DescriptionIt is common linguistic knowledge that the phonology and orthography of a word in a language say nothing about the meaning of that word. This is supposedly true in two ways. First, there is no natural link between things and words used to refer to them: different languages name different things differently, accounting for philogenetic links. Second, within a same language, there is no systematicity between form and meaning at the morpheme level. I will focus on contradicting the second statement, showing that it is indeed possible to guess something about the meaning of a word by only knowing its form, without any referential or contextual knowledge, using simple computational methods.
I am going to review three studies which exploit simple linear mappings to estimate the semantic embedding of a word from its form. The first study successfully used this technique to account for behavioral patterns in an experiment where children were asked to match pseudo-words to pictures. The second shows that the mismatch between observed word embeddings (learnt using distributional semantic methods) and word embeddings estimated from form only accounts for a significant portion of variance in age of acquisition norms. Specifically, words with lower mismatches are learned earlier. The last study shows that the same mismatch also accounts for a significant portion of variance in reaction times in lexical decision tasks. Taken together, these results suggest that the form of words conveys a lot more information than typically thought and shows that this information affects language acquisition and processing.
|Period||8 Jan 2020|
|Held at||CIMeC, University of Trento, Italy|
|Degree of Recognition||Local|