Identifying diachronic change in semantic relations

test the hypothesis that words which are similar in meaning occur in similar contexts, this being a corollary of the initial hypothesis that the meaning of a word is defined by the collocational company which it keeps. This similarity is measured by creating 'collocational profiles'-summaries of the words typically occurring in the context of a given word-for each unique word form (type) in a large corpus of newspaper data. By comparing the profile of the target word, the 'headword' in thesaurus or dictionary terms, with the profiles of the other words in the corpus, a set of words can be established which are semantically linked to the target word; these are referred to as 'nyms', since they could be in a traditional thesaural relationship with each other, for instance, as synonyms or antonyms, or may stand in a previously undefined relationship to the target word. The findings from ACRONYM will help to overcome several of the difficulties which have hitherto been involved in the creation of both printed and electronic thesauri. One of these, which forms the topic of this paper, is the problem of keeping the thesaurus up-to-date with the evolution of the language and the changes in real-world referents-the incumbents of government posts, as an example. This issue affects the thesaural facilities incorporated in many software applications, such as word processors, but is particularly problematic in database search engines accessing databases which contain material which is regularly updated. An example of this would be the textual databases held by Financial Times Electronic Publishing, one of the industrial partners of ACRONYM, who provide online search access to past issues of national daily newspapers spanning back several years. It is not hard to imagine how quickly the semantic equivalents described in a thesaurus in such a system can become outdated. In order to achieve the greatest possible accuracy in the identification of nyms, as much data as is available is generally used in deriving the collocational profiles. The ACRONYM corpus database system was designed in such a way,