The state-of-the-art in NLP

Grammar formalisms

The semantic systems of the 1970s had quite deliberately avoided the extensive use of syntactic processing and some workers tried to purge all syntactic information from their systems. Syntax is seen by many theoretical linguists as a fundamental part of human language processing, while language engineers see it as a very useful way of doing a lot of disambiguation with relatively small amounts of knowledge (at least when compared with the amount of knowledge semantic processing requires). Theoretical linguists were also reacting against the then prevalent Transformational/Generative Grammar as propounded by Noam Chomsky and almost universally accepted.

The result as a family of grammars that encoded bundles of syntactic information (eg [category: noun; person: third; definite: positive]) rather than single, atomic categories (eg noun, verb). These grammars could give far more discriminating analyses of sentences, but there were some prices to pay. The main point is that bundles of features need far more complicated methods of matching in search algorithms, and it is from this that these grammars get their family name of Unification Grammar. The matching of bundles of features can be achieved by the use of unification.

There are several grammars that can use unification as their main information combining operation. Of these, Generalised Phrase Structure Grammar (GPSG) was at one time very popular, but Lexical Functional Grammar (LFG) is probably the most widely used of these formalisms.

Extended lexicons

The lexicon (or dictionary) used to be the repository of very simple information. The use of Unification Grammars with their feature bundles has meant that the lexicon has become far more complex. It is now the primary source of feature information about words and this information is combined during parsing (using unification) to produce the final parse trees which represent the structure of the sentences being analysed.

Logic for semantics

If syntactic information is to be widely used, then it is necessary to have a way of bringing in semantic information at a later stage, when syntactic processing has contributed as much to disambiguation as possible. An approach that has proved interesting has been the use of predicate logic to represent semantic information. The semantic information can be combined in much the same way that syntactic information is combined with unification. Logical reasoning methods can be used to infer and deduce more information about the input sentences. Such logics don't have to be classical (ie they don't have to have strict true or false values) and there has been much work on alternatives and logics to express time and space relations.

Text generation

An area not covered in this course includes outputting texts in natural language. The idea here is that an application that can output solutions can be made to output that information in a natural language, thus making easy for a user to assimilate. One development has been work on producing texts of paragraph length. There are significant problems here, because sentences within a paragraph are usually (and perhaps should) be related one to another, perhaps referring back and forwards to other sentences.

Multilingual applications

The growth of multilingual communities has meant an increased demand for translation, especially of mundane texts that human translators find very boring. While full high-quality machine translation is no longer considered a reasonable short or medium-term objective, there is room for limited machine translation systems (eg the Canadian Méteo system which translates weather forecasts from English to French) and for aids for translators. Other applications include systems which generate text in several languages.

Transportability

One of the drawbacks of impressive systems such as LIFER/LADDER was the difficulty of transferring the system from one domain (eg US Navy ships) to another (eg personnel records). Transportability attempts to develop systems that can be readily applied to new domains. The developments in Unification Grammars and logic for semantics are trends toward generalisation in systems and so make the portability easier.