|
Conversational Grounding in Dialog Systems| title | Conversational Grounding in Dialog Systems |
|---|
| start_date | 2023/10/20 |
|---|
| schedule | 11h |
|---|
| online | no |
|---|
| location_info | Visioconférence Big Blue Button |
|---|
| summary | In linguistics, Clark and Brennan propose the concept of "common ground" - the mutual knowledge and mutual assumptions that are essential for successful communication. This common ground is accumulated over the course of a conversation and is built via words, of course, but also through the use of other modalities: pointing to objects in the environment, nodding to indicate that one has understood, and staring at the speaker to indicate that one needs more information. This interactive process of building a common ground during a conversation by making sure that the interlocutors have understood the information that is being exchanged is called conversational grounding. The utterances have an underlying uncertainty which is negotiated and removed by the participants before getting added to the shared information. Today’s dialog systems use language models extensively for processing and generating utterances. However, previous work has shown a lack of conversational grounding capabilities in the language models. In this talk, I'll provide an overview of conversational grounding, highlighting its significance and the challenges it presents. I will share our efforts in annotating datasets to study and model Grounding Acts and Grounding Units. I'll also unveil our methodology in crafting test cases utilizing these annotations and share our critical observations. I will then talk about our work of creating test cases using these annotated data and discuss our findings. The talk will conclude with a brief mention of our ongoing work on storing and representing grounded information. |
|---|
| responsibles | Bawden |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2023/10/18 11:25 UTC |
| |
|