Three Knowledge-Free Methods for Automatic Lexical Chain Extraction

Proceedings of NAACL-HLT 2013, pages 989–999, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Three Knowledge-Free Methods for Automatic Lexical Chain Extraction Steffen Remus and Chris Biemann FG Language Technology Department of Computer Science Technische Universit¨at Darmstadt remus@kdsl.informatik.tu-darmstadt.de, biem@cs.tu-darmstadt.de Abstract We present three approaches to lexical chain- ing based on the LDA topic model and eval- uate them intrinsically on a manually anno- tated set of German documents. After motivat- ing the choice of statistical methods for lexi- cal chaining with their adaptability to different languages and subject domains, we describe our new two-level chain annotation scheme, which rooted in the concept of cohesive har- mony. Also, we propose a new measure for direct evaluation of lexical chains. Our three LDA-based approaches outperform two knowledge-based state-of-the art methods to lexical chaining by a large margin, which can be attributed to lacking coverage of the knowl- edge resource. Subsequent analysis shows that the three methods yield a different chaining behavior, which could be utilized in tasks that use lexical chaining as a component within NLP applications. 1 Introduction A text that is understandable by its nature exhibits an underlying structure which makes the text co- herent; that is, the structure is responsible for mak- ing the text “hang” together (Halliday and Hasan, 1976). The theoretic foundation of this structure is defined as coherence and cohesion. While the for- mer is concerned with the meaning of a text, the lat- ter can be seen as a collection of devices for cre- ating it. Cohesion and coherence build the basis for most of the current natural language processing problems that deal with text understanding. Lex- ical cohesion ties together words or phrases that are semantically related. Once all the cohesive ties are identified the involved items can be grouped to- gether to form so-called lexical chains, which form a theoretically well-founded building block in various natural language processing applications, such as word sense disambiguation (Okumura and Honda, 1994), summarization (Barzilay and Elhadad, 1997), malapropism detection and correction (Hirst and St- Onge, 1998), document hyperlinking (Green, 1996), text segmentation (Stokes et al., 2004), topic track- ing (Carthy, 2004), and others. The performance of the individual task heavily depends on the quality of the identified lexical chains. 1.1 Motivation for Corpus-driven Approach Previous approaches mainly focus on the use of knowledge resources like lexical semantic databases (Hirst and St-Onge, 1998) or thesauri (Morris and Hirst, 1991) as background information in order to resolve possible semantic relations. A major draw- back of this strategy is the dependency on the cov- erage of the resource, which has a direct impact on the lexical chains. Their quality can be expected to be poor for resource-scarce languages or specialized application domains. Statistical methods to modeling language seman- tics have proven to deliver good results in many nat- ural language processing applications. In particu- lar, probabilistic topic models have been success- fully used for tasks such as summarization (Gong and Liu, 2001; Hennig, 2009), text segmentation (Misra et …

Content truncated to 500 words