On the Importance of Distinguishing Word Meaning Representations: A Case Study on Reverse Dictionary Mapping

Proceedings of NAACL-HLT 2019, pages 2151–2156 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2151 On the Importance of Distinguishing Word Meaning Representations: A Case Study on Reverse Dictionary Mapping Mohammad Taher Pilehvar Tehran Institute for Advanced Studies (TeIAS), Tehran, Iran DTAL, University of Cambridge, Cambridge, UK mp792@cam.ac.uk Abstract Meaning conflation deficiency is one of the main limiting factors of word representations which, given their widespread use at the core of many NLP systems, can lead to inaccu- rate semantic understanding of the input text and inevitably hamper the performance. Sense representations target this problem. How- ever, their potential impact has rarely been in- vestigated in downstream NLP applications. Through a set of experiments on a state-of- the-art reverse dictionary system based on neu- ral networks, we show that a simple adjust- ment aimed at addressing the meaning con- flation deficiency can lead to substantial im- provements. 1 Meaning Conflation Deficiency Words are often the most fine-grained meaning bearing components of NLP systems. As a stan- dard practise, particularly for neural models, the input text is treated as a sequence of words and each word in the sequence is represent with a dense distributional representation (word embed- ding). Importantly, this setting ignores the fact that a word can be polysemous, i.e., it can take multi- ple (possibly unrelated) meanings. Representing a word with all its possible meanings as a sin- gle point (vector) in the embedding space, the so- called meaning conflation deficiency (Camacho- Collados and Pilehvar, 2018), can hinder system’s semantic effectiveness. To address this deficiency, many techniques have been put forward over the past few years, the most prominent of which is sense repre- sentation or multi-prototype embedding (Sch¨utze, 1998; Reisinger and Mooney, 2010). However, as a general trend, these representations are usually evaluated either on generic benchmarks, such as word similarity, or on sense-centered tasks such as Word Sense Disambiguation, leaving their po- tential impact on downstream word-based systems unknown. In this paper, we provide an analy- sis to highlight the importance of addressing the meaning conflation deficiency. Specifically, we show how distinguishing different meanings of a word can facilitate a more accurate semantic un- derstanding of a state-of-the-art reverse dictionary system, reflected by substantial improvements in recall and generalisation power. 2 Reverse Dictionary Reverse dictionary, conceptual dictionary, or con- cept lookup is the task of returning a word given its description or definition (Brown and McNeill, 1966; Zock and Bilac, 2004). For example, given “a crystal of snow”, the system has to return the word snowflake. The task is closely related to the “tip of the tongue” problem where an individ- ual recalls some general features about a word but cannot retrieve that from memory. Therefore, a re- verse dictionary system can be particularly useful to writers and translators when they cannot recall a word in time or are unsure how to express an idea they want to convey. 2.1 Evaluation framework Our experiments are based on the reverse dictio- nary …

Content truncated to 500 words