Tables 2 and 3. Table 8 gives the comparison of the linguistic, statistic and LFSL approaches in this study. From Table 8 , it shows that the CWAC agent using the technique of LFSL achieves the best optimization of precision-and-recall with the greatest F-measure. Table 9 is the overall experimental results of the CWAC agent for n-gram frequencies of \u2265 3 to \u2265 10. From Table 9 , it indicates the precisions, recalls and F-measures of the CWAC agent are close for different n-gram frequency conditions. The linguistic approaches include auto-confirmation types K3, D6, D7, D8 and D9; the statistical approaches include auto-confirmation types K1, K5, D1 and D5; the LFSL (linguistic approach first, statistical approach last) approaches include auto-confirmation types K2, K4 as shown in Fig. 3 2 \"large\" means large scale corpus mode and \"fixed\" means fixed size corpus mode ", "cite_spans": [], "ref_spans": [ { "start": 146, "end": 153, "text": "Table 7", "ref_id": "TABREF6" }, { "start": 227, "end": 234, "text": "Table 7", "ref_id": "TABREF6" }, { "start": 622, "end": 636, "text": "Tables 8 and 9", "ref_id": "TABREF7" }, { "start": 688, "end": 712, "text": "Tables 2 and 3. Table 8", "ref_id": "TABREF2" }, { "start": 803, "end": 810, "text": "Table 8", "ref_id": "TABREF7" }, { "start": 954, "end": 961, "text": "Table 9", "ref_id": "TABREF8" }, { "start": 1060, "end": 1067, "text": "Table 9", "ref_id": "TABREF8" }, { "start": 1478, "end": 1484, "text": "Fig. 3", "ref_id": "FIGREF2" } ], "eq_spans": [], "section": "Third Experiment", "sec_num": "3.5" }, { "text": "UWI is the most important problem in handling unknown lexicons in NLP systems. A lexicon consists of words, POSs, word senses and word pronunciations. As shown in [Lin et al. 1993 , Lai et al. 2000 , Chen et al. 2002 and Sun et al. 2002 , UWI is still a very difficult task for Chinese NLP systems. One important trend toward resolving unknown word problems is to follow a hybrid approach by combining the advantages of statistical and linguistic approaches. One of the most critical issues in identifying unknown words is to overcome the problem of precision-and-recall trade-off.", "cite_spans": [ { "start": 163, "end": 179, "text": "[Lin et al. 1993", "ref_id": "BIBREF12" }, { "start": 180, "end": 197, "text": ", Lai et al. 2000", "ref_id": null }, { "start": 198, "end": 216, "text": ", Chen et al. 2002", "ref_id": null }, { "start": 217, "end": 236, "text": "and Sun et al. 2002", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Directions for Future Research", "sec_num": "4." }, { "text": "In this paper, we create a CWAC agent adopting a hybrid method to auto-confirm n-gram input. Our experiment shows that the LFSL (linguistic approach first, statistical approach last) approach achieves the best precision-and-recall optimization. Our results demonstrate that, for n-gram frequency \u2265 4 in large corpus mode, our CWAC agent can achieve 97.82% precision, 77.11% recall, and 86.24% F-measure. Thus, it satisfies that two criteria. Moreover, we discover that the use of large scale corpus in this method increases recall but not precision. On the other hand, we find that the precision of using either a large scale corpus or a fixed size corpus is not statistical significantly different at the 5% level.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Directions for Future Research", "sec_num": "4." }, { "text": "This method is our first attempt to create a CWAC agent. We have also considered a building-block approach to construct a multi-CWAC agent. We believe a multi-CWAC agent could be able to maintain the 98% precision rate and increase recall rate by integrating more CWAC agents.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Directions for Future Research", "sec_num": "4." }, { "text": "In the future, we will continue addressing agent-oriented and service-oriented approaches for handling unknown lexicons, such as unknown word POS auto-tagging agent and unknown word-sense auto-determining agent. Furthermore, the method to achieve corpus-independent recall will also be considered.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Directions for Future Research", "sec_num": "4." } ], "back_matter": [ { "text": "We are grateful to the our colleagues in the Intelligent Agent Systems Lab., Gladys Hsieh, Yi-Fan Chang and Yin-Ru Chen, who helped us to manually-confirm n-grams and create tools for this study.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgements", "sec_num": "5." }, { "text": "I. Begining stop word list /\u5152/\u5440/\u55ce/\u5427/\u5462/\u547c/\u4e86/\u662f/\u4f60/\u6211/\u4ed6/\u53c8/\u7b49/\u65e2/\u6216/\u6709/\u5230/\u53bb/\u5728/\u70ba//\u600e/\u65e2/\u751a/ /\u7686/\u76f8/\u82e5/\u5537/\u54fc/\u54e9/\u5509/\u54e6/\u554a/\u5f97/\u90fd/\u6700/\u5582/\u5594/\u55b3/\u55b2/\u7b49/\u8457/\u55ce/\u55e8/ /\u55da/\u55e1/\u6108/\u8ddf/\u8f03/\u904e/\u561b/\u560e/\u561f/\u563b/\u563f/\u5653/\u5657/\u7f77/\u5679/\u566f/\u9084/\u96d6/\u5695/", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Appendix A. Stop Words List", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "means large scale corpus mode and \"fixed\" means fixed size corpus mode References Chen", "authors": [ { "first": "K", "middle": [ "J" ], "last": "", "suffix": "" }, { "first": "W", "middle": [ "Y" ], "last": "Ma", "suffix": "" } ], "year": 2002, "venue": "Proceedings of 19 th COLING", "volume": "", "issue": "", "pages": "169--175", "other_ids": {}, "num": null, "urls": [], "raw_text": "1 \"large\" means large scale corpus mode and \"fixed\" means fixed size corpus mode References Chen, K.J. and W.Y. Ma, \"Unknown Word Extraction for Chinese Documents,\" Pro- ceedings of 19 th COLING 2002, Taipei, pp.169-175", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Named Entity Recognition: A Maximum Entropy Approach Using Global Information", "authors": [ { "first": "H", "middle": [ "L" ], "last": "Chieu", "suffix": "" }, { "first": "H", "middle": [ "T" ], "last": "Ng", "suffix": "" } ], "year": 2002, "venue": "Proceedings of 19 th COLING", "volume": "", "issue": "", "pages": "190--196", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chieu, H.L. and H.T. Ng, \"Named Entity Recognition: A Maximum Entropy Ap- proach Using Global Information,\" Proceedings of 19 th COLING 2002, Taipei, pp.190-196", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "An Unsupervised Iterative Method for Chinese New Lexicon Extraction", "authors": [ { "first": "J", "middle": [ "S" ], "last": "Chang", "suffix": "" }, { "first": "K", "middle": [ "Y" ], "last": "Su", "suffix": "" } ], "year": 1997, "venue": "International Journal of Computational Linguistics & Chinese language Processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chang, J.S. and K.Y. Su, \"An Unsupervised Iterative Method for Chinese New Lexi- con Extraction,\" International Journal of Computational Linguistics & Chinese language Processing, 1997", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Word Identification for mandarin Chinese Sentences", "authors": [ { "first": "K", "middle": [ "J" ], "last": "Chen", "suffix": "" }, { "first": "S", "middle": [ "H" ], "last": "Liu", "suffix": "" } ], "year": null, "venue": "Proceedings of 14 th COLING", "volume": "", "issue": "", "pages": "101--107", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chen, K.J. and S.H. Liu, \"Word Identification for mandarin Chinese Sentences,\" Proceedings of 14 th COLING, pp. 101-107", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Empirical Estimates of Adaptation: The Chance of Two Noriegas is Closer to p/2 than p*p", "authors": [ { "first": "K", "middle": [ "W" ], "last": "Church", "suffix": "" } ], "year": 2000, "venue": "Proceedings of 18 th COLING", "volume": "", "issue": "", "pages": "180--186", "other_ids": {}, "num": null, "urls": [], "raw_text": "Church, K.W., \"Empirical Estimates of Adaptation: The Chance of Two Noriegas is Closer to p/2 than p*p,\" Proceedings of 18 th COLING 2000, pp.180-186", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "the content and illustration of Sinica corpus of Academia Sinica", "authors": [], "year": 1995, "venue": "A study of Chinese Word Boundaries and Segmentation Standard for Information processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "CKIP (Chinese Knowledge Information processing Group), Technical Report no. 95-02, the content and illustration of Sinica corpus of Academia Sinica. Institute of Information Science, Taiwan, Taipei, Academia Sinica, 1995. http://godel.iis.sinica.edu.tw/CKIP/r_content.html CKIP (Chinese Knowledge Information processing Group), A study of Chinese Word Boundaries and Segmentation Standard for Information processing (in Chinese). Technical Report, Taiwan, Taipei, Academia Sinica, 1996.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "What computers still can't do: a critique of artificial reason", "authors": [ { "first": "H", "middle": [ "L" ], "last": "Dreyfus", "suffix": "" } ], "year": 1992, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dreyfus, H.L., What computers still can't do: a critique of artificial reason, Cam- bridge, Mass. : MIT Press, 1992", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Accurate Methods for the Statistics of Surprise and Coincidence", "authors": [ { "first": "Z", "middle": [], "last": "Dong", "suffix": "" }, { "first": "Q", "middle": [], "last": "Dong", "suffix": "" }, { "first": "Hownet", "middle": [], "last": "", "suffix": "" } ], "year": 1993, "venue": "Computational Linguistics", "volume": "19", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dong, Z. and Q. Dong, Hownet, 1999, http://www.keenage.com/ Dunnin, T., \"Accurate Methods for the Statistics of Surprise and Coincidence,\" Computational Linguistics, vol. 19, n 1., 1993", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Dynamic nonlocal language modeling via hierarchical topic-based adaptation", "authors": [ { "first": "R", "middle": [], "last": "Florian", "suffix": "" }, { "first": "D", "middle": [], "last": "Yarowsky", "suffix": "" } ], "year": 1999, "venue": "Proceedings of ACL99", "volume": "", "issue": "", "pages": "167--174", "other_ids": {}, "num": null, "urls": [], "raw_text": "Florian, R. and D. Yarowsky, \"Dynamic nonlocal language modeling via hierarchical topic-based adaptation,\" Proceedings of ACL99, 1999, pp. 167-174", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Segmentation Standard for Chinese natural language Processing", "authors": [ { "first": "C", "middle": [ "R" ], "last": "Huang", "suffix": "" }, { "first": "K", "middle": [ "C" ], "last": "Chen", "suffix": "" }, { "first": "F", "middle": [ "Y" ], "last": "Chen", "suffix": "" }, { "first": "L", "middle": [ "L" ], "last": "Chang", "suffix": "" } ], "year": 1997, "venue": "Computational Linguistics and Chinese Language Processing", "volume": "2", "issue": "", "pages": "47--62", "other_ids": {}, "num": null, "urls": [], "raw_text": "Huang, C.R., Chen, K.C., Chen, F.Y. and Chang, L.L., \"Segmentation Standard for Chinese natural language Processing,\" Computational Linguistics and Chinese Language Processing, 2(2), Aug., 1997, pp.47-62", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Unknown Word and Phrase Extraction Using a Phrase-Like-Unit-based Likelihood Ratio", "authors": [ { "first": "Y", "middle": [ "S" ], "last": "Lai", "suffix": "" }, { "first": "C", "middle": [ "H" ], "last": "Wu", "suffix": "" } ], "year": null, "venue": "International Journal of Computer Processing Oriental Language", "volume": "13", "issue": "1", "pages": "83--95", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lai, Y.S. and Wu, C.H., \"Unknown Word and Phrase Extraction Using a Phrase-Like-Unit-based Likelihood Ratio,\" International Journal of Computer Processing Oriental Language, 13(1), pp.83-95", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "A preliminary Study on Unkown Word Problem in Chinese Word Segmentation", "authors": [ { "first": "M", "middle": [ "Y" ], "last": "Lin", "suffix": "" }, { "first": "T", "middle": [ "H" ], "last": "Chiang", "suffix": "" }, { "first": "K", "middle": [ "Y" ], "last": "Su", "suffix": "" } ], "year": 1993, "venue": "Proceedings of ROCLING VI", "volume": "", "issue": "", "pages": "119--137", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lin, M.Y., T.H. Chiang and K.Y. Su, \"A preliminary Study on Unkown Word Problem in Chinese Word Segmentation,\" Proceedings of ROCLING VI, 1993, pp. 119-137", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Fundations of Statistical Natural Language Processing", "authors": [ { "first": "C", "middle": [ "D" ], "last": "Manning", "suffix": "" }, { "first": "H", "middle": [], "last": "Schuetze", "suffix": "" } ], "year": 1999, "venue": "", "volume": "", "issue": "", "pages": "534--538", "other_ids": {}, "num": null, "urls": [], "raw_text": "Manning, C.D. and Schuetze, H., Fundations of Statistical Natural Language Proc- essing, MIT Press, 1999, pp.534-538", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "On the Definition of Word", "authors": [ { "first": "", "middle": [], "last": "Openfind", "suffix": "" }, { "first": "A", "middle": [ "M D" ], "last": "Sciullo", "suffix": "" }, { "first": "E", "middle": [], "last": "Williams", "suffix": "" } ], "year": 1987, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "OPENFIND, OPENDFIN Chinese Search Web Site, http://www.openfind.com.tw/ Sciullo, A.M.D. and Williams, E., On the Definition of Word, MIT press, 1987", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Retrieving Collocations from Text: Xtract", "authors": [ { "first": "F", "middle": [], "last": "Smadja", "suffix": "" } ], "year": null, "venue": "Computational Linguistics", "volume": "22", "issue": "1", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Smadja, F., \"Retrieving Collocations from Text: Xtract,\" Computational Linguistics, 22(1)", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "A Stochastic Finite-State Word-Segmentation Algorithm for Chinese", "authors": [ { "first": "R", "middle": [], "last": "Sproat", "suffix": "" }, { "first": "C", "middle": [], "last": "Shih", "suffix": "" }, { "first": "W", "middle": [], "last": "Gale", "suffix": "" }, { "first": "N", "middle": [], "last": "Chang", "suffix": "" } ], "year": 1996, "venue": "Computational Linguistics", "volume": "22", "issue": "3", "pages": "377--404", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sproat, R., C. Shih, W. Gale and N. Chang, \"A Stochastic Finite-State Word-Segmentation Algorithm for Chinese,\" Computational Linguistics, 22(3), 1996, pp.377-404", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Chinese Named Entity Identification Using Class-based Language Model", "authors": [ { "first": "J", "middle": [], "last": "Sun", "suffix": "" }, { "first": "J", "middle": [], "last": "Gao", "suffix": "" }, { "first": "L", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "M", "middle": [], "last": "Zhou", "suffix": "" }, { "first": "C", "middle": [], "last": "Huang", "suffix": "" } ], "year": 2002, "venue": "Proceedings of 19 th COLING", "volume": "", "issue": "", "pages": "967--973", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sun, J., J. Gao, L. Zhang, M. Zhou and C. Huang, \"Chinese Named Entity Identifica- tion Using Class-based Language Model,\" Proceedings of 19 th COLING 2002, Taipei, pp.967-973", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "The Segmentation Guidelines for the Penn Chinese Treebank (3.0)", "authors": [ { "first": "F", "middle": [], "last": "", "suffix": "" } ], "year": 2000, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "UDN, On-Line United Daily News , http://udnnews.com/NEWS/ Xia, F., The Segmentation Guidelines for the Penn Chinese Treebank (3.0), October 17, 2000", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Description of the Kent Ridge Digital Labs System Used for MUC-7", "authors": [ { "first": "S", "middle": [], "last": "Yu", "suffix": "" }, { "first": "S", "middle": [], "last": "Bai", "suffix": "" }, { "first": "P", "middle": [], "last": "Wu", "suffix": "" } ], "year": 1998, "venue": "Proceedings of the 7th Message Understanding Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yu, S., S. Bai, and P. Wu, \"Description of the Kent Ridge Digital Labs System Used for MUC-7,\" Proceedings of the 7th Message Understanding Conference, 1998", "links": null } }, "ref_entries": { "FIGREF0": { "uris": null, "num": null, "type_str": "figure", "text": "An illustration of a multi-CWAC agent system" }, "FIGREF1": { "uris": null, "num": null, "type_str": "figure", "text": "An illustration of n-gram extractor and CWAC agent" }, "FIGREF2": { "uris": null, "num": null, "type_str": "figure", "text": "Flow chart for the CWAC agent" }, "TABREF1": { "content": "
RL-LWF)
Auto-ConfExamples
irming ResultsTypesBrief DescriptionsInputOutput
K0N-gram exists in system dictionary\u8a08\u7a0b\u8eca\u8a08\u7a0b\u8eca1
K1Both polysyllabic words exist in online dic-tionary\u63a5\u99c1\u516c\u8eca\u63a5\u99c1/\u516c\u8eca1
K2Two polysyllabic word compounds\u98df\u54c1\u516c\u53f8\u98df\u54c1/\u516c\u53f81
WordK3Both first and last word of segmented N-gram are polysyllabic words and N \u2267 3\u6771\u6e2f\u9ed1\u9baa\u9b5a \u6771\u6e2f/\u9ed1/\u9baa\u9b5a1
K4Segmentation ambiguity is \u2266 50%\u8178\u75c5\u6bd2\u8178/\u75c5\u6bd21
K5N-gram frequency exceeds 10\u963f\u723e\u5df4\u5c3c\u4e9e \u88d4\u963f/\u723e/\u5df4/\u5c3c/\u4e9e\u88d41
D1Two polysyllabic word compounds with at least function word\u554f\u984c\u4e00\u76f4\u554f\u984c/\u4e00\u76f4 2
D2N-gram contains function word\u5e02\u5834\u6307\u51fa\u5e02\u5834/\u6307\u51fa 2
D5Segmentation ambiguity is > 50%\u53f0\u5317\u5e02\u7acb\u53f0\u5317\u5e02/\u7acb 2
Not WordD6 D7Suffix Chinese digit string Digits suffix polysyllabic word\u968a\u4f0d 5 \u706b\u934b\u968a/\u4f0d 1 5/\u706b\u934b 2
D8N-gram is a classifier-noun phrase\u540d\u5b78\u751f\u540d/\u5b78\u751f 2
D9N-gram includes unknown symbol\uff20\u516c\u53f8\uff20/\u516c\u53f8 2
D0Unknown reason33
", "num": null, "text": "Confirming results, types, descriptions and examples of the CWAC agent (The symbol / stands for word boundary according to system dictionary using", "html": null, "type_str": "table" }, "TABREF2": { "content": "
n-grams frequency \u2265 3Large scale Corpus n-grams frequency \u2265 4
ClassPRFPRF
\u5730\u65b997.72% 76.37% 85.74%98.54% 76.27% 85.99%
\u80a1\u5e0294.32% 74.40% 83.19%95.32% 75.51% 84.26%
\u79d1\u628096.51% 76.33% 85.24%97.64% 76.54% 85.81%
\u65c5\u904a97.51% 77.80% 86.55%98.13% 78.09% 86.97%
\u6d88\u8cbb97.85% 79.41% 87.67%98.56% 78.72% 87.53%
\u8ca1\u7d9395.68% 74.63% 83.86%97.32% 75.74% 85.18%
\u570b\u969b96.41% 78.64% 86.62%97.26% 78.36% 86.79%
\u904b\u52d594.17% 78.99% 85.92%95.08% 78.66% 86.10%
\u91ab\u85e596.80% 78.09% 86.44%98.60% 76.85% 86.38%
\u85dd\u658796.94% 76.87% 85.75%98.20% 76.44% 85.96%
Avg.96.31% 77.18% 85.69%97.82% 77.11% 86.24%
Table 3. The first experimental results of the CWAC agent in fixed size corpus mode
n-grams frequency \u2265 3Fixed size Corpus n-grams frequency \u2265 4
ClassPRFPRF
\u5730\u65b997.93% 73.46% 83.95%98.37% 73.91% 84.41%
\u80a1\u5e0295.76% 69.60% 80.61%96.63% 70.30% 81.39%
\u79d1\u628097.70% 69.01% 80.89%98.15% 68.99% 81.03%
\u65c5\u904a97.95% 70.09% 81.71%98.61% 70.49% 82.21%
\u6d88\u8cbb98.20% 74.76% 84.89%98.79% 74.73% 85.09%
\u8ca1\u7d9397.02% 67.41% 79.55%97.76% 68.56% 80.60%
\u570b\u969b97.06% 73.56% 83.69%97.81% 73.00% 83.60%
\u904b\u52d595.77% 74.03% 83.51%97.02% 74.96% 84.57%
\u91ab\u85e597.68% 71.72% 82.71%98.26% 71.64% 82.87%
\u85dd\u658798.22% 70.20% 81.88%99.02% 69.40% 81.61%
Avg.97.32% 71.44% 82.39%98.11% 71.61% 82.79%
", "num": null, "text": "The first experimental results of the CWAC agent in large scale corpus mode", "html": null, "type_str": "table" }, "TABREF3": { "content": "
CWAC modeP-value (H1a)P-value (H1b)
Large scale & Frequency \u2265 30.0018 (accept H 0 )0.3927 (reject H 0 )
Large scale & Frequency \u2265 40.1114 (reject H 0 )0.3842 (reject H 0 )
Fixed size & Frequency \u2265 30.0023 (accept H 0 )0.0 (accept H 0 )
Fixed size & Frequency \u2265 40.4306 (reject H 0 )0.0 (accept H 0 )
", "num": null, "text": "The p-values of the hypothesis tests, H1a and H1b, for four CWAC modes", "html": null, "type_str": "table" }, "TABREF4": { "content": "
SystemMethodTargetTest sizePRF
[Our CWAC]Hybridn-gram word100,000 sentences94.3274.5083.25
[Chen et al. 2002] Hybridn-gram word100 documents896877.10
[Sun et al. 2002]Statisticalname entityMET2 (Chen et al. 1997) 77.8986.0981.79
[Chang et al. 1997] Statisticalbi-gram word1,000 sentences72.3982.8376.38
", "num": null, "text": "Comparison of works on UWI", "html": null, "type_str": "table" }, "TABREF5": { "content": "
CWAC mode\u00b5 0 (F-measure)P-value
Large scale & Frequency \u2265 386%0.4898 (accept H 0 )
Large scale & Frequency \u2265 486%0.7466 (accept H 0 )
Fixed size & Frequency \u2265 383%0.2496 (accept H 0 )
Fixed size & Frequency \u2265 483%0.6190 (accept H 0 )
", "num": null, "text": "The p-values of the hypothesis test H2 for four CWAC modes", "html": null, "type_str": "table" }, "TABREF6": { "content": "
Frequency modeP-value (H3a)P-value (H3b)
Frequency \u2265 3 Frequency \u2265 40.079392 (accept H 0 ) 0.238017 (accept H 0 )0.0000107 (reject H 0 ) 0.0000045 (reject H 0 )
", "num": null, "text": "The p-values of the hypothesis tests, H3a and H3b, for two frequency modes", "html": null, "type_str": "table" }, "TABREF7": { "content": "
N-grams frequency \u2265 3 \u2265 3 \u2265 3Approach 1 Linguistic (L) 92.44%, 93.71% Precision (large, fixed) 2 Statistical (S) 89.15%, 100.00% LFSL 96.72%, 97.43%Recall (large, fixed) 67.41%, 48.96% 4.67%, 3.39% 98.27%, 97.24%F-measure (large, fixed) 77.96%, 64.31% 8.88%, 6.56% 97.49%, 97.34%
1
", "num": null, "text": "Comparison of the linguistic, statistical and LFSL approaches results", "html": null, "type_str": "table" }, "TABREF8": { "content": "
experiment results
N-grams frequency# of n-gramsPrecision (large, fixed) 1Recall (large, fixed)F-measure (large, fixed)
", "num": null, "text": "Overall", "html": null, "type_str": "table" } } } }