Hi 👋 I’m a 3rd year PhD student at the LISN Lab, Université Paris-Saclay, working on Dialogue Systems and NLP.
My research focuses on the intersection between open-domain dialogue (a.k.a. chitchat) and task-oriented dialogue (TOD), which are typically regarded as two distinct modes. I aim to assess and integrate useful open-domain qualities into TOD systems, enhancing response appropriateness and system resilience to open-domain inputs.
My latest projects include:
Evaluation of the robustness of a Few-shot TOD Bot to inter-mode inputs (task requests that mix in chitchat or that are based on previous chitchat turns). Casting dialogue state tracking as a function call generation task heavily improves JGA and end-to-end performance compared with an existing baseline prompting strategy, and stands out as more robust to inter-mode inputs. Under Review
Dialogue state tracking for inter-mode inputs using TOD data only. Combining a TOD-only state tracker with 0-shot prompting for query reformulation leads to superior and comparable results on two benchmarks (respectively) compared with an LLM fine-tuned on inter-mode data. Under Review
Novel chitchat augmentation for MultiWOZ using in-context learning with Llama-2-70B. Provides a testbed for building more chitchat-resilient TOD systems and enables a system to respond with chitchat and TOD all in the same turn. LREC-COLING 2024 link
A unified approach to user emotion detection and TOD modeling, using a mixture of LoRA fine-tuning and CoT prompting on Llama-2-7B. Leads to improved task success rates and more empathetic responses. IWSDS 2024 link
Comparative study based on entropy of recent Task-Oriented Dialogue datasets enhanced with chitchat. ASRU 2023 link