Afsaneh Fazly -- Research

My research over the past ten years has aimed at advancing the linguistic and psychological theories, by building, and studying the behaviour of, cognitive models of language acquisition and processing; and also by building 'predictive' statistical models of language given large amounts of textual data.

Computational modeling of language acquisition greatly contributes to the understanding of the observed patterns in the course of language development, in particular because it allows a systematic verification of the suggested explanations by separating out the different factors and studying each independently, and by manipulating the quantity and quality of the input.

Predictive models of language in use can be used not just as 'descriptive' statistics summarizing observations about a particular construction, but also for testing the generalizability and predictiveness of contextual and linguistic cues and properties attributed to the construction.

Computational Psycholinguistics

  • Early Word Learning: Word learning is among the most difficult tasks children face early in life, especially since the input they receive is very complex, noisy, and ambiguous. One strand of my research over the past five years has focused on the development of a computational model of word-to-meaning mapping that is cognitively plausible, and hence can provide explanations for a range of issues related to children's vocabulary development. (The model, first presented at the 30th Annual Conference of the Cognitive Science Society, received the 2008 Prize for Best Paper on Language Modeling.) Specifically, we have used the model to study the role of context in the acquisition of low-frequency words, to better understand the complex interaction between cross-situational evidence and guidance from the sentential context in an adult language learning scenario, and to understand factors contributing to an observed delay in the acquisition of vocabulary in late-talking children who are at risk for SLI. My ongoing work focuses on developing a unified model of early language acquisition, which will make it possible to investigate various aspects of language acquisition in a more naturalistic context. Clearly, a comprehensive study of language acquisition must examine: (i) the interactions of various acquisition processes, e.g., speech segmentation, word and concept learning, and syntax acquisition; and (ii) how these interactions are situated within the learning environment, and how they are affected by the strengths and limitations in the human cognitive system.

  • Language and the Mind: One of my recent modeling projects focuses on the acquisition of mental state verbs (MSVs), such as think, know, and wish. Understanding how children acquire MSVs is of importance for the following reasons: (i) MSVs are frequent and polysemous, yet they are produced later than less frequent but more concrete physical action verbs, such as throw. (ii) there is evidence for a developmental connection between producing MSVs in their true mental meaning, and the understanding of Theory of Mind in young children. Our initial modeling of the acquisition of MSVs has provided a precise explanation for the role of the complex syntax of sentential complementation often associated with MSVs: a strong association between the sentential complement syntax and MSVs is not sufficient to explain some of the observed patterns in children and adults' experimental data. Instead, our findings suggest that the data can be simulated as a result of an erroneous association between physical action verbs that appear within the complement of an MSV usage (e.g., drink in "She thinks she is drinking tea") and sentential-complement syntax.

  • Categorization: Categorization is an important aspect of language acquisition. Grouping of linguistic knowledge into meaningful categories enables children perform abstraction and generalization over their learned knowledge. Part of my research has focused on studying the role of morphological, phonological, and contextual cues in children’s acquisition of syntactic categories through computational modeling. I have further investigated the role of categorization in normal and delayed vocabulary acquisition, as well as in adult language learning.

Statistical Modeling of Language in Use

  • Another strand of my research focuses on developing usage-based statistical models for identifying similar-on-the-surface expressions that fall on the literal-to-idiomatic continuum of meaning. E.g., make a cake has a literal meaning, make a decision is a light verb construction (with a non-literal, non-idiomatic meaning), and make a killing is an idiom that means "to have a great financial success". In particular, I have devised statistical models that tap into the lexical, syntactic, and semantic properties of idioms and light verb constructions (LVCs). The models examine the distributional behaviour of such expressions in a large corpus, and summarize them into quantitative measures that can be used to predict the semantic class of a new (previously unseen) expression.

    LVCs are a particularly interesting class of non-literal expressions because: (i) they are prevalent in many (genetically-unrelated) languages, and (ii) it is not clear whether they should be treated as lexical units or as grammatical constructions. My ongoing work concerns with the expansion of the statistical models to the identification of LVCs in other languages (e.g., Persian), and to enrich a lexical resource (VerbNet) with appropriate information about LVCs.

Learning from Multimodal (Visual/Linguistic) Data

  • Object recognition and image annotation by using linguistic labels.

  • Word sense disambiguation by drawing on both linguistic and visual contexts.