Universiteit Leiden

nl en

Lecture | LACG Meetings

The pervasive role of style and the surprising inefficacy of informativity in lexical choice

Date
Thursday 13 April 2023
Time
Series
LACG Meetings
Location
Online via MS Teams

Abstract

Language production is a difficult process. Producers have at their disposal a vast (potentially infinite) array of possible names and descriptions when they set out to convey a message when writing, signing, or speaking, alone or in dialogue. In the case of choosing between closely related words that differ primarily in their length, such as chimp and chimpanzee, a number of factors are at play. Some accounts of language production, such as information theoretic approaches, posit that producers choose what to say depending on how surprising the message is, with prior corpus and experimental work supporting these claims. However, in a recent set of studies (Jacobs & MacDonald, 2023) I have tested another account in which lexical choice does not explicitly optimize the wordforms that people select, but rather is constrained by more global factors, such as the formality of the register in which a word appears. Using cutting edge computational approaches from natural language processing in conjunction with machine learning methods, I will show that we can explain both individual behavior -- judgments of appropriateness in context -- and corpus data. The results suggest a high-level role of the broader utterance on lexical choice than information theoretic approaches have historically admitted. Toward the end of the talk, I will discuss consequences for computational modeling of language production, and specifically highlight the need for modelers to think about contextual factors such as register or socioindexicality that drive utterance planning.

                                                                                                        

To join the meeting, please follow this link. To be added to the LACG group on Teams, please follow this link

This website uses cookies.  More information.