Combining Relational and Attributional Similarity for Semantic Relation Classification

Proceedings of Recent Advances in Natural Language Processing, pages 323–330, Hissar, Bulgaria, 12-14 September 2011. Combining Relational and Attributional Similarity for Semantic Relation Classification Preslav Nakov Department of Computer Science National University of Singapore 13 Computing Drive Singapore 117417 nakov@comp.nus.edu.sg Zornitsa Kozareva University of Southern California Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695, USA kozareva@isi.edu Abstract We combine relational and attributional similarity for the task of identifying instances of semantic relations, such as PRODUCT-PRODUCER and ORIGIN- ENTITY, between nominals in text. We use no pre-existing lexical resources, thus simulating a realistic real-world situation, where the coverage of any such resource is limited. Instead, we mine the Web to au- tomatically extract patterns (verbs, prepo- sitions and coordinating conjunctions) ex- pressing the relationship between the rela- tion arguments, as well as hypernyms and co-hyponyms of the arguments, which we use in instance-based classifiers. The eval- uation on the dataset of SemEval-1 Task 4 shows an improvement over the state-of- the-art for the case where using manually annotated WordNet senses is not allowed. 1 Introduction Recently, the natural language processing (NLP) community has shown renewed interest in the problem of deep language understanding, which was inspired by the notable progress in this impor- tant research direction in the last few years. To- day, lexical semantics tasks such as word sense disambiguation, semantic role labeling, and tex- tual entailment are already well-established and are gradually finding their way in real NLP ap- plications, while a number of new semantic tasks are emerging. One such example is the task of ex- tracting semantic relations between nominals from text, which has attracted a lot of research attention following the creation of two benchmark datasets as part of SemEval-1 Task 4 (Girju et al., 2007) and SemEval-2 Task 8 (Hendrickx et al., 2010). The ability to recognize semantic relations in text could potentially help many NLP applica- tions. For example, a question answering system facing the question What causes tumors to shrink? would need to identify the CAUSE-EFFECT rela- tion between shrinkage and radiation in order to be able to extract the answer from the following sentence: The period of tumor shrinkage after ra- diation therapy is often long and varied. One can also imagine a relational search engine that can serve queries such as “find all X such that X causes wrinkles”, asking for all entities that are in a particular relation with a given entity (Cafarella et al., 2006). Finally, modeling semantic relations has been shown to help statistical machine trans- lation (Nakov, 2008a). The task of identifying semantic relations in text is complicated by their heterogeneous nature. Thus, it is often addressed using non-parametric instance-based classifiers like the k nearest neigh- bors (kNN), which effectively reduce it to mea- suring the relational similarity between a testing and each of the training examples. The latter is studied in detail by Turney (2006), who distin- guishes between attributional similarity or corre- spondence between attributes, and relational sim- ilarity or correspondence between …

Content truncated to 500 words