Extracting Relation Descriptors with Conditional Random Fields

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 392–400, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Extracting Relation Descriptors with Conditional Random Fields Yaliang Li†, Jing Jiang†, Hai Leong Chieu‡, Kian Ming A. Chai‡ †School of Information Systems, Singapore Management University, Singapore ‡DSO National Laboratories, Singapore {ylli,jingjiang}@smu.edu.sg, {chaileon,ckianmin}@dso.org.sg Abstract In this paper we study a novel relation extraction problem where a general rela- tion type is defined but relation extrac- tion involves extracting specific relation descriptors from text. This new task can be treated as a sequence labeling problem. Although linear-chain conditional random fields (CRFs) can be used to solve this problem, we modify this baseline solution in order to better fit our task. We propose two modifications to linear-chain CRFs, namely, reducing the space of possible la- bel sequences and introducing long-range features. Both modifications are based on some special properties of our task. Using two data sets we have annotated, we eval- uate our methods and find that both modi- fications to linear-chain CRFs can signif- icantly improve the performance for our task. 1 Introduction Relation extraction is the task of identifying and characterizing the semantic relations between en- tities in text. Depending on the application and the resources available, relation extraction has been studied in a number of different settings. When relation types are well defined and labeled relation mention instances are available, supervised learn- ing is usually applied (Zelenko et al., 2003; Zhou et al., 2005; Bunescu and Mooney, 2005; Zhang et al., 2006). When relation types are known but little training data is available, bootstrapping has been used to iteratively expand the set of seed ex- amples and relation patterns (Agichtein and Gra- vano, 2000). When no relation type is pre-defined but there is a focused corpus of interest, unsu- pervised relation discovery tries to cluster en- tity pairs in order to identify interesting relation types (Hasegawa et al., 2004; Rosenfeld and Feld- man, 2006; Shinyama and Sekine, 2006). More recently, open relation extraction has also been proposed where there is no fixed domain or pre- defined relation type, and the goal is to identify all possible relations from an open-domain cor- pus (Banko and Etzioni, 2008; Wu and Weld, 2010; Hoffmann et al., 2010). These different relation extraction settings suit different applications. In this paper, we focus on another setting where the relation types are defined at a general level but a more specific relation de- scription is desired. For example, in the widely used ACE1 data sets, relation types are defined at a fairly coarse granularity. Take for instance the “employment” relation, which is a major relation type defined in ACE. In ACE evaluation, extrac- tion of this relation only involves deciding whether a person entity is employed by an organization en- tity. In practice, however, we often also want to find the exact job title or position this person holds at the organization if this information is mentioned in the text. Table 1 gives some …

Content truncated to 500 words