Identifying Semantic Relations in Context: Near-misses and Overlaps

International Conference RANLP 2009 - Borovets, Bulgaria, pages 381–387 Identifying Semantic Relations in Context: Near-misses and Overlaps Alla Rozovskaya and Roxana Girju University of Illinois at Urbana-Champaign {rozovska,girju}@illinois.edu Abstract This paper addresses the problem of semantic relation identification for a set of relations difficult to differen- tiate: near-misses and overlaps. Based on empirical observations on a fairly large dataset of such exam- ples we provide an analysis and a taxonomy of such cases. Using this taxonomy we create various contin- gency sets of relations. These semantic categories are automatically identified by training and testing three state-of-the-art semantic classifiers employing various feature sets. The results show that in order to identify such near-misses and overlaps accurately, a seman- tic relation identification system needs to go beyond the ontological information of the two nouns and rely heavily on contextual and pragmatic knowledge. Keywords lexical semantics; semantic relations; machine learning 1 Introduction Although semantic relations have been studied for a long time both in linguistics and natural language processing, they received special attention recently due to research done in various knowledge-rich tasks such as question an- swering [3, 17], information retrieval [11], and textual en- tailment [24]. The identification of semantic relations between nomi- nals is the task of recognizing the relationship between two nouns in context. For example, the noun pair (cycling, hap- piness) encodes a Cause-Effect relation in the sentence He derives great joy and happiness from cycling. This task re- quires several local and global decisions needed for relation identification. This involves the meaning of the two noun entities along with the meaning of other words in context. The problem, while simple to state is hard to solve. The reason is that the meaning encoded by the two nouns is not always explicitely stated in context. Despite the en- couraging results obtained by the participating systems at the SemEval-2007 - Task 4: Classification of Semantic Re- lations between Nominals [9], the problem needs further analysis. For example, the set of semantic relations consid- ered for this problem needs to be better understood. Thus, a more thorough analysis of semantic relations needs to be done before building systems capable of recognizing them automatically in context. Particular attention should be given to those semantic relations that are difficult to dif- ferentiate (near-misses) and those relations that coexist in some particular contexts (overlaps). Consider for example the following sentences: (1) a. I got home and big ⟨e1⟩branches⟨/e1⟩ had fallen off the ⟨e2⟩tree⟨/e2⟩ into the driveway. b. He fell off the ⟨e1⟩tree⟨/e1⟩ and hit every ⟨e1⟩branch⟨/e1⟩ on the way down. (2) Whisk together the mustard, vinegar and two ⟨e1⟩teaspoons⟨/e1⟩ of the remaining ⟨e2⟩lemon juice⟨/e2⟩. (1)a and (1)b are near-misses since the same noun-noun concept pair branch - tree encodes Origin-Entity in (1)a and Part-Whole in (1)b (the branches are still part of the tree). In example (2) the noun-noun concept pair teaspoon - lemon juice encodes Part-Whole (in particular Portion- Mass, a subtype of Part-Whole relations), but also Measure and Content-Container, so these three …

Content truncated to 500 words