Description
In this talk I will review research on lexical representations as learned by BERT and CLIP, focusing on four main aspects from a psycholinguistic point of view. First, I will present empirical studies on representations of isolated words and the degree to which they account for semantic effects in isolated word processing tasks, such as (primed) lexical decision and semantic relatedness ratings. Humans exhibit consistent semantic effects for words outside of context, and I contend that LLMs should too, if we are to take them seriously as models of semantic memory.I will then consider how linguistic context affects and shapes lexical representations, considering discrete senses as well as subtler modulations: a tomato growing in the garden and a tomato boiling in a pasta sauce may be the same tomato but arguably differ in many semantic aspects and associations. Contextualised embeddings learned by LLMs should be able to capture such variations that extend beyond polysemy. From here I will branch out to multimodal models which incorporate images, and assess whether the multimodal training improves the underlying semantic representations.
Finally, I present work on novel words, probing their linguistic and visual associations under the assumption that sub-word tokens carry non-arbitrary information in their co-occurrence distributions that can be used to gauge the meaning of a previously unheard string, even outside of context. I will conclude by drawing attention on the flaws and promises these models hold for the study of the mental lexicon and semantic memory.
| Period | 6 Sept 2023 |
|---|---|
| Held at | Chinese Univ Hong Kong, Chinese University of Hong Kong, Hong Kong |
| Degree of Recognition | International |