Acquiring Strongly-related Events using Predicate-argument Co-occurring Statistics and Case Frames

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1028–1036, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Acquiring Strongly-related Events using Predicate-argument Co-occurring Statistics and Case Frames Tomohide Shibata and Sadao Kurohashi Graduate School of Informatics, Kyoto University Yoshida-honmachi, Sakyo-ku, Kyoto, 606-8501, Japan {shibata,kuro}@i.kyoto-u.ac.jp Abstract This paper proposes a method for automat- ically acquiring strongly-related events from a large corpus using predicate- argument co-occurring statistics and case frames. The strongly-related events are acquired in the form of strongly-related two predicates with their relevant ar- guments. First, strongly-related events are acquired from predicate-argument co- occurring statistics. Then, the remaining argument alignment is performed by us- ing case frames. We conducted experi- ments using a Web corpus consisting of 1.6G sentences. The accuracy for the ex- tracted event pairs was 96%, and the accu- racy of the argument alignment was 79%. The number of acquired event pairs was about 20 thousands. 1 Introduction Natural language understanding requires a wide variety of knowledge. One is the relation be- tween predicate and argument. This relation has been automatically acquired in the form of case frames from a large corpus, and is utilized for pars- ing (Kawahara and Kurohashi, 2006). Another is the relation between events. The relation between events includes temporal relation, causality, and so on, and is useful for coreference resolution (Bean and Riloff, 2004) and anaphora resolution (Gerber and Chai, 2010). This paper extracts two strongly-related events. Since the meaning of a predicate itself is often am- biguous, an event is treated as predicate-argument structure, namely the predicate with their relevant arguments. An example of two strongly-related events is shown below1: 1nom, acc, dat denotes nominative, accusative, dative, respectively. PA1 PA2 P1: pick up P2: bring nom acc A1:{man, person} A2:{purse} ⇒ nom acc dat A1:{man, person} A2:{purse} A3:{police} In the above example, while the argument A1 and A2 appear both in PA1 and PA2, the argu- ment A3 appears only in PA2. The argument A3 works for specifying the meaning of the predicate P2. The method that automatically extracts sets of events from unlabeled corpora (Chambers and Ju- rafsky, 2008; Chambers and Jurafsky, 2009) relies on the coreference relation of arguments, and thus cannot extract an argument such as A3. In languages where an argument is often omit- ted, such as Japanese, sentences illustrating the above two events usually occur in the following form (for simplicity, the sentences are explained in English): (1) a. A man picked up a purse and brought (φ) to the police. b. (φ) picked up a purse and brought (φ) to the police. In the sentence (1-a), the argument A1 and A2 are omitted in PA2. Moreover, as an agent is specifi- cally omitted, in the sentence (1-b), the argument A1 in PA1 is also omitted. The coreference-based method (Chambers and Jurafsky, 2008; Chambers and Jurafsky, 2009) is hard to be applied to such a language since an argument rarely appears in both PA1 and PA2. Our proposed method extracts …

Content truncated to 500 words