PhD Thesis

My PhD Thesis is titled "Exploring the nature of associations: Semantic factors in the formation of word associations". I completed it under the supervision of Prof. Shlomo Bentin, in the Department of Psychology in the Hebrew University of Jerusalem. In my work I undertook a critical examination of underlying notions regarding processes leading to the naturalistic formation of word associations. You can find a copy of the thesis here, and of the appendices here.

Words in the lexicon are associated with one another. Although the consequences of an associative lexical network have been investigated extensively, the mechanisms by which words become associated were investigated to a lesser degree. It is often assumed that words that tend to co-occur frequently in language will become associated and consequently will activate each other in the lexicon. However, the structure of natural language, and the ways in which human speakers use it, cause co-occurrence statistics to be greatly confounded with semantic factors. Words co-occurring in a sentence are purposefully put there by a human speaker, with the goal of expressing a specific idea. The present study was motivated by the hypothesis that semantic context plays a significant or even essential role in the formation of associations between words. In order to test this hypothesis I adopted two complementary lines of research. First, I ran a series of five experiments studying the involvement of semantic contextual factors in the formation of novel associations. Second, I analyzed existing association norms, in order to deduce semantic factors constraining or influencing the formation of associations.

My experimental work demonstrated that during natural language processing, associative links are established between words by virtue of them being processed jointly in a meaningful manner in the context of a sentence. Specifically, direct co-occurrence of the word in a pair might be a pre-requisite, but is not the driving force of association formation. The results further support the claim that sentence-final integrative processes are, indeed, crucial for the incidental formation of associations. Thus, incidental associations are, most likely, formed between sentence constituents during the construction of a mental model, when links are established between the incoming information and existing schemas and world knowledge. These results are also in line with previous findings of semantic influence on associative processes, and further attest to the overlap of the associative domain and the semantic domain.

In the second part of my work I sought to validate the two main conclusions arising from my experimental work. To this end, I carried out a computational analysis of two sets of free-association norms, in Hebrew and in English, with a total of 6,000 associated word pairs. The first experimental finding, regarding the role of sentence integration in association formation, was not unequivocally supported by the computational results. Conversely, the computational results largely corroborate the notion of overlap between semantic an associative links. I conclude that free association norms are not solely determined by co-occurrence, even when it is collected within sentence boundaries. I suggest, therefore, that free association norms can be seen as being simultaneously determined by several forces including sentential co-occurrence, semantic similarity, word class, word frequency, semantic density, concreteness and category typicality. Thus, free association norms reflect to a large degree the semantic structure of the lexicon, but also the dynamics of word activation and retrieval.

Based on the results of both the computational and experimental work, as well as a theoretical analysis of the free association task, I call for the introduction of more precise terminology and a fractionation of the term "association" to several distinct phenomena. Such a fractionation will, in all probability, allow researchers in the domains of language and memory to evaluate empirical results more easily. More precise terms will also foster the formulation of better specified theoretical accounts of how each type of association is acquired, how it can be measured, and what consequence it might have for cognition.