A Survey on Automatically-Constructed WordNets and their Evaluation: Lexical and Word Embedding-based Approaches

A Survey on Automatically-Constructed WordNets and their Evaluation: Lexical and Word Embedding-based Approaches Steven Neale School of English, Communication and Philosophy Cardiff University, Wales, United Kingdom NealeS2@cardiff.ac.uk Abstract WordNets – lexical databases in which groups of synonyms are arranged according to the semantic relationships between them – are crucial resources in semantically-focused natural language processing tasks, but are extremely costly and labour intensive to produce. In languages besides English, this has led to growing interest in constructing and extending WordNets automatically, as an alternative to producing them from scratch. This paper describes various approaches to constructing WordNets automatically – by leveraging traditional lexical resources and newer trends such as word embeddings – and also offers a discussion of the issues affecting the evaluation of automatically constructed WordNets. Keywords: WordNet, evaluation, automatic synset extraction, bilingual dictionaries, word embeddings 1. Introduction WordNets (Fellbaum, 1998) are lexical databases in which open-class words – nouns, verbs, adjectives and adverbs – are stored as sets of synonyms or ‘synsets’ and linked to each other by various semantic relationships. These re- lationships include antonymy (opposites, i.e. ‘wet’ and ‘dry’), hypernymy and hyponymy (types and types-of, i.e. ‘animal’ and ‘cat’), and meronymy (parts-of, i.e. ‘toe’ as a meronym of ‘foot’), among others. The Princeton Word- Net (PWN)1 is the original and pioneering English language WordNet – now totalling over 117,000 synsets in version 3.0 – and its format has become the gold standard for lexi- cal databases representing meanings and concepts. Besides its adoption as a standard for lexical databases, the PWN has become a vital resource across a range of seman- tic processing tasks – it has been used to produce gold- standard semantically-annotated corpora such as SemCor2 and to determine the correct meanings of words in natu- ral language processing tasks such as word-sense disam- biguation, text summarization, and semantic textual sim- ilarity. Naturally, its extensive usage in English has in- spired the construction of new WordNets in many other languages. The Global WordNet Association (GWA)3, for example, was set up to provide a platform for discussing, sharing and connecting WordNets in any language, while large-scale research projects such as EuroWordNet (Vossen, 2004) and MultiWordNet (Pianta et al., 2002) have focused on aligning synsets from multiple WordNets in different languages. However, while the PWN has been constructed, extended and refined over decades, building new WordNets from scratch is an enormous undertaking. Manual construction of course ensures accurate synsets covering as many con- cepts as possible, but it requires lexicographers to spend hours crafting synsets and is thus far too labour-intensive to be feasible in most cases. Inevitably, this has led to a 1http://wordnet.princeton.edu/wordnet/ 2http://web.eecs.umich.edu/˜mihalcea/ downloads.html#semcor 3http://globalwordnet.org/ growing and improving body of research into techniques and approaches for constructing and extending WordNets automatically. In this paper, a number of approaches to automatically con- struct and extend WordNets are presented, taking into ac- count both lexical and word embedding-based approaches. In addition, a discussion of the evaluation of these resources – and particular bottlenecks that a lack …

Content truncated to 500 words