- Diversity, Equity, and Inclusion
- Research News
- Department Life
- Oral History of Cornell CS
- CS 40th Anniversary Booklet
- ABC Book for Computer Science at Cornell by David Gries
- Department Timeline
- Job Postings
- Ithaca Info
- Internal info
- Graduation Information
- Cornell Tech Colloquium
- Student Colloquium
- Spring 2024 Colloquium
- Conway-Walker Lecture Series
- Salton 2023 Lecture Series
- Seminars / Lectures
- Big Red Hacks
- Cornell University - High School Programming Contests 2024
- Game Design Initiative
- CSMore: The Rising Sophomore Summer Program in Computer Science
- Explore CS Research
- ACSU Research Night
- Cornell Junior Theorists' Workshop
Reinforced Target-driven Conversational Promotion (via Zoom)
Abstract: The ability to promote items and proactively engage with users is highly desired for conversational virtual assistants, which mimic a salesperson to nudge the users towards accepting a designated item. Under such a target-driven conversational promotion setting, existing methods fall short of hands as they only focus on a conventional recommendation scheme, where they first acquire user preferences through the multi-turn conversations and then produce recommendations. The new proactive nudging setting brings in two new challenges: (1) how to carefully plan a sequence of goals or topics which lead to the target while can also keep the user interested; and (2) how to generate proper responses that are in line with the plan while can also quickly adapt to new scenarios.
In this work, we hence propose a Reinforced Target-driven Conversational Promotion framework to address these two challenges. Specifically, RTCP first integrates the long-term planning (achieving the predefined target) and short-term planning (engaging with user) via a balanced gating mechanism. Inside which, the dialogue actions are predicted via a knowledge-integrated multi-head attention and guided via estimated reinforcement learning rewards. RTCP then employs action-guided prefix tuning to generate relevant responses. It pushes action plan factors into different prefix parameters, while leaves the main generation task to the pre-trained language model.
Experimental results demonstrate that our model outperforms state-of-the-art models on both automatic metrics and human evaluation. Moreover, RTCP has a strong capability in quickly adapting to unseen dialogue settings just by adding some prefixes and tuning its parameters without re-training the whole model.
Bio: Le Duy Dung (Andrew) is an Assistant Professor at College of Engineering and Computer Science, VinUniversity. Previously, he was a senior data scientist in Ads and Personalization team, Grab Holdings Inc. and a research scientist in School of Information Systems, Singapore Management University (SMU). He earned his PhD in Data Science and Engineering from SMU, under the supervision of Associate Professor Hady W. Lauw. Formerly, he earned his Degree of Engineer in Mathematics and Informatics from Hanoi University of Science and Technology, Vietnam in 2014.