Towards a Standardized Dataset for Noun Compound Interpretation

Towards a Standardized Dataset for Noun Compound Interpretation Girishkumar Ponkiya∗, Kevin Patel∗, Pushpak Bhattacharyya∗, Girish K Palshikar† ∗Indian Institute of Technology Bombay, Mumbai, India {girishp,kevin.patel,pb}@cse.iitb.ac.in †TRDDC, Pune, India gk.palshikar@tcs.com Abstract Noun compounds are interesting constructs in Natural Language Processing (NLP). Interpretation of noun compounds is the task of uncovering a relationship between component nouns of a noun compound. There has not been much progress in this field due to lack of a standardized set of relation inventory and associated annotated dataset which can be used to evaluate suggested solutions. Available datasets in the literature suffer from two problems. Firstly, the approaches to creating some of the relation inventories and datasets are statistically motivated, rather than being linguistically motivated. Secondly, there is little overlap among the semantic relation inventories used by them. We attempt to bridge this gap through our paper. We present a dataset that is (a) linguistically grounded by using Levi (1978)’s theory, and (b) uses frame elements of FrameNet as its semantic relation inventory. The dataset consists of 2,600 examples created by an automated extraction from FrameNet annotated corpus, followed by a manual investigation. These attributes make our dataset useful for noun compound interpretation in a general-purpose setting. Keywords: noun compounds, compounds, interpretation, semantic relations. 1. Introduction Noun compounds are continuous sequences of nouns that act as a single semantic construct. They raise interest- ing challenges in Natural Language Processing. Without proper interpretation and paraphrasing of noun compounds, NLP methods may fail miserably at different tasks. The meaning of a noun compound is composed of the meanings of the individual constituents and the way they are seman- tically related. Noun compound interpretation is the task of detecting this underlying semantic relation (e.g., student protest: student ← AGENT ← protest). It is an important submodule for various NLP tasks such as machine translation (Baldwin and Tanaka, 2004; Balyan and Chatterjee, 2014), question answering (Ahn et al., 2005), etc. Noun compound interpretation can manifest itself in two settings: out-of-context interpretation and context- dependent interpretation. In out-of-context interpretation, given the noun compound, the task is either to annotate it with a relation from a semantic relation inventory (e.g., student protest: AGENT), or to produce a paraphrase (e.g., student protest: “protest carried out by student”). Any automated approaches for noun compound interpre- tation need a semantic relation inventory of noun-noun relations and an annotated dataset on which models can be trained. However, there is little agreement among re- searchers regarding the set of relations that can hold be- tween the constituents of a noun compound. None of the proposed semantic relation inventories has been accepted by the community as complete and appropriate for general- purpose text. Some are coarse-grained, while others are fine grained. There is little overlap among them. Also, some of these inventories and their accompanying dataset are cre- ated from another application’s perspective, and not for the sake of creating a noun compound dataset. Thus they can- not be used for learning noun compound interpretation in a general-purpose setting. …

Content truncated to 500 words