Traditionally, work in natural language processing has tended to view the process of language analysis as being decomposable into a number of stages, mirroring the theoretical linguistic distinctions drawn between SYNTAX, SEMANTICS, and PRAGMATICS. The simple view is that the sentences of a text are first analyzed in terms of their syntax; this provides an order and structure that is more amenable to an analysis in terms of semantics, or literal meaning; and this is followed by a stage of pragmatic analysis whereby the meaning of the utterance or text in context is determined. This last stage is often seen as being concerned with DISCOURSE, whereas the previous two are generally concerned with sentential matters. This attempt at a correlation between a stratificational distinction (syntax, semantics, and pragmatics) and a distinction in terms of granularity (sentence versus discourse) sometimes causes some confusion in thinking about the issues involved in natural language processing; and it is widely recognized that in real terms it is not so easy to separate the processing of language neatly into boxes corresponding to each of the strata. However, such a separation serves as a useful pedagogic aid, and also constitutes the basis for architectural models that make the task of natural language analysis more manageable from a software engineering point of view. Nonetheless, the tripartite distinction into syntax, semantics, and pragmatics only serves at best as a starting point when we consider the processing of real natural language texts. A finer-grained decomposition of the process is useful whenwe take into account the current state of the art in combination with the need to deal with real language data. We identify here the stage of tokenization and sentence segmentation as a crucial first step. Natural language text is generally not made up of the short, neat, well-formed, and well-delimited sentences we find in textbooks; and for languages such as Chinese, Japanese, or Thai, which do not share the apparently easy space-delimited tokenization we might believe to be a property of languages like English, the ability to address issues of tokenization is essential to getting off the ground at all. We also treat lexical analysis as a separate step in the process. To some degree this finer-grained decomposition reflects our current state of knowledge about language processing: we know quite a lot about general techniques for tokenization, lexical analysis, and syntactic analysis, but much less about semantics and discourse-level processing. But it also reflects the fact that the known is the surface text, and anything deeper is a representational abstraction that is harder to pin down; so it is not so surprising that we have better-developed techniques at the more concrete end of the processing spectrum. Of course, natural language analysis is only one-half of the story. We also have to consider natural language generation, where we are concerned with mapping from some (typically nonlinguistic) internal representation to a surface text. In the history of the field so far, there has been much less work on natural language generation than there has been on natural language analysis. One sometimes hears the suggestion that this is because natural language generation is easier, so that there is less to be said. This is far from the truth: there are a great many complexities to be addressed in generating fluent and coherent multi-sentential texts from an underlying source of information. A more likely reason for the relative lack of work in generation is precisely the correlate of the observation made at the end of the previous paragraph: it is relatively straightforward to build theories around the processing of something known (such as a sequence of words), but much harder when the input to the process is more or less left to the imagination. This is the question that causes researchers in natural language generation to wake in the middle of the night in a cold sweat: what does generation start from? Much work in generation is concerned with addressing these questions head-on; work in natural language understanding may eventually see benefit in taking generation’s starting point as its end goal.
As we have already noted, not all languages deliver text in the form of words neatly delimited by spaces. Languages such as Chinese, Japanese, and Thai require first that a segmentation process be applied, analogous to the segmentation process that must first be applied to a continuous speech stream in order to identify the words that make up an utterance. There are significant segmentation and tokenization issues in apparently easier-to-segment languages—such as English—too. Fundamentally, the issue here is that of what constitutes a word; as Palmer shows, there is no easy answer here. The problem of sentence segmentation: since so much work in natural language processing views the sentence as the unit of analysis, clearly it is of crucial importance to ensure that, given a text, we can break it into sentence-sized pieces. This turns out not to be so trivial either. Palmer offers a catalog of tips and techniques that will be useful to anyone faced with dealing with real raw text as the input to an analysis process, and provides a healthy reminder that these problems have tended to be idealized away in much earlier, laboratory-based work in natural language processing.
The previous addressed the problem of breaking a stream of input text into the words and sentences that will be subject to subsequent processing. The words, of course, are not atomic, and are themselves open to further analysis. Here we enter the realms of computational morphology. By taking words apart, we can uncover information that will be useful at later stages of processing. The combinatorics also mean that decomposing words into their parts, and maintaining rules for how combinations are formed, is much more efficient in terms of storage space than would be the case if we simply listed every word as an atomic element in a huge inventory. And, once more returning to our concern with the handling of real texts, there will always be words missing from any such inventory; morphological processing can go some way toward handling such unrecognized words. Hippisley provides a wide-ranging and detailed review of the techniques that can be used to carry out morphological processing, drawing on examples from languages other than English to demonstrate the need for sophisticated processing methods; along the way he provides some background in the relevant theoretical aspects of phonology and morphology.
A presupposition in most work in natural language processing is that the basic unit of meaning analysis is the sentence: a sentence expresses a proposition, an idea, or a thought, and says something about some real or imaginary world. Extracting the meaning from a sentence is thus a key issue. Sentences are not, however, just linear sequences of words, and so it is widely recognized that to carry out this task requires an analysis of each sentence, which determines its structure in one way or another. In NLP approaches based on generative linguistics, this is generally taken to involve the determining of the syntactic or grammatical structure of each sentence. This area is probably the most well established in the field of NLP, enabling the authors here to provide an inventory of basic concepts in parsing, followed by a detailed catalog of parsing techniques that have been explored in the literature.
Identifying the underlying syntactic structure of a sequence of words is only one step in determining the meaning of a sentence; it provides a structured object that is more amenable to further manipulation and subsequent interpretation. It is these subsequent steps that derive a meaning for the sentence in question. It is here that we begin to reach the bounds of what has so far been scaled up from theoretical work to practical application. As pointed out earlier in this introduction, the semantics of natural language have been less studied than syntactic issues, and so the techniques described here are not yet developed to the extent that they can easily be applied in a broad-coverage fashion. After setting the scene by reviewing a range of existing approaches to semantic interpretation, Goddard and Schalley provide a detailed exposition of Natural Semantic Metalanguage, an approach to semantics that is likely to be new to many working in natural language processing. They end by cataloging some of the challenges to be faced if we are to develop truly broad coverage semantic analyses.
Natural Language Generation (NLG)
At the end of the day, determining the meaning of an utterance is only really one-half of the story of natural language processing: in many situations, a response then needs to be generated, either in natural language alone or in combination with other modalities. For many of today’s applications, what is required here is rather trivial and can be handled by means of canned responses; increasingly, however, we are seeing natural language generation techniques applied in the context of more sophisticated back-end systems, where the need to be able to custom-create fluent multi-sentential texts on demand becomes a priority. The Applications part bear testimony to the scope here. A far-reaching survey of work in the field of natural language generation. McDonald begins by lucidly characterizing the differences between natural language analysis and natural language generation. He goes on to show what can be achieved using natural language generation techniques, drawing examples from systems developed over the last 35 years. With laying out a picture of the component processes and representations required in order to generate fluent multi-sentential or multi-paragraph texts, built around the nowstandard distinction between text planning and linguistic realization.