Open Information Extraction from Conjunctive Sentences

Proceedings of the 27th International Conference on Computational Linguistics, pages 2288–2299 Santa Fe, New Mexico, USA, August 20-26, 2018. 2288 Open Information Extraction from Conjunctive Sentences Swarnadeep Saha∗ IBM Research - India swarnads@in.ibm.com Mausam Indian Institute of Technology, Delhi mausam@cse.iitd.ac.in Abstract We develop CALM, a coordination analyzer that improves upon the conjuncts identified from dependency parses. It uses a language model based scoring and several linguistic constraints to search over hierarchical conjunct boundaries (for nested coordination). By splitting a conjunctive sentence around these conjuncts, CALM outputs several simple sentences. We demonstrate the value of our coordination analyzer in the end task of Open Information Extraction (Open IE). State-of-the-art Open IE systems lose substantial yield due to ineffective processing of conjunc- tive sentences. Our Open IE system, CALMIE, performs extraction over the simple sentences identified by CALM to obtain up to 1.8x yield with a moderate increase in precision compared to extractions from original sentences. 1 Introduction Open Information Extraction (Open IE) (Etzioni et al., 2008) extracts relational tuples from text in an unsupervised domain-independent manner, by identifying relational phrases and arguments from the sentences themselves. Recent work (Saha et al., 2017) has highlighted the lack of proper conjunction processing as the most significant source of missed yield in Open IE. We found Open IE 4.2 (Christensen et al., 2011; Pal and Mausam, 2016) and ClausIE (Corro and Gemulla, 2013) to frequently miss important extractions due to conjunctive relation phrases (see Table 1), and occasionally output conjunctive argu- ments, which are not ideal for readability or downstream applications (Angeli et al., 2015; Stanovsky et al., 2016a). Most modern Open IE systems process dependency parses to obtain extractions. However, dependency parsers frequently make errors in resolving coordination ambiguity. Predicting the correct conjunct span is considered to be one of the biggest challenges for parsers (Ficler and Goldberg, 2016). The state of the art approach by Ficler and Goldberg (2016) trains an LSTM-based network for predicting the boundaries for the two conjuncts on either side of the coordinating conjunction, but does not handle cases where a conjunction coordinates more than two conjuncts. Contributions: We propose a novel coordination analyzer called CALM (Coordination Analyzer us- ing Language Model), which corrects the typical errors made by dependency parsers (specifically Clear parser, which is used in Open IE 4.2) using additional features and linguistic constraints. Under the in- tuition that one can split a conjunction to form multiple coherent simple sentences (see Table 2), CALM scores each simple sentence using a (modified) language model. It additionally employs several linguis- tic constraints to reduce errors further. An important linguistic constraint is that multiple coordination structures in a sentence must either be disjoint or completely nested. A key contribution of our work is operationalizing this constraint through a novel hierarchical coordination tree, which is helpful in sentences with multiple coordinations. Experiments on 577 conjunctive sentences in British News Cor- pus demonstrate that CALM improves upon the conjuncts from the parser, with significant benefits for sentences with multiple coordinations. …

Content truncated to 500 words