SmartWrite: Extracting Chinese Lexical Grammar Patterns Using Dependency Parsing

SmartWrite: Extracting Chinese Lexical Grammar Patterns Using Dependency Parsing Cheng-Cyuan Peng1, Ching-Yu Yang2, Jhih-Jie Chen2, Jason S. Chang2 1Institute of Information Systems and Applications National Tsing Hua University 2Department of Computer Science National Tsing Hua University {zeno, chingyu, jjc, jason}@nlplab.cc Abstract This paper presents an interactive writing pro- totype system, SmartWrite, which provides Chinese word usage in the form of grammar patterns, collocations, and examples to assist language learners in writing. We propose a method of inducing common Chinese gram- mar patterns from a large-scale Chinese cor- pus. First, we try to identify grammar pat- terns based on the dependency relations of a given sentence parsed with existing language parsers. Then, we calculate words in a depen- dency relation with a target word to automat- ically generate the collocations. Finally, we find good example sentences that best exem- plify each pattern. At run-time, the last con- tent word an user typed is used for obtaining grammar patterns, collocations, and example sentences as continuous writing suggestions. Evaluation on the suggested word usage of real verbs extracted from a Web search engine log shows that the method significantly out- performs underlying search engines, and even prevails over a commercial search engine by providing structural information of verbs. 1 Introduction Many queries are submitted by language learn- ers to online dictionaries or other reference re- sources everyday, and an increasing number of services on the Web specifically target language learning. For example, VOA special English (learningenglish.voanews.com) provides grade-level-appropriate news articles for learning how to read and speak, while Corpus of Contem- porary American English (COCA) (corpus.byu. edu/coca/) and Corpus of Contemporary Man- darin in Taiwan (COCT) (coct.naer.edu.tw/ cqpweb/) use an information retrieval system to display collocations and example sentences from a corpus. Typical linguistic search engines such as COCA and COCT usually fashion a bottom-up approach by providing a wealth of examples for inductive learn- ing. However, users may be overwhelmed by the examples and are slow to grasp the deep vocabulary knowledge including the grammar patterns of word usage. Such deep lexical knowledge can be acquired more effectively, if a system automatically can in- duce and display grammar patterns with examples on demand. Consider the query word “公佈” submitted to find lexical information. The best way to support lan- guage learning and assist writing is probably not just providing a random list of collocations and exam- ples related to “公佈”, but rather a list of grammar patterns and corresponding collocations ordered by frequency counts: • 公佈：V n [結果 名單 報告] 教育部公 公 公佈 佈 佈歐盟獎學金獲獎名 名 名單 單 單 • 公佈：V clause [為 達] 經濟部公 公 公佈 佈 佈3月外銷訂單為 為 為歷史新高 These patterns and collocations can be induced from a very large corpus by parsing and selecting a sample of prototypical sentences transforming their depen- dency relations into grammar patterns. Intuitively, by selecting a small representative sentences, we can cope with parsing errors and induce a set of reason- PACLIC 32 541 32nd Pacific Asia Conference on Language, Information and Computation Hong …

Content truncated to 500 words