![]() |
|||
![]()
|
![]() |
![]() |
![]() |
4. STATE OF THE ART AND FUTURE TRENDS IN THE NLP DOMAINWe will mention here some recent developments in the NLP domain that are, or could be, of a particular importance for applications in the ES or KBS style. The speech recognition area has experienced significant progress in the past decade: for speaker-independent systems, performances vary now from an error rate of 0.3% in the recognition of digits to about 7% in dictation applications, and word error rates continue to be reduced by a factor of two every 2 years. This success is largely due to the resolute adoption of pure statistical methods, see the so-called Hidden Markov Modeling (HMM), for generating a sequence of word hypotheses from an acoustic signal. A model like HMM is particularly powerful given that, in the presence of training data, the parameters of the model can be adjusted automatically to produce optimal performance. Other factors that have contributed to the outstanding achievements obtained in this field are, e.g., the development of large speech corpora (up to tens of thousands of sentences) for system development, training and testing, the establishment of sound standards for performance evaluation, etc. There is room, however, for many other sorts of improvements. They concern robustness, portability (at present, speech recognition systems tend to experiment significant degradation when they are moved to a news task, which implies that they must be trained again), progress in the modeling of language by incorporation syntactic and semantic constraints in the pure statistical models, the possibility of handling out-of-vocabulary words (presently, the systems are designed for being used with a particular set of words that can include, however, more than 20,000 words), etc. From an ES/KBS point of view, Spoken Language Understanding (SLU) systems -- where (as we have seen in the previous subsection) speech recognition techniques are associated with natural language comprehension -- should normally be more interesting than simple speech recognition systems. SLU technology constitutes, in fact, the only way of building up ES/KBS (not only database systems) where the access can be realized in spoken natural language. Nevertheless, there are, many problems to solve before being able to achieve really integrated systems. To the contrary (at least partially), of the SpeechActs system examined in the previous subsection, many SLU systems still adhere to a very elementary architectural model that consists simply of the concatenation of an existing speech recognition system with an existing NL understanding system. This sort of concatenation, however, does not tend to work very well, mainly because existing NL understanding systems have been normally built up for dealing with written texts, i.e., according to modalities very different from those occurring in a spoken language environment. Moreover, the two components cannot really collaborate, etc. To surmount this sort of difficulties, two tendencies have recently appeared. The first consists, as in SpeechActs, of a revitalization of the semantic grammars (task-oriented grammars) role: because semantic grammars focus on finding an interpretation of the input that is based only on the semantic properties of the domain, without requiring a very precise "grammaticality" (syntactic correctness) of this input, they can be more robust in face of the grammatical deviations that are typical in a spoken language context. The second option, which is now becoming popular, consists of the adoption of the so-called "n-best" approach. In this case, the effects of a strictly serial connection of two existing systems are mitigated by the fact that the speech recognition system sends to the NL component not just the best word hypothesis obtained from the acoustic signal, but the n-best, where n may be between 10 and 100. The NL component makes use of its syntactic and semantic tools to narrow progressively the search space to determine the best-scoring hypothesis.
|
![]() |
|
Use of this site is subject certain Terms & Conditions. Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details. |