Article Preview
Top1. Introduction
Human language understanding is generally characterized as a logico-deductive process with an underlying inductive learning procedure. For instance, syntactic and semantic analyses, among others, rely heavily on deductive logical structures to infer meaning. As well, the observation of utterances by children is believed to configure the neural language apparatus inductively. However, learning theories cannot explain how children can learn language so well and so rapidly given the inherently small amount of data available to them. This lack of learning data has resulted in the proposal that there may be innate language structures in the brain (Chomsky, 1969, 2006).
Despite this known deficiency of learning data, with the large quantities of electronic texts and increasing computational power, computational linguists have in the last two decades sought meaning in the enormous amount of co-occurrence information found in corpora. Applications of this idea can be found, for instances, in corpus-based statistical methods like Latent Dirichlet Allocation (LDA) (Blei et al., 2003), Latent Semantic Analysis (LSA) (Landauer & Dumais, 2003; McNamara, 2011), text clustering (Feldman & Sanger, 2006; Jain, 1999; Massey, 2003) and document classification (Qi & Davison, 2009; Sebastiani, 2002).
Although interesting from a technological application point of view, these techniques are unlikely to be representative of the human cognitive processes of language understanding. Indeed, it is implausible that humans achieve understanding by relying on the presence of large quantities of documents in some sort of mental corpora, or learn by observing and generalizing on thousands of previously classified exemplars. One might claim that reproduction of natural processes or biological plausibility is not a requirement for Artificial Intelligence (AI) or Natural Language Understanding (NLU). On the contrary, one might argue that biological and human cognitive processes might be a valuable source of inspiration for research. It is the position taken in this paper. In that regard, instead of investigating artificial intelligence, we rather study natural intelligence and the underlying cognitive processes involved in language understanding.
There are also real and important practical reasons to motivate our alternative approach. The first one is that existing techniques, by relying on large datasets, are tightly coupled with this data. Coupling is generally recognized as an undesirable situation in software engineering, because it causes modules inter-dependence and greatly increases the complexity of maintenance (Pressman, 2004). Similarly in machine learning and artificial intelligence, dependence on data makes intelligent systems less robust and adaptive when facing changing situations, since they are built according to a static specification and specific view of the world represented by the data (Massey, 2005, 2008). However, human language and conceptualizations are constantly changing (Steels, 2006). It therefore appears futile to statistically induce concepts and meaning from large language datasets (Steels, 2007). The second (and related) problem is autonomy (the ability to function without human intervention). Existing systems, by depending on learning from large collections of language data, are not autonomous since this data must be collected, organized and otherwise prepared by humans. Moreover, following deployment, such systems must be maintained – that is, re-trained on new datasets - when the environment changes. Supervised learning systems are particularly susceptible to this situation, but unsupervised methods are not immune either since they require all sorts of human interventions, for example parameter settings and labeling of extracted clusters. It doesn’t seem to make much sense that so-called intelligent systems would be lacking both autonomy and adaptability, two qualities most people would surely expect to observe from anything qualified of ‘intelligent.’