Humans are very good at disambiguation; natural language users quickly discard ridiculous readings of a given sentence by taking into account the context and situation of the utterance. In fact, humans typically are unaware of the alternative readings of an utterance. Only in special circumstances (for instance in jokes) the ambiguity property of natural language becomes obvious. A computational theory of language use therefore will have to explain how it is that the appropriate reading in a given situation is selected from the space of possible readings provided by the grammar.
We propose to investigate techniques which are capable to augment the knowledge of language (modelled in the form of a grammar) with a theory of how this knowledge of language is applied in a given situation. We refer to such techniques as grammar specialisation in order to stress the fact that we take a linguistic grammar (the model of the knowledge of language) as an important point of departure for such techniques.
Problems of disambiguation are often avoided in natural language processing by focusing on a limited domain. In such a domain, utterances can often be understood unambiguously. The grammar used in such an approach is heavily tuned towards the domain. This tuning is performed by human experts. It is unclear what the relation is between this tuned version of the grammar and the underlying, more general grammar. Moreover, such a tuned grammar is only useful for the intended application and cannot be used without major investments for other domains: the grammar is not extendible.
The investigation will focus on possibilities to perform such grammar tuning automatically. Such automated grammar specialisation techniques would constitute a potential answer to the question how the desiderata of suitability for a specific domain (on the one hand) and ease of portability (on the other hand) can be combined. Such techniques incorporate some elements from the data-driven parsing approach, yet it differs from pure data-driven parsing in that a general and abstract linguistically inspired grammar is still essential as the starting point of development of the parser.
The expected practical gain from such techniques is that the grammar
can be made more general, i.e. suited for several domains, without
loss of parsing accuracy for a specific domain of application. This
is because the grammar can be automatically specialised for such an
application, omitting from consideration all language phenomena that
do not occur in corpora for the corresponding domain.
It may be worthwhile to explain more carefully the role of such specialisation techniques. We assume that a grammar provides for a set of linguistically possible meanings for a given utterance. This is the set of meanings that could be assigned to a given utterance if only linguistic constraints are taken into account. In real life, almost every utterance receives only a single meaning: the situation or context typically is such that the other readings do not make sense. Only in the case of jokes or other language games is ambiguity intended.
The purpose of grammar specialisation techniques is to formalise at least certain aspects of the interaction between what is linguistically possible, and what is appropriate in a given situation.
It may be argued that, rather than specialising the grammar, some extra-grammatical reasoning mechanism should be employed. In such an argument, it is assumed that knowledge of the situation and context is available, and that a given candidate meaning is somehow first checked for consistency with the available information. Apart from the consideration that the sophisticated knowledge representation and automated reasoning techniques that would be required are not yet readily available, there are a number of theoretical reasons why we believe that it is worthwhile to consider specialisation techniques (perhaps as a supplement to such reasoning techniques).
Firstly, in many cases not enough information will be available to base the relevant inference on. It is typically extremely hard to spell out precisely why a certain reading should be avoided in a certain context or situation, because typically not enough information is available to base the decision on. Therefore, the specialisation techniques that we propose to investigate are are a means to aid reasoning in circumstances where not enough information is available.
Secondly, it seems that it is quite unlikely that extra-grammatical inference techniques are employed to solve all disambiguation tasks. Such an hypothesis seems to predict that all of the linguistically possible readings of a given sentence are to be constructed, before inference filters most of these; that would lead to the expectation that we should be somehow conscious of the other possible readings. Consider the example:
People tend to be surprised if the alternative reading where saw is understood as the present tense of the verb to saw is explained to them, suggesting that these readings were never actually constructed. Moreover, given the fact that there can be exponentially many different readings for a given utterance, such a setup would also lead to efficiency problems (both from a practical and a theoretical point of view).
In the grammar specialisation approach that we propose here, the
integration of disambiguation techniques into the grammar, in
combination with standard best-first search techniques, will in
general avoid the computation of all linguistically possible
meanings, but instead return only the most plausible one.
Obviously, humans quickly adapt their language use to varying situations and domains. In an appropriate context our examples of ambiguous sentences might be understood differently. It is thus clear that disambiguation techniques should be able to model these dynamic aspects of natural language. In this proposal, however, we will not immediately concentrate on these dynamic aspects but we will start with the assumption that useful insights can already be gained with a simpler static approach. Given this limitation we will focus mostly on examples which can be explained under fairly stable domain characteristics; we will focus less on dynamic aspects such as pronoun resolution beyond sentence boundaries.