{ "paper_id": "O06-4001", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:07:07.020377Z" }, "title": "An Approach to Using the Web as a Live Corpus for Spoken Transliteration Name Access", "authors": [ { "first": "Ming-Shun", "middle": [], "last": "Lin", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Taiwan University", "location": { "postCode": "106", "settlement": "Taipei", "country": "Taiwan" } }, "email": "" }, { "first": "Chia-Ping", "middle": [], "last": "Chen", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Sun Yat-Sen University", "location": { "addrLine": "70, Lien-Hai Road", "postCode": "804", "settlement": "Kaohsiung", "country": "Taiwan" } }, "email": "cpchen@cse.nsysu.edu.tw" }, { "first": "Hsin-Hsi", "middle": [], "last": "Chen", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Taiwan University", "location": { "postCode": "106", "settlement": "Taipei", "country": "Taiwan" } }, "email": "hhchen@csie.ntu.edu.tw" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Recognizing transliteration names is challenging due to their flexible formulation and lexical coverage. In our approach, we employ the Web as a giant corpus. The patterns extracted from the Web are used as a live dictionary to correct speech recognition errors. The plausible character strings recognized by an Automated Speech Recognition (ASR) system are regarded as query terms and submitted to Google. The top N snippets are entered into PAT trees. The terms of the highest scores are selected. Our experiments show that the ASR model with a recovery mechanism can achieve 21.54% performance improvement compared with the ASR only model on the character level. The recall rate is improved from 0.20 to 0.42, and the MRR from 0.07 to 0.31. For collecting transliteration names, we propose a named entity (NE) ontology generation engine, called the X NE-Tree engine, which produces relational named entities by a given seed. The engine incrementally extracts high co-occurring named entities with the seed. A total of 7,642 named entities in the ontology were initiated by 100 seeds. When the bi-character language model is combined with the NE ontology, the ASR recall rate and MRR are improved to 0.48 and 0.38, respectively.", "pdf_parse": { "paper_id": "O06-4001", "_pdf_hash": "", "abstract": [ { "text": "Recognizing transliteration names is challenging due to their flexible formulation and lexical coverage. In our approach, we employ the Web as a giant corpus. The patterns extracted from the Web are used as a live dictionary to correct speech recognition errors. The plausible character strings recognized by an Automated Speech Recognition (ASR) system are regarded as query terms and submitted to Google. The top N snippets are entered into PAT trees. The terms of the highest scores are selected. Our experiments show that the ASR model with a recovery mechanism can achieve 21.54% performance improvement compared with the ASR only model on the character level. The recall rate is improved from 0.20 to 0.42, and the MRR from 0.07 to 0.31. For collecting transliteration names, we propose a named entity (NE) ontology generation engine, called the X NE-Tree engine, which produces relational named entities by a given seed. The engine incrementally extracts high co-occurring named entities with the seed. A total of 7,642 named entities in the ontology were initiated by 100 seeds. When the bi-character language model is combined with the NE ontology, the ASR recall rate and MRR are improved to 0.48 and 0.38, respectively.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Named entities [MUC 1998 ], which denote persons, locations, organizations, etc., are common foci of searchers. Thompson and Dozier [1997] showed that named entity recognition (NER) could improve the performance of information retrieval systems. Capturing named entities is challenging due to their flexible formulation and novelty. The issues behind speech recognition make named entity recognition more challenging on the spoken level than on the written level. This paper focuses on a special type of named entities, called transliteration names. They describe foreign people, places, etc. Spoken transliteration name recognition is useful for many applications. For example, cross language image retrieval via spoken queries aims to employ the latter in one language to retrieve images with captions in another language [Lin et al. 2004] .", "cite_spans": [ { "start": 15, "end": 24, "text": "[MUC 1998", "ref_id": null }, { "start": 112, "end": 138, "text": "Thompson and Dozier [1997]", "ref_id": "BIBREF17" }, { "start": 824, "end": 841, "text": "[Lin et al. 2004]", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "In the past, Appelt and Martin [1999] adapted the TextPro system to process transcripts generated by a speech recognizer. Miller et al. [2000] analyzed the effects of out-of-vocabulary errors and the loss of punctuation on name finding in automatic speech recognition. Huang and Waibel [2002] proposed an adaptive method of named entity extraction for the meeting understanding. Chen [2003] dealt with spoken cross-language access to image collections. The coverage of a lexicon is one of the major issues in spoken transliteration name access. Recently, researchers are interested in using the Web, which provides a huge collection of up-to-date data, as a corpus. Keller and Lapata [2003] employed the Web to obtain frequencies for bigrams that are unseen in a given corpus.", "cite_spans": [ { "start": 13, "end": 37, "text": "Appelt and Martin [1999]", "ref_id": "BIBREF0" }, { "start": 122, "end": 142, "text": "Miller et al. [2000]", "ref_id": "BIBREF14" }, { "start": 269, "end": 292, "text": "Huang and Waibel [2002]", "ref_id": "BIBREF9" }, { "start": 379, "end": 390, "text": "Chen [2003]", "ref_id": "BIBREF2" }, { "start": 666, "end": 690, "text": "Keller and Lapata [2003]", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Named entities are important objects in web documents. Building named entity relationship chains from the web is an important task. Matsuo et al. [2004] found social networks of trust from related web pages. Google sets 1 extracts named entity from web pages by inputting a few named entities. For some emerging applications like personal name disambiguation [Fleischman and Hovy 2004] [Mann and Yarowsky 2003] , social chain finding [Bekkerman and McCallum 2005] [Culotta et al. 2004] [Raghavan et al. 2004] , etc., glossary-based representations of named entities are not enough. For collecting transliteration names and building a bi-character language model, we propose a named entity (NE) ontology generation engine, called the X NE -Tree engine. This engine produces relational named entities by given a seed. The engine uses Google to incrementally extract high co-occurrence named entities from related web pages and those named entities have similar relational properties with the seed. In each iterative step, the seed will be replaced by its siblings or descendants, which form new seeds. In this way, the X NE -Tree engine will build a tree structure as follows with the original seed as a root.", "cite_spans": [ { "start": 132, "end": 152, "text": "Matsuo et al. [2004]", "ref_id": "BIBREF13" }, { "start": 359, "end": 385, "text": "[Fleischman and Hovy 2004]", "ref_id": "BIBREF7" }, { "start": 386, "end": 410, "text": "[Mann and Yarowsky 2003]", "ref_id": "BIBREF12" }, { "start": 434, "end": 463, "text": "[Bekkerman and McCallum 2005]", "ref_id": "BIBREF1" }, { "start": 464, "end": 485, "text": "[Culotta et al. 2004]", "ref_id": "BIBREF6" }, { "start": 486, "end": 508, "text": "[Raghavan et al. 2004]", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "In this paper, we discuss using the Web as a live dictionary for recognizing spoken transliteration names and employ the fuzzy search capability of Google to retrieve relevant web page summaries. In section 2, we sketch the steps in our method. In section 3, we discuss using PAT trees to learn patterns from the Web dynamically and to correct recognition errors. Section 4 shows the experiments, which are the ASR model with/without the recovery mechanism. Section 5 presents the X NE -Tree named entity ontology engine and our experimental results. In section 6, we make concluding remarks.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Spoken Transliteration Name Access", "sec_num": null }, { "text": "The spoken transliteration name recognition system shown in Figure 1 accepts a speech signal denoting a foreign named entity and converts it into a character string. It is composed of the following four major stages. Stages (1) and (2) consist of the fundamental tasks in speech recognition. In the Stages (3) and (4), speech-to-text errors are corrected by using the Web.", "cite_spans": [], "ref_spans": [ { "start": 60, "end": 68, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Spoken Transliteration Name Recognition System", "sec_num": "2." }, { "text": "(1) First, we employ the HTK 2 and SRILM 3 toolkits to build speech recognition models. For each speech signal, we use the model to get a syllable lattice.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 1. Stage in transliteration name recognition", "sec_num": null }, { "text": "(2) Then, the syllable lattice is mapped into a character lattice by using a mapping table. The mapping table is a syllable-to-character mapping. Top-N character strings are selected from the character lattice by using Viterbe algorithm and a bi-character model which is trained from a transliteration name corpus. Such character strings will be called ASR strings in the following.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 1. Stage in transliteration name recognition", "sec_num": null }, { "text": "(3) Next, each ASR string is regarded as a query and is submitted to a web search engine like Google. From the top-N search result, we select higher frequency patterns from a PAT tree structure. The PAT tree [Chien 1997 ] [Gonnet et al. 1992] , which was derived from the Patricia tree, can be employed to extract word boundary and key phrases automatically. Because we employ the PAT tree to extract patterns, the patterns will be called PAT candidates in the following. A PAT tree example, \"\u6e6f\u59c6\u6f22\u514b\u65af\u6e6f\u59c6\u514b\uf939\u65af\u55ac\u6cbb\u514b\uf939\u5c3c\" in MS950 encoding, is shown in Figure 2 . The circles represent semi-infinite string numbers. The number above each circle denotes the length, which indicates the first different bit of the character strings recorded in the sub-trees. In this example, the longest patterns are for \"\u514b\uf939\" and \"\u6e6f\u59c6\" on nodes (7, 12) and (0, 5), with lengths of 33 and 34 bits, respectively. The second longest patterns are for \"\u514b\", \"\uf939\", \"\u59c6\" and \"\u65af\" on nodes (3, 7, 12), (8, 13), (1, 6) and (4, 9), with lengths of 16, 17, 18 and 18 bits, respectively.", "cite_spans": [ { "start": 208, "end": 219, "text": "[Chien 1997", "ref_id": "BIBREF5" }, { "start": 222, "end": 242, "text": "[Gonnet et al. 1992]", "ref_id": "BIBREF8" } ], "ref_spans": [ { "start": 541, "end": 549, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Figure 1. Stage in transliteration name recognition", "sec_num": null }, { "text": "(4) Finally, the PAT candidates of all the ASR strings are merged together and ranked based on their number of occurrences and similarity scores. Candidates with the highest ranks are regarded as the recognition results for a spoken transliteration name.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 2. An example of extracting longest length pattern and its frequency", "sec_num": null }, { "text": "Consider the example shown in Figure 3 . The Chinese speech signal is a transliteration name, \"\u6e6f\u59c6\u514b\uf939\u65af\", in Chinese, which denotes the name of the movie star \"Tom Cruise.\" The lattice shows different combinations of syllables. Each syllable corresponds to several Chinese characters. For example, \"ke\" is converted into \"\u514b\", \"\u67ef\", \"\u79d1\", \"\u53ef\", \"\u5580\", \"\u523b\", etc. The ASR strings \"\u5854\u83ab\u514b\uf939\u65af\", \"\u5854\u9580\u514b\uf939\u65af\", \"\u5854\u83ab\u67ef\uf939\u65af\", etc. are selected from the Spoken Transliteration Name Access character lattice. Through Google fuzzy search using the query \" \u5854 \u83ab \u514b \uf939 \u65af \", some summaries of Chinese web pages are obtained and shown in Figure 4 . Although the common transliteration of \"Tom Cruise\" in Chinese is \"\u6e6f\u59c6\u514b\uf939\u65af\", which is different from the query \"\u5854\u83ab\u514b\uf939\u65af\", fuzzy matching using Google can still identify relevant snippets containing the correct transliteration. We will call this operation \"recognition error recovery using the Web\" in the following. ", "cite_spans": [], "ref_spans": [ { "start": 30, "end": 38, "text": "Figure 3", "ref_id": null }, { "start": 598, "end": 606, "text": "Figure 4", "ref_id": null } ], "eq_spans": [], "section": "Figure 2. An example of extracting longest length pattern and its frequency", "sec_num": null }, { "text": "In the above examples, each partial matching part is enclosed in a rectangle symbol and the correct transliteration name is underlined. Summaries (1), (4) and (5) mention the movie star \"\u6e6f\u59c6\u514b\uf939\u65af\" (Tom Cruise) and summaries (2) and (3) mention a football star, \"\u514b\uf939\u65af\" (Cruz). Figure 3 shows that PAT patterns like \"\u8056\u5854\u514b\uf939\u65af\", \"\u6e6f\u59c6\u514b\uf939\u65af\", \"\u59c6\u514b\uf939\u65af\u6f14\", etc. are proposed. After merging and ranking are performed, the possible recognition results are \"\u6e6f\u59c6\u514b\uf939\u65af\", \"\uf915\u666e\u514b\uf939\u8332\", \"\u8056\u5854\u514b\uf939\u8332\", etc. ", "cite_spans": [], "ref_spans": [ { "start": 272, "end": 280, "text": "Figure 3", "ref_id": null } ], "eq_spans": [], "section": "Figure 4. Summaries obtained through fuzzy search for the query \"\u5854\u83ab\u514b\uf939\u65af\"", "sec_num": null }, { "text": "\u2026 \u2026 \u2026 \u2026 \u2026 \u2026 \u2026 \u2026 \u2026 \u4ed6\u83ab\u514b\uf939\u65af \u5854\u6469\u67ef\uf937\u601d \u7d0d\u9ed8\u79d1\uf933\u53f8 \u90a3\u58a8\u53ef\uf9d3\u7d72 \u62ff\u6a21\u5580\uf93c\u8cdc \uf952\u523b\uf938\u8332 \uf914\u606a\u5982\u5b50 \u723e \u5112\u7386 \u8033 \u8339\u6cbb \uf92f \uf90f\u65e5 \u9580 \uf915 \u8499 \u5b5f \u2026 \u2026 \u2026 \u2026 \u2026", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Speech Input", "sec_num": null }, { "text": "The error recovery module tries to select higher frequency patterns from the Web search results and substitute the speech recognition results of Stages 1 and 2 (shown in Section 2) with the pattern. In this approach, Web search results obtained with an ASR string are placed in a PAT tree, and PAT candidates are selected from the tree. Two points are worth noting. A PAT candidate should occur many times in the PAT tree and should be similar to the ASR string.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Recognition Error Recovery Using the Web", "sec_num": "3." }, { "text": "The frequency, Freq, of a PAT candidate can be computed easily based on the PAT tree structure. The similarity between a PAT candidate and an ASR string is modeled by edit distance, which is the minimum number of insertions, deletions and substitutions needed to transform one character string (ASR) into another string (PAT). The smaller number is, the more similar they are. The similarity score, between an ASR string and a PAT string, is the frequency of the PAT string minus their number of edit operations. Finally, the score of a PAT string relative to an ASR string is defined as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Recognition Error Recovery Using the Web", "sec_num": "3." }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( , ) ( ) tan ( , ) Score ASR PAT Freq PAT Dis ce ASR PAT \u03b1 \u03b2 = \u00d7 \u2212 \u00d7 .", "eq_num": "(1)" } ], "section": "Recognition Error Recovery Using the Web", "sec_num": "3." }, { "text": "It is computed through weighted merging of the frequency of the PAT string and by using the similarity between the ASR string and PAT string. This value determines if the ASR string will be replaced by the PAT string. In the above example, Freq(\u6e6f\u59c6\u514b\uf939\u65af)=43 and Distance (\u5854\u83ab\u514b\uf939\u65af, \u6e6f\u59c6\u514b\uf939\u65af)=2.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Recognition Error Recovery Using the Web", "sec_num": "3." }, { "text": "The speech input to the transliteration name recognition system is a Chinese utterance. We employed 51,111 transliteration names [Chen et al. 2003 ] to train the bi-character language model discussed in Section 2. In the experiments, the test data include 50 American state names, 29 names of movie stars from the 31 st Annual People's Choice Awards (http://www.pcavote.com), and 21 names of NBA stars from the 2005 NBA All Star Team (http://www.nba.com/allstar2005/). The 50 American state names are not very active on the Web. In contrast, the 50 names of stars are very active. The test set is different from the training data set, so it is an open test. Because there may be more than one transliteration for a foreign named entity, the answer keys are manually prepared. For example, \"Arizona\" has four possible transliterations in Chinese: \"\u4e9e\uf9dd\u6851\u7d0d\", \"\u4e9e\uf98c\u6851\u7d0d\", \"\u4e9e\uf9dd\u6851\u90a3\", and \"\u4e9e\uf98c\u6851 \u90a3\". On average, there are 1.9 Chinese transliterations for a foreign name in our test set.", "cite_spans": [ { "start": 129, "end": 146, "text": "[Chen et al. 2003", "ref_id": "BIBREF3" } ], "ref_spans": [], "eq_spans": [], "section": "Experimental Results", "sec_num": "4." }, { "text": "Appendix A lists the name test set and its answer keys. As explained in Section 2, the transliteration name recognition system is composed of four major stages. Stages 1 and 2 include the fundamental speech recognition tasks, and Stages 3 and 4 comprise the error Spoken Transliteration Name Access recovery task. To examine the effects of these two parts, we will evaluate them separately in the following two subsections.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Results", "sec_num": "4." }, { "text": "We assume that correct syllables have been identified in the speech recognition task. We simulate this assumption by transforming all the characters in the answer keys into syllables. Then, in Stage 2, we map the syllable lattice to obtain a character lattice. A total of 50 ASR strings are extracted from the character lattice in Stage 2 and submitted to Google. Finally, the best 10 PAT candidates are selected. We use the MRR (Mean Reciprocal Rank) [Voorhees 1999 ] and recall rate to evaluate the performance. The MRR represents the average rank of the correctly identified transliteration names among in the proposed candidates and it is defined as follow:", "cite_spans": [ { "start": 452, "end": 466, "text": "[Voorhees 1999", "ref_id": "BIBREF19" } ], "ref_spans": [], "eq_spans": [], "section": "Performance in the Error Recovery Task", "sec_num": "4.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "1 1 M i i M R R r M = = \u2211 ,", "eq_num": "(2)" } ], "section": "Performance in the Error Recovery Task", "sec_num": "4.1" }, { "text": "where M is the total number of test cases ; r i equals 1/rank i if rank i > 0 and r i is 0 if no answer is found. The rank i is the rank of the first correct answer for the i th test case. That is, if the first correct answer is ranked 1, then the score is 1/1; if it is ranked 2, the score is 1/2, and so on. The MRR value is between 0 and 1. The inverse of the MRR denotes the average position of the correct answer in the proposed candidate list. The higher the MRR value is, the better the performance is. The recall rate is the number of correct references divided by M. It indicates how many transliteration names are correctly recognized. Table 1 summarizes the experimental results obtained with models without/with the error recovery procedure. With the \"ASR only\" model, the top 10 ASR strings produced in Stage 2 are regarded as answers. This model does not employ the error recovery procedure. The recall rate is 0.79 and the MRR is 0.50. That is, 79 of 100 transliteration names are recognized correctly, and they appear in the first 2 (=1/0.50) position. In contrast, the \"ASR + Web\" model utilizes the error recovery procedure. PAT candidates extracted from the Web are selected in Stage 4. The recall rate is 0.90 and the MRR is 0.88. A total of 90 transliteration names are recognized correctly, and they appear in the first 1.13 (=1/0.88) position on average. In other words, when they are recognized correctly, they are always the top 1. Compared with the first model, the recall rate is increased 13.92%. As for the third model, i.e., the \"ASR/Pre-removed + Web\" model, we try to evaluate the error recovery ability. The correct transliteration names appearing in the set of ASR strings are removed. That is, all of the ASR strings contain at least one incorrect character. In such cases, the recall rate is 0.59 and the MRR is 0.48. This means that 59 transliteration names are recovered, and they appeared in the first 2.08 (=1/0.48) position on average. We further examine the number of errors produced by the \"ASR/Pre-removed + Web\" model to study the error tolerance when using the Web. Table 2 shows the lengths of the transliteration names (in the rows), and the number of matching characters (in the columns). For a transliteration name of length l, the number of matching characters is 0 to l. Each cell denotes how many strings belong to the specific category. For example, before error recovery, there are 6, 25, 90, 184, and 0 strings of length 4, which have 0, 1, 2, 3, and 4 characters matching the corresponding answer keys, respectively. After error recovery, there are 19, 52, 66, 62, and 106 strings of length 4, which have 0, 1, 2, 3, and 4 characters matching the answer keys, respectively. In other words, the recovery procedure corrects some wrong characters. The number of 1-character (2-character) errors decreased from 184 (90) to 62 (66), and total number of correct strings are increased from 0 to 106. Table 3 shows the effects of the error position (in the rows) and the string length (in the columns). A total of 0, 2, 106, 137, and 22 utterances recover 1 character with length 2, 3, 4, 5, and 6, respectively. A total of 0, 0, 3, 4, and 5 utterances recover 2 characters with length 2, 3, 4, 5, and 6, respectively. No utterances can recover over 3 characters. The cell denotes how many strings can be recovered under the specific position and length. For example, a total of 37, 35, 20, and 17 errors for strings of length 4 appearing at positions 1, 2, 3, and 4, respectively, can be recovered by using the Web. In the experiments, 0% (=0/34), 1.80% (=2/111), 35.74% (=109/305), 46.84% (=141/301), and 39.71% (=27/68) of the strings of length 2, 3, 4, 5, and 6 can be recovered, respectively. The 34 is the number of the PAT candidates with length 2. Similarly, the 111, 305, 301, and 68 are the number of the PAT candidates with length 3, 4, 5, and 6. As for length, the longer strings facilitate better recovery than the shorter strings. Another results show that 30.82% (=86/279), 29.75% (=83/279), 17.20% (=48/279), 15.77% (=44/279), 6.09% (=17/279), and 0.36% (=1/279) of the strings with incorrect character appearing at positions 1, 2, 3, 4, 5 and 6 can be recovered, respectively. The 279 is the number of characters on which the 100 test data. Because the bi-character language model proceeds from the left side to the right side, the errors occurring at the beginning are easier to recover than those at the end.", "cite_spans": [], "ref_spans": [ { "start": 646, "end": 653, "text": "Table 1", "ref_id": "TABREF2" }, { "start": 2112, "end": 2120, "text": "Table 2", "ref_id": "TABREF3" }, { "start": 2951, "end": 2958, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Performance in the Error Recovery Task", "sec_num": "4.1" }, { "text": "The set of 100 transliteration names discussed in Section 4.1 are spoken by 2 males and 1 female, so 300 transliteration names are recorded. We employ HTK and SRILM to get the best 100 syllable lattices (N-Best, N=100). The TCC-300 dataset for Mandarin is used to train the acoustic models. There are 417 HMM models, and each one has 39 feature vectors. The syllable accuracy is computed as follows: (M-I-D-S)/M * 100%, where M is the number of correct syllables; I, D, and S denote the number of insertion, deletion, and substitution errors, respectively. The syllable accuracy is 76.57%. To estimate the character recovery ability, we consider the correct character number. The accuracy of the ASR only and ASR+Web models on the character level are computed as follows, respectively: ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Performance in the Speech Recognition Task", "sec_num": "4.2" }, { "text": "where M is the total test number and K is the answer key number for test name i. A total of 50 ASR strings are extracted from the character lattice, and the best 50 PAT candidates are selected. Table 4 shows the character level results. The \"ASR+Web\" model achieves 21.54% better performance than the \"ASR Only\" model on average. Table 5 shows the word level results. The \"ASR+Web\" model using error recovery procedure improves the recall rate and the MRR of the \"ASR Only\" model from 0.20 and 0.07 to 0.42 and 0.31, respectively. In other words, the average ranks of the correct transliteration names move from the 14 th position (=1/0.07) to the 3 rd position (=1/0.31) after error recovery. Web fuzzy search produces useful patterns for error recovery. Our fault tolerance experiments show that longer transliteration names have stronger tolerance than shorter transliteration names and that the incorrect characters appearing at the beginning of a transliteration name are relatively easier to correct than those appearing at the end. Thus, the improvement in the character level accuracy is helpful for the recovery mechanism, and vice versa.", "cite_spans": [], "ref_spans": [ { "start": 194, "end": 201, "text": "Table 4", "ref_id": "TABREF4" }, { "start": 330, "end": 337, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "Performance in the Speech Recognition Task", "sec_num": "4.2" }, { "text": "For collecting transliteration names to build a bi-character language model, we propose using a named entity (NE) ontology generation mechanism, called the X NE -Tree engine. Given a seed, the engine incrementally extracts relational named entities with the seed from related web pages and the output is a tree structure. Each node in the structure is a named entity (NE).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Re-training the Bi-Character Language Model", "sec_num": "5." }, { "text": "Recognizing a named entity and calculating the relational property score with a seed are two crucial tasks. Firstly, we submit the given seed to a search engine and select the top N returned snippets. Then, we use the suffix tree to extract possible patterns automatically. The patterns, which are extracted based on the global statistic, may be impacted by the frequency variance of patterns with the same substrings [Yang and Li 2002] . Because our aim is to generate named entities, most of the max-duplicated strings can be filtered out by using a named entity Spoken Transliteration Name Access recognition (NER) system [Chen et al. 1998 ]. The NER system will re-segment a candidate pattern to obtain some substrings and give each substring a part of speech (POS) and a possible name tag. If any substring is tagged as a location, an organization, or a person by using an NER-POS server [Chen et al. 1998 ], the candidate pattern is considered to be a named entity. Because prepositions frequently occur before/after a named entity, the suffix tree approach may introduce an incorrect boundary. Thus, we filter out substrings that have a preposition tag.", "cite_spans": [ { "start": 418, "end": 436, "text": "[Yang and Li 2002]", "ref_id": "BIBREF18" }, { "start": 625, "end": 642, "text": "[Chen et al. 1998", "ref_id": "BIBREF4" }, { "start": 893, "end": 910, "text": "[Chen et al. 1998", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "A Named Entity Ontology Generation Engine", "sec_num": "5.1" }, { "text": "Secondly, we calculate a relational property score, called the Co-Occurrence Double-Check score (CODC, for each extracted name entity (denoted Y i ) with a seed (denoted X). We postulate that X and Y i have a strong relationship if we can find Y i from X (a forward process) and find X from Y i (a backward process). The forward and backward processes form a double check operation. CODC(X, Y) is defined as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Named Entity Ontology Generation Engine", "sec_num": "5.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( @ ) ( @ ) log ( ) ( ) ( , ) f Y X f X Y f X f Y CODC X Y e \u03b1 \u239b \u239e \u00d7 \u239c \u239f \u239d \u23a0 = ,", "eq_num": "(5)" } ], "section": "A Named Entity Ontology Generation Engine", "sec_num": "5.1" }, { "text": "where f(X@Y i ) is the total number of occurrences of X in the top N snippets when query Y i is submitted to the search engine. Similarly, f(Y i @X) is the total number of occurrences of Y i in the top N snippets for query X; f(X) is the total number of occurrences of X in the top N snippets for query X, and f(Y) is the total number of occurrences of Y in the top N snippets of query Y. In each iterative step, Y i is added to a queue when the CODC (X,Y i ) value is larger than a threshold \u03b8. Then, we get a new seed X from the queue. The CODC measure is best when \u03b1=0.15. The overall process is shown in Figure 5 . ", "cite_spans": [], "ref_spans": [ { "start": 608, "end": 616, "text": "Figure 5", "ref_id": "FIGREF3" } ], "eq_spans": [], "section": "A Named Entity Ontology Generation Engine", "sec_num": "5.1" }, { "text": "When building a bi-character language model, we choose 100 seeds, which are the same 100 utterances described in Section 4. Here, we set a condition to control generation of the ontology. Each initial seed can derive at most four layers, and no more than 15 children are allowed in the first layer. The maximum number of children of a named entity in layer i is bounded by the number in layer (i-1) multiplying a decreasing rate. In the experiments, we set the decreasing rate to be 0.7, so that at most 15, 15\u00d70.7, 15\u00d70.7 2 , and 15\u00d70.7 3 children can be expanded by a named entity in layers 0-3, respectively. We set the threshold \u03b8 at 0.1. Those named entities with CODC scores larger than the predefined threshold are sorted, and a sufficient number of named entities are selected in a sequence for expansion. In this way, a total of 7,642 nodes are generated by the 100 seeds. We employ Touch-Graph (http://www.touchgraph.com) to represent the named entity ontology. Figure 6 shows an example by using \"\u6e6f\u59c6\u514b\uf939\u65af\" as a seed, which is a Mandarin transliteration name of an actor \"Tom Cruise\", to build an ontology. To evaluate the performance, we consider the following four types.", "cite_spans": [], "ref_spans": [ { "start": 972, "end": 980, "text": "Figure 6", "ref_id": null } ], "eq_spans": [], "section": "Constructing a Named Entity Ontology", "sec_num": "5.2" }, { "text": "( 1)Named Entity (NE) type: In this case, the proposed candidate should be a named entity and should not have incorrect boundary. A personal name with a title or a first name with more than 4 characters is regarded as being correct. In contrast, patterns with a last name only are considered incorrect.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Constructing a Named Entity Ontology", "sec_num": "5.2" }, { "text": "(2) Relational property of NE (RNE) type: For those acceptable strings in (1), which have the same relational property with the initial seed or its parents are considered to be correct. The remaining nodes are incorrect.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Constructing a Named Entity Ontology", "sec_num": "5.2" }, { "text": "(3) Partial Named Entity (PNE) type: We relax the restriction on boundary errors specified in (1). Patterns consisting of partial named entities are regarded as being correct. The remaining nodes are incorrect.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Constructing a Named Entity Ontology", "sec_num": "5.2" }, { "text": "(4) Relational property of PNE (RPNE) type: For those acceptable strings in (3), which have the same relational property with the initial seed or its parents are considered to be correct. The remaining nodes are incorrect. Table 6 shows the performance in ontology generation. Of those 7,642 nodes, the error rates for the NE type, the RNE type, the PNE type, and the PRNE type are 19.60%, 34.20%, 12.62%, and 29.82%, respectively. Figure 6 . A snapshot of the named entity ontology of \"\u6e6f\u59c6\u514b\uf939\u65af\" (\"Tom Cruise\")", "cite_spans": [], "ref_spans": [ { "start": 223, "end": 230, "text": "Table 6", "ref_id": "TABREF5" }, { "start": 432, "end": 440, "text": "Figure 6", "ref_id": null } ], "eq_spans": [], "section": "Constructing a Named Entity Ontology", "sec_num": "5.2" }, { "text": "In the previous experiments, we employed 51,111 transliteration names (BaselineTN) to build the bi-character language model. However, these transliteration names might not be active on the Web. We submitted these transliteration names to a search engine (i.e., Google). For a transliteration name, if the search engine does not return any web pages, we filter it out. Finally, we filter out 14,933 named entities and get 36,178 transliteration names (FilterTN) with this method. Refer to Table 6 . Of the 7,642 named entities (Total-Ontology) reported by X ne -engine, 6,146 named entities (NE-Ontology) are of the correct NE type, and 5,023 named entities (RNE-Ontology) are of the correct RNE type.", "cite_spans": [], "ref_spans": [ { "start": 488, "end": 495, "text": "Table 6", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Combining the Bi-Character Language Model with the NE Ontology", "sec_num": "5.3" }, { "text": "In the experiments, we consider word level accuracy only. Two basic transliteration name corpora, i.e., BaselineTN and FilterTN, are employed to build bi-character language models. In ideal case, correct syllables have been identified in the ASR (ASR_Perfect). Table 7 shows that FilterTN is a little better than BaselineTN. We further combine FilterTN with the NE ontology derived by the X NE -Tree engine to perform evaluation. In this way, we employ the FilterTN+RNE-Ontology, FilterTN+NE-Ontology, and FilterTN+Total-Ontology to build bi-character language models. Table 7 summarizes the experimental results obtained with the language model with the NE ontology. The three models with the NE ontology outperform the baseline model. In particular, the NE ontology improve the recall rate and the MRR from 0.79 and 0.50 (BaselineTN) to 0.84 and 0.55 (FilterTN+RNE-Ontology), respectively. Table 8 lists the results obtained using both the NE ontology and error recovery procedure. The NE ontology is still helpful, in particular for the recall rate. In the best case, it improves the recall rate from 0.90 (BaselineTN) to 0.94 (FilterTN+RNE-Ontology). In summary, the model using NE ontology resources, the recall rate is improved 13.92%. On comparing the \"FilterTN+RNE-Ontology\" model with the ASR model without the error recovery procedure and NE ontology resources, the recall rate is improved 18.98%. Table 9 summarizes the experimental results obtained with language models that use the NE ontology without/with error recovery procedure in the complete transliteration name ASR system. The system without the error recovery procedure (ASR Only), the NE ontology still improves the performance. Comparing the \"FilterTN+RNE-Ontology\" with BaselineTN, the recall rate is increased 15%. When the ASR system incorporates the error recovery procedure (ASR+Web), the recall rate is increased 14.28% (FilterTN+RNE-Ontology vs. BaselineTN).", "cite_spans": [], "ref_spans": [ { "start": 261, "end": 269, "text": "Table 7", "ref_id": "TABREF6" }, { "start": 570, "end": 577, "text": "Table 7", "ref_id": "TABREF6" }, { "start": 893, "end": 900, "text": "Table 8", "ref_id": null }, { "start": 1409, "end": 1416, "text": "Table 9", "ref_id": null } ], "eq_spans": [], "section": "Combining the Bi-Character Language Model with the NE Ontology", "sec_num": "5.3" }, { "text": "In this study, we employ the Web as a giant corpus to correct transliteration name recognition errors. Web fuzzy search produces useful patterns for error recovery. In the ideal case, we input the correct syllable sequences, convert them into text strings, and test the recovery capability by using the Web corpus. On comparing with the model without the web recovery procedure, the recall rate is improved 13.92%. For collecting transliteration names beforehand, we propose using a named entity (NE) ontology generation engine, called the X NE -Tree engine. The engine automatically creates named entity ontology for a given seed. In the experiments, a total of 7,642 named entities in the ontology were initiated by 100 seeds. After the language model for speech recognition combined the named entity ontology, the recall rate is improved 18.98%. With a complete transliteration name ASR system, the error recovery experiments show that the recall rate is increased from 0.20 to 0.42 and the MRR from 0.07 to 0.31. When the RNE-Ontology is incorporated, the recall rate and the MRR is increased 0.48 and 0.38, respectively. Thus, we conclude that the error recovery procedure and NE ontology can helpful to the ASR model.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "http://htk.eng.cam.ac.uk/ 3 http://www.speech.sri.com/projects/srilm/", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "This research was partially supported by the National Science Council, Taiwan, under contracts NSC94-2752-E-001-001-PAE and NSC95-2752-E-001-001-PAE.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgements", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Named Entity Extraction from Speech: Approach and Results Using the TextPro System", "authors": [ { "first": "D", "middle": [ "E" ], "last": "Appelt", "suffix": "" }, { "first": "D", "middle": [], "last": "Martin", "suffix": "" } ], "year": 1999, "venue": "Proceedings of DARPA Broadcast News Workshop", "volume": "", "issue": "", "pages": "51--54", "other_ids": {}, "num": null, "urls": [], "raw_text": "Appelt, D. E., and D. Martin, \"Named Entity Extraction from Speech: Approach and Results Using the TextPro System,\" In Proceedings of DARPA Broadcast News Workshop, 1999, pp. 51-54.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Disambiguating Web Appearances of People in a Social Network", "authors": [ { "first": "R", "middle": [], "last": "Bekkerman", "suffix": "" }, { "first": "A", "middle": [], "last": "Mccallum", "suffix": "" } ], "year": 2005, "venue": "Proceedings of WWW", "volume": "", "issue": "", "pages": "463--470", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bekkerman, R., and A. McCallum, \"Disambiguating Web Appearances of People in a Social Network,\" In Proceedings of WWW, 2005, pp. 463-470.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Spoken Cross-Language Access to Image Collection via Captions", "authors": [ { "first": "H", "middle": [ "H" ], "last": "Chen", "suffix": "" } ], "year": 2003, "venue": "Proceedings of 8 th Eurospeech", "volume": "", "issue": "", "pages": "2749--2752", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chen, H. H., \"Spoken Cross-Language Access to Image Collection via Captions,\" In Proceedings of 8 th Eurospeech, 2003, pp. 2749-2752.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Learning Formulation and Transformation Rules for Multilingual Named Entities", "authors": [ { "first": "H", "middle": [ "H" ], "last": "Chen", "suffix": "" }, { "first": "C", "middle": [ "H" ], "last": "Yang", "suffix": "" }, { "first": "Y", "middle": [], "last": "Lin", "suffix": "" } ], "year": 2003, "venue": "Proceedings of the Association for Computational Linguistics on Multilingual and Mixed-language Named Entity Recognition", "volume": "", "issue": "", "pages": "1--8", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chen, H. H., C. H. Yang, and Y. Lin, \"Learning Formulation and Transformation Rules for Multilingual Named Entities,\" In Proceedings of the Association for Computational Linguistics on Multilingual and Mixed-language Named Entity Recognition, 2003, pp. 1-8.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Named Entity Extraction for Information Retrieval", "authors": [ { "first": "H", "middle": [ "H" ], "last": "Chen", "suffix": "" }, { "first": "Y", "middle": [ "W" ], "last": "Ding", "suffix": "" }, { "first": "S", "middle": [ "C" ], "last": "Tsai", "suffix": "" } ], "year": 1998, "venue": "Computer Processing of Oriental Languages, Special Issue on Information Retrieval on Oriental Languages", "volume": "", "issue": "", "pages": "75--85", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chen, H. H., Y. W. Ding, and S. C. Tsai, \"Named Entity Extraction for Information Retrieval,\" Computer Processing of Oriental Languages, Special Issue on Information Retrieval on Oriental Languages, 1998, pp. 75-85", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "PAT-Tree-Based Keyword Extraction for Chinese Information Retrieval", "authors": [ { "first": "L", "middle": [ "F" ], "last": "Chien", "suffix": "" } ], "year": 1997, "venue": "Proceedings of 20th ACM SIGIR Conference", "volume": "", "issue": "", "pages": "50--58", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chien, L. F., \"PAT-Tree-Based Keyword Extraction for Chinese Information Retrieval,\" In Proceedings of 20th ACM SIGIR Conference, 1997, pp. 50-58.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Extracting Social Networks and Contact Information from email and the Web", "authors": [ { "first": "A", "middle": [], "last": "Culotta", "suffix": "" }, { "first": "R", "middle": [], "last": "Bekkerman", "suffix": "" }, { "first": "A", "middle": [], "last": "Mccallum", "suffix": "" } ], "year": 2004, "venue": "Proceedings of the First Conference on Email and Anti-Spam", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Culotta, A., R. Bekkerman, and A. McCallum, \"Extracting Social Networks and Contact Information from email and the Web,\" In Proceedings of the First Conference on Email and Anti-Spam, 2004.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Multi-document Person Name Resolution", "authors": [ { "first": "M", "middle": [ "B" ], "last": "Fleischman", "suffix": "" }, { "first": "E", "middle": [], "last": "Hovy", "suffix": "" } ], "year": 2004, "venue": "Proceedings of the Association for Computational Linguistics (ACL), Reference Resolution Workshop", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Fleischman, M. B., and E. Hovy, \"Multi-document Person Name Resolution,\" In Proceedings of the Association for Computational Linguistics (ACL), Reference Resolution Workshop, 2004. (presentation order: second)", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "New Indices for Text: PAT Trees and PAT Arrays", "authors": [ { "first": "G", "middle": [ "H" ], "last": "Gonnet", "suffix": "" }, { "first": "R", "middle": [ "A" ], "last": "Baeza-Yates", "suffix": "" }, { "first": "T", "middle": [], "last": "Snider", "suffix": "" } ], "year": 1992, "venue": "In Information Retrieval Data Structures Algorithms", "volume": "", "issue": "", "pages": "66--82", "other_ids": {}, "num": null, "urls": [], "raw_text": "Gonnet, G. H., R. A. Baeza-Yates, and T. Snider, \"New Indices for Text: PAT Trees and PAT Arrays,\" In Information Retrieval Data Structures Algorithms, 1992, pp. 66-82.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "An Adaptive Approach of Name Entity Extraction for Meeting Application", "authors": [ { "first": "F", "middle": [], "last": "Huang", "suffix": "" }, { "first": "A", "middle": [], "last": "Waibel", "suffix": "" } ], "year": 2002, "venue": "Proceedings Of`Human Language Technology Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Huang, F., and A. Waibel, \"An Adaptive Approach of Name Entity Extraction for Meeting Application,\" In Proceedings Of`Human Language Technology Conference, 2002.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Using the Web to Obtain Frequencies for Unseen Bigrams", "authors": [ { "first": "F", "middle": [], "last": "Keller", "suffix": "" }, { "first": "M", "middle": [], "last": "Lapata", "suffix": "" } ], "year": 2003, "venue": "Computational Linguistics", "volume": "", "issue": "", "pages": "459--484", "other_ids": {}, "num": null, "urls": [], "raw_text": "Keller, F., and M. Lapata, \"Using the Web to Obtain Frequencies for Unseen Bigrams,\" Computational Linguistics, 2003, pp. 459-484.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Cross-Language Image Retrieval via Spoken Query", "authors": [ { "first": "W", "middle": [ "C" ], "last": "Lin", "suffix": "" }, { "first": "M", "middle": [ "S" ], "last": "Lin", "suffix": "" }, { "first": "H", "middle": [ "H" ], "last": "Chen", "suffix": "" } ], "year": 2004, "venue": "Proceedings of RIAO", "volume": "", "issue": "", "pages": "524--536", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lin, W. C., M. S. Lin, and H. H. Chen, \"Cross-Language Image Retrieval via Spoken Query,\" In Proceedings of RIAO, 2004, pp. 524-536.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Unsupervised Personal Name Disambiguation", "authors": [ { "first": "G", "middle": [ "S" ], "last": "Mann", "suffix": "" }, { "first": "D", "middle": [], "last": "Yarowsky", "suffix": "" } ], "year": 2003, "venue": "Proceedings of Conference on Computational Natural Language Learning", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mann, G. S., and D. Yarowsky, \"Unsupervised Personal Name Disambiguation,\" In Proceedings of Conference on Computational Natural Language Learning, 2003.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Finding Social Network for Trust Calculation", "authors": [ { "first": "Y", "middle": [], "last": "Matsuo", "suffix": "" }, { "first": "H", "middle": [], "last": "Tomobe", "suffix": "" }, { "first": "K", "middle": [], "last": "Hasida", "suffix": "" }, { "first": "M", "middle": [], "last": "Ishizuka", "suffix": "" } ], "year": 2004, "venue": "Proceedings of 16 th European Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "510--514", "other_ids": {}, "num": null, "urls": [], "raw_text": "Matsuo, Y., H. Tomobe, K. Hasida, and M. Ishizuka, \"Finding Social Network for Trust Calculation,\" In Proceedings of 16 th European Conference on Artificial Intelligence, 2004, pp. 510-514.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Named Entity Extraction from Noisy Input: Speech and OCR", "authors": [ { "first": "D", "middle": [], "last": "Miller", "suffix": "" }, { "first": "S", "middle": [], "last": "Boisen", "suffix": "" }, { "first": "R", "middle": [], "last": "Schwartz", "suffix": "" }, { "first": "R", "middle": [], "last": "Stone", "suffix": "" }, { "first": "R", "middle": [], "last": "Weischedel", "suffix": "" } ], "year": 2000, "venue": "Proceedings of 6th Applied Natural Language Processing Conference", "volume": "", "issue": "", "pages": "316--324", "other_ids": {}, "num": null, "urls": [], "raw_text": "Miller, D., S. Boisen, R. Schwartz, R. Stone, and R. Weischedel, \"Named Entity Extraction from Noisy Input: Speech and OCR,\" In Proceedings of 6th Applied Natural Language Processing Conference, 2000, pp. 316-324.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "MUC Message Understanding Competition", "authors": [], "year": 1998, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "MUC Message Understanding Competition, http://www.itl.nist.gov/iaui/894.02/related_projects/muc/index.html, 1998.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "An Exploration of Entity Models, Collective Classification and Relation Description", "authors": [ { "first": "H", "middle": [], "last": "Raghavan", "suffix": "" }, { "first": "J", "middle": [], "last": "Allan", "suffix": "" }, { "first": "A", "middle": [], "last": "Mccallum", "suffix": "" } ], "year": 2004, "venue": "Proceedings of LinkKDD", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Raghavan, H., J. Allan, and A. McCallum, \"An Exploration of Entity Models, Collective Classification and Relation Description,\" In Proceedings of LinkKDD, 2004.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Name Searching and Information Retrieval", "authors": [ { "first": "P", "middle": [], "last": "Thompson", "suffix": "" }, { "first": "C", "middle": [], "last": "Dozier", "suffix": "" } ], "year": 1997, "venue": "Proceedings of Second Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "134--140", "other_ids": {}, "num": null, "urls": [], "raw_text": "Thompson, P., and C. Dozier, \"Name Searching and Information Retrieval,\" In Proceedings of Second Conference on Empirical Methods in Natural Language Processing, 1997, pp. 134--140.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Chinese Keyword Extraction Based on Max-Duplicated Strings of the Documents", "authors": [ { "first": "W", "middle": [], "last": "Yang", "suffix": "" }, { "first": "X", "middle": [], "last": "Li", "suffix": "" } ], "year": 2002, "venue": "Proceedings of 25th ACM SIGIR Conference", "volume": "", "issue": "", "pages": "439--440", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yang, W., and X. Li, \"Chinese Keyword Extraction Based on Max-Duplicated Strings of the Documents,\" In Proceedings of 25th ACM SIGIR Conference, 2002, pp. 439-440.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "The TREC-8 Question Answering Track Evaluation", "authors": [ { "first": "E", "middle": [], "last": "Voorhees", "suffix": "" } ], "year": 1999, "venue": "Proceedings of the 8 th TREC", "volume": "", "issue": "", "pages": "23--37", "other_ids": {}, "num": null, "urls": [], "raw_text": "Voorhees, E., \"The TREC-8 Question Answering Track Evaluation,\" In Proceedings of the 8 th TREC, 1999, pp. 23-37.", "links": null } }, "ref_entries": { "FIGREF1": { "type_str": "figure", "text": "Figure 3. An example of recognizing a transliteration name: \"\u6e6f\u59c6\u514b\uf939\u65af\" (\"Tom Cruise\") (1) ... \u81f3\u65bc\u8d0f\u5bb6\u90e8\u4efd\uff0c\u5247\u9084\u662f\u6e6f\u59c6\u6f22\u514b\u65af\u3001\u6e6f\u59c6\u514b\uf939\u65af\u3001\u55ac\u6cbb\u514b\uf939\u5c3c\u9019\u4e9b\uf934\u9762\u5b54\uff0c\u2026 (2) ... \u7b2c 76 \u5206\u9418\uff0c\u514b\uf939\u65af\u63db\u4e0b\u6885\u958b\u4e8c\ufa01\u7684\u7dad\u57c3\u88cf\u3002 (3) ... \u570b\u969b\u7c73\uf91f (4-4-2)\uff1a\u8c50\u5854\u7d0d/\u79d1\u723e\u591a\u74e6\uff0c\u5e03\u723e\u8fea\uf96a\uff0c\u99ac\u7279\uf925\u9f4a\uff0c\u6cd5\u74e6\uf9dd/\u65af\u5766\u79d1\u7dad\u5947\uff0c \u8c9d\uf9dc\uff0c\u624e\u5167\u8482\uff0c\u57c3\u59c6\uf949/\u514b\uf939\u65af\uff0c\u99ac\u4e01\u65af\u3002 (4) ... \u63d0\u8d77\u59ae\u53ef\u5373\u767c\u706b\u6e6f\u59c6\u514b\uf939\u65af\"\u60f3\uf970\u6b7b\u8a18\u8005\". (5) ... \u96fb\u5f71\u7bc0\u6700\u5177\u770b\u9ede\u7684\u660e\u661f\u7576\u7136\u975e\u59ae\u53ef\u57fa\u5fb7\u66fc\u8207\u6e6f\u59c6\u514b\uf939\u65af\u6709\u671b\u5728\u6c34\u57ce\u7684\u6232\u5287\u6027\u91cd\u9022", "num": null, "uris": null }, "FIGREF3": { "type_str": "figure", "text": "Named entity ontology generation process", "num": null, "uris": null }, "TABREF2": { "num": null, "type_str": "table", "text": "", "html": null, "content": "
ModelsRecallMRR
ASR only0.790.50
ASR + Web0.900.88
ASR/Pre-removed + Web0.590.48
" }, "TABREF3": { "num": null, "type_str": "table", "text": "", "html": null, "content": "
Before Error RecoveryAfter Error Recovery
LengthNumber of Matching CharactersNumber of Matching Characters
of NEs01234560123456
211230----13210----
3629760---639642---
4625 90 1840--1952 66 62 106--
591012771930-1123364153137-
60018203900319 127522
Table 3. Effects of error positions and string lengths
Error PositionsLength=2 Length=3 Length=4 Length=5 Length=6Total
Position 1003742786
Position 2023542483
Position 3-02019948
Position 4--1724344
Position 5---14317
Position 6----11
Total0210914127279
" }, "TABREF4": { "num": null, "type_str": "table", "text": "", "html": null, "content": "
ASR Only (Character Level Accuracy)ASR + Web (Character Level Accuracy)
Top 1Top 2Top 3Top 4Top 5Top 1Top 2Top 3Top 4Top 5
38.01% 43.34% 47.30% 49.07% 50.93% 48.18% 54.01% 55.93% 58.03% 59.48%
Table 5. Performance on the word level
ASR Only (Word Level)ASR + Web (Word Level)
RecallMRRRecallMRR
0.200.070.420.31
" }, "TABREF5": { "num": null, "type_str": "table", "text": "", "html": null, "content": "
Size of Seed Size of OntologyNERNEPNERPNE
1007,64219.60%34.20%12.62%29.82%
" }, "TABREF6": { "num": null, "type_str": "table", "text": "", "html": null, "content": "
Language ModelSize of TNASR_Perfect Only (Word Level) Recall MRR
BaselineTN51,1110.790.50
FilterTN36,1780.800.50
FilterTN + RNE-Ontology41,2010.840.55
FilterTN + NE-Ontology42,3240.830.57
FilterTN + Total-Ontology43,8200.820.57
Table 8. Bi-character language models with both the NE ontology and error
recovery procedure
Language ModelASR_Perfect + Web (Word Level) Recall MRR
BaselineTN0.900.88
FilterTN0.900.87
FilterTN + RNE-Ontology0.940.88
FilterTN + NE-Ontology0.930.88
FilterTN + Total-Ontology0.930.90
Table 9. Combining the bi-character language model with the NE ontology
without/with the error recovery procedure in ASR systems
Language ModelASR Only (Word Level) ASR+Web (Word Level) Recall MRR Recall MRR
BaselineTN0.200.070.420.31
FilterTN0.200.060.410.32
FilterTN + RNE-Ontology0.230.110.480.38
FilterTN + NE-Ontology0.240.110.480.37
FilterTN + Total-Ontology0.240.120.470.39
" } } } }