Neural Models of Selectional Preferences for Implicit Semantic Role Labeling

Neural Models of Selectional Preferences for Implicit Semantic Role Labeling Minh Le and Antske Fokkens Computational Lexicology and Terminology Lab Vrije Universiteit Amsterdam, the Netherlands {m.n.le, antske.fokkens}@vu.nl Abstract Implicit Semantic Role Labeling is a challenging task: it requires high-level understanding of the text while annotated data is very limited. Due to the lack of training data, most researches either resort to simplistic machine learning methods or focus on automatically acquiring training data. In this paper, we explore the possibilities of using more complex and expressive machine learning models trained on a large amount of explicit roles. In addition, we compare the impact of one-way and multi-way selectional preference with the hypothesis that the added information in multi-way models are beneficial. Although our models surpass a baseline that uses prototypical vectors for SemEval-2010, we otherwise face mostly negative results. Selectional preference models perform lower than the baseline on ON5V, a dataset of five ambiguous and frequent verbs. They are also outperformed by the Na¨ıve Bayes model of Feizabadi and Pado (2015) on both datasets. We conclude that, even though multi-way selectional preference improves results for predicting explicit semantic roles compared to one-way selectional preference, it harms performance for implicit roles. We release our source code, including the reimplementation of two previously unavailable systems to enable further experimentation. Keywords: neural network, implicit semantic role labeling, selectional preferences 1. Introduction Defined as the recovery of semantic roles beyond imme- diate syntactic structure, implicit Semantic Roles Labeling (iSRL) can contribute valuable information for obtaining complete semantic interpretations of text. Yet, it has been elusive since its first shared task eight years ago (Ruppen- hofer et al., 2010). The main difficulty faced by researchers is the small size of training data. Compared to traditional SRL datasets, SemEval-2010 is hundreds-fold smaller, containing only slightly more than a hundred of training examples (Ta- ble 1). Early work applying traditional semantic role la- beling (SRL) techniques to iSRL was met with deflating results. Therefore, researchers limited themselves to sim- plistic machine learning models such as Na¨ıve Bayes (Feiz- abadi and Pado, 2015, among others) or abandoned ma- chine learning altogether (Laparra and Rigau, 2013). Sev- eral studies were devoted to the automatic expansion of training data (see Section 2 for an overview). This paper presents an attempt to recover implicit semantic roles using neural networks. We take advantage of the fact that OntoNotes contains a vast amount of manually anno- tated explicit semantic roles from which we can learn the selectional preference of frames (e.g. look.01 prefers ani- mate fillers for role A0 (looker)). A neural network is used to capture complex interactions between a predicate, a tar- get role and its co-occurring roles. In addition, we com- pare the impact of one-way selectional preference, taking only the selectional preference of the predicate for the tar- get role into account, to multi-way selectional preference, which uses information from all semantic roles related to the predicate. The contribution of this paper is twofold: First, we experi- mented with a …

Content truncated to 500 words