Collocation Extraction Using Monolingual Word Alignment Method

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 487–495, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Collocation Extraction Using Monolingual Word Alignment Method Zhanyi Liu1,2, Haifeng Wang2, Hua Wu2, Sheng Li1 1Harbin Institute of Technology, Harbin, China 2Toshiba (China) Research and Development Center, Beijing, China {liuzhanyi,wanghaifeng,wuhua}@rdc.toshiba.com.cn lisheng@hit.edu.cn Abstract Statistical bilingual word alignment has been well studied in the context of machine trans- lation. This paper adapts the bilingual word alignment algorithm to monolingual scenario to extract collocations from monolingual cor- pus. The monolingual corpus is first repli- cated to generate a parallel corpus, where each sentence pair consists of two identical sentences in the same language. Then the monolingual word alignment algorithm is employed to align the potentially collocated words in the monolingual sentences. Finally the aligned word pairs are ranked according to refined alignment probabilities and those with higher scores are extracted as colloca- tions. We conducted experiments using Chi- nese and English corpora individually. Com- pared with previous approaches, which use association measures to extract collocations from the co-occurring word pairs within a given window, our method achieves higher precision and recall. According to human evaluation in terms of precision, our method achieves absolute improvements of 27.9% on the Chinese corpus and 23.6% on the English corpus, respectively. Especially, we can ex- tract collocations with longer spans, achiev- ing a high precision of 69% on the long-span (>6) Chinese collocations. 1 Introduction Collocation is generally defined as a group of words that occur together more often than by chance (McKeown and Radev, 2000). In this pa- per, a collocation is composed of two words oc- curring as either a consecutive word sequence or an interrupted word sequence in sentences, such as "by accident" or "take … advice". The collo- cations in this paper include phrasal verbs (e.g. "put on"), proper nouns (e.g. "New York"), idi- oms (e.g. "dry run"), compound nouns (e.g. "ice cream"), correlative conjunctions (e.g. "either … or"), and the other commonly used combinations in following types: verb+noun, adjective+noun, adverb+verb, adverb+adjective and adjec- tive+preposition (e.g. "break rules", "strong tea", "softly whisper", "fully aware", and "fond of"). Many studies on collocation extraction are carried out based on co-occurring frequencies of the word pairs in texts (Choueka et al., 1983; Church and Hanks, 1990; Smadja, 1993; Dun- ning, 1993; Pearce, 2002; Evert, 2004). These approaches use association measures to discover collocations from the word pairs in a given win- dow. To avoid explosion, these approaches gen- erally limit the window size to a small number. As a result, long-span collocations can not be extracted1. In addition, since the word pairs in the given window are regarded as potential col- locations, lots of false collocations exist. Al- though these approaches used different associa- tion measures to filter those false collocations, the precision of the extracted collocations is not high. The above problems could be partially solved by introducing more resources into collo- cation extraction, such as chunker (Wermter and Hahn, 2004), parser (Lin, …

Content truncated to 500 words