Retrieving Bilingual Verb-Noun Collocations by Integrating Cross-Language Category Hierarchies

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 233–240 Manchester, August 2008 Retrieving Bilingual Verb–noun Collocations by Integrating Cross-Language Category Hierarchies Fumiyo Fukumoto Yoshimi Suzuki Kazuyuki Yamashita∗ Interdisciplinary Graduate School of Medicine and Engineering Univ. of Yamanashi {fukumoto, ysuzuki}@yamanashi.ac.jp ∗The Center for Educational Research Faculty of Education and Human Sciences Univ. of Yamanashi kazuyuki@yamanashi.ac.jp Abstract This paper presents a method of retriev- ing bilingual collocations of a verb and its objective noun from cross-lingual docu- ments with similar contents. Relevant doc- uments are obtained by integrating cross- language hierarchies. The results showed a 15.1% improvement over the baseline non- hierarchy model, and a 6.0% improvement over use of relevant documents retrieved from a single hierarchy. Moreover, we found that some of the retrieved colloca- tions were domain-specific. 1 Introduction A bilingual lexicon is important for cross-lingual NLP applications, such as CLIR, and multilingual topic tracking. Much of the previous work on find- ing bilingual lexicons has made use of comparable corpora, which exhibit various degrees of paral- lelism. Fung et al. (2004) described corpora rang- ing from noisy parallel, to comparable, and finally to very non-parallel. Obviously, the latter are easy to collect because very non-parallel corpora con- sist of sets of documents in two different languages from the same period of dates. However, a good solution is required to produce a higher quality of lexicon retrieval. In this paper, we focus on English and Japanese bilingual verb–objective noun collocations which we call verb–noun collocations and retrieve them using very non-parallel corpora. The method first finds cross-lingual relevant document pairs with similar contents from non-parallel corpora, and c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. then we estimate bilingual verb–noun collocations within these relevant documents. Relevant doc- uments are defined here as pairs of English and Japanese documents that report identical or closely related contents, e.g., a pair of documents describ- ing an aircraft crash and the ensuing investigation to compensate the victims’ families or any safety measures proposed as a result of the crash. In the task of retrieving cross-lingual relevant docu- ments, it is crucial to identify an event as some- thing occurs at some specific place and time asso- ciated with some specific action. One solution is to use a topic, i.e., category in the hierarchical struc- ture, such as Internet directories. Although a topic is not an event, it can be a broader class of event. Therefore, it is helpful for retrieving relevant docu- ments, and thus bilingual verb–noun collocations. Consider the Reuters’96 and Mainichi newspaper documents shown in Figure 1. The documents re- port on the same event, “Russian space station col- lides with cargo craft,” were published within two days of each other, and have overlapping content. Moreover, as indicated by the double-headed ar- rows in the figure, there are a number of bilingual collocations. However, as shown in Figure 1, the Reuters document is classified into “Science and Technology,” while …

Content truncated to 500 words