Where's the Natural in Natural Language Processing?

Robert Dale
Macquarie University

Abstract

After some 30+ years of symbolic approaches to natural language processing, the 1990s saw a shift to statistical approaches. Although that shift was driven in part by advances in processing power and storage capacity, it was also driven to some extent by disillusionment with the results that had been achieved so far. I'm not convinced that the shift has gotten us any closer to natural language processing, but I suspect that's because the goal has never really made sense in the first place. In this talk I'll explore this particularly from the point of view of speech recognition, but time permitting I'll also try to say something about other forms of language processing too. Advances in speech recognition over the last 15 years mean that we can now talk to machines, and many of us do so on a daily basis. In my home city, there are widely used systems that allow anyone with a telephone to talk to a machine in order to ask for directory assistance, book a taxi, place a bet or buy and sell shares, amongst many other tasks. Vendors in this space now frequently use the term "natural language speech recognition". I will look at how people talk to machines, and ask how people should talk to machines. The question that interests me is the extent to which this kind of communication can be said to be carried out in natural language. I will argue that, at least in terms of the current behaviour of such systems, the medium of communication that is used is not natural language. I will also suggest that, before we try to give such systems the ability to communicate in natural language, we first need to have a better understanding of what this really means; one possible consequence is that we may not want our machines to communicate in natural language after all.