Conversational Recommendation Systems(CRSs) explore user preferences through interactive conversation to make recommendations and generate responses. In this paper, we propose user Perference Conversational Recommendation (UPCR) for CRSs, which captures long-term and short-term user preferences from history conversations and the current conversation, respectively.
python==3.8, pytorch==1.9.0, torch-geometric==2.0.3
We implement our models on TG-ReDial and REDIAL.
- On the TG-ReDial dataset, we directly use annotated topic path, and link them to Conceptnet.
- On the REDIAL dataset, we extract entities from context as topic path, and link them to DBpedia (the linker is provided by KGSF).
- For the topic prediction task :
python /TG-Redial/topic.py - For the recommendation task :
python /TG-Redial/recommendation.py - For the response generation task:
python /TG-Redial/generation.py
- For the recommendation task :
python /REDIAL/recommendation.py - For the response generation task:
python /REDIAL/generation.py
For convenience, our model will report the result on test data automatically after covergence.
Variational Reasoning about User Preferences for Conversational Recommendation