How RAGE Boosts E‑Commerce QA with Fast, Accurate Answer Generation

Introduction

With the rapid growth of e‑commerce, users increasingly rely on online reviews to make purchasing decisions, but filtering massive comment data is time‑consuming. Existing community Q&A services alleviate this partially, yet still suffer from low efficiency and recall. To address these challenges, the authors propose RAGE, a review‑driven answer generation model that integrates overall structure, comment extraction, representation, and fusion to enhance both response speed and answer quality.

Model

2.1 Basic Architecture

2.1.1 Question Encoder

The model replaces recurrent encoders with a multi‑layer gated convolutional neural network (GCNN). Position embeddings and part‑of‑speech (POS) tags are added to the input matrix to preserve positional information and syntactic cues. Stacked GCNN layers expand the receptive field, and residual connections prevent gradient vanishing. The final question encoding combines the original word vector with the highest‑layer GCNN output.

2.1.2 Decoder

The decoder mirrors the encoder, using GCNN layers to generate each token based on previously generated words. Hierarchical attention incorporates relevant question information at each layer, ensuring the decoder focuses on appropriate context while maintaining causality.

2.2 Comment Extraction, Representation, and Fusion

2.2.1 Extraction

Word Mover’s Distance (WMD) measures semantic similarity between a QA pair and comment fragments. By constructing bag‑of‑words representations and computing optimal transport costs, the model selects the most relevant comment snippets for each question.

2.2.2 Representation

Selected comment fragments are weighted using term frequency and inverse document frequency, then combined with word embeddings to form a weighted lexical dictionary.

2.2.3 Fusion

A hierarchical attention mechanism fuses the weighted comment information with the question encoding at each decoder layer. A gating network selectively integrates external comment cues with the generated state, improving relevance and reducing redundancy.

Experiments

The authors evaluate RAGE on two real‑world Taobao “Ask Everyone” datasets (mobile phones and large appliances). Baselines include Seq2seq with attention, TA‑Seq2seq, ConvSeq2seq, and variants that restrict generation to comment vocabularies. Metrics comprise Embedding‑based Similarity (ES), Distinct, and human ratings (0‑3). RAGE consistently outperforms baselines in both objective and subjective scores, especially on the more diverse large‑appliance dataset.

Efficiency tests on a Tesla K40 GPU show that convolution‑based models (including RAGE) train and infer significantly faster than recurrent‑based counterparts.

Conclusion

RAGE demonstrates that incorporating non‑structured review information via gated convolutional networks and hierarchical attention yields faster, more fluent, and information‑rich answers for e‑commerce QA tasks, surpassing existing generation methods.