Natural Language Processing as a Source of Linguistic Knowledge

1 The paper discusses a number of specific problems of natural text parsing that emerge during the operation of a highly developed rule-based machine translation system, ETAP-3. Emphasis is laid on two classes of problems: 1) adequacy of linguistic description of the working languages of the MT system and 2) means of resolving lexical and syntactic ambiguity of the source text. It is claimed that no parser, however sophisticated or advanced, can be made entirely free of lacunae and gaps. The reason is that many of the linguistic facts, including those critical for parser operation, have never come into view of researchers simply because they have not had at their disposal mass material of unexpected or incorrect parsing. It is exactly such material that is amply provided by a highly developed NLP system. If handled properly, this feedback helps the researcher to find the gaps of scientific descriptions and eliminate them. Consequently, linguistic experimentation with NLP systems becomes a rightful and very promising scientific method. In a way, linguistic applications start to stimulate theoretical research, thus inverting the situation that has existed ever since NLP came to life. The paper deals with a number of instructive cases that opened up in the course of experimental operation of the Russian-to-English automatic translation module of a high-level multifunctional NLP system, ETAP-3, developed by a Moscow research team [1-3]. The system, largely based on the Meaning Û Text theory by Igor Melčuk [4], makes use of dependency syntax: the syntactic structure of any sentence is represented as a dependency tree whose nodes correspond to all words of the sentence and whose arcs are labeled with names of one of several dozens syntactic relations. The method of syntactic representation will be essential in the following account. All cases were evolving in a fairly similar way: the MT module was offered Russian sentences for translation, for which it yielded unsatisfactory English equivalents. Normally, the sentences came from current Russian Internet news sites. The system's operation was subsequently subjected to severe scrutiny, which enabled the experimenters to locate the errors and correct them if at all possible. In the reverse case, when any of the errors proved incorrigible, the experimenters could make a step towards establishing natural limits to machine translation performance. Below, I will consider some representative situations in which unsatisfactory performance of the NLP system has led to important theoretical findings. It can readily …