Using the Distributional Hypothesis to Derive Cooccurrence Scores from the British National Corpus

In this paper I examine a number of cooccurrence-based scoring systems using the British National corpus to measure word association over wide contexts. The principal aim of this paper is to address the question of how to evaluate a given scoring system, or how to compare two scoring systems, without relying on a small list of example pairs and a ‘feel’ for the results. I evaluate these systems using i) a list of noun-noun pairs and ii) a simple test on aligned and misaligned sets of nouns. I also consider why noun-noun pairs are deemed appropriate for such mechanisms and explore the prospects for determining which words or lemmas will be appropriate for a distributional scoring approach. For my specific application an algorithm similar to MI-score operating across multiple windows offers the best results.