PP-Attachment Disambiguation Boosted by a Gigantic Volume of Unambiguous Examples

PP-Attachment Disambiguation Boosted by a Gigantic Volume of Unambiguous Examples Daisuke Kawahara and Sadao Kurohashi Graduate School of Information Science and Technology, University of Tokyo, 7-3-1 Hongo Bunkyo-ku, Tokyo, 113-8656, Japan {kawahara, kuro}@kc.t.u-tokyo.ac.jp Abstract. We present a PP-attachment disambiguation method based on a gigantic volume of unambiguous examples extracted from raw cor- pus. The unambiguous examples are utilized to acquire precise lexical preferences for PP-attachment disambiguation. Attachment decisions are made by a machine learning method that optimizes the use of the lexical preferences. Our experiments indicate that the precise lexical preferences work effectively. 1 Introduction For natural language processing (NLP), resolving various ambiguities is a fun- damental and important issue. Prepositional phrase (PP) attachment ambigu- ity is one of the structural ambiguities. Consider, for example, the following sentences [1]: (1) a. Mary ate the salad with a fork. b. Mary ate the salad with croutons. The prepositional phrase in (1a) “with a fork” modifies the verb “ate”, because “with a fork” describes how the salad is eaten. The prepositional phrase in (1b) “with croutons” modifies the noun “the salad”, because “with croutons” de- scribes the salad. To disambiguate such PP-attachment ambiguity, some kind of world knowledge is required. However, it is currently difficult to give such world knowledge to computers, and this situation makes PP-attachment disambigua- tion difficult. Recent state-of-the-art parsers perform with the practical accuracy, but seem to suffer from the PP-attachment ambiguity [2, 3]. For NLP tasks including PP-attachment disambiguation, corpus-based ap- proaches have been the dominant paradigm in recent years. They can be divided into two classes: supervised and unsupervised. Supervised methods automati- cally learn rules from tagged data, and achieve good performance for many NLP tasks, especially when lexical information, such as words, is given. Such methods, however, cannot avoid the sparse data problem. This is because tagged data are not sufficient enough to discriminate a large variety of lexical information. To deal with this problem, many smoothing techniques have been proposed. R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 188–198, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 PP-Attachment Disambiguation Boosted by Unambiguous Examples 189 The other class for corpus-based approaches is unsupervised learning. Unsu- pervised methods take advantage of a large number of data that are extracted from large raw corpora, and thus can alleviate the sparse data problem. How- ever, the problem is their low performance compared with supervised methods, because of the use of unreliable information. For PP-attachment disambiguation, both supervised and unsupervised meth- ods have been proposed, and supervised methods have achieved better perfor- mance (e.g., 86.5% accuracy by [1]). Previous unsupervised methods tried to ex- tract reliable information from large raw corpora, but the extraction heuristics seem to be inaccurate [4, 5]. For example, Ratnaparkhi extracted unambiguous word triples of (verb, preposition, noun) or (noun, preposition, noun), and re- ported that their accuracy was 69% [4]. This means that the extracted triples are not truly unambiguous, and this inaccurate treatment may have led to low PP-attachment performance (81.9%). This …

Content truncated to 500 words