Too Many Mammals: Improving the Diversity of Automatically Recognized Terms

International Conference RANLP 2009 - Borovets, Bulgaria, pages 490–495 Too Many Mammals: Improving the Diversity of Automatically Recognized Terms Ziqi Zhang, Lei Xia, Mark A. Greenwood and Jos´e Iria The Department of Computer Science The University of Sheffield, United Kingdom {initial.surname}@dcs.shef.ac.uk Abstract Automatic Term Recognition systems extract domain-specific terms from text corpora. Un- fortunately current systems fail to capture the whole of the domain covered by a corpus. To address this problem, we present a novel term re-ranking method that generates term lists con- taining terms that are not only individually salient, but also contribute to a globally diverse list that is truly representative of the corpus. We show that, even without any prior knowl- edge about the domain, our proposed method improves the diversity of the results produced by two popular automatic term recognition algo- rithms. Keywords Automatic Term Recognition, Diversity in Ranking, Random Walk, Semantic Similarity 1 Introduction Automatic Term Recognition (ATR) is an important research area that deals with the recognition and ex- traction of technical terms from domain-specific cor- pora. ATR is often a processing step preceding more complex tasks, such as semantic search or ontology en- gineering [14, 3]. It can also be used as an end-user tool to, for instance, generate a list of terms that sum- marizes a text corpus provided by the user. Whilst state-of-the-art ATR algorithms are rea- sonably successful in identifying the most relevant domain-specific terms from a corpus, the analysis we carried out of the output of these algorithms over sev- eral corpora led us to conclude that the ranking of terms rarely reflects the whole domain. Having carefully studied the experimental outputs from [25], we have observed that terms from a subset of the sub-domains (of the domain covered by the corpus) tend to dominate the results, pushing other character- istic terms, which are perhaps not so globally relevant but nevertheless fundamental to get a comprehensive coverage of the domain, far down the ranking. For example, on the Wikipedia animal corpus described in Section 5, which contains 1051 random Wikipedia articles describing animals across roughly 30 scientific classes, in the top ranked 50 terms by the C-Value [10] algorithm there are the names of 13 mammals, 3 fish, 2 birds and 2 insects. Among these, 3 of the 13 mammals are species of whale, and 2 of the 3 fish are species of shark. Likewise, for the TF-IDF algorithm applied on the same corpus, in the top ranked 50 terms we obtain the names of 18 mammals, 6 birds and 4 in- sects. Hence terms belonging to these 3 or 4 classes dominate the results, preventing the remaining classes from being properly represented. Therefore, with cur- rent methods, taking the top-ranked terms is unlikely to produce a diverse list of terms that is fully represen- tative of the entire domain. Unfortunately, for certain applications of ATR this is not a desirable behaviour, e.g., generating a list of terms that best summarizes a text corpus. In this paper, we …

Content truncated to 500 words