Improving Word Association Measures in Repetitive Corpora with Context Similarity Weighting

Aleksi Sahala, Krister Lindén


Although word association measures are useful for deciphering the semantic nuances of long extinct languages, they are very sensitive to excessively formulaic narrative patterns and full or partial duplication caused by different copies, edits, or fragments of historical texts. This problem is apparent in the corpora of the ancient Mesopotamian languages such as Sumerian and Akkadian. When word associations are measured, vocabulary from repetitive passages tends to dominate the top-ranks and conceal more interesting and descriptive use of the language. We propose an algorithmic way to reduce the impact of repetitiveness by weighting the co-occurrence probabilities by a factor based on their contextual similarity. We demonstrate that the proposed approach does not only effectively reduce the impact of distortion in repetitive corpora, but that it also slightly improves the performance of several PMI-based association measures in word relatedness tasks in non-repetitive corpora. Additionally, we propose normalization for PMI2, a commonly-used association measure, and show that the normalized variant can outperform the base measure in both, repetitive and non-repetitive corpora.


Paper Citation