Voice Communication Between Humans and Machines (1994) / Chapter Skim
Currently Skimming:
Models of Natural Language Understanding
Models of Natural Language Understanding
Pages 238-253
The Chapter Skim interface presents what we've algorithmically identified as the most significant single chunk of text within every page in the chapter.
Select key terms on the right to highlight them within pages of the chapter.
From page 238... ...
Of particular importance are techniques that can be tuned to such requirements as full versus partial understanding and spoken language versus text. Portability (the ease with which one can configure an NL system for a particular application)
|
From page 239... ...
Some aspects of language understanding seem tantalizingly similar to problems that have been solved (or at least attacked) in speech recognition, but other aspects seem to emphasize differences that may never allow the same solutions to be used for both problems.
|
From page 240... ...
That analysis would then be used to retrieve the answer to the question from a database. In the 1970s interest broadened from database interfaces to other kinds of application systems, but the focus was still on the kinds of natural language that would be produced by a person interacting with a computer system typed queries or commands issued one at a time by the person, each of which needed to be understood completely in order to produce the correct response.
|
From page 241... ...
Spoken language systems (SLSs) (which combine speech recognition with language understanding)
|
From page 242... ...
and on metrics for evaluating the quality of SR systems. The word error rate, which incorporates insertions, deletions, and substitutions, has been the generally accepted metric for many years; it is widely accepted, easy to apply, and works so well that there is little reason for the speech research community to change it.
|
From page 243... ...
Understanding what Calvin meant by "Verbing weirds language" stretches the limits of human language performance. One of the reasons that NL is challenging to computational linguists is its variety.
|
From page 244... ...
244 Typed input Message text Speech Speech Recogn~zer _ O r u | L meaning ~ | C P s r 0 P r FIGURE 2 A generic NL system. Words | Parser 1 - ~ Syntactic Analysis Meaning Representation MADELEINE BATES Answer output _ Database update Spoken response Other Semantic Processor Pragmatic & Discourse Processor Plan | Reasoner t~ of a Response Meaning Representation Action and l Response Generator !
|
From page 245... ...
A second use of syntactic analysis is to help detect new or unusual meanings. Without syntactic analysis it might be possible to use semantic probabilities to determine that a string containing "boy" and "dog" and "bit" means that a dog bit a boy, but syntax makes it easy to determine who bit whom in the input "boy bit dog." Calvin's observation that "Verbing weirds language" can be understood only by using morphological and syntactic cues, not by semantics alone.
|
From page 246... ...
hundreds of words. It is the difficult task of the discourse and pragmatics component to determine the referents of pronouns and definite noun phrases and to try to understand elliptical sentence fragments, dropped articles, false starts, misspellings, and other forms of nonstandard language, as well as a host of other long-range language phenomena that have not even been adequately characterized much less conquered.
|
From page 247... ...
or on the output side (e.g., extract from multiple paragraphs of newspaper text just three pieces of information about company mergers; extract from a single spoken utterance one of six possible commands to an underlying display system)
|
From page 248... ...
This architecture is illustrated in Figure 4. In this view the lexical processor would use a dictionary to help it transform the input words into a structure with more meaning; the syntactic processor would use a grammar of the language; the semantic processor would use semantic interpretation rules and a domain model of concepts and relationships that defines the domain the system can understand; and discourse and pragmatics might use a task model that specifies the user's goals and the portions of those goals that have been achieved by previous inputs.
|
From page 249... ...
process and an appropriately annotated corpus, as shown in Figure 5. The use of a single common understanding search process provides the framework for using all of the knowledge sources in ways that are similar enough for the results to be combined; in the old pipelined architecture (Figure 2)
|
From page 250... ...
and the semantic links between those concepts, given a set of local linguistic structures (called grammatical relations) and the a priori likelihood of semantic links between concepts.
|
From page 251... ...
For example, the methodology for the SLS program can be used for both spoken language systems (with speech input) or just NL systems (by omitting the SR component and giving the NL system a word string as input)
|
From page 252... ...
achieve an understanding error rate of about 6 percent, which appears to be quite close to the threshold of real utility for applications. Detailed descriptions of the methodology, as well as the underlying databases and annotated corpora for the ATIS domain (as well as many other NL corpora)
|
From page 253... ...
Hirschman, L., et al., "Multisite Data Collection and Evaluation in Spoken Language Understanding," in Bates (ed.) , Proceedings of the ARPA Workshop on Human Language Technology, Morgan Kaufmann, 1993.
|
Key Terms
This material may be derived from roughly machine-read images, and so is provided only to facilitate research.
More
information on Chapter Skim is available.