Collocation Extraction Based on Syntactic Parsing

Pervasive across texts of different genres and domains, collocations (typical lexical associations like to wreak havoc, to meet a condition, to believe firmly, a deep concern, highly controversial) constitute a large proportion of the multi-word expressions in a language. Due to their encoding idiomaticity, collocations are of paramount importance to text production tasks. Their recognition and appropriate usage is essential, for instance, in Foreign Language Learning or in Natural Language Processing applications such as machine translation and natural language generation. At the same time, collocations have a wide applicability to tasks concerned with the opposite

process of text analysis.

The problem that is tackled in this thesis is the automatic acquisition of accurate collocational information from text corpora. More specifically, the thesis provides a methodological framework for the syntax-based identification of collocation candidates in the source text, prior to the statistical computation step. The development of syntax-based approaches to collocation extraction, which has traditionally been hindered by the absence of appropriate linguistic tools, is nowadays possible thanks to the advances achieved in parsing. Until now, the absence of sufficiently robust parsers was typically circumvented by applying linear proximity constraints in order to detect syntactic links between words. This method is relatively successful for English, but for languages with a richer morphology and a freer word order, parsing is a prerequisite for a good performance.

The thesis proposes (and fully evaluates on data in four different languages, English, French, Spanish and Italian) a core extraction procedure for discovering binary collocations, which is based on imposing syntactic constraints on the component items instead of linear proximity constraints. This procedure is further employed in several methods of advanced extraction, whose aim is to cover a broader spectrum of collocational phenomena in text. Three distinct but complementary extension directions have been considered in this thesis: extraction of n-ary collocations (n > 2), data-driven induction of collocationally relevant syntactic configurations, and collocation mining from an alternative source corpus, the World Wide Web. The possibility to abstract away from the surface text form and to recover, thanks to parsing, the syntactic links between discontinuous elements in text, plays a crucial role in achieving highly efficient results.

The methods proposed in this study were adopted in the development of an integrated system of collocation extraction and visualization in parallel corpora, a system which was intended to enrich the workbench of translators or other users (e.g., terminologists, lexicographers, language learners) wanting to exploit their text archives. Finally, the thesis gives an example of a practical application that builds on this system in order to further process the extracted collocations, by automatically translating them when parallel corpora are available.