Towards Deep Conversational Recommendations

TL;DR

<3-5 sentence high-level summary> The paper addresses the challenge of building neural dialogue systems for goal-directed conversational recommendations by introducing ReDial, a large real-world dataset of movie-recommendation conversations. It proposes a modular neural architecture that combines a GenSen-based hierarchical encoder, a switching decoder, dynamic per-movie sentiment analyses, and an autoencoder recommender pretrained on MovieLens to tackle cold-start scenarios. The approach enables isolated study of subcomponents (sentiment, recommendation, dialogue) while enabling end-to-end dialogue generation through a switching mechanism that injects concrete movie recommendations into conversations. Empirical results on subtask evaluations, coupled with human judgments, show improvements over a standard HRED baseline and demonstrate the value of pretraining and modularization for conversational recommendations. The work provides a practical dataset and a flexible framework applicable to other product domains beyond movies.

Abstract

There has been growing interest in using neural networks and deep learning techniques to create dialogue systems. Conversational recommendation is an interesting setting for the scientific exploration of dialogue with natural language as the associated discourse involves goal-driven dialogue that often transforms naturally into more free-form chat. This paper provides two contributions. First, until now there has been no publicly available large-scale dataset consisting of real-world dialogues centered around recommendations. To address this issue and to facilitate our exploration here, we have collected ReDial, a dataset consisting of over 10,000 conversations centered around the theme of providing movie recommendations. We make this data available to the community for further research. Second, we use this dataset to explore multiple facets of conversational recommendations. In particular we explore new neural architectures, mechanisms, and methods suitable for composing conversational recommendation systems. Our dataset allows us to systematically probe model sub-components addressing different parts of the overall problem domain ranging from: sentiment analysis and cold-start recommendation generation to detailed aspects of how natural language is used in this setting in the real world. We combine such sub-components into a full-blown dialogue system and examine its behavior.

Figures (5)

• Figure 1: Our proposed model for conversational recommendations.
• Figure 2: Confusion matrices for movie sentiment analysis on the validation set.
• Figure 3: Results of human assessment of dialogue quality. The percentages are relative to the total number of ranking tasks, so that bars of the same color sum to 1.
• Figure 4: Data collection interface.
• Figure 5: 2D embedding of movies in our conversation database. The edge weight in the similarity matrix is proportional to the number of co-occurrences in the same dialogue. Left: all movies, colored by number of occurrences from light blue (low) to red (high). Right: names of movies with highest number of occurrences. Embedding via jacomy2014forceatlas2.