Validating a mental lexicon derived from a large-scale word association project

Semantics is at the core of any language-based activity in our society but arguably one of the most complex aspects to study. To gain a deeper knowledge in how word meaning is processed and stored in human memory, several lexico-semantic models have been proposed. These models have been proven to be instrumental in a number of lines of research, ranging from purely theoretical questions, such as atypical word processing, the structure and acquisition of the mental lexicon in children, to more pragmatic issues, including second language learning, text processing and expert system development.
In studying the meaning of words, several approaches have been undertaken up to this day. In the last decades, distributional semantics models that derive meaning from word co-occurrence patterns in text corpora have been very popular and successful. Older models (e.g., Latent Semantic Analysis or LSA) count co-occurrences within a given context and apply dimension reduction techniques on that count-matrix to reduce dimensions. Newer models (e.g., word2vec) learn word embeddings using a neural network model. Another way of studying semantics is making use of word associations. Although having a long history in psychology, it is only recent that word association data have been collected on a large scale. In all the chapters included in this dissertation, the use of these association data are being validated as a means to study different aspects of semantics. In three of the chapters a direct comparison with distributional semantics models is made.
Chapter 2 deals with the extent to which a word association-based approach can account for lexcico-semantic dimensions like concreteness and valence. It deals with a direct comparison between both type of models in predicting these semantic dimensions in Dutch and English. In Chapter 3, newly collected ratings for the semantic dimension of gender in Dutch are presented. Reliability and validity are assessed and the relationship with grammatical gender is explored. To judge the semantic relevance of semantic gender, word association data are used to test how predictable these ratings are and whether this dimension can be found in a semantic space based on those data. Chapter 4 and 5 take a different approach and use a word association model and a distributional semantic model to create different sets of stimuli to be used in different types of experiments. In Chapter 4 these two types of stimuli are used in a word guessing game to investigate which type of hints lead participants to faster guessing of the correct word. In Chapter 5, the Deese–Roediger–McDermott (DRM) experimental paradigm to elicit false memories is used to see which type of stimuli give rise to more false recognition of non-presented items. The second study in that chapter focusses on several types of semantic relatedness derived from word associations in eliciting false recognition, including backward associative strength as this variable is crucial for many of the theories in the DRM-literature.