|
Multiword Compositionality as a Lens on Language Models| title | Multiword Compositionality as a Lens on Language Models |
|---|
| start_date | 2026/04/10 |
|---|
| schedule | 14h-15h |
|---|
| online | no |
|---|
| location_info | A118 |
|---|
| summary | Multiword expressions (MWEs) are a source of difficulty in Natural Language Processing, as their meaning often depends on convention rather than simple composition. Expressions such as cheese knife illustrate that combining words does not necessarily yield additive meaning, while others, such as loan shark or kick the bucket, demonstrate varying degrees of semantic opacity. This talk examines how MWE meaning can be characterised, focusing on the contribution of individual components and the variability observed across expressions. Compositionality is treated as a continuum, reflecting asymmetric contributions among MWE components and the role of conventionalised usage. The discussion also considers how computational models handle MWEs in comparison with more regular linguistic patterns, highlighting where their behaviour remains sensitive to non-compositionality and variation. Recent reasoning-based models are considered in this context, with particular attention to whether more explicit reasoning is associated with improved handling of multiword meaning. The perspective is then extended beyond text, considering multimodal settings where meaning depends on grounding in visual context. Overall, the talk invites reflection on the role of MWEs as a controlled setting for examining how computational models process meaning. |
|---|
| responsibles | Vanzeveren, Gao |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2026/04/07 08:00 UTC |
| |
|