Leveraging Lexical Resources and Constraint Grammar for Rule-Based Part-of-Speech Tagging in Welsh

Leveraging Lexical Resources and Constraint Grammar for Rule-Based Part-of-Speech Tagging in Welsh Steven Neale1, Kevin Donnelly2, Gareth Watkins1, Dawn Knight1 1School of English, Communication and Philosophy Cardiff University, Wales, United Kingdom {NealeS2, WatkinsG13, KnightD5}@cardiff.ac.uk 2Independent Researcher kevin@dotmon.com Abstract As the quantity of annotated language data and the quality of machine learning algorithms have increased over time, statistical part-of-speech (POS) taggers trained over large datasets have become as robust or better than their rule-based counterparts. However, for lesser-resourced languages such as Welsh there is simply not enough accurately annotated data to train a statistical POS tagger. Furthermore, many of the more popular rule-based taggers still require that their rules be inferred from annotated data, which while not as extensive as that required for training a statistical tagger must still be sizeable. In this paper we describe CyTag, a rule-based POS tagger for Welsh based on the VISL Constraint Grammar parser. Leveraging lexical information from Eurfa (an extensive open-source dictionary for Welsh), we extract lists of possible POS tags for each word token in a running text and then apply various constraints – based on various features of surrounding word tokens – to prune the number of possible tags until the most appropriate tag for a given token can be selected. We explain how this approach is particularly useful in dealing with some of the specific intricacies of Welsh - such as morphological changes and word mutations - and present an evaluation of the performance of the tagger using a manually checked test corpus of 611 Welsh sentences. Keywords: Part-of-speech, lexical resources, open-source dictionaries, Constraint Grammar 1. Introduction POS tagging is a well-explored problem in NLP, and highly-accurate taggers have been built using statistical and probabilistic methods for decades. These taggers are typi- cally trained using already-annotated text, from which the probabilities of POS tags being appropriate for certain word tokens are calculated based on features such as the lexical properties of tokens (capitalisation, common prefixes and suffixes etc.) or the POS tags of their n-neighbouring to- kens. However, the amount of pre-annotated data that these taggers require to be properly trained is considerable, and usually in the region of many hundreds of thousands to mil- lions of word tokens. For languages such as Welsh – for whom resources are typically more scarce – pre-annotated data in context in these kinds of quantities is very difficult either to create or to obtain. The traditional alternative to probabilistic POS tagging is the rule-based approach, whereby tags are as- signed based on pre-defined rules concerning which syn- tactic categories can be co-located together. Crafting and refining the rules by hand is, however, almost as costly and labour-intensive as producing manually-annotated data from which to train a probabilistic POS tagger. This paper introduces CyTag1, a rule-based tagger that leverages an open source dictionary and uses the VISL Constraint Grammar parser to assign POS tags to Welsh words in context, using a minimal set of easily-adaptable rules and without the need for millions of tokens of …

Content truncated to 500 words