Automatic induction of FrameNet lexical units

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 457–465, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Automatic induction of FrameNet lexical units Marco Pennacchiotti(†), Diego De Cao(‡), Roberto Basili(‡), Danilo Croce(‡), Michael Roth(†) (†) Computational Linguistics Saarland University Saarbr¨ucken, Germany {pennacchiotti,mroth}@coli.uni-sb.de (‡) DISP University of Roma Tor Vergata Roma, Italy {decao,basili,croce}@info.uniroma2.it Abstract Most attempts to integrate FrameNet in NLP systems have so far failed because of its lim- ited coverage. In this paper, we investigate the applicability of distributional and WordNet- based models on the task of lexical unit induc- tion, i.e. the expansion of FrameNet with new lexical units. Experimental results show that our distributional and WordNet-based models achieve good level of accuracy and coverage, especially when combined. 1 Introduction Most inference-based NLP tasks require a large amount of semantic knowledge at the predicate- argument level. This type of knowledge allows to identify meaning-preserving transformations, such as active/passive, verb alternations and nominal- izations, which are crucial in several linguistic in- ferences. Recently, the integration of NLP sys- tems with manually-built resources at the predi- cate argument-level, such as FrameNet (Baker et al., 1998) and PropBank (Palmer et al., 2005) has received growing interest. For example, Shen and Lapata (2007) show the potential improvement that FrameNet can bring on the performance of a Ques- tion Answering (QA) system. Similarly, several other studies (e.g. (Bar-Haim et al., 2005; Garoufi, 2007)) indicate that frame semantics plays a central role in Recognizing Textual Entailment (RTE). Un- fortunately, most attempts to integrate FrameNet or similar resources in QA and RTE systems have so far failed, as reviewed respectively in (Shen and La- pata, 2007) and (Burchardt and Frank, 2006). These studies indicate limited coverage as the main reason of insuccess. Indeed, the FrameNet database only contains 10,000 lexical units (LUs), far less than the 210,000 entries in WordNet 3.0. Also, frames are based on more complex information than word senses, so that their manual development is much more demanding (Burchardt et al., 2006; Subirats and Petruck, 2003). Therefore, there is nowadays a pressing need to adopt learning approaches to extend the coverage of the FrameNet lexicon by automatically acquiring new LUs, a task we call LU induction, as recently proposed at SemEval-2007 (Baker et al., 2007). Un- fortunately, research in this area is still somehow limited and fragmentary. The aim of our study is to pioneer in this field by proposing two unsuper- vised models for LU induction, one based on dis- tributional techniques and one using WordNet as a support; and a combined model which mixes the two. The goal is to investigate to what extent distri- butional and WordNet-based models can be used to induce frame semantic knowledge in order to safely extend FrameNet, thus limiting the high costs of manual annotation. In Section 2 we introduce the LU induction task and present related work. In Sections 3, 4 and 5 we present our distributional, WordNet-based and com- bined models. Then, in Section 6 we report …

Content truncated to 500 words