A Convolution Kernel Approach to Identifying Comparisons in Text

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 376–386, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics A Convolution Kernel Approach to Identifying Comparisons in Text Maksim Tkachenko School of Information Systems Singapore Management University maksim.tkatchenko@gmail.com Hady W. Lauw School of Information Systems Singapore Management University hadywlauw@smu.edu.sg Abstract Comparisons in text, such as in online re- views, serve as useful decision aids. In this paper, we focus on the task of iden- tifying whether a comparison exists be- tween a specific pair of entity mentions in a sentence. This formulation is trans- formative, as previous work only seeks to determine whether a sentence is com- parative, which is presumptuous in the event the sentence mentions multiple en- tities and is comparing only some, not all, of them. Our approach leverages not only lexical features such as salient words, but also structural features expressing the re- lationships among words and entity men- tions. To model these features seamlessly, we rely on a dependency tree representa- tion, and investigate the applicability of a series of tree kernels. This leads to the de- velopment of a new context-sensitive tree kernel: Skip-node Kernel (SNK). We fur- ther describe both its exact and approxi- mate computations. Through experiments on real-life datasets, we evaluate the effec- tiveness of our kernel-based approach for comparison identification, as well as the utility of SNK and its approximations. 1 Introduction When weighing various alternatives, users in- creasingly turn to the social media, by scouring online reviews, discussion forums, etc. Our goal is to extract from such corpora those text snip- pets where users make direct comparisons of en- tities. While sentiment analysis (Pang and Lee, 2008) may be helpful in evaluating individual en- tities, comparison by the same author within a sen- tence provides an unambiguous and more equi- table basis for the relative positions of two enti- ties on some aspect. For example, the sentence s1 in Table 1, taken from an Amazon review about a digital camera, makes two distinct comparisons: #1) between “A630” and “A-series cameras” and #2) between “A630” and “its competition”, with a clear sense of which entity mention is the greater on some aspect (“larger”). Moreover, comparisons may be objective (e.g., larger) or subjective (e.g., better), while sentiments are primarily subjective. Problem Given a sentence and a specific pair of entity mentions, we seek to determine if a com- parison exists between those two mentions. In pre- vious work, the problem was formulated as identi- fying comparative sentences, i.e., those containing at least one comparison (Jindal and Liu, 2006a). This is not ideal because a sentence may contain more than two entity mentions, and may be com- paring only some of them. For instance, s1 is com- parative with respect to the pair (A630, A-series cameras) and the pair (A630, its competition), but not the pair (A-series cameras, its competition). We therefore postulate that the more appropri- …

Content truncated to 500 words