Investigating a Generic Paraphrase-Based Approach for Relation Extraction

Investigating a Generic Paraphrase-based Approach for Relation Extraction Lorenza Romano ITC-irst via Sommarive, 18 38050 Povo (TN), Italy romano@itc.it Milen Kouylekov ITC-irst via Sommarive, 18 38050 Povo (TN), Italy kouylekov@itc.it Idan Szpektor Department of Computer Science Bar Ilan University Ramat Gan, 52900, Israel szpekti@cs.biu.ac.il Ido Dagan Department of Computer Science Bar Ilan University Ramat Gan, 52900, Israel dagan@cs.biu.ac.il Alberto Lavelli ITC-irst via Sommarive, 18 38050 Povo (TN), Italy lavelli@itc.it Abstract Unsupervised paraphrase acquisition has been an active research field in recent years, but its effective coverage and per- formance have rarely been evaluated. We propose a generic paraphrase-based ap- proach for Relation Extraction (RE), aim- ing at a dual goal: obtaining an applicative evaluation scheme for paraphrase acquisi- tion and obtaining a generic and largely unsupervised configuration for RE. We an- alyze the potential of our approach and evaluate an implemented prototype of it using an RE dataset. Our findings reveal a high potential for unsupervised paraphrase acquisition. We also identify the need for novel robust models for matching para- phrases in texts, which should address syn- tactic complexity and variability. 1 Introduction A crucial challenge for semantic NLP applica- tions is recognizing the many different ways for expressing the same information. This seman- tic variability phenomenon was addressed within specific applications, such as question answering, information extraction and information retrieval. Recently, the problem was investigated within generic application-independent paradigms, such as paraphrasing and textual entailment. Eventually, it would be most appealing to apply generic models for semantic variability to concrete applications. This paper investigates the applica- bility of a generic “paraphrase-based” approach to the Relation Extraction (RE) task, using an avail- able RE dataset of protein interactions. RE is highly suitable for such investigation since its goal is to exactly identify all the different variations in which a target semantic relation can be expressed. Taking this route sets up a dual goal: (a) from the generic paraphrasing perspective - an objective evaluation of paraphrase acquisition performance on a concrete application dataset, as well as identi- fying the additional mechanisms needed to match paraphrases in texts; (b) from the RE perspective - investigating the feasibility and performance of a generic paraphrase-based approach for RE. Our configuration assumes a set of entailing templates (non-symmetric “paraphrases”) for the target relation. For example, for the target rela- tion “X interact with Y” we would assume a set of entailing templates as in Tables 3 and 7. In addi- tion, we require a syntactic matching module that identifies template instances in text. First, we manually analyzed the protein- interaction dataset and identified all cases in which protein interaction is expressed by an entailing template. This set a very high idealized upper bound for the recall of the paraphrase-based ap- proach for this dataset. Yet, obtaining high cover- age in practice would require effective paraphrase acquisition and lexical-syntactic template match- ing. Next, we implemented a prototype that uti- lizes a state-of-the-art method for learning en- tailment relations from the web (Szpektor et al., 2004), the …

Content truncated to 500 words