How RAGE Accelerates Accurate Answer Generation for E‑Commerce Q&A

Introduction

With the rapid growth of online shopping, users increasingly rely on product details and reviews to make purchase decisions, but browsing and filtering massive comment data is time‑consuming and degrades the shopping experience. Existing community Q&A (CQA) services alleviate this partially, yet waiting for human answers remains inefficient. To address these challenges, e‑commerce platforms are exploring intelligent QA systems that can automatically and promptly provide accurate answers.

2 Model

2.1 Basic Structure

2.1.1 Question Encoder

To overcome the limitations of recurrent networks, a gated convolutional neural network (GCNN) is employed to encode the question sequence. Position embeddings and part‑of‑speech (POS) tags are incorporated into the input matrix to preserve positional information and syntactic cues.

Given a tokenized question Q of length Q, each word is represented by concatenating its word embedding w, position vector t, and POS vector p (Equation 2‑1).

Stacked GCNN layers expand the receptive field, and residual connections ensure stable gradient propagation (Equation 2‑2). The final encoded representation combines the original word vector with the highest‑layer output via a residual connection (Equation 2‑3).

2.1.2 Basic Decoder

The decoder mirrors the encoder, using a multi‑layer GCNN to generate the answer token by token. At time step j, the decoder receives the previously generated word a_{j‑1} and encodes it with word, POS, and position embeddings.

Unlike the encoder, the convolutional kernel in the decoder is masked to prevent access to future tokens (Equations 2‑4 and 2‑5).

Hierarchical attention integrates relevant question information into the generation state (Equations 2‑6 and 2‑7).

The final word distribution is computed from the top‑layer state (Equation 2‑9).

2.2 Review Extraction, Representation, and Fusion

2.2.1 Review Extraction

Word Mover’s Distance (WMD) measures semantic similarity between a QA pair and review fragments. After tokenizing and removing stop words, each text is represented as a bag‑of‑words, and the optimal transport cost yields the distance (Equation 2‑10).

2.2.2 Review Representation

Relevant review fragments are weighted by term frequency and cosine similarity with the QA pair, producing a weighted dictionary (Equations 2‑11 to 2‑13).

2.2.3 Review Fusion

A hierarchical attention mechanism combines the encoded question vector with the weighted review dictionary, allowing the decoder to selectively incorporate external information (Equations 2‑14 to 2‑17).

3 Experiments

3.1 Baseline Models

Seq2seq with Attention (Bahdanau 2016)

TA‑Seq2seq (Xing 2017)

ConvSeq2seq (Gehring 2017)

ConvSeq2seq‑RV (adds review‑based word constraints)

RAGE/POS (RAGE without POS information)

3.2 Datasets

Two real “Ask Everyone” datasets from Taobao are used: a small “mobile” set (4,457 items) and a larger “large‑appliance” set (47,979 items). Test sets are created by holding out a subset of items per category.

3.3 Evaluation Metrics

Objective metrics: Embedding‑based Similarity (ES) and Distinct‑n. Subjective scoring ranges from 0 (nonsensical) to 3 (fluent, relevant, and information‑rich). Inter‑annotator agreement is measured with Cohen’s κ.

3.4 Results and Analysis

RAGE consistently achieves higher ES and Distinct scores than baselines, especially on the more diverse large‑appliance dataset. Introducing POS tags further improves fluency. Training and inference speed are significantly faster for convolution‑based models compared to recurrent ones.

4 Conclusion

The RAGE model demonstrates that leveraging non‑structured review information with a multi‑layer gated convolutional architecture yields fast, fluent, and semantically accurate answers for e‑commerce QA tasks, outperforming existing seq2seq and attention‑based methods.