Robust Tagging System for Lexicon Creation

This paper presents a robust rule-based system of shallow parsing for part-of-speech (PoS) recognition and tagging. Unlike previous work the system uses parsing to tagging based on unsupervised learning methods with no prior knowledge, nor training or pre-tagged corpora. START (System of Textual Analysis Recognition and Tagging) has been evaluated on both French and Greek non-annotated corpora, proving portability and adaptability to languages with similar syntactic features. Its accuracy rate exceeds 92% for recognition of noun and verb phrases and 99% for disambiguation of ambiguous cases –such as the definite article and the personal pronoun. After an auto-evaluation, the low error rate-less than 1%-makes trustworthy the knowledge acquisition and the parser proceeds to a lexicon creation, with entries having morpho-syntactic features for further automatic annotation on plain text corpora.