In this chapter we will consider the practical applications of quantum computation in natural language processing. In the conventional approach to natural language processing, a ``rule-based'' linguistic framework [5,42,43,6] is usually employed in designing NLP systems. These approaches analyze the data and the regularities of natural languages in order to formalize the results as explicit statements about how to manipulate symbols. These statements are called linguistic rules. Accompanied and motivated by the symbol manipulating power of digital computers, many formalisms have been developed. Interestingly, at one time this was assumed to be the only right way for NLP and was considered a synonym for computational linguistics by many researchers in this field. This is referred to in the following sections as a rationalist view of computational linguistics. A schematic description of rationalist natural language processing is illustrated in Figure 7.1.
As shown in the figure, the ``raw'' linguistic data (either as phonetic transcripts or orthographic expressions) are subject to a rule-based analyzer and transformed into a well-defined abstract structure. All the subsequent processing is done on the abstract structures. For example, in a machine translation task the structure can be a parse-tree or some other graph. The translation is a series of rule-based symbol-manipulations done on the graph. The resulting structure is subject to another rule-based system to generate the result either as phonetic transcript or orthographic expression. All these rules are mostly hand-coded by experts who may work closely with conventional linguists.
However, a glance at natural language data from everyday life shows that there are many anomalies that are difficult, if not impossible, to be accommodated in a rationalist framework7.1. These would have been regarded as ``pathological'' from the rationalist viewpoint and treated exceptionally. The use of language is strongly influenced by the sociocultural environment of the speakers7.2, as well as non-linguistic factors7.3. These turn out to be very difficult to model in a rationalist framework. Instead of blaming the speakers for not using the language correctly, a sophisticated NLP system should at least take these issues into serious account. In fact, all these ``anomalies'' may be just as normal as other ``authentic'' usages. In this regard, the rationalist linguists are incomplete at best.
Owing to these shortcomings, there is recent surge in the number of ``bottom-up'' approaches to NLP, most of which are motivated by speech recognition research. In these approaches one tries to shift the burden of gathering empirical data from human experts to computer programs. Moreover, one emphasizes the method of describing the language per se instead of using abstract grammar rules which are essentially knowledge about the language. These approaches are called empiricist in the following.
The philosophy of an empiricist computational linguistic NLP is to keep the linguistic formalism minimal and let a carefully designed mechanism gather rules by itself, although these rules might be unintelligible to a human. Particularly in a very practical application of NLP such as machine translation, the example-based [44,45,46,47] and statistical approaches [48,49,50] both assume this thesis and have achieved certain success. Specifically, an example-based machine translation approach acknowledges the need to extract empirical rules from corpora. These rules are, however, ``shallow'' in comparison with those of a rationalist approach. Thus an example-based machine translation system does not inherently exclude the possibility of applying a theoretical linguistic framework. In this regard, it deserves to be called a hybrid approach. A purist statistical machine translation, on the other hand, assumes a totally empirical modeling of natural language and rejects any top-down knowledge about the language as a whole. From a statistical NLP's point of view, the mechanism is just a hidden Markov model (HMM) [51] or a frequency/position distribution estimator, and that is by no means linguistic in its conventional sense7.4.
The framework of an empiricist approach to NLP (e.g. machine translation) is illustrated in Figure 7.2. As shown in the figure, the task of a NLP system is to gather reliable parameters from natural language corpora. The ``raw'' linguistic data are then modeled by these parameters and subject to other ``parameter-manipulations.''
While the aforementioned empiricist approaches emphasize the learning capacity of an NLP system, they do not assume that the underlying hardware of an NLP system should resemble our brain. This is mostly due to practical considerations. In fact, research on the neurological bases of language shows that the ``hardware'' of human language is very complicated and very likely works according to a principle totally different from that of a Turing machine (today's digital computer) [52,53] or a hidden Markov model mechanism. Interestingly, while many aforementioned NLP frameworks have the ability to ``learn,'' the process of acquiring a first and second language reveals quite a lot properties which have been overlooked in these approaches [54,55,56].
Recently there have been attempts to employ connectionist techniques for modeling cognition in general and NLP in particular [57,58,59]. Connectionism carries naive empiricism one step further. For instance, connectionism provides an alternative and a convenient method of gathering linguistic rules and generating the representations of the symbols implicitly. Those symbols can then be manipulated by the underlying computational agent. Few connectionists will dispute the strength of classical symbols. Most connectionists are convinced, however, that connectionism implements and extends the classical symbolic approach. As a result, considerable effort has been devoted to establish the correspondence between the ``rules'' and the ``mechanism'' of an artificial neural network [60,61,62,63,64,65]. That is, how can neural mechanism implement symbols and rules? While promising in some limited areas, such as pattern recognition, a connectionist approach is nevertheless based on the assumption of classical physics. This undermines its ability to account for many interesting aspects of human language phenomena.
From a pure engineering point of view, all these approaches are in some way productive. In fact, one might argue that this is what NLP as engineering is about. However, owing to the theoretic weakness of these approaches (see the discussion in Chapter 4), they are not very plausible scientific accounts. History teaches us that a correct scientific theory usually leads to a more fruitful engineering application. In this regard, it is the aim of this chapter to show that a quantum mechanical approach to natural language processing is also efficacious in engineering.
