Fish Transporters and Miracle Homes: How Compositional Distributional Semantics can Help NP Parsing

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1908–1913, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Fish transporters and miracle homes: How compositional distributional semantics can help NP parsing Angeliki Lazaridou, Eva Maria Vecchi and Marco Baroni Center for Mind/Brain Sciences University of Trento, Italy first.last@unitn.it Abstract In this work, we argue that measures that have been shown to quantify the degree of semantic plausibility of phrases, as obtained from their compositionally-derived distributional seman- tic representations, can resolve syntactic am- biguities. We exploit this idea to choose the correct parsing of NPs (e.g., (live fish) trans- porter rather than live (fish transporter)). We show that our plausibility cues outperform a strong baseline and significantly improve performance when used in combination with state-of-the-art features. 1 Introduction Live fish transporter: A transporter of live fish or rather a fish transporter that is not dead? While our intuition, based on the meaning of this phrase, prefers the former interpretation, the Stan- ford parser, which lacks semantic features, incor- rectly predicts the latter as the correct parse.1 The correct syntactic parsing of sentences is clearly steered by semantic information (as formal syn- tacticians have pointed out at least since Fillmore (1968)), and consequently the semantic plausibil- ity of alternative parses can provide crucial evidence about their validity. An emerging line of parsing research capitalizes on the advances of compositional distributional se- mantics (Baroni and Zamparelli, 2010; Guevara, 2010; Mitchell and Lapata, 2010; Socher et al., 2012). Information related to compositionally- derived distributional representations of phrases is 1http://nlp.stanford.edu:8080/parser/ index.jsp integrated at various stages of the parsing process to improve overall performance.2 We are aware of two very recent studies exploiting the semantic in- formation provided by distributional models to re- solve syntactic ambiguity: Socher et al. (2013) and Le et al. (2013). Socher et al. (2013) present a recursive neural net- work architecture which jointly learns semantic rep- resentations and syntactic categories of phrases. By annotating syntactic categories with their distribu- tional representation, the method emulates lexical- ized approaches (Collins, 2003) and captures sim- ilarity more flexibly than solutions based on hard clustering (Klein and Manning, 2003; Petrov et al., 2006). Thus, their approach mainly aims at improv- ing parsing by capturing a richer, data-driven cate- gorial structure. On the other hand, Le et al. (2013) work with the output of the parser. Their hypothesis is that parses that lead to less semantically plausible interpreta- tions will be penalized by a reranker that looks at the composed semantic representation of the parse. Their method achieves an improvement of 0.2% in F-score. However, as the authors also remark, be- cause of their experimental setup, they cannot con- clude that the improvement is truly due to the se- mantic composition component, a crucial issue that is deferred to further investigation. This work aims at corroborating the hypothesis that the semantic plausibility of a phrase can in- deed determine its correct parsing. We develop a system based on simple …

Content truncated to 500 words