Author:
Chien-Cheng Wu, and Jason S. Chang

Abstract:
In this paper, we describe an algorithm that employs syntactic and statistical analysis to extract bilingual collocations from a parallel corpus. Collocations are pervasive in all types of writing and can be found in phrases, chunks, proper names, idioms, and terminology. Therefore, automatic extraction of monolingual and bilingual collocations is important for many applications, including natural language generation, word sense disambiguation, machine translation, lexicography, and cross language information retrieval.

Collocations can be classified as lexical or grammatical collocations. Lexical collocations exist between content words, while a grammatical collocation exists between a content word and function words or a syntactic structure. In addition, bilingual collocations can be rigid or flexible in both languages. Rigid collocation refers to words in a collocation must appear next to each other, or otherwise (flexible/elastic). We focus in this paper on extracting rigid lexical bilingual collocations. In our method, the preferred syntactic patterns are obtained from idioms and collocations in a machine-readable dictionary. Collocations matching the patterns are extracted from aligned sentences in a parallel corpus. We use a new alignment method based on punctuation statistics for sentence alignment. The punctuation-based approach is found to outperform the length-based approach with precision rates approaching 98%. The obtained collocations are subsequently matched up based on cross-linguistic statistical association. Statistical association between the whole collocations as well as words in collocations is used to link a collocation with its counterpart collocation in the other language. We implemented the proposed method on a very large Chinese-English parallel corpus and obtained satisfactory results.

Author:
Tyne Liang and Dian-Song Wu

Abstract:
Anaphora is a common phenomenon in discourses as well as an important research issue in the applications of natural language processing. In this paper, anaphora resolution is achieved by employing WordNet ontology and heuristic rules. The proposed system identifies both intra-sentential and inter-sentential antecedents of anaphors. Information about animacy is obtained by analyzing the hierarchical relations of nouns and verbs in the surrounding context. The identification of animacy entities and pleonastic-it usage in English discourses are employed to promote resolution accuracy.

Traditionally, anaphora resolution systems have relied on syntactic, semantic or pragmatic clues to identify the antecedent of an anaphor. Our proposed method makes use of WordNet ontology to identify animate entities as well as essential gender information. In the animacy agreement module, the property is identified by the hypernym relation between entities and their unique beginners defined in WordNet. In addition, the verb of the entity is also an important clue used to reduce the uncertainty. An experiment was conducted using a balanced corpus to resolve the pronominal anaphora phenomenon. The methods proposed in [Lappin and Leass, 94] and [Mitkov, 01] focus on the corpora with only inanimate pronouns such as “it” or “its”. Thus the results of intra-sentential and inter-sentential anaphora distribution are different. In an experiment using Brown corpus, we found that the distribution proportion of intra-sentential anaphora is about 60%. Seven heuristic rules are applied in our system; five of them are preference rules, and two are constraint rules. They are derived from syntactic, semantic, pragmatic conventions and from the analysis of training data. A relative measurement indicates that about 30% of the errors can be eliminated by applying heuristic module.

Author:
Jia-Lin Tsai, Gladys Hsieh, and Wen-Lian Hsu

Abstract:
Noun-verb event frame (NVEF) knowledge in conjunction with an NVEF word-pair identifier [Tsai et al. 2002] comprises a system that can be used to support natural language processing (NLP) and natural language understanding (NLU). In [Tsai et al. 2002a], we demonstrated that NVEF knowledge can be used effectively to solve the Chinese word-sense disambiguation (WSD) problem with 93.7% accuracy for nouns and verbs. In [Tsai et al. 2002b], we showed that NVEF knowledge can be applied to the Chinese syllable-to-word (STW) conversion problem to achieve 99.66% accuracy for the NVEF related portions of Chinese sentences. In [Tsai et al. 2002a], we defined a collection of NVEF knowledge as an NVEF word-pair (a meaningful NV word-pair) and its corresponding NVEF sense-pairs. No methods exist that can fully and automatically find collections of NVEF knowledge from Chinese sentences. We propose a method here for automatically acquiring large-scale NVEF knowledge without human intervention in order to identify a large, varied range of NVEF-sentences (sentences containing at least one NVEF word-pair). The auto-generation of NVEF knowledge (AUTO-NVEF) includes four major processes: (1) segmentation checking; (2) Initial Part-of-Speech (IPOS) sequence generation; (3) NV knowledge generation; and (4) NVEF knowledge auto-confirmation.

Our experimental results show that AUTO-NVEF achieved 98.52% accuracy for news and 96.41% for specific text types, which included research reports, classical literature and modern literature. AUTO-NVEF automatically discovered over 400,000 NVEF word-pairs from the 2001 United Daily News (2001 UDN) corpus. According to our estimation, the acquired NVEF knowledge from 2001 UDN helped to identify 54% of the NVEF-sentences in the Academia Sinica Balanced Corpus (ASBC), and 60% in the 2001 UDN corpus.

We plan to expand NVEF knowledge so that it is able to identify more than 75% of NVEF-sentences in ASBC. We will also apply the acquired NVEF knowledge to support other NLP and NLU researches, such as machine translation, shallow parsing, syllable and speech understanding and text indexing. The auto-generation of bilingual, especially Chinese-English, NVEF knowledge will be also addressed in our future work.

Keyword:
natural language understanding, verb-noun collection, machine learning, HowNet

Author:
Tzong-Han Tsai, Shih-Hung Wu, Cheng-Wei Lee, Cheng-Wei Shih, and Wen-Lian Hsu

Abstract:
This paper presents a Chinese named entity recognizer (NER): Mencius. It aims to address Chinese NER problems by combining the advantages of rule-based and machine learning (ML) based NER systems. Rule-based NER systems can explicitly encode human comprehension and can be tuned conveniently, while ML-based systems are robust, portable and inexpensive to develop. Our hybrid system incorporates a rule-based knowledge representation and template-matching tool, called InfoMap [Wu et al. 2002], into a maximum entropy (ME) framework. Named entities are represented in InfoMap as templates, which serve as ME features in Mencius. These features are edited manually, and their weights are estimated by the ME framework according to the training data. To understand how word segmentation might influence Chinese NER and the differences between a pure template-based method and our hybrid method, we configure Mencius using four distinct settings. The F-Measures of person names (PER), location names (LOC) and organization names (ORG) of the best configuration in our experiment were respectively 94.3%, 77.8% and 75.3%. From comparing the experiment results obtained using these configurations reveals that hybrid NER Systems always perform better performance in identifying person names. On the other hand, they have a little difficulty identifying location and organization names. Furthermore, using a word segmentation module improves the performance of pure Template-based NER Systems, but, it has little effect on hybrid NER systems.

Author:
Yu-Fang Tsai, and Keh-Jiann Chen

Abstract:
In order to achieve fast, high quality Part-of-speech (pos) tagging, algorithms should achieve high accuracy and require less manually proofreading. This study aimed to achieve these goals by defining a new criterion of tagging reliability, the estimated final accuracy of the tagging under a fixed amount of proofreading, to be used to judge how cost-effective a tagging algorithm is. In this paper, we also propose a new tagging algorithm, called the context-rule model, to achieve cost-effective tagging. The context rule model utilizes broad context information to improve tagging accuracy. In experiments, we compared the tagging accuracy and reliability of the context-rule model, Markov bi-gram model and word-dependent Markov bi-gram model. The result showed that the context-rule model outperformed both Markov models. Comparing the models based on tagging accuracy, the context-rule model reduced the number of errors 20% more than the other two Markov models did. For the best cost-effective tagging algorithm to achieve 99% tagging accuracy, it was estimated that, on average, 20% of the samples of ambiguous words needed to be rechecked. We also compared tradeoff between the amount of proofreading needed and final accuracy for the different algorithms. It turns out that an algorithm with the highest accuracy may not always be the most reliable algorithm.

Keyword:
part-of-speech tagging, corpus, reliability, ambiguous resolution

Author:
林頌堅 (Sung-Chen Lin)

Abstract:
本論文針對主題抽取的問題，提出一系列以自然語言處理為基礎的技術，應用這些技術可以從學術論文抽取重要的術語，並將這些術語依據彼此間的共現關係進行叢集，以叢集所得到的術語集合表示領域中重要的主題，提供研究人員學術領域的梗概並釐清他們的資訊需求。我們將所提出的方法應用到ROCLING研討會的論文資料上，結果顯示這個方法可以同時抽取出計算語言學領域的中文和英文術語，所得到的術語叢集結果也可以表示領域中重要的主題。這個初步的研究驗證了本論文所提出方法的可行性。重要的主題包括機器翻譯、語音處理、資訊檢索、語法模式與剖析、斷詞和統計式語言模型等等。從研究結果中，我們也發現計算語言學研究與實務應用有密切的關係。

In this paper, we propose a series of natural language processing techniques to be used to extract important topics in a given research field. Topics as defined in this paper are important research problems, theories, and technical methods of the examined field, and we can represent them with groups of relevant terms. The terms are extracted from the texts of papers published in the field, including titles, abstracts, and bibliographies, because they convey important research information and are relevant to knowledge in that field. The topics can provide a clear outline of the field for researchers and are also useful for identifying users’ information needs when they are applied to information retrieval. To facilitate topic extraction, key terms in both Chinese and English are extracted from papers and are clustered into groups consisting of terms that frequently co-occur with each other. First, a PAT-tree is generated that stores all possible character strings appearing in the texts of papers. Character strings are retrieved from the PAT-tree as candidates of extracted terms and are tested using the statistical information of the string to filter out impossible candidates. The statistical information for a string includes (1) the total frequency count of the string in all the input papers, (2) the sum of the average frequency and the standard deviation of the string in each paper, and (3) the complexity of the front and rear adjacent character of the string. The total frequency count of the string and the sum of its average frequency and standard deviation are used to measure the importance of the corresponding term to the field. The complexity of adjacent characters is a criterion used to determine whether the string is a complete token of a term. The less complexity the adjacent characters, the more likely the string is a partial token of other terms. Finally, if the leftmost or rightmost part of a string is a stop word, the string is also filtered out. The extracted results are clustered to generate term groups according to their co-occurrences. Several techniques are used in the clustering algorithm to obtain multiple clustering results, including the clique algorithm and a group merging procedure. When the clique algorithm is performed, the latent semantic indexing technique is used to estimate the relevance between two terms to improve the deficiency of term co-occurrences in the papers. Two term groups are further merged into a new one when their members are similar because it is possible that the clusters represent the same topic. The above techniques were applied to the proceedings of ROCLING to uncover topics in the field of computational linguistics. The results show that the key terms in both Chinese and English were extracted successfully, and that the clustered groups represented the topics of computational linguistics. Therefore, the initial study proved the feasibility of the proposed techniques. The extracted topics included “machine translation,” “speech processing,” “information retrieval,” “grammars and parsers,” “Chinese word segmentation,” and “statistical language models.” From the results, we can observe that there is a close relation between basic research and applications in computational linguistics.

Keyword:
主題抽取、術語抽取、術語叢集 (Topic extraction, term extraction, term clustering)

Author:
Yih-Jeng Lin, and Ming-Shing Yu

Abstract:
This paper reveals some important properties of CFSs and applications in Chinese natural language processing (NLP). We have previously proposed a method for extracting Chinese frequent strings that contain unknown words from a Chinese corpus [Lin and Yu 2001]. We found that CFSs contain many 4-character strings, 3-word strings, and longer n-grams. Such information can only be derived from an extremely large corpus using a traditional language model(LM). In contrast to using a traditional LM, we can achieve high precision and efficiency by using CFSs to solve Chinese toneless phoneme-to-character conversion and to correct Chinese spelling errors with a small training corpus. An accuracy rate of 92.86% was achieved for Chinese toneless phoneme-to-character conversion, and an accuracy rate of 87.32% was achieved for Chinese spelling error correction. We also attempted to assign syntactic categories to a CFS. The accuracy rate for assigning syntactic categories to the CFSs was 88.53% for outside testing when the syntactic categories of the highest level were used.

Keyword:
Chinese frequent strings, unknown words, Chinese toneless phoneme-to-character, Chinese spelling error correction, language model.