Word sense disambiguation
WordNet may well be criticized for arbitrariness in its division of words into different polysemous senses, but no one doubts that a division of some kind is a practical necessity. Any automatic language processing system will have to be able to ‘individuate’ (differentiate between/disambiguate) the different senses of polysemous or homonymous word-forms if it is to correctly interpret the meaning of a sentence. For instance, the failure to distinguish between the different senses of the highlighted words in (18) would be disastrous for any automated translation of the sentences:
Since in French, for instance, the two senses of bank are translated by two different words (rive for the ‘edge of river’ sense and banque for ‘financial institution’), a necessary first step in any automatic translation is the selection of the appropriate sense for the context. If this is not done, we risk obtaining translations of the following meanings:
In natural language, sentences (18a–e) do not typically create confusion. Why not is still a mystery. We do not yet know how humans succeed in picking the right senses of ambiguous words: the relevant psychological processes are simply not at all understood (see Taylor 2003: Chapter 8 for discussion). This is quite a remarkable state of affairs. As discussed in 5.3, polysemy is a universal feature of natural language. If we do not understand how the correct polysemous sense is selected in a particular context, we clearly lack a fundamental part of the explanation of linguistic processes.
Developing a successful sense-individuation procedure for computers may therefore give us some clues about the relative advantages of different solutions to the problem. It will not, of course, show us what procedures the human mind does in fact use, but it may give us some insight into the nature of the task and point us in the direction of likely hypotheses. Given the complexity of the task of developing a sense-individuation procedure, we fortunately do not need to aim for complete accuracy. After all, people do not themselves correctly identify the intended sense one hundred per cent of the time in ambiguous contexts: as an initial goal, we should not expect to develop programs which achieve full accuracy either.
There are currently two main approaches to word-sense disambiguation in computational linguistics. The first, the selectional restriction approach (Hirst 1987), generates complete semantic representations for all the words in a sentence, and then eliminates those which violate selectional restrictions coded in the component words. For instance, consider (20), taken from Jurafsky and Martin (2000: 632):
Dish is ambiguous between the senses ‘piece of crockery’ and ‘course in a meal’. Selection restrictions on the patient argument of the verbs in (20) would be used to exclude the wrong reading of dish. The lexical entry for wash would include the specification that the verb could only take a physical object as its patient argument, which would guarantee the correct reading in (20a), and the entry for stir-fry would specify that the argument had to belong to the category of food in (20b). These selectional restrictions would be encoded reasonably straightforwardly by adopting categories from WordNet: thus, the possible object of wash could be limited to any noun which had ‘artefact’ or ‘natural object’ as a hypernym, and the possible object of stir-fry to nouns with ‘food’ as a hypernym. In both these cases, selectional restrictions on the verbs eliminate the wrong sense of the nouns.
Similarly, serve in the following sentences has three different senses (Jurafsky and Martin 2000: 633):
Once again, selectional restrictions can be used to ensure the correct interpretation is reached: the sense in (21a) is only found with food objects; the (21b) sense only with place objects, and (21c) only with meal objects. Giving the program access to this information would allow the wrong interpretations to be avoided.
QUESTION What are some problems and limitations of this procedure?
One major problem with this approach is that selection restrictions will often not be available to resolve ambiguities. ‘What kind of dishes would you recommend?’ is one such example: either the ‘course in meal’ or the ‘piece of crockery’ sense could be intended. While the non-linguistic con text would probably resolve the ambiguity, nothing in the linguistic con text itself does. Another difficulty is caused by the fact that selection restrictions may often be violated. For instance, the representation of the meaning of eat would presumably contain the restriction that only food could be selected as object. Jurafsky and Martin cite (22) as examples of perfectly well-formed sentences which violate this restriction:
More generally, the ubiquity of metaphor and metonymy in language poses significant problems for sense-individuation based on selectional restrictions, since these create contexts in which words’ typical co-occurrence restrictions do not hold.
A possible solution to these problems might be to take selection restrictions not as absolute criteria which must be met, but as probabilistic preferences (Resnik 1998). There is not time to discuss this development here, beyond noting that this way of thinking is a semantic analogue to the optimality theory approach to syntax and phonology, which describes grammatical phenomena not as the results of absolute rules, but as the most optimal results of the satisfaction of competing constraints. We should also note that Resnik (1998) only reported fairly low success rates with this approach, an average of 44 per cent for verb–object relationships. This is actually an improvement over the average success rate of 28 per cent (Jurafsky and Martin 2000: 633–634). The extreme modesty of these figures highlights the difficulty of the task. The other approach uses the immediate context of the target word as a clue in identifying the intended sense. The computer assesses the words sur rounding the target word, and chooses the appropriate sense on the basis of the other words in this immediate context (the size of this ‘window’ varies from program to program). For instance, if the word to be disambiguated is pen, the ‘writing implement’ sense will be chosen if other words in the environment include paper, write, etc., whereas ‘sty’ will be chosen if the words are ‘pig’, ‘feed’, ‘farm’ and so on. Similarly, Manning and Schütze (1999: 238) give the following contextual clues for sense disambiguation of drug:
QUESTION Suggest possible contextual clues for the different senses of tank, match and interest.
There are many contexts, however, which pose a problem for these con text-based disambiguation procedures. Wilks et al. (1996: 202) discuss the particular problem posed by a sentence like The young page put the goat in the pen. This will be a challenge for any disambiguation procedure that chooses the relevant word sense on the basis of the immediate context, since the cooccurrence of pen and page might lead to the wrong sense of each being chosen. Other sentences which might raise a similar problem are the following:
QUESTION Are there other contextual features of the sentence which would allow these problems to be avoided?
These studies are in their infancy, and we can certainly expect improvements in the efficiency of disambiguation programs. Nevertheless, the fact that most current disambiguation processes have been tested on a very small number of lexical items, or report low success rates, means that the question of polysemy and sense-disambiguation will be the focus of intense theoretical and applied research for some time to come.
