Knowledge-Poor Approach to Shallow Parsing: Contribution of Unsupervised Part-of-Speech Induction

Proceedings of Recent Advances in Natural Language Processing, pages 33–40, Hissar, Bulgaria, 12-14 September 2011. Knowledge-Poor Approach to Shallow Parsing: Contribution of Unsupervised Part-of-Speech Induction Marie Guégan Syllabs 15, rue Jean-Baptiste Berlier 75013 Paris, France guegan@syllabs.com Claude de Loupy Syllabs 15, rue Jean-Baptiste Berlier 75013 Paris, France loupy@syllabs.com Abstract Natural language processing tasks often rely on part-of-speech (POS) tagging as a prepro- cessing step. However it is not clear how the absence of any part-of-speech tagger should hamper the development of other natural lan- guage processing tools. In this paper we in- vestigate the contribution of fully unsuper- vised part-of-speech induction to a common natural language processing task. We focus on the supervised English shallow parsing task and compare systems relying either on POS induction, on POS tagging, or on lexical features only as a baseline. Our experiments on the English CoNLL'2000 dataset show a significant benefit from POS induction over the baseline, with performances close to those obtained with a traditional POS tagger. Re- sults demonstrate a great potential of POS in- duction for shallow parsing which could be applied to resource-scarce languages. 1 Introduction Shallow parsing is a specific type of phrase chunking which is often used for different Natu- ral Language Processing (NLP) tasks like text mining or question answering. The goal of the task is to divide a text into syntactically related non-overlapping groups of words (Tjong Kim Sang and Buchholz, 2000). These include noun, verb, or adjective phrases. It usually requires a part-of-speech (POS) tagger and a training cor- pus annotated with shallow parsing tags. Unfortunately, one is often constrained by the lack of resources, tools or language experts, for instance when dealing with resource-scarce lan- guages. In particular, the elaboration of a POS tagger is a delicate issue. Without any linguistic expert, the only possible approaches are statistic- al. Training POS taggers requires the manual constitution of either a large annotated corpus or a large morphosyntactic lexicon. These resources are very costly, both in time and in terms of lin- guistic knowledge required from the annotator. By contrast, we notice that the concept of shal- low parsing is relatively easily understandable by native speakers, even if they are not linguists. Relative to POS tagging, its annotation does not require a prohibitive amount of time and effort1. This is especially the case when the full shallow parsing task is reduced to a certain chunk type, as noun phrases for instance. Hence we think the most difficult requirement for the task is the POS tagging preprocessing step. This observation drew our attention to the fol- lowing question: is the POS tagging step neces- sary to shallow parsing? In this paper we intend to show how shallow parsing may benefit from fully unsupervised POS induction methods, as an alternative to accurate POS tagging. Section 2 introduces related work. Despite the popularity of shallow parsing and POS induction, we found only one paper related to POS induction for shal- low parsing. Section 3 describes the models, tools and …

Content truncated to 500 words