Contextual Conversational Recommendation in the Movie Domain

Abstract

The purpose of a Context-aware Conversational Recommender System is to address the limitations of traditional recommender systems, which often overlook the broader context in favor of past user preferences and movie qualities. The current landscape shows that users express their preferences more explicitly through contextual information, adding a new dimension to recommendation accuracy. Our research combines the strong sides of both the Conversational Recommender System and the Context-aware Recommender System for a movie recommendation, which can be used for research purposes.

The main objective is to incorporate contextual information into a movie recommender system. Our goal is to enhance an already existing Conversational Recommender System with contextual information to improve the recommendation process and create a contextual dataset for movies from movie reviews. To achieve this goal we pre-processed the movie reviews dataset and analyzed how users talk about contexts. Extracted contextual information, generated candidate sentences, and controlled candidate phrases with few-shot prompting. Enhanced NLU and User Model modules with contextual information and finally updated the movies database with contextual information.

To evaluate the success of our system, we used a dataset containing movie recommenda- tion dialogues, enriched with contextual information. We analyzed the strengths and weaknesses of natural language understanding (NLU) in two different approaches (NLU based on BERT and keyword extraction).

The results demonstrated that extracting contextual information from the movie reviews dataset was successful, with large language models (LLMs) significantly aiding the context detection process. The enhanced system successfully detects context related to time, companion, and location, recommends movies accordingly from the enriched movie database, and provides explanations based on this contextual information.