Introduction
In Community Question Answering (CQA) platforms like Stack Overflow or Quora, identifying the most relevant answer from a pool of candidates is a critical challenge. This task, known as Answer Selection, is complicated by two primary factors: the lexical gap—where questions and answers use different vocabulary for the same concepts—and the omission of user expertise, where models treat all contributors as anonymous entities regardless of their historical performance.
This paper introduces UIA-LSTM-CNN (User-Interactive-Attentive LSTM-CNN), a novel architecture that addresses these hurdles by combining a hybrid attention mechanism with deep user modeling based on historical answer content.
The UIA-LSTM-CNN Architecture
The proposed model integrates three distinct components to create a comprehensive representation of the question, the answer, and the user's expertise.
1. Hybrid Attention Mechanism
Standard attention mechanisms often focus on a sentence in isolation. UIA-LSTM-CNN employs a dual-strategy approach:
- Individual Attention: Identifies locally important keywords within a single sentence.
- Interactive Attention: Uses Information Entropy to measure how well a word in one sentence aligns with the other. Low entropy indicates a strong, specific alignment, helping the model filter out "stop words" that appear frequently but lack discriminative power.
2. Deep User Modelling
To overcome the "cold-start" problem of identifying expertise, the model represents users through the content of their past contributions rather than simple IDs. All previous answers by a user are concatenated and processed via a CNN to extract a latent expertise vector. An attention layer then determines the relevance of this expertise to the current question.
3. Latent Representation and Fusion
Sequential dependencies are captured using LSTMs. The final representations of the question, answer, and user expertise are fused into a single vector for binary classification (Correct vs. Incorrect).
Experimental Results
The model was evaluated on the SemEval-2016 Task 3 dataset, which includes over 5,000 questions and 40,000 QA pairs. The UIA-LSTM-CNN achieved a Mean Average Precision (MAP) of 0.8086, significantly outperforming baselines like AI-CNN (0.788) and AMRNL (0.704).
Hyperparameter tuning revealed that the model's performance is sensitive to the depth of the recurrent layers. As shown in the figure above, both MAP and Accuracy peak when using two LSTM layers, after which performance slightly degrades due to increased complexity or overfitting.
Conclusion
By leveraging both the semantic interaction between sentences and the historical expertise of users, the UIA-LSTM-CNN provides a robust framework for answer selection. Future work may involve incorporating broader metadata, such as reputation scores or social network connections, to further refine user modeling in increasingly large and diverse online communities.
Summary updated