ICCM: Linguistic Phenomena

Even when talking about novel things and without a fully shared vocabulary, people can come to understand each other through communicative turn taking (what we call communicative alignment). State-of-the-art computational models cannot yet explain this capacity, because (1) empirically corroborated models only work under shared knowledge and vocabularies, and leave out interactive processes needed to overcome misalignment; (2) models that do include misalignment and interactive processes cannot account for communicative successes under real-world conditions; and (3) models that overcome the limits in (2) use a theoretical ‘hack’. In this paper, we add a challenge to the list: the interactive processes in both models of type (2) and (3) are intractable. We explore the robustness and implications of this theoretical challenge for models of communicative alignment in general.

This is an in-person presentation on July 20, 2024 (10:00 ~ 10:20 CEST).

We identify theoretical challenges for developing a computational explanation of flexible linguistic inference. Specifically, the human ability to interpret a novel linguistic expression (like mask-shaming), where inferring plausible meanings requires integrating relevant background knowledge (e.g., COVID-19 pandemic). We lay out (i) the core properties of the phenomenon that together make up our construal of the explanandum, (ii) explanatory desiderata to help make sure a theory explains the explanandum, and (iii) cognitive constraints to ensure a theory can be plausibly realised by human cognition and the brain. By doing so, we lay bare the ‘force field’ that theories of this explanandum will have to navigate, and we give examples of tensions that arise between different components of this force field. This is an important step in theory-development because it allows researchers who aim to solve one part of the puzzle of flexible linguistic inference to keep in clear view the other parts.

This is an in-person presentation on July 20, 2024 (10:20 ~ 10:40 CEST).

During conversations, speakers tend to reuse the lexical expressions of their interlocutors. This is called “lexical alignment,” and it facilitates the listener’s understanding of the speaker’s intention. Branigan (2011) has shown that this tendency increases when speakers believe that their partner is a computer agent rather than a human. Memory activation for the expressions used by the interlocutors and the strategy preference whereby speakers attempt to use their partners’ expressions rather than those that first come to mind have been shown to be the causes of lexical alignment. For this study, we constructed an ACT–R model for which we could adjust the parameter values related to these two features. Through parameter adjustment, we simulated lexical alignment with both human and computer agents. For both partner conditions, additional activation was added to the knowledge of the partners’ expressions. The computer–partner model preferred trying to retrieve the partners’ expression rather than using the knowledge that had a strong association with the stimulus and was easy to retrieve. In contrast, the human–partner model had no specific preference; that is, it displayed equal utility for both. A comparison of these parameter values revealed that the computer–partner model preferred to retrieve the partner’s knowledge; in addition, it also kept the knowledge’s activation sufficiently high so that it could be available for a longer duration.

This is an in-person presentation on July 20, 2024 (10:40 ~ 11:00 CEST).

Language development is supported by phonological awareness, which is related to attention to phonological aspects of spoken language. We aim to develop a system that supports phonological awareness formation using cognitive models. Estimating the state of a user's phonological awareness is a kind of identification of the user's "auditory filter." This paper reports on an experiment with typically developed native speakers by setting up an audio filter that is applied to the system's output sound. The user's phonological awareness is estimated as a relative preference for two computational models presented by the system. Using the system with audio filters, we test the hypothesis that there is a difference in participants' selection behavior depending on the characteristics of the model under the application of the audio filter. The results of the experiment showed that there was a difference in the selection probability between models only when a specific audio filter was applied.

This is an in-person presentation on July 20, 2024 (11:00 ~ 11:20 CEST).