Word Sense Disambiguation based on Constrained Random Walks in Linked Semantic Networks

Proceedings of Recent Advances in Natural Language Processing, pages 516–525, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_061 516 Word Sense Disambiguation Based on Constrained Random Walks in Linked Semantic Networks Arkadiusz Janz, Maciej Piasecki Wrocław University of Science and Technology, Poland {arkadiusz.janz|maciej.piasecki}@pwr.edu.pl Abstract Word Sense Disambiguation remains a challenging NLP task. Due to the lack of annotated training data, especially for rare senses, the supervised approaches are usually designed for specific subdomains limited to a narrow subset of identified senses. Recent advances in this area have shown that knowledge-based approaches are more scalable and obtain more promis- ing results in all-words WSD scenarios. In this work we present a faster WSD algo- rithm based on the Monte Carlo approxi- mation of sense probabilities given a con- text using constrained random walks over linked semantic networks. We show that the local semantic relatedness is mostly sufficient to successfully identify correct senses when an extensive knowledge base and a proper weighting scheme are used. The proposed methods are evaluated on English (SenseEval, SemEval) and Polish (Składnica, KPWr) datasets. 1 Introduction Semantic analysis is the process of understanding the underlying meaning in text. Building a rich se- mantic representation is a crucial element of mod- ern Natural Language Processing NLP. It allows us to build comprehensive text representations in order to analyse the underlying text meaning more effectively. Semantic information helps us to re- solve the issue of textual ambiguity. The main aim of semantic analysis is to differentiate texts which use the same vocabulary yet present differ- ent ideas about the same topic. One of the most important building blocks of semantic analysis is word sense disambiguation (WSD). WSD aims at solving the problem of lexical ambiguity, i.e. two or more meanings being represented by one word e.g. English words space, shape, or Polish words tło (background color, a context or an incidental music) wn˛etrze (a room, a soul or a set). This lex- ical ambiguity is still an open issue for semantic analysis due to the lack of required training data and the computational power required to process the growing number of possible classes. The task of word sense disambiguation is usually solved by mapping words onto their senses given a particular sense inventory. WSD can be performed in a supervised way, i.e. on the basis of the annotated lexical meanings in training texts, or in an unsupervised way, i.e. sense induction from texts. Supervised machine learning approaches can achieve a very good performance when trained on a large training data annotated with good inter- annotator agreement and well designed features. However, it is difficult to acquire large train- ing data for WSD because training data is usu- ally manually annotated and manual annotation is labour-intensive and thus costly1. Semi-automatic approaches on the other hand often result in lower quality data blurred with statistical noise. Thus, the best performing supervised ap- proaches to WSD task are usually limited to a spe- cific narrow subset of training vocabulary. More- over, supervised approaches are …

Content truncated to 500 words