The company that words keep

The extraction of large-scale lexical co-occurrence statistics from huge corpora makes it possible to explore the syntagmatic and paradigmatic relationships between words. Statistics of this type can be computed automatically without human supervision. This technique has been used in the literature to solve certain semantic tasks. This paper reports on an experiment in which context vectors storing co-occurrence data were built for 30 target words, then the system chose the most similar word for each of the target words automatically. After training the model, semantically related words, including synonyms, antonyms and hypernyms/hyponyms were found by the system automatically with 90% precision.