Corpus-based Techniques for Word Sense Disambiguation

Consider the task of building a speech-to-speech translation system. One signi cant problem confronting the designer is the absence of a one-to-one mapping from word sounds to text strings to word meanings. The following examples reveal the ubiquity of this problem . In a highly homophonous language like Chinese the single sound sequence 'shi' maps to 56 di erent characters, each of which in turn has at least one meaning. In English, not only are there many text strings with context-dependent pronunciations and meanings (\record": the verb \re-c ord" and the noun \r e-cord"), but there are also many words like \bank" which have only one pronunciation but take on numerous meanings. For example, \bank" can be used as \the bank of a river", \bank account", and \bank a plane". The most extreme form of this ambiguity appears in pronous like \it", which take meaning only by reference to another element of the discourse. These mismatches multiply across languages , where in English the word \sentence" has two meanings, but in French, these meanings must be realized as two di erent words peine, in the criminal sense, and phrase in the grammatical. So, performing dictation,speech recognition, machine translation, or Web-search document retrieval, all require the ability to correctly select word senses.