| { |
| "paper_id": "C14-1028", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T12:23:55.076887Z" |
| }, |
| "title": "Chinese Word Ordering Errors Detection and Correction for Non-Native Chinese Language Learners", |
| "authors": [ |
| { |
| "first": "Shuk-Man", |
| "middle": [], |
| "last": "Cheng", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taiwan University", |
| "location": { |
| "settlement": "Taipei", |
| "country": "Taiwan" |
| } |
| }, |
| "email": "smcheng@nlg.csie.ntu.edu.tw" |
| }, |
| { |
| "first": "Chi-Hsin", |
| "middle": [], |
| "last": "Yu", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taiwan University", |
| "location": { |
| "settlement": "Taipei", |
| "country": "Taiwan" |
| } |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Hsin-Hsi", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taiwan University", |
| "location": { |
| "settlement": "Taipei", |
| "country": "Taiwan" |
| } |
| }, |
| "email": "hhchen@ntu.edu.tw" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "Word Ordering Errors (WOEs) are the most frequent type of grammatical errors at sentence level for non-native Chinese language learners. Learners taking Chinese as a foreign language often place character(s) in the wrong places in sentences, and that results in wrong word(s) or ungrammatical sentences. Besides, there are no clear word boundaries in Chinese sentences. That makes WOEs detection and correction more challenging. In this paper, we propose methods to detect and correct WOEs in Chinese sentences. Conditional random fields (CRFs) based WOEs detection models identify the sentence segments containing WOEs. Segment point-wise mutual information (PMI), inter-segment PMI difference, language model, tag of the previous segment, and CRF bigram template are explored. Words in the segments containing WOEs are reordered to generate candidates that may have correct word orderings. Ranking SVM based models rank the candidates and suggests the most proper corrections. Training and testing sets are selected from HSK dynamic composition corpus created by Beijing Language and Culture University. Besides the HSK WOE dataset, Google Chinese Web 5gram corpus is used to learn features for WOEs detection and correction. The best model achieves an accuracy of 0.834 for detecting WOEs in sentence segments. On the average, the correct word orderings are ranked 4.8 among 184.48 candidates.", |
| "pdf_parse": { |
| "paper_id": "C14-1028", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "Word Ordering Errors (WOEs) are the most frequent type of grammatical errors at sentence level for non-native Chinese language learners. Learners taking Chinese as a foreign language often place character(s) in the wrong places in sentences, and that results in wrong word(s) or ungrammatical sentences. Besides, there are no clear word boundaries in Chinese sentences. That makes WOEs detection and correction more challenging. In this paper, we propose methods to detect and correct WOEs in Chinese sentences. Conditional random fields (CRFs) based WOEs detection models identify the sentence segments containing WOEs. Segment point-wise mutual information (PMI), inter-segment PMI difference, language model, tag of the previous segment, and CRF bigram template are explored. Words in the segments containing WOEs are reordered to generate candidates that may have correct word orderings. Ranking SVM based models rank the candidates and suggests the most proper corrections. Training and testing sets are selected from HSK dynamic composition corpus created by Beijing Language and Culture University. Besides the HSK WOE dataset, Google Chinese Web 5gram corpus is used to learn features for WOEs detection and correction. The best model achieves an accuracy of 0.834 for detecting WOEs in sentence segments. On the average, the correct word orderings are ranked 4.8 among 184.48 candidates.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "Detection and correction of grammatical errors are practical for many applications such as document editing and language learning. Non-native language learners usually encounter problems in learning a new foreign language and are prone to generate ungrammatical sentences. Sentences with various types of errors are written by language learners of different backgrounds. In the HSK corpus, which contains compositions of students from different countries who study Chinese in Beijing Language and Culture University (http://nlp.blcu.edu.cn/online-systems/hsk-language-lib-indexing-system.html), there are 35,884 errors at sentence level. The top 10 error types and their occurrences are listed below: Word Ordering Errors (WOE) (8,515), Missing Component (Adverb) (3,244), Missing Component (Predicate) (3,018), Grammatical Error (\"Is \u2026 DE\") (2,629), Missing Component (Subject) (2,405), Missing Component (Head Noun) (2364), Grammatical Error (\"Is\" sentence) (1,427), Redundant Component (Predicate) (1,130), Uncompleted Sentence (1,052), and Redundant Component (Adverb) (1,051). WOEs are the most frequent type of errors (Yu and Chen, 2012) .", |
| "cite_spans": [ |
| { |
| "start": 1124, |
| "end": 1143, |
| "text": "(Yu and Chen, 2012)", |
| "ref_id": "BIBREF21" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "The types of WOEs in Chinese are different from those in English. A Chinese character has its own meaning in text, while individual characters are meaningless in English. Learners taking Chinese as a foreign language often place character(s) in the wrong places in sentences, and that results in wrong word(s) or ungrammatical sentences. Besides, there are no clear word boundaries in Chinese sentences.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Word segmentation is fundamental in Chinese language processing (Huang and Zhao, 2007) . WOEs may result in wrong segmentation. That may make WOEs detection and correction more challenging.", |
| "cite_spans": [ |
| { |
| "start": 64, |
| "end": 86, |
| "text": "(Huang and Zhao, 2007)", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "This paper aims at identifying the positions of WOEs in the text written by non-native Chinese language learners, and proposes candidates to correct the errors. It is organized as follows. Section 2 surveys the related work. Section 3 gives an overview of the study. Section 4 introduces the dataset used for training and testing. Sections 5 and 6 propose models to detect and correct Chinese WOEs, respectively. Section 7 concludes this study and propose some future work.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "There are only a few researches on the topic of detection and correction of WOEs in Chinese language until now. We survey the related work from the four aspects: (1) grammatical errors made by nonnative Chinese learners, (2) word ordering errors in Chinese language, (3) computer processing of grammatical errors in Chinese language, and (4) grammatical error correction in other languages. Leacock et al. (2014) give thorough surveys in automated grammatical error detection for language learners. Error types, available corpora, evaluation methods, and approaches for different types of errors are specified. Several shared tasks on grammatical error correction in English have been organized in recent years, including HOO 2011 (Dale and Kilgarriff, 2011) , HOO 2012 (Dale et al., 2012) and CoNLL 2013 (Ng et al., 2013 . Different types of grammatical errors are focused: (1) HOO 2011: article and preposition errors, (2) HOO 2012: determiner and preposition errors, and (3) CoNLL 2013: article or determiner errors, preposition errors, noun number errors, verb form errors, and subject-verb agreement errors. In Chinese, spelling check evaluation was held at SIGHAN Bake-off 2013 . However, none of the above evaluations deals with word ordering errors.", |
| "cite_spans": [ |
| { |
| "start": 391, |
| "end": 412, |
| "text": "Leacock et al. (2014)", |
| "ref_id": "BIBREF9" |
| }, |
| { |
| "start": 731, |
| "end": 758, |
| "text": "(Dale and Kilgarriff, 2011)", |
| "ref_id": "BIBREF1" |
| }, |
| { |
| "start": 770, |
| "end": 789, |
| "text": "(Dale et al., 2012)", |
| "ref_id": "BIBREF0" |
| }, |
| { |
| "start": 794, |
| "end": 804, |
| "text": "CoNLL 2013", |
| "ref_id": null |
| }, |
| { |
| "start": 805, |
| "end": 821, |
| "text": "(Ng et al., 2013", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2" |
| }, |
| { |
| "text": "Wang 2011focuses on the Chinese teaching for native English-speaking students. He shows the most frequent grammatical errors made by foreigners are missing components, word orderings and sentence structures. One major learning problem of foreign learners is the influence of negative transfer of mother tongue. Lin (2011) studies the biased errors of word order in Chinese written by foreign students in the HSK corpus. Sun (2011) compares the word orderings between English and Chinese to figure out the differences in sentence structures. Yu and Chen (2012) propose classifiers to detect sentences containing WOEs, but they do not deal with where WOEs are and how to correct them. Wagner et al. (2007) deal with common grammatical errors in English. They consider frequencies of POS n-grams and the outputs of parsers as features. Gamon et al. (2009) identify and correct errors made by non-native English writers. They first detect article and preposition errors, and then apply different techniques to correct each type of errors. Huang et al. (2010) propose a correction rule extraction model trained from 310,956 sets of erroneous and corrected pairwise sentences. Some studies related to word orderings are specific to the topic of pre-processing or post-processing of statistical machine translation, such as Galley and Manning (2008) , Setiawan et al. (2009) , and DeNero and Uszkoreit (2011).", |
| "cite_spans": [ |
| { |
| "start": 420, |
| "end": 430, |
| "text": "Sun (2011)", |
| "ref_id": "BIBREF16" |
| }, |
| { |
| "start": 541, |
| "end": 559, |
| "text": "Yu and Chen (2012)", |
| "ref_id": "BIBREF21" |
| }, |
| { |
| "start": 683, |
| "end": 703, |
| "text": "Wagner et al. (2007)", |
| "ref_id": "BIBREF17" |
| }, |
| { |
| "start": 833, |
| "end": 852, |
| "text": "Gamon et al. (2009)", |
| "ref_id": "BIBREF4" |
| }, |
| { |
| "start": 1035, |
| "end": 1054, |
| "text": "Huang et al. (2010)", |
| "ref_id": "BIBREF5" |
| }, |
| { |
| "start": 1317, |
| "end": 1342, |
| "text": "Galley and Manning (2008)", |
| "ref_id": "BIBREF3" |
| }, |
| { |
| "start": 1345, |
| "end": 1367, |
| "text": "Setiawan et al. (2009)", |
| "ref_id": "BIBREF15" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2" |
| }, |
| { |
| "text": "The major contributions of this paper cover the following aspects: (1) application aspect: detecting and correcting a common type of Chinese written errors of foreign learners with HSK corpus; (2) language aspect: considering the effects of words and segments in Chinese sentences; and (3) resource aspect: exploring the feasibility of using a Chinese web n-gram corpus in WOE detection/correction. Figure 1 sketches an overview of our Chinese WOE detection and correction system. It is composed of three major parts, including dataset preparation, WOE detection, and WOE correction. At first, a corpus is prepared. Sentences containing WOEs are selected from the corpus and corrected by two Chinese native speakers. This corpus will be used for training and testing. Then, a sentence is segmented into a sequence of words, and chunked into several segments based on punctuation marks. Regarding words and segments as fundamental units reduce the number of reordering and limit the reordering scope. The segments containing WOEs are identified by using CRF-based models. Finally, the candidates are generated by reordering and ranked by Ranking SVM-based models. To examine the performance of WOE correction, two datasets, C ans and C sys , consisting of error segments labelled by human and detected by our system, respectively, are employed. The example shown below demonstrates the major steps. This sentence is composed of three segments. The second segment contains a WOE, i.e., \u4eca\u5e74\u590f\u5929\u6bd5\u4e1a\u4e86\u5927\u5b66 (Graduated college this summer). The correct sentence should be \u4eca\u5e74\u590f\u5929\u5927\u5b66\u6bd5\u4e1a\u4e86 (Graduated from college this summer).", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 399, |
| "end": 407, |
| "text": "Figure 1", |
| "ref_id": "FIGREF0" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2" |
| }, |
| { |
| "text": "(1) Reduce the number of reordering units in a sentence by using word segmentation. We filter out sentences with multiple error types and remove duplicate sentences. Total 1,150 error sentences with WOEs remain for this study. Two Chinese native speakers are asked to correct the 1,150 sentences. Only reordering operation is allowed during correction. A dataset composed of 1,150 sets of original sentence S and its two corrections A1 and A2 is formed for training and testing in the experiments. A1 may be different from A2. The following shows an example. Without context, either A1 or A2 is acceptable.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Overview of a Chinese Word Ordering Detection and Correction System", |
| "sec_num": "3" |
| }, |
| { |
| "text": "(She we encouraged to study music and foreign languages.) A1: \u6211\u4eec\u5144\u5f1f\u59ca\u59b9\u9f13\u52b1\u5979\u5b66\u97f3\u4e50\u548c\u5916\u8bed\u3002 (We encouraged her to study music and foreign languages.) A2: \u5979\u9f13\u52b1\u6211\u4eec\u5144\u5f1f\u59ca\u59b9\u5b66\u97f3\u4e50\u548c\u5916\u8bed\u3002 (She encouraged us to study music and foreign languages.) In some cases, A1 and/or A2 may be equal to S. That is, the annotators may think S is correct. That may happen when context is not available. Finally, 327 of 1,150 sets contain different corrections. Both A1 and A2 are equal to S in 27 sets. Total 47 sentences corrected by one annotator are the same as the original sentences, and total 65 sentences corrected by another annotator are the same as the original sentences. This corpus is available at http://nlg.csie.ntu.edu.tw/nlpresource/woe_corpus/. Figure 2 shows the Damerau Levenshtein distance between the original sentences S and the corrections A1 and A2. It counts the minimum number of operations needed to transform a source string into a target one. Here the operation is the transposition of two adjacent characters. Total 823 sets of A1 and A2 have a distance of 0. It means 71.5% of sentences have the same corrections by the two Chinese native speakers. The distances between S and A1 are similar to those between S and A2. Total 850 sets of original sentences and the corrections have a distance below 10 characters and 1,014 sets of sentences have a distance below 20. We can also observe that the number of sentences with even distances is larger than that of sentences with odd distances because most of the Chinese words are composed of two characters. ", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 718, |
| "end": 726, |
| "text": "Figure 2", |
| "ref_id": "FIGREF1" |
| } |
| ], |
| "eq_spans": [], |
| "section": "S: \u5979\u6211\u4eec\u5144\u5f1f\u59ca\u59b9\u9f13\u52b1\u5b66\u97f3\u4e50\u548c\u5916\u8bed\u3002", |
| "sec_num": null |
| }, |
| { |
| "text": "This section first defines the fundamental units for error detection, then introduces the error detection models along with their features, and finally presents and discusses the experimental results.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Detection of Word Ordering Errors", |
| "sec_num": "5" |
| }, |
| { |
| "text": "Permutation is an intuitive way to find out the correct orderings, but its cost is very high. Unrestrictive permutation will generate too many candidates to be acceptable in computation time. What units to be reordered in what range under what condition has be considered. Chinese is different from English in that characters are the smallest meaningful units, and there are no clear word boundaries. Computation cost and segmentation performance is a trade-off to select character or word as a reordering unit. On the one hand, using words as the reordering units will reduce the number of candidates generated. On the other hand, word segmentation results will affect the performance of WOE detection and correction. The following two examples show that reordering the words cannot generate the correct answers. In these two examples, a word in the original sentence (S) is segmented into two words in the correct sentence (A). These words are underlined. Because a word is regarded as a unit for reordering, the correct sentence cannot be generated by word reordering only in these two cases. S: \u4ed6 / \u6559\u7ed9 / \u5b66\u751f\u4eec / \u82f1\u8bed /\u3002 (He / teach to / students / English / .) A:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Fundamental Units for Reordering", |
| "sec_num": "5.1" |
| }, |
| { |
| "text": "\u4ed6 / \u7ed9 / \u5b66\u751f\u4eec / \u6559 / \u82f1\u8bed / \u3002 (He / for / students / teach / English / .) S: \u6700\u8fd1 / \u6211 / \u5f00\u59cb / \u5b66 / \u4e2d\u56fd / \u7684 / \u505a\u83dc\u3002 (Recently / I / start to / learn / China / 's / cooking cuisine.) A: \u6700\u8fd1 / \u6211 / \u5f00\u59cb / \u5b66 / \u505a / \u4e2d\u56fd / \u7684 / \u83dc\u3002 (Recently / I / start to / learn / cooking / China / 's /cuisine.)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Fundamental Units for Reordering", |
| "sec_num": "5.1" |
| }, |
| { |
| "text": "Total 76 sets of sentences belong to such cases. They occupy 6% of the experimental dataset. Considering the benefits of words, we still adopt words as reordering units in the following experiments.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Fundamental Units for Reordering", |
| "sec_num": "5.1" |
| }, |
| { |
| "text": "To prevent reordering all the words in the original sentences, we further divide a sentence into segments based on comma, caesura mark, semi-colon, colon, exclamation mark, question mark, and full stop. Sentence segments containing WOEs will be detected and words will be reordered within the segments to generate the candidates for correction. In our dataset, there are only 31 sets of sentences (i.e., 2.7%) with WOEs across segments. The following shows two examples. The underlined words are moved to other segments.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Fundamental Units for Reordering", |
| "sec_num": "5.1" |
| }, |
| { |
| "text": "(In fact, when I am still working, I am not honest.) A: \u5176\u5b9e\uff0c\u6211\u505a\u4e8b\u60c5\u7684\u65f6\u5019\uff0c\u8fd8\u662f\u4e0d\u600e\u4e48\u8001\u5b9e\u3002 (In fact, when I am working, I am still not honest.) S: \u6240\u4ee5\u6709\u7edd\u5bf9\u7684\u5bfc\u6e38\u5de5\u4f5c\u7ecf\u9a8c\uff0c\u65e0\u987b\u518d\u57f9\u8bad\u3002 (Therefore we have absolute guide work experience, we do not need retraining.) A: \u6709\u7edd\u5bf9\u7684\u5bfc\u6e38\u5de5\u4f5c\u7ecf\u9a8c\uff0c\u6240\u4ee5\u65e0\u987b\u518d\u57f9\u8bad\u3002 (We have absolute guide work experience, therefore we do not need retraining.) In summary, the upper bound of the correction performance would be 91.3%. That is, 6%+2.7% of sentences cannot be resolved.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "S: \u5176\u5b9e\uff0c\u6211\u8fd8\u662f\u505a\u4e8b\u60c5\u7684\u65f6\u5019\uff0c\u4e0d\u600e\u4e48\u8001\u5b9e\u3002", |
| "sec_num": null |
| }, |
| { |
| "text": "Conditional random fields (CRFs) (Lafferty, 2001) are used to implement the WOE detection in sentence segments. Segments with WOEs are labelled with answer tags before training. The original sentence S written by non-native Chinese learner is compared with the annotated correct sentence A. Characters are compared from the start and the end of sentences, respectively. The positions are marked ERR start and ERR end once the characters are different. All words within ERR start and ERR end are marked ERR range . The longest common subsequence (LCS) within ERR range of S and ERR range of A are excluded from ERR range and the remaining words are marked ERR words . Figure 3 shows an example. We use BIO encoding (Ramshaw and Marcus, 1995) to label segments with WOEs. Segments contain-ing words in ERR words are defined to be segments with WOEs. The leftmost segment with WOEs is tagged B, and the following segment with WOEs are tagged I. Those segments without WOEs are tagged O. Feature f B is a bigram template given by SGD-CRF tool 1 . Bigram template combines the tags of the previous segment and current segment, and generates T*T*N feature functions, where T is number of tags and N is number of strings expanded with a macro.", |
| "cite_spans": [ |
| { |
| "start": 33, |
| "end": 49, |
| "text": "(Lafferty, 2001)", |
| "ref_id": "BIBREF8" |
| }, |
| { |
| "start": 714, |
| "end": 740, |
| "text": "(Ramshaw and Marcus, 1995)", |
| "ref_id": "BIBREF14" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 667, |
| "end": 675, |
| "text": "Figure 3", |
| "ref_id": "FIGREF2" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Word Ordering Errors Detection Models", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "WOE detection models will annotate the segments of a sentence with labels B, I or O. These labels will determine which segments may contain WOEs. In the experiments, we use 5-fold cross-validation to evaluate the proposed models. Performance for detecting WOEs is measured at the segment and the sentence levels, respectively. The metrics at the segment level are defined as follows. Here set notation is adopted. The symbol |S| denotes the number of elements in the set S which is derived by the logical formula after vertical bar. TAG pred (SEG) and TAG ans (SEG) mean the labels of segment SEG tagged by WOE detection model and human, respectively. The symbol m denotes total number of segments in the test set.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results and Discussion", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "The metrics at the sentence level are defined as follows:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results and Discussion", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "Accuracy and F1-score measure whether the models can find out segments with WOEs. Correctable Rate of sentences measures whether it is possible that the candidates of the correct word order can be generated by the WOE correction models. If a segment without WOEs is misjudged to be erroneous, the word order still has a chance to be kept by the WOE correction models. However, if a segment with WOEs is misjudged to be correct, words in the misjudged segment will not be reordered in the correction part because the error correction module is not triggered. A sentence is said to be \"correctable\" if no segments in it are misjudged as \"correct\". The ratio of the \"correctable\" sentences is considered as a metric at the sentence level. Table 2 shows the performance of WOE detection. Five models are compared. We regard tagging all the segments with the labels B and O respectively as two baselines. Clearly, the recall at the segment level and the correctable rate at the sentence level are 1 by the all-tag-B baseline. However, its accuracy at the segment and the sentence levels are low. The all-tag-O baseline has better accuracy at the segment level than the all-tag-B baseline, but has very bad F1-score, i.e., 0. The proposed models are much better than the two baselines. Among the feature combinations, f PMI f Diff f Tag f B show the best performance. The accuracy at the segment level is 0.834, and the correctable rate is 0.883. The best detection result will be sent for further correction. ", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 736, |
| "end": 743, |
| "text": "Table 2", |
| "ref_id": "TABREF3" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Results and Discussion", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "This section deals with generating and ranking candidates to correct WOEs. Two datasets, C ans and C sys , are explored in the experiments. We evaluate the optimal performance of the WOE correction models with the C ans dataset, and evaluate WOE detection and correction together with the C sys dataset.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Correction of Word Ordering Errors", |
| "sec_num": "6" |
| }, |
| { |
| "text": "Instead of direct permutation, we consider three strategies shown as follows to correct the error sentences. The complexity of generating candidates by permutation is O(n!). The complexity of using these three strategies decreases to O(n 2 ).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Candidate Generation", |
| "sec_num": "6.1" |
| }, |
| { |
| "text": "(1) Reorder single unit (R single ) R single strategy reorders only one reordering unit (i.e., a word) to n-1 positions within a segment containing n words. Total (n-1) 2 candidates can be generated by this strategy. The following shows an example.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Candidate Generation", |
| "sec_num": "6.1" |
| }, |
| { |
| "text": "S: \u4eca\u5929 / \u5b66\u6821 / \u53bb (Today / school / go to) A: \u4eca\u5929 / \u53bb / \u5b66\u6821 (Today / go to / school) (2) Reorder bi-word (R bi-word ) R bi-word is similar to R single , but two reordering units are merged into a new word before reordering. Because n-1 bi-words can be generated in a segment and n-2 positions are available for each merged bi-word, (n-1)(n-2) candidates are generated by R bi-word . The following shows an example.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Candidate Generation", |
| "sec_num": "6.1" |
| }, |
| { |
| "text": "S: \u65e9/\u5c31/\u4e00\u5bb6/\u516c\u53f8/\u627e/\u6211/\u5de5\u4f5c (before / already / one / company / employ / me / work) A: \u4e00\u5bb6/\u516c\u53f8/\u65e9/\u5c31/\u627e/\u6211/\u5de5\u4f5c (one / company / before / already / employ / me / work) (3) Reorder tri-word (R tri-word ) R tri-word works similarly to R bi-word, but three reordering units are merged before reordering. Total (n-2)(n-3) candidates are generated by R tri-word . The following shows an example.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Candidate Generation", |
| "sec_num": "6.1" |
| }, |
| { |
| "text": "S: \u6211/\u9700\u8981/\u5de5\u4f5c/\u7684/\u7ecf\u9a8c/\u5728/\u60a8/\u7684/\u516c\u53f8\u3002 (I / need / working / (de) / experience / in / your / (de) / company.) A: \u6211/\u9700\u8981/\u5728/\u60a8/\u7684/\u516c\u53f8/\u5de5\u4f5c/\u7684/\u7ecf\u9a8c\u3002 (I / need / in / your / (de) / company / working / (de) / experience.) Table 3 shows the recall rate of each candidate generation strategy. With the C ans dataset, correct word ordering can be generated for 85.8% of the original sentences by fusing R single , R bi-word and R tri-word . The candidates generated by using the C sys dataset cover 69.7% of the correct word orderings. The difference would probably be due to the error propagation of word ordering error detection specified in Section 5.3. Furthermore, 6% of correct word orderings are unable to be generated by using the reordering units due to the word segmentation issue as mentioned in Section 5.1. We can also find that 72.3% of sentences with WOEs can be corrected by the R single strategy using the C ans dataset. It means most of the WOEs made by non-native Chinese learners can be corrected by moving only one word. We use Ranking SVM (Joachims, 2002) for candidates ranking. Because WOEs may produce abnormal POS sequence, POS bigrams and POS trigrams are considered as features for Ranking SVM. We use a k-tuple feature vector for each candidate sentence, where k is the number of features. In each dimension, binary weight is assigned: 1 if the feature exists in a candidate, and 0 otherwise. Score for each candidate is assigned by a binary classifier: 1 if the candidate is the same as either of the annotated corrections, and 0 otherwise.", |
| "cite_spans": [ |
| { |
| "start": 1030, |
| "end": 1046, |
| "text": "(Joachims, 2002)", |
| "ref_id": "BIBREF7" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 194, |
| "end": 201, |
| "text": "Table 3", |
| "ref_id": "TABREF5" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Candidate Generation", |
| "sec_num": "6.1" |
| }, |
| { |
| "text": "Mean Reciprocal Rank (MRR) defined below is used for performance evaluation. The reciprocal rank is the multiplicative inverse of the rank of the first correct answer. MRR is the mean of reciprocal rank for all sentences S, value from 0 to 1. The larger MRR means the correct answer more closes to the top ranking.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results and Discussion", |
| "sec_num": "6.3" |
| }, |
| { |
| "text": "Percentage of answers having rank 1 is another metric. Five-fold cross-validation is used for training and testing. In the C ans and C sys datasets, 182.03 and 184.48 candidates are proposed by the approach of fusing the results of R single , R bi-word , and R tri-word on the average. Experimental results are listed in Table 4 . The proposed candidate ranking method achieves an MRR of 0.270 in the C ans dataset. It means the correct candidates are ranked 3.7 on the average. In contrast, the MRR by using the C sys dataset is 0.208. It means the correct candidates are ranked 4.8 on the average when error detection and correction are performed in pipelining. There are some major types of errors shown as follows in WOE correction.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 321, |
| "end": 328, |
| "text": "Table 4", |
| "ref_id": "TABREF7" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Results and Discussion", |
| "sec_num": "6.3" |
| }, |
| { |
| "text": "(1) Word ordering errors across segments Section 5.1 mentions there are 31 sets of sentences (i.e., 2.7%) with WOEs across segments. Our algorithm cannot capture such kinds of sentences. (2) Propagation errors from candidate generation Table 3 shows the recall of word ordering error detection using the C ans dataset is 0.858. Besides, 6% of sentences mentioned in Section 5.1 cannot be reordered to correct word ordering due to word segmentation issue. (3) Limitation of our models", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 236, |
| "end": 243, |
| "text": "Table 3", |
| "ref_id": "TABREF5" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Results and Discussion", |
| "sec_num": "6.3" |
| }, |
| { |
| "text": "In the fused n-gram models, only one n-gram can be moved. It reduces the number of candidates to be generated, but some types of reorderings are missed. An example is shown as follows. The 2-gram \u51fa\u751f / \u4e8e (was born in) and the unigram \u4e8e (on) have to be exchanged.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results and Discussion", |
| "sec_num": "6.3" |
| }, |
| { |
| "text": "S\uff1a\u6211 / \u51fa\u751f / \u4e8e / 1968 \u5e74 10 \u6708 25 \u65e5 / \u5728 / \u7ef4\u4e5f\u7eb3\u3002 (I / was born / in / 25 October 1968 / on / Vienna.) A\uff1a\u6211 / \u5728 / 1968 \u5e74 10 \u6708 25 \u65e5 / \u51fa\u751f / \u4e8e / \u7ef4\u4e5f\u7eb3\u3002 (I / on / 25 October 1968 / was born / in / Vienna.)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results and Discussion", |
| "sec_num": "6.3" |
| }, |
| { |
| "text": "In this paper, we consider words as the reordering units in WOE detection and correction. Sentences are chunked into segments based on punctuation marks and the CRF technique is used to detect segments that possibly contain WOEs. The best error detection model achieves an accuracy of 0.834. Three reordering strategies are further proposed to generate candidates with correct word ordering and reduce the numerous number of candidates generated by permutation. If the segments containing WOEs are known, 85.8% of correct sentences can be generated by our approach. Finally, Ranking SVM orders the generated candidates based on POS bigrams and POS trigrams features, and achieves an MRR of 0.270 when all erroneous segments are given and an MRR of 0.208 when both detection and correction modules are considered.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusion", |
| "sec_num": "7" |
| }, |
| { |
| "text": "Using words as the reordering unit reduces the cost to generate numerous candidates, but 6% of sentences are unable to reorder due to the word segmentation issue. How to balance the trade-off has to be investigated further. In the candidate ranking, selection of proper weights for POS bigram and trigram features may improve the ranking performance. Since the corpus of WOEs in Chinese is still in a limited size, expanding the related corpus for further research is also indispensable.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusion", |
| "sec_num": "7" |
| }, |
| { |
| "text": "http://leon.bottou.org/projects/sgd", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "This research was partially supported by National Taiwan University and Ministry of Science and Technology, Taiwan under grants 103R890858, 101-2221-E-002-195-MY3 and 102-2221-E-002-103-MY3. We are also very thankful to the anonymous reviewers for their helpful comments to revise this paper.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Acknowledgements", |
| "sec_num": null |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "HOO 2012: A Report on the Preposition and Determiner Error Correction Shared Task", |
| "authors": [ |
| { |
| "first": "Robert", |
| "middle": [], |
| "last": "Dale", |
| "suffix": "" |
| }, |
| { |
| "first": "Ilya", |
| "middle": [], |
| "last": "Anisimoff", |
| "suffix": "" |
| }, |
| { |
| "first": "George", |
| "middle": [], |
| "last": "Narroway", |
| "suffix": "" |
| } |
| ], |
| "year": 2012, |
| "venue": "Proceedings of The 7th Workshop on the Innovative Use of NLP for Building Educational Applications", |
| "volume": "", |
| "issue": "", |
| "pages": "54--62", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Robert Dale, Ilya Anisimoff and George Narroway. 2012. HOO 2012: A Report on the Preposition and Deter- miner Error Correction Shared Task. In Proceedings of The 7th Workshop on the Innovative Use of NLP for Building Educational Applications, pages 54-62, Montre\u00b4al, Canada.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Helping Our Own: The HOO 2011 Pilot Shared Task", |
| "authors": [ |
| { |
| "first": "Robert", |
| "middle": [], |
| "last": "Dale", |
| "suffix": "" |
| }, |
| { |
| "first": "Adam", |
| "middle": [], |
| "last": "Kilgarriff", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Proceedings of the 13th European Workshop on Natural Language Generation (ENLG)", |
| "volume": "", |
| "issue": "", |
| "pages": "242--249", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Robert Dale and Adam Kilgarriff. 2011. Helping Our Own: The HOO 2011 Pilot Shared Task. In Proceedings of the 13th European Workshop on Natural Language Generation (ENLG), pages 242-249, Nancy, France.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Inducing Sentence Structure from Parallel Corpora for Reordering", |
| "authors": [ |
| { |
| "first": "John", |
| "middle": [], |
| "last": "Denero", |
| "suffix": "" |
| }, |
| { |
| "first": "Jakob", |
| "middle": [], |
| "last": "Uszkoreit", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing", |
| "volume": "", |
| "issue": "", |
| "pages": "193--203", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "John DeNero and Jakob Uszkoreit. 2011. Inducing Sentence Structure from Parallel Corpora for Reordering. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 193-203, Edinburgh, Scotland, UK.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "A Simple and Effective Hierarchical Phrase Reordering Model", |
| "authors": [ |
| { |
| "first": "Michel", |
| "middle": [], |
| "last": "Galley", |
| "suffix": "" |
| }, |
| { |
| "first": "Christopher", |
| "middle": [ |
| "D" |
| ], |
| "last": "Manning", |
| "suffix": "" |
| } |
| ], |
| "year": 2008, |
| "venue": "Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing", |
| "volume": "", |
| "issue": "", |
| "pages": "848--856", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Michel Galley and Christopher D. Manning. 2008. A Simple and Effective Hierarchical Phrase Reordering Mod- el. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 848-856, Honolulu.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Using Statistical Techniques and Web Search to Correct ESL Errors", |
| "authors": [ |
| { |
| "first": "Michael", |
| "middle": [], |
| "last": "Gamon", |
| "suffix": "" |
| }, |
| { |
| "first": "Claudia", |
| "middle": [], |
| "last": "Leacock", |
| "suffix": "" |
| }, |
| { |
| "first": "Chris", |
| "middle": [], |
| "last": "Brockett", |
| "suffix": "" |
| }, |
| { |
| "first": "William", |
| "middle": [ |
| "B" |
| ], |
| "last": "Dolan", |
| "suffix": "" |
| }, |
| { |
| "first": "Jianfeng", |
| "middle": [], |
| "last": "Gao", |
| "suffix": "" |
| }, |
| { |
| "first": "Dmitriy", |
| "middle": [], |
| "last": "Belenko", |
| "suffix": "" |
| }, |
| { |
| "first": "Alexandre", |
| "middle": [], |
| "last": "Klementiev", |
| "suffix": "" |
| } |
| ], |
| "year": 2009, |
| "venue": "CALICO Journal", |
| "volume": "26", |
| "issue": "3", |
| "pages": "491--511", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Michael Gamon, Claudia Leacock, Chris Brockett, William B. Dolan, Jianfeng Gao, Dmitriy Belenko, and Alex- andre Klementiev. 2009. Using Statistical Techniques and Web Search to Correct ESL Errors. CALICO Jour- nal, 26(3):491-511.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "Discovering Correction Rules for Auto Editing", |
| "authors": [ |
| { |
| "first": "An-Ta", |
| "middle": [], |
| "last": "Huang", |
| "suffix": "" |
| }, |
| { |
| "first": "Tsung-Ting", |
| "middle": [], |
| "last": "Kuo", |
| "suffix": "" |
| }, |
| { |
| "first": "Ying-Chun", |
| "middle": [], |
| "last": "Lai", |
| "suffix": "" |
| }, |
| { |
| "first": "Shou-De", |
| "middle": [], |
| "last": "Lin", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "Computational Linguistics and Chinese Language Processing", |
| "volume": "15", |
| "issue": "", |
| "pages": "219--236", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "An-Ta Huang, Tsung-Ting Kuo, Ying-Chun Lai, and Shou-De Lin. 2010. Discovering Correction Rules for Auto Editing. Computational Linguistics and Chinese Language Processing, 15(3-4):219-236.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Chinese Word Segmentation: A Decade Review", |
| "authors": [ |
| { |
| "first": "Chang-Ning", |
| "middle": [], |
| "last": "Huang", |
| "suffix": "" |
| }, |
| { |
| "first": "Hai", |
| "middle": [], |
| "last": "Zhao", |
| "suffix": "" |
| } |
| ], |
| "year": 2007, |
| "venue": "Journal of Chinese Information Processing", |
| "volume": "21", |
| "issue": "3", |
| "pages": "8--19", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Chang-ning Huang and Hai Zhao. 2007. Chinese Word Segmentation: A Decade Review. Journal of Chinese Information Processing, 21(3):8-19.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Optimizing Search Engines using Clickthrough Data", |
| "authors": [ |
| { |
| "first": "Thorsten", |
| "middle": [], |
| "last": "Joachims", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining", |
| "volume": "", |
| "issue": "", |
| "pages": "133--142", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Thorsten Joachims. 2002. Optimizing Search Engines using Clickthrough Data. In Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 133-142, Edmon- ton, Alberta, Canada.", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Conditional Random Fields: Probabilistic Models for Segmenting and Labelling Sequence Data", |
| "authors": [ |
| { |
| "first": "John", |
| "middle": [], |
| "last": "Lafferty", |
| "suffix": "" |
| }, |
| { |
| "first": "Andrew", |
| "middle": [], |
| "last": "Mccallum", |
| "suffix": "" |
| }, |
| { |
| "first": "Fernando", |
| "middle": [ |
| "C N" |
| ], |
| "last": "Pereira", |
| "suffix": "" |
| } |
| ], |
| "year": 2001, |
| "venue": "Proceedings of the 18th International Conference on Machine Learning", |
| "volume": "", |
| "issue": "", |
| "pages": "282--289", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "John Lafferty, Andrew McCallum, and Fernando C.N. Pereira. 2001. Conditional Random Fields: Probabilistic Models for Segmenting and Labelling Sequence Data. In Proceedings of the 18th International Conference on Machine Learning (ICML 2001), pages 282-289, San Francisco, CA, USA.", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Automated Grammatical Error Detection for Language Learners. 2 nd Edition", |
| "authors": [ |
| { |
| "first": "Claudia", |
| "middle": [], |
| "last": "Leacock", |
| "suffix": "" |
| }, |
| { |
| "first": "Martin", |
| "middle": [], |
| "last": "Chodorow", |
| "suffix": "" |
| }, |
| { |
| "first": "Michael", |
| "middle": [], |
| "last": "Gamon", |
| "suffix": "" |
| }, |
| { |
| "first": "Joel", |
| "middle": [], |
| "last": "Tetreault", |
| "suffix": "" |
| } |
| ], |
| "year": 2014, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Claudia Leacock, Martin Chodorow, Michael Gamon, and Joel Tetreault. 2014. Automated Grammatical Error Detection for Language Learners. 2 nd Edition. Morgan and Claypool Publishers.", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "Analysis on the Biased Errors of Word Order in Written Expression of Foreign Students", |
| "authors": [ |
| { |
| "first": "Jia-Na", |
| "middle": [], |
| "last": "Lin", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Jia-Na Lin. 2011. Analysis on the Biased Errors of Word Order in Written Expression of Foreign Students. Mas- ter Thesis. Soochow University.", |
| "links": null |
| }, |
| "BIBREF11": { |
| "ref_id": "b11", |
| "title": "Chinese Web 5-gram Version 1. Linguistic Data Consortium, Philadelphia", |
| "authors": [ |
| { |
| "first": "Fang", |
| "middle": [], |
| "last": "Liu", |
| "suffix": "" |
| }, |
| { |
| "first": "Meng", |
| "middle": [], |
| "last": "Yang", |
| "suffix": "" |
| }, |
| { |
| "first": "Dekang", |
| "middle": [], |
| "last": "Lin", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Fang Liu, Meng Yang, Dekang Lin. 2010. Chinese Web 5-gram Version 1. Linguistic Data Consortium, Phila- delphia. http://catalog.ldc.upenn.edu/LDC2010T06.", |
| "links": null |
| }, |
| "BIBREF13": { |
| "ref_id": "b13", |
| "title": "Shared Task on Grammatical Error Correction", |
| "authors": [], |
| "year": null, |
| "venue": "Proceedings of the Seventeenth Conference on Computational Natural Language Learning: Shared Task", |
| "volume": "", |
| "issue": "", |
| "pages": "1--12", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Shared Task on Grammatical Error Correction. In Proceedings of the Seventeenth Conference on Computa- tional Natural Language Learning: Shared Task, pages 1-12, Sofia, Bulgaria.", |
| "links": null |
| }, |
| "BIBREF14": { |
| "ref_id": "b14", |
| "title": "Text Chunking Using Transformation-based Learning", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Lance", |
| "suffix": "" |
| }, |
| { |
| "first": "Mitchell", |
| "middle": [ |
| "P" |
| ], |
| "last": "Ramshaw", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Marcus", |
| "suffix": "" |
| } |
| ], |
| "year": 1995, |
| "venue": "Proceedings of Third Workshop on Very Large Corpora", |
| "volume": "", |
| "issue": "", |
| "pages": "82--94", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Lance A. Ramshaw and Mitchell P. Marcus. 1995. Text Chunking Using Transformation-based Learning. In Proceedings of Third Workshop on Very Large Corpora. Pages 82-94.", |
| "links": null |
| }, |
| "BIBREF15": { |
| "ref_id": "b15", |
| "title": "Topological Ordering of Function Words in Hierarchical Phrase-based Translation", |
| "authors": [ |
| { |
| "first": "Hendra", |
| "middle": [], |
| "last": "Setiawan", |
| "suffix": "" |
| }, |
| { |
| "first": "Min-Yen", |
| "middle": [], |
| "last": "Kan", |
| "suffix": "" |
| }, |
| { |
| "first": "Haizhou", |
| "middle": [], |
| "last": "Li", |
| "suffix": "" |
| }, |
| { |
| "first": "Philip", |
| "middle": [], |
| "last": "Resnik", |
| "suffix": "" |
| } |
| ], |
| "year": 2009, |
| "venue": "Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP", |
| "volume": "", |
| "issue": "", |
| "pages": "324--332", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Hendra Setiawan, Min-Yen Kan, Haizhou Li, and Philip Resnik. 2009. Topological Ordering of Function Words in Hierarchical Phrase-based Translation. In Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 324-332, Suntec, Singapore.", |
| "links": null |
| }, |
| "BIBREF16": { |
| "ref_id": "b16", |
| "title": "Comparison of Chinese and English Word Ordering and Suggestion of Chinese Teaching for Foreign Learners", |
| "authors": [ |
| { |
| "first": "Li-Li", |
| "middle": [], |
| "last": "Sun", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Li-Li Sun. 2011. Comparison of Chinese and English Word Ordering and Suggestion of Chinese Teaching for Foreign Learners. Master Thesis. Heilongjiang University.", |
| "links": null |
| }, |
| "BIBREF17": { |
| "ref_id": "b17", |
| "title": "A Comparative Evaluation of Deep and Shallow Approaches to the Automatic Detection of Common Grammatical Errors", |
| "authors": [ |
| { |
| "first": "Joachim", |
| "middle": [], |
| "last": "Wagner", |
| "suffix": "" |
| }, |
| { |
| "first": "Jennifer", |
| "middle": [], |
| "last": "Foster", |
| "suffix": "" |
| }, |
| { |
| "first": "Josef", |
| "middle": [], |
| "last": "Van Genabith", |
| "suffix": "" |
| } |
| ], |
| "year": 2007, |
| "venue": "Proceedings of the", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Joachim Wagner, Jennifer Foster, and Josef van Genabith. 2007. A Comparative Evaluation of Deep and Shal- low Approaches to the Automatic Detection of Common Grammatical Errors. In Proceedings of the 2007", |
| "links": null |
| }, |
| "BIBREF18": { |
| "ref_id": "b18", |
| "title": "Natural Language Processing and Computational Natural Language Learning", |
| "authors": [], |
| "year": null, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "112--121", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Lan- guage Learning, pages 112-121, Prague, Czech Republic.", |
| "links": null |
| }, |
| "BIBREF19": { |
| "ref_id": "b19", |
| "title": "A Study on the Teaching of Unique Syntactic Pattern in Modern Chinese for Native English-Speaking Students", |
| "authors": [ |
| { |
| "first": "Zhuo", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Zhuo Wang. 2011. A Study on the Teaching of Unique Syntactic Pattern in Modern Chinese for Native English- Speaking Students. Master Thesis. Northeast Normal University.", |
| "links": null |
| }, |
| "BIBREF20": { |
| "ref_id": "b20", |
| "title": "Chinese Spelling Check Evaluation at SIGHAN Bakeoff 2013", |
| "authors": [ |
| { |
| "first": "Shih-Hung", |
| "middle": [], |
| "last": "Wu", |
| "suffix": "" |
| }, |
| { |
| "first": "Chao-Lin", |
| "middle": [], |
| "last": "Liu", |
| "suffix": "" |
| }, |
| { |
| "first": "Lung-Hao", |
| "middle": [], |
| "last": "Lee", |
| "suffix": "" |
| } |
| ], |
| "year": 2013, |
| "venue": "Proceedings of the Seventh SIGHAN Workshop on Chinese Language Processing (SIGHAN-7)", |
| "volume": "", |
| "issue": "", |
| "pages": "35--42", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Shih-Hung Wu, Chao-Lin Liu, and Lung-Hao Lee. 2013. Chinese Spelling Check Evaluation at SIGHAN Bake- off 2013. In Proceedings of the Seventh SIGHAN Workshop on Chinese Language Processing (SIGHAN-7), pages 35-42, Nagoya, Japan.", |
| "links": null |
| }, |
| "BIBREF21": { |
| "ref_id": "b21", |
| "title": "Detecting Word Ordering Errors in Chinese Sentences for Learning Chinese as a Foreign Language", |
| "authors": [ |
| { |
| "first": "Chi-Hsin", |
| "middle": [], |
| "last": "Yu", |
| "suffix": "" |
| }, |
| { |
| "first": "Hsin-Hsi", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| } |
| ], |
| "year": 2012, |
| "venue": "Proceedings of the 24th International Conference on Computational Linguistics", |
| "volume": "", |
| "issue": "", |
| "pages": "3003--3018", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Chi-Hsin Yu and Hsin-Hsi Chen. 2012. Detecting Word Ordering Errors in Chinese Sentences for Learning Chi- nese as a Foreign Language. In Proceedings of the 24th International Conference on Computational Linguis- tics, pages 3003-3018, Mumbai, India.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "text": "Overview of word ordering error detection and correction.", |
| "num": null, |
| "uris": null, |
| "type_str": "figure" |
| }, |
| "FIGREF1": { |
| "text": "Transposition distance among the original sentences and two corrections.", |
| "num": null, |
| "uris": null, |
| "type_str": "figure" |
| }, |
| "FIGREF2": { |
| "text": "An example for ERR range and ERR words .", |
| "num": null, |
| "uris": null, |
| "type_str": "figure" |
| }, |
| "TABREF0": { |
| "text": "dynamic composition corpus created by Beijing Language and Culture University is adopted. It contains the Chinese composition articles written by non-native Chinese learners. There are 11,569 articles and 4.24 million characters in 29 composition topics. Composition articles are scanned into text and annotated with tags of error types ranging from character level, word level, sentence level, to discourse level. There are 35,884 errors at sentence level, and WOEs are the most frequent type at this level. Total 8,515 sentences are annotated with WOEs.", |
| "content": "<table><tr><td>HSK</td><td/><td/><td/><td/><td/><td/></tr><tr><td>\u6211 \u53eb \u738b\u5927\u5b89 \uff0c \u4eca\u5e74</td><td>\u590f\u5929</td><td>\u6bd5\u4e1a</td><td colspan=\"2\">\u4e86 \u5927\u5b66</td><td>\uff0c \u73b0\u5728</td><td>\u627e \u5de5\u4f5c</td><td>\u3002</td></tr><tr><td>( I / am /Wang Daan/ , /this</td><td colspan=\"5\">/summer /graduated/le /college /, /now</td><td colspan=\"2\">/look for/job /.)</td></tr><tr><td colspan=\"4\">(2) Chunk a sentence into segments by punctuation marks.</td><td/><td/><td/></tr><tr><td>\u6211 \u53eb \u738b\u5927\u5b89 \uff0c \u4eca\u5e74</td><td>\u590f\u5929</td><td>\u6bd5\u4e1a</td><td colspan=\"2\">\u4e86 \u5927\u5b66</td><td>\uff0c \u73b0\u5728</td><td>\u627e \u5de5\u4f5c</td><td>\u3002</td></tr><tr><td>\u6211 \u53eb \u738b\u5927\u5b89 \uff0c \u4eca\u5e74</td><td>\u590f\u5929</td><td>\u6bd5\u4e1a</td><td colspan=\"2\">\u4e86 \u5927\u5b66</td><td>\uff0c \u73b0\u5728</td><td>\u627e \u5de5\u4f5c</td><td>\u3002</td></tr><tr><td colspan=\"5\">(4) Reorder words in an erroneous segment and generate candidates.</td><td/><td/></tr><tr><td>\u6211 \u53eb \u738b\u5927\u5b89 \uff0c \u4eca\u5e74</td><td>\u590f\u5929</td><td>\u6bd5\u4e1a</td><td>\u5927\u5b66</td><td colspan=\"2\">\u4e86 \uff0c \u73b0\u5728</td><td>\u627e \u5de5\u4f5c</td><td>\u3002</td></tr><tr><td>\u2026</td><td/><td/><td/><td/><td/><td/></tr><tr><td>\u6211 \u53eb \u738b\u5927\u5b89 \uff0c \u4eca\u5e74</td><td>\u590f\u5929</td><td>\u5927\u5b66</td><td>\u6bd5\u4e1a</td><td colspan=\"2\">\u4e86 \uff0c \u73b0\u5728</td><td>\u627e \u5de5\u4f5c</td><td>\u3002</td></tr><tr><td colspan=\"7\">(5) Rank candidates and suggest correct word ordering by Ranking SVM-based methods.</td></tr><tr><td>\u6211 \u53eb \u738b\u5927\u5b89 \uff0c \u4eca\u5e74</td><td>\u590f\u5929</td><td>\u5927\u5b66</td><td>\u6bd5\u4e1a</td><td colspan=\"2\">\u4e86 \uff0c \u73b0\u5728</td><td>\u627e \u5de5\u4f5c</td><td>\u3002</td></tr></table>", |
| "html": null, |
| "num": null, |
| "type_str": "table" |
| }, |
| "TABREF1": { |
| "text": "Yang and Lin, 2010) is adopted to get the frequencies of Chinese words for f PMI , f Diff and f LM . for normalization, where n denotes the segment length. The segment PMI values are partitioned into intervals by equal frequency discretization. Feature f PMI of the segment Seg i reflects the label of the interval to which PMI(Seg i ) belongs. Diff ) Feature f Diff captures the PMI difference between two segments Seg j-1 and Seg j . It aims to measure the coherence between segments. The feature setting is also based on equal frequency discretization.(3) Language Model (f LM )Feature f LM uses bigram language model to measure the log probability of the words in a segment defined below. Labels of interval are also determined by equal frequency discretization.", |
| "content": "<table><tr><td colspan=\"8\">lists the distribution of B, I and O segments. Recall that two Chinese native speakers are</td></tr><tr><td colspan=\"6\">asked to correct the 1,150 sentences, thus we have two sets of B-I-O tagging.</td><td/><td/></tr><tr><td>Tagging\uf0ae</td><td>B Tag</td><td/><td>I Tag</td><td/><td>O Tag</td><td/><td>Total</td></tr><tr><td colspan=\"8\">Statistics\uf0ae #Segments Percentage #Segments Percentage #Segments Percentage Segments</td></tr><tr><td>Annotator 1</td><td>1111</td><td>40.6%</td><td>53</td><td>1.9%</td><td>1572</td><td>57.5%</td><td>2736</td></tr><tr><td>Annotator 2</td><td>1097</td><td>40.1%</td><td>59</td><td>2.2%</td><td>1580</td><td>57.7%</td><td>2736</td></tr><tr><td/><td colspan=\"5\">Table 1: Distribution of B, I, and O segments.</td><td/><td/></tr></table>", |
| "html": null, |
| "num": null, |
| "type_str": "table" |
| }, |
| "TABREF3": { |
| "text": "", |
| "content": "<table/>", |
| "html": null, |
| "num": null, |
| "type_str": "table" |
| }, |
| "TABREF5": { |
| "text": "", |
| "content": "<table/>", |
| "html": null, |
| "num": null, |
| "type_str": "table" |
| }, |
| "TABREF7": { |
| "text": "Performance of candidate ranking", |
| "content": "<table/>", |
| "html": null, |
| "num": null, |
| "type_str": "table" |
| } |
| } |
| } |
| } |
