Text generation [2]: product reviews

Building data-driven models that can generate reviews for the given products/ratings helps understand how a specific user comments. A product could thus be promoted to users who have not bought it by generating novel and personalized recommendations. Also a review writing assistant for E-commerce websites can be built from which users could select one candidate review and refine it, which makes the procedure more user-friendly. This attribute-conditioned generation problem is very challenging due to the variety of candidate reviews that satisfy the input attributes; unknown or latent factors influence the generated reviews thus making the process non-deterministic. Moreover, although some attributes (e.g., rating) explicitly determine usage of sentiment words, others (e.g., user information) implicitly influence word usage. Additionally, interactions between attributes are important to obtain hidden factors used for generation, e.g., different users tend to describe different aspects of a product and use different sentiment words to express a rating score [Dong 17].

Contextual (data-to-text) natural language generation helps estimate personalized reviews that a user would write about a product, i.e., to discover their nuanced opinions about each of its individual aspects. A successful model could work (for instance) as (a) a highly-nuanced recommender system that tells users their likely reaction to a product in the form of text fragments; (b) a tool that helps users brainstorm the review-writing process; or (c) a querying system that facilitates personalized natural language queries (i.e., to find items about which a user would be most likely to write a particular phrase). Existing works treat the user and item identity as input; [Ni 18] seeks a system with more nuance/precision by allowing users to guide the model via short phrases, or auxiliary data such as item specifications, e.g., an assistant might allow users to write short phrases and expand these key points into a plausible review. Aspect-aware representation reflects user-aspect preferences/item-aspect relationships as well as helps discover what each user is likely to discuss about each item.

Automatic review generation takes product information/user behavior as input and generates user reviews following arbitrarily given users’ sentiment designation and writing style personalized towards specific product/user. Existing works did not consider inner hierarchical word-sentence-paragraph structure within user reviews, thus making their generation results significantly limited in length and coherence; including production descriptions in the generation process further improves credibility/ diversity? with more words from the user’s history corpus [Li 19a].

Focus of existing work on review generation is word-level, neglecting the importance of topical and syntactic characteristics from natural languages. [Li 19b] in contrast models first aspect transitions to capture overall content flow. Then, to generate a sentence, an aspect-aware sketch will be predicted after which semantic slots are filled by generating corresponding words. Thus, aspect semantics, syntactic sketch, and context information are jointly utilized.

Existing work does not explicitly model factual description of products, tending to generate uninformative content. Moreover, they mainly focus on word-level generation, but cannot accurately reflect more abstractive user preference in multiple aspects. Generation process proposed in [Li 20] contains two major steps, namely aspect sequence generation and sentence generation. First, based on graph capsules, aspect capsules are adaptively learned for inferring the aspect sequence. Then, conditioned on the inferred aspect label, a graph-based copy mechanism is designed to generate sentences by incorporating related entities or words from a hierarchical knowledge graph (HKG). The incorporated KG information is able to enhance user preference at both aspect and word levels.?

[Dong 17] Learning to Generate Product Reviews from Attributes

[Ni 18] Personalized Review Generation by Expanding Phrases and Attending on Aspect-Aware Representations

[Li 19a] Towards Controllable and Personalized Review Generation

[Li 19b] Generating Long and Informative Reviews with Aspect-Aware Coarse-to-Fine Decoding

[Li 20] Knowledge-Enhanced Personalized Review Generation with Capsule Graph Neural Network