No call today

No call today.

I am busy working on my analysis of NLU and NLG.  A useful metaphor is that of a production line where each stage of the line involves applying different processes, which occur concurrently at all stages just like in a car factory. For NLU the production line stages are:  phonological, morphological, syntactic, semantic and pragmatic. This is reversed for NLG.

The analysis considers what processes are needed to handle a set of example utterances. For now, I will ignore phonology and start from a sequence of words. Processing is incremental, i.e. word by word. This can be contrasted with approaches that work on complete utterances, however, that is cognitively implausible given evidence from eye tracking and limitations on working memory.

For each word, it is possible to describe the processing needed at each stage in the production line, in terms of the application of condition-action rules. The target for NLU is a graph that models the meaning of an utterance in terms of its semantics and pragmatics. This, moreover, should include all of the information needed for NLG. The analysis also needs to cover the corresponding processing at each stage of the production line for NLG.

Whilst there is an extensive body of work on linguistics, there is little work on an integrated processing model for each of the stages, and even less on representing meaning in a practical way.  The work on linguistics provides a rich vocabulary, e.g. tense, mood, modal operators, part of speech, and so forth. My conjecture is that concurrent processing at each of the stages can avoid the need for backtracking except where processing breaks down, e.g. for words or phrasing that the system hasn’t seen before. Moreover, the overall processing time will be roughly linear with the number of words in an utterance.

If anyone wants to help with the details of the analysis, I would be happy to arrange a call to progress the work. The aim is to follow the analysis with a web-based proof of concept implementation that combines NLU and NLG.

Best regards,

Dave Raggett <dsr@w3.org> http://www.w3.org/People/Raggett
W3C Data Activity Lead & W3C champion for the Web of things