Source-linked AI summary
Towards Deep Conversational Recommendations
Raymond Li, Samira Kahou, Hannes Schulz, Vincent Michalski, Laurent Charlin, Chris Pal
TL;DR
Conversational recommendation lacks a large public corpus of real-world recommendation dialogues, limiting systematic study of neural dialogue components. The paper collects REDIAL and uses it to explore modular neural architectures for movie recommendation, sentiment analysis, and dialogue generation. Its modular design supports pre-training recommendation components on larger specialized data and integrating recommendations into recurrent dialogue generation, while interactive human evaluation remains future work.
Problem
Conversational recommendation lacks a publicly available large-scale dataset of real-world recommendation dialogues, despite involving both goal-driven and free-form natural-language interaction.
Method
The paper collects REDIAL, pairs workers as movie recommendation seekers and recommenders, and evaluates modular neural components within a complete dialogue system.
Results
The modular architecture supports pre-training its recommendation engine on larger recommendation-specific data and switching recommendations into a recurrent decoder.
Takeaways & Limitations
REDIAL supports separate or integrated study of sentiment analysis, movie recommendation, and natural-language dialogue in conversational recommendation.
Takeaways & Limitations
Interactive evaluation with human conversational partners remains an open challenge and is left for future work.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper studies conversational recommendation as a natural-language dialogue problem and introduces REDIAL, a large real-world movie-recommendation corpus for evaluating neural components and complete systems.
- Model exploration: Deep-learning-based dialogue systems offer flexible underlying algorithms, which the paper examines through a suite of neural architectures for conversational recommendation sub-problems.
- Dataset contribution: REDIAL provides over 10,000 movie-recommendation dialogues to support research on natural-language conversational recommendation.The corpus addresses the data needs of deep-learning approaches and is publicly released for further research.
- Research setting: Movie discussions combine chit-chat, goal-directed dialogue, and question answering, making conversational recommendation a setting for studying their continuum.
- Model exploration: The paper systematically proposes and evaluates neural models for recommendation, sentiment classification, and other sub-components of a complete conversational recommendation system.The contribution includes cold-start recommendation generation and sentiment prediction within the broader system.
2 Related Work
Prior movie-dialogue resources are either synthetic, self-dialogue, small, or not recommendation-focused, while related recommender systems commonly use modular goal-oriented approaches.
- Movie dialogue corpora: Existing Facebook Movie Dialog datasets include synthetic QA and recommendation corpora derived from ratings data, plus a Reddit corpus whose free-form exchanges do not target recommendations.The recommendation dataset typically produces a single movie-name answer, whereas Reddit dialogue lacks recommendation as its discourse goal.
- Movie dialogue corpora: The Alexa Prize corpus uses workers who simulate both sides of conversations on topics including movies, music, and sport, rather than focusing specifically on movie recommendations.
- Movie dialogue corpora: An older movie-recommendation corpus contained 24 dialogues and 2,684 utterances, compared with over 10,000 conversations and 160,000 utterances in REDIAL.
- Conversational recommender systems: Conversational recommender systems often combine independently trained modules, track slot-value dialogue states, optimize query counts, or use reinforcement learning with simulated users.The paper instead proposes a conditional generative model of conversational recommendation.
3 REDIAL dataset collection
REDIAL pairs workers as recommendation seekers and recommenders in English movie conversations, using structured movie annotations and dialogue constraints to support systematic evaluation.
- Collection setup: Workers are paired through Amazon Mechanical Turk, assigned seeker or recommender roles, and exchange movie tastes and suggestions using natural language.The seeker explains preferences and requests suggestions, while the recommender interprets those tastes and recommends movies.
- Collection setup: Collection instructions require roughly ten messages, at least four movie mentions, formal language, and discussion restricted to movies.
- Entity annotation: Movie mentions are tagged with ‘@’ and matched against DBpedia film names, with release dates helping distinguish movies sharing a title.
- Movie dialogue forms: For each movie, workers label whether it was suggested, whether the seeker had seen it, and whether the seeker liked it.Both workers complete the forms, and the released dataset preserves both sets of answers.
- Movie dialogue forms: The movie dialogue forms enable separate or integrated evaluation of sub-components such as sentiment analysis and movie recommendation.
- Dataset characteristics: The liked/disliked/did-not-say labels are highly imbalanced because participants are more likely to discuss movies they like and recommenders target movies the seeker may like.
- Collection constraints: Pairing workers slows collection because both participants must be online simultaneously, while the demanding task requires qualified workers and appropriate compensation.The reported qualification requirements and reward were decisive for conversation quality.
4 Our Approach
The approach combines hierarchical neural dialogue representations with sentiment analysis, cold-start recommendation, and a switching decoder for conversational movie recommendations.
- System overview: The complete system combines a hierarchical recurrent encoder, sentiment analysis, cold-start recommendation, and a movie-recommendation switching decoder.The architecture is designed to chat with a partner, ask about movie tastes, and make recommendations.
- Hierarchical recurrent encoder: A hierarchical encoder represents utterances and their dialogue context, while appending a role indicator for the seeker or recommender.Utterances use bidirectional GRU representations, which are then processed by a conversation-level GRU.
- Movie sentiment analysis: The sentiment module conditions a hierarchical encoder on a movie name and predicts suggested, seen, and liked labels for that movie.Its output has 14 dimensions, with sigmoid and softmax outputs for the binary and categorical labels.
- Autoencoder recommender: The recommender uses dialogue-inferred movie preferences as input to an autoencoder that produces a full movie-rating vector.The autoencoder is pretrained with MovieLens data and trained as a denoising autoencoder to improve early cold-start recommendations.
- Recommendation switching decoder: During decoding, a switching mechanism combines word generation with explicit movie recommendations conditioned on the dialogue context and decoder state.The movie distribution remains fixed throughout an utterance because no new preference information is gathered while decoding.
- Limitations: The decoder is limited because recommendations depend on mentioned movies rather than language, and the method relies on marking movie mentions with the '@' symbol.Conditioning directly on dialogue language caused overfitting in initial experiments; entity recognition is proposed as a possible remedy.
5 Experiments
The experiments evaluate sentiment analysis, recommendation quality, and response quality using established metrics and utterance-level human comparisons. The authors also identify the absence of fully interactive human evaluation as an unresolved limitation.
- Evaluation setup: The evaluation separately measures recommendation and sentiment-analysis modules and compares generated responses from different models in similar contexts.The paper treats these module-level metrics as a proxy for overall dialogue quality and performs utterance-level human evaluation.
- Limitations: Fully interactive conversations with human users remain unevaluated and are left for future work.
- Movie sentiment analysis: 81% of movies were liked versus 4.9% disliked, so sentiment-analysis loss weighting compensates for class imbalance.
- Movie recommendation quality: RMSE evaluates movie recommendations on a 0–1 rating scale, with MovieLens results also reported on a 0.5–5 scale.
- Movie recommendation quality: Denoising training slightly improves standard training on MovieLens, while MovieLens pre-training significantly improves performance on the authors’ data.
- Dialogue quality: 349 true responses, 267 model responses, and 223 HRED responses were ranked first in the human assessment.
6 Discussion and Conclusions
The paper contributes REDIAL and a modular neural architecture for conversational movie recommendation. The architecture outperforms HRED in utterance-level comparisons, but whole human–model conversations were not evaluated.
- Contributions: REDIAL is a real-world dataset of human-generated conversations about movie recommendations, containing 10,000 conversations.
- Contributions: The modular formulation supports separate or complete-system exploration of subtasks such as sentiment analysis and movie recommendation.
- Architecture: The architecture combines sentence representations, hierarchical encoder-decoder models, dynamically instantiated RNNs, and an autoencoder-based recommendation engine.
- Architecture: Pre-training the recommendation engine on larger recommendation-specific data sources is enabled by the model’s modularization.
- Evaluation and limitations: The model outperforms the HRED baseline in utterance-level evaluation, but whole conversations between the model and a human user were not evaluated.
A Data collection interface
The supplied passage identifies Figure 4 as the data collection interface.
- Data collection interface: Figure 4 presents the interface used for data collection.
B Training Procedure
Training freezes selected pre-trained components, fine-tunes the recommender with the dialogue network, and generates responses with beam search.
- Training Procedure: The first utterance-encoder layer and sentiment-analysis module are frozen after pre-training, while the recommender is pre-trained on MovieLens and fine-tuned jointly.
- Training Procedure: Adam optimization uses a learning rate of 0.001, and beam search of size 10 generates test responses.
C Conversation samples
The validation-set examples compare responses from HRED and the authors’ model while conditioning on ground-truth previous turns. The section also visualizes movie relationships using co-occurrence-based similarity embeddings.
- Tables 3–7 present randomly selected validation conversations alongside responses generated by HRED and the authors’ model.
- Each generated response is conditioned on the ground-truth previous turns labeled “SEEKER” and “HUMAN.”
- In sample conversations, HRED asks about preferred movie types while the authors’ model opens with a general offer of help.
- Figure 5 embeds movies using a similarity matrix whose edge weights reflect co-occurrences within the same dialogue.