Measuring Frame Instance Relatedness

Proceedings of the 7th Joint Conference on Lexical and Computational Semantics (*SEM), pages 245–254 New Orleans, June 5-6, 2018. c⃝2018 Association for Computational Linguistics Measuring Frame Instance Relatedness Valerio Basile University of Turin Italy basile@di.unito.it Roque Lopez Condori Université Côte d’Azur, Inria, CNRS, I3S, France roque.lopez-condori@inria.fr Elena Cabrio Université Côte d’Azur, Inria, CNRS, I3S, France elena.cabrio@unice.fr Abstract Frame semantics is a well-established framework to represent the meaning of nat- ural language in computational terms. In this work, we aim to propose a quantita- tive measure of relatedness between pairs of frame instances. We test our method on a dataset of sentence pairs, highlighting the correlation between our metric and hu- man judgments of semantic similarity. Fur- thermore, we propose an application of our measure for clustering frame instances to extract prototypical knowledge from natu- ral language. 1 Introduction Frame Semantics has been a staple of artificial intelligence and cognitive linguistics since its first formulation in the ’70s (Fillmore, 1976). In particular, frame semantics has been widely adopted as a theoretical backbone for the inter- pretation of natural language, in order to repre- sent its meaning with formal structures suited for computation. In a nutshell, according to frame semantics, the meaning of a sentence can be represented as a set of situations (frames) and the entities involved in them (frame ele- ments), each with their own role. Several approaches have proposed in the past years to automatically interpret natural language in terms of frame semantics (Gildea and Jurafsky, 2002; Thompson et al., 2003; Erk and Padó, 2006, among others). How- ever, the vast majority of these approaches fo- cuses on the extraction of the structure of the frames evoked in the natural language fragment (frames and roles), while leaving the frame el- ements either underspecified or simply repre- senting them as spans of the original text. In this work, we propose to fully represent the meaning of a natural language sentence with instantiated frames, where the frame elements are nodes in a knowledge graph. Moreover, while a great deal of effort has been directed towards the extraction of frames from natural language, not many systems pro- cess frames further, to solve downstream tasks in NLP and AI — an example is Sentilo (Recu- pero et al., 2015), a sentiment analysis system built on top of the frame-based machine read- ing tool FRED by Presutti et al. (2012). In this paper we define a quantitative mea- sure to compute the semantic relatedness of a pair of frame instances, and apply it to the task of creating a commonsense knowledge base. The main contributions of this paper are: • A novel measure of relatedness between frame instances (Section 3). • A high-quality data set of natural lan- guage sentences aligned to the frame in- stances evoked by them (Sections 4 and 5). • A pilot study on the extraction of pro- totypical knowledge based on frame in- stance clustering (Section 6). Before introducing the novel contributions, we describe related work (Section 2), while …

Content truncated to 500 words