Abstract
Traditional collocation extraction systems often rely on a limited, arbitrary set of pre-defined syntactic patterns, such as Adjective-Noun or Verb-Noun. This research presents a bottom-up, data-driven methodology that utilizes a syntactic parser to identify all syntactically related word pairs within a corpus. By applying minimal syntactic constraints rather than restrictive pattern filters, the approach induces a significantly broader range of collocation patterns, including those involving function words. Experimental results on English and French corpora demonstrate the discovery of approximately 60 distinct part-of-speech (POS) combinations, suggesting that standard extraction methods under-represent linguistic complexity.
Introduction
In linguistics, collocations are word pairs that co-occur with a frequency greater than chance, forming "natural" sounding expressions like "heavy smoker." Automated extraction of these pairs typically involves searching for specific syntactic patterns. However, the selection of these patterns is often arbitrary, leading to inconsistency between systems and the omission of valid collocations that do not fit narrow criteria. This paper addresses the need for a more exhaustive and less biased method of identifying productive syntactic patterns for collocation extraction.
Methodology
The proposed approach shifts from top-down pattern matching to a bottom-up induction process using the Fips syntactic parser. Instead of filtering for specific POS combinations, the system extracts all word pairs identified by the parser as sharing a generic syntactic relationship. These relationships include:
- Head-Modifier: Such as an adjective modifying a noun.
- Head-Complement: Such as a verb and its direct object.
- Verb-Argument: Including subject-verb and verb-object relations.
By grouping these structurally linked pairs by their POS tags, the researchers can identify which combinations are truly productive in natural language. This method provides a "minimal syntactic constraint" that is more precise than simple proximity-based "window" methods while remaining more flexible than traditional pattern-based systems.
Experimental Setup and Datasets
The methodology was evaluated using two distinct corpora consisting of newspaper articles in English and French. The English corpus contained 0.5 million words, while the French corpus was larger, at 1.6 million words.
As shown in the experimental data, the system extracted a substantial number of word-pair tokens and unique types, leading to the identification of dozens of distinct POS combinations in both languages.
Results and Discussion
The induction process revealed that the complexity of syntactic collocations is far greater than typically assumed. While many systems focus on 5-10 "standard" patterns, this study identified 60 distinct POS combinations in English and 57 in French.
The Role of Function Words
A significant finding was the productivity of "closed-class" or function words (e.g., prepositions, conjunctions). While often discarded as noise in other systems, these words form essential collocation patterns. Examples include:
- Verb-Conjunction: judge whether
- Noun-Preposition: alliance between
- Adverb-Preposition: together with
Structural Precision
By utilizing a parser, the system successfully identified long-distance collocations that window-based methods would likely miss. For instance, in a sentence where a subject and verb are separated by a relative clause (e.g., "a proposal which addressed various topics has been submitted"), the parser maintains the structural link between the core components.
Conclusion
This research demonstrates that a data-driven, parser-based approach can exhaustively identify the syntactic patterns relevant to collocation extraction. While the system successfully induces a wide variety of potential patterns, future work involves refining these results through in-depth analysis to distinguish between general syntactic structures and truly significant lexical associations. Although computationally more intensive than window-based methods, this approach offers a more robust foundation for building comprehensive dictionaries and translation tools.
Summary updated