Convolution Kernels for Opinion Holder Extraction

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 795–803, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Convolution Kernels for Opinion Holder Extraction Michael Wiegand and Dietrich Klakow Spoken Language Systems Saarland University D-66123 Saarbr¨ucken, Germany {Michael.Wiegand|Dietrich.Klakow}@lsv.uni-saarland.de Abstract Opinion holder extraction is one of the impor- tant subtasks in sentiment analysis. The ef- fective detection of an opinion holder depends on the consideration of various cues on vari- ous levels of representation, though they are hard to formulate explicitly as features. In this work, we propose to use convolution kernels for that task which identify meaningful frag- ments of sequences or trees by themselves. We not only investigate how different levels of information can be effectively combined in different kernels but also examine how the scope of these kernels should be chosen. In general relation extraction, the two candidate entities thought to be involved in a relation are commonly chosen to be the boundaries of se- quences and trees. The definition of bound- aries in opinion holder extraction, however, is less straightforward since there might be sev- eral expressions beside the candidate opinion holder to be eligible for being a boundary. 1 Introduction In recent years, there has been a growing interest in the automatic detection of opinionated content in natural language text. One of the more impor- tant tasks in sentiment analysis is the extraction of opinion holders. Opinion holder extraction is one of the critical components of an opinion question- answering system (i.e. systems which automatically answer opinion questions, such as “What does [X] like about [Y]?”). Such systems need to be able to distinguish which entities in a candidate answer sen- tence are the sources of opinions (= opinion holder) and which are the targets. On other NLP tasks, in particular, on relation extrac- tion, there has been much work on convolution ker- nels, i.e. kernel functions exploiting huge amounts of features without an explicit feature representa- tion. Previous research on that task has shown that convolution kernels, such as sequence and tree ker- nels, are quite effective when compared to manual feature engineering (Moschitti, 2008; Bunescu and Mooney, 2005; Nguyen et al., 2009). In order to effectively use convolution kernels, it is often nec- essary to choose appropriate substructures of a sen- tence rather than represent the sentence as a whole structure (Bunescu and Mooney, 2005; Zhang et al., 2006; Moschitti, 2008). As for tree kernels, for ex- ample, one typically chooses the syntactic subtree immediately enclosing two entities potentially ex- pressing a specific relation in a given sentence. The opinion holder detection task is different from this scenario. There can be several cues within a sen- tence to indicate the presence of a genuine opinion holder and these cues need not be member of a par- ticular word group, e.g. they can be opinion words (see Sentences 1-3), communication words, such as maintained in Sentence 2, or other lexical cues, such as according in Sentence …

Content truncated to 500 words