Introduction
Community-based Question Answering (CQA) platforms like Yahoo! Answers and Baidu Zhidao are vital knowledge-sharing hubs. However, they frequently suffer from the "lexical gap"—a phenomenon where users ask semantically identical questions using different vocabularies. Traditional keyword-matching algorithms, such as Okapi BM25, often fail to bridge this gap. To address this, researchers at Fudan University developed the Convolutional Neural Tensor Network (CNTN), a unified architecture designed to capture deep semantic interactions between questions and answers.
Methodology: The CNTN Architecture
The CNTN integrates two sophisticated deep learning components into a single end-to-end pipeline: sentence modeling and semantic matching.
Sentence Modeling with CNNs
The model uses a Convolutional Neural Network (CNN) to encode variable-length text into fixed-length "semantic vectors." This process begins with a lookup layer for word embeddings, followed by convolutional filters that capture local context. A $k$-max pooling layer is then applied to extract the most salient features, ensuring the representation is invariant to the specific position of key terms.
Semantic Matching via Neural Tensor Layers
The core innovation of the CNTN is the Neural Tensor Layer. Unlike standard approaches that rely on simple cosine similarity or basic Multi-Layer Perceptrons (MLPs), the tensor layer uses a 3D tensor to model multiple, complex interactions between the question and answer vectors. This allows the model to learn a highly expressive "metric" for relevance, capturing "strong interactions" that simpler models miss.
Experimental Evaluation
The CNTN was rigorously tested on two large-scale real-world datasets:
- English: 312,000 question-answer pairs from Yahoo! Answers.
- Chinese: 423,000 question-answer pairs from Baidu Zhidao.
The model was trained using a contrastive max-margin criterion, which penalizes the model if the score for a correct answer is not significantly higher than the score for a randomly sampled "corrupted" answer.
Key Results
The CNTN demonstrated state-of-the-art performance in both QA Matching and Question Retrieval tasks.
QA Matching (Precision@1)
The CNTN achieved a P@1 of 75.8% for English and 68.5% for Chinese, significantly outperforming standard CNN+MLP baselines (68.5% and 66.1%, respectively).
Question Retrieval (Mean Average Precision)
In the more challenging retrieval task, the CNTN outperformed both traditional methods and translation-based models. For English, it achieved a MAP of 43.9, compared to 38.6 for TransLM and 32.5 for Okapi BM25.
Conclusion and Future Directions
The CNTN successfully bridges the lexical gap by jointly learning sentence representations and matching metrics. Its language-agnostic nature allows it to perform robustly across different linguistic contexts without the need for complex preprocessing. Future work includes addressing segmentation errors in Chinese and incorporating category-specific information to further refine semantic understanding.
Summary updated