Towards Controllable and Personalized Review Generation

Pan Li, Alexander TuzhilinEMNLP 2019 · DOI

TL;DR

This paper introduces RevGAN, a novel generative model that creates controllable and personalized user reviews. By leveraging self-attentive autoencoders, conditional discriminators, and personalized decoders, RevGAN significantly outperforms existing methods in generating coherent, contextually relevant, and human-like reviews.

What Problem Does This Paper Address?

The proliferation of online marketplaces has led to an explosion of user-generated reviews. While these reviews are invaluable for consumers, the sheer volume and the effort required to write them mean that only a fraction of users contribute. This creates a need for automated review generation systems that can produce high-quality, informative reviews. Existing methods, while achieving some success, often struggle with generating reviews that are not only coherent and grammatically correct but also personalized to specific users and products, and controllable in terms of sentiment and style. Furthermore, many models neglect the hierarchical structure of reviews (word-sentence-paragraph) and fail to incorporate crucial product information, leading to less credible and diverse outputs.

Methods

RevGAN tackles these challenges through a three-stage process:

  • Stage 1: Self-Attentive Recursive Autoencoder: This component maps discrete user reviews and product descriptions into a continuous latent space. It uses bidirectional GRUs for encoding and decoding, capturing both word-level and sentence-level information. A self-attention mechanism is incorporated to weigh the importance of different words within a sentence, and a recursive structure builds paragraph embeddings from sentence embeddings, effectively modeling the hierarchical nature of reviews.
  • Stage 2: Conditional Discriminator: To control the sentiment of generated reviews, RevGAN employs a novel conditional discriminator. Instead of directly feeding sentiment labels to the generator, the sentiment label is conditioned on the discriminator. This forces the generator to adapt its output to match the desired sentiment more effectively, improving controllability.
  • Stage 3: Personalized Decoder: To imbue reviews with a user's unique writing style, a personalized decoder is used. It takes the user's historical reviews, calculates a "writing style vector" based on word frequencies, and concatenates this vector with the generated review embeddings. This guides the decoder to generate words that align with the user's typical vocabulary and expression.

Datasets

The authors evaluated RevGAN on three subsets of the Amazon Review Dataset: Musical Instrument, Automotive, and Patio. These datasets comprise a significant number of reviews and users, providing a robust basis for evaluation.

Key Insights and Ideas

  • Hierarchical Structure Matters: Explicitly modeling the word-sentence-paragraph hierarchy within reviews is crucial for generating coherent and longer text, a limitation in many prior works.
  • Conditional Discriminator for Control: Conditioning sentiment on the discriminator, rather than directly on the generator's input, proves to be a more effective strategy for achieving controllable sentiment in generated reviews.
  • Personalization Enhances Credibility: Incorporating a user's writing style significantly improves the perceived quality and helpfulness of generated reviews, making them more akin to human-written content.
  • Linguistic Law Adherence: The generated reviews not only mimic human writing in style and content but also adhere to established linguistic laws like Zipf's Law and Heap's Law, suggesting a deeper level of linguistic fidelity.

Key Results

RevGAN significantly outperforms state-of-the-art baseline models across various metrics, including Log-Likelihood, WMD, BLEU, and ROUGE. Specifically, the full RevGAN model (RevGAN+CD+SA+PD) achieved the highest scores, demonstrating the effectiveness of combining all three novel components. Human evaluations indicated that generated reviews were statistically indistinguishable from original human-written reviews, with comparable helpfulness scores. Controllable generation accuracy was also superior to existing methods like Attr2Seq.

Limitations and Future Work

The paper acknowledges the "cold start" problem for users with no historical reviews, for which they propose setting the writing style vector to an identity matrix. Future work includes exploring methods to incorporate specific keywords as input for review generation and developing novel techniques to better distinguish between machine-generated and organic reviews.

Summary updated

BibTeX
@inproceedings{li2019towards,
  author = {Pan Li and Alexander Tuzhilin},
  title = {{Towards Controllable and Personalized Review Generation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2019},
  doi = {10.18653/v1/d19-1319},
}
Figure