Set Expansion using Sibling Relations between Semantic Categories

Copyright 2012 by Sho Takase, Naoaki Okazaki, and Kentaro Inui 26th Pacific Asia Conference on Language,Information and Computation pages 525–534 Set Expansion using Sibling Relations between Semantic Categories Sho Takase Naoaki Okazaki Kentaro Inui Graduate School of Information Sciences Tohoku University, Japan {takase, okazaki, inui}@ecei.tohoku.ac.jp Abstract Most set expansion algorithms assume to acquire new instances of different seman- tic categories independently even when we have seed instances of multiple semantic cat- egories. However, in the setting of set ex- pansion with multiple semantic categories, we might leverage other types of prior knowl- edge about semantic categories. In this paper, we present a method of set expansion when ontological information related to target se- mantic categories is available. More specifi- cally, the proposed method makes use of sib- ling relations between semantic categories as an additional type of prior knowledge. We demonstrate the effectiveness of sibling rela- tions in set expansion on the dataset in which instances and sibling relations are extracted from Wikipedia in a semi-automatic manner. 1 Introduction Set expansion is the task of expanding a list of named entities from a few named entities (seed in- stances). For example, given a few instances of car vehicles “Prius”, “Lexus” and “Insight”, the task outputs new car instances such as “Corolla”, “Civic”, and “Fit”. Set expansion has many ap- plications in NLP including named entity recogni- tion (Collins and Singer, 1999), word sense disam- biguation (Pantel and Lin, 2002), document cate- gorization (Pantel et al., 2009), and query sugges- tion (Cao et al., 2008). Set expansion is often implemented as bootstrap- ping algorithms (Hearst, 1992; Yarowsky, 1995; Ab- ney, 2004; Pantel and Ravichandran, 2004; Pantel and Pennacchiotti, 2006). A bootstrapping algo- rithm iteratively acquires new instances of the tar- get category using seed instances. First, a bootstrap- ping algorithm mines phrasal patterns that co-occur frequently with seed instances in a corpus. Given the words “Prius” and “Lexus” as seed instances of the car category, the algorithm finds patterns such as “Toyota produce X” and “X is a hybrid car” (X is a variable filled with a noun phrase). Next, the bootstrapping algorithm acquires instances that co- occur with patterns, i.e., noun phrases that appear frequently in the variable slots in the patterns. For example, the pattern “Toyota produce X” might re- trieve vehicles manufactured by Toyota. Bootstrap- ping algorithms repeat these steps, expanding pat- terns using newly acquired instances. However, bootstrapping algorithms often suffer from patterns that retrieve instances not only of the target category but also of other categories. For example, given the seed instances “Prius” and “Lexus”, a bootstrapping algorithm might choose the pattern “new type of X”, which might extract un- related instances such as “iPhone” and “ThinkPad”. The semantic drift problem (Curran et al., 2007), the phenomenon by which a bootstrapping algorithm deviates from the target category, has persisted as the major impediment of bootstrapping algorithms. Bootstrapping algorithms assume prior knowl- edge about a semantic category in the form of seed instances. Recently, researchers have been …

Content truncated to 500 words