[Paper Review] Attention-based Pairwise Multi-Perspective Convolutional Neural Network for Answer Selection in Question Answering

Jamshid Mozafari, Mohammad Ali Nematbakhsh+1 more|arXiv (Cornell University)|Sep 3, 2019

Topic Modeling4 citations

TL;DR

This paper proposes an attention-based pairwise multi-perspective convolutional neural network for answer selection in question answering, leveraging context-sensitive interactions between questions and answers using attention mechanisms and wide convolutions to enhance interrogative word focus. The model achieves state-of-the-art performance on TrecQA Raw, improving MAP by 1.4% and MRR by 1.1% without external syntactic parsers or tools.

ABSTRACT

Over the past few years, question answering and information retrieval systems have become widely used. These systems attempt to find the answer of the asked questions from raw text sources. A component of these systems is Answer Selection which selects the most relevant from candidate answers. Syntactic similarities were mostly used to compute the similarity, but in recent works, deep neural networks have been used, making a significant improvement in this field. In this research, a model is proposed to select the most relevant answers to the factoid question from the candidate answers. The proposed model ranks the candidate answers in terms of semantic and syntactic similarity to the question, using convolutional neural networks. In this research, Attention mechanism and Sparse feature vector use the context-sensitive interactions between questions and answer sentence. Wide convolution increases the importance of the interrogative word. Pairwise ranking is used to learn differentiable representations to distinguish positive and negative answers. Our model indicates strong performance on the TrecQA Raw beating previous state-of-the-art systems by 1.4% in MAP and 1.1% in MRR while using the benefits of no additional syntactic parsers and external tools. The results show that using context-sensitive interactions between question and answer sentences can help to find the correct answer more accurately.

Motivation & Objective

• To improve answer selection in open-domain question answering by capturing deep semantic and syntactic similarities between questions and candidate answers.
• To address the limitations of traditional syntactic similarity methods by leveraging deep neural networks for better representation learning.
• To eliminate dependency on external tools like syntactic parsers by designing a self-contained neural architecture.
• To enhance model performance through context-sensitive attention mechanisms and multi-perspective convolutional interactions.
• To achieve state-of-the-art results on the TrecQA Raw benchmark using only question-answer pairs and no additional linguistic resources.

Proposed method

• The model employs a pairwise ranking framework to learn differentiable representations that distinguish positive (correct) from negative (incorrect) answers.
• Multi-perspective convolutional networks are applied to capture diverse interactions between question and answer embeddings at multiple levels of abstraction.
• An attention mechanism is integrated to dynamically weigh important words in the question and answer, especially interrogative words, enhancing context-sensitive alignment.
• Wide convolutions are used to expand receptive fields, increasing the model's sensitivity to key terms such as 'who', 'what', 'when', etc.
• Sparse feature vectors are incorporated to preserve salient linguistic patterns and improve generalization.
• The architecture is end-to-end trainable, eliminating the need for pre-processing with external tools like dependency parsers or POS taggers.

Experimental results

Research questions

• RQ1 Can an attention-based, multi-perspective convolutional network improve answer selection performance without relying on external linguistic tools?
• RQ2 How effectively can context-sensitive attention mechanisms enhance the alignment between questions and candidate answers?
• RQ3 To what extent do wide convolutions and pairwise ranking contribute to capturing semantic and syntactic similarities in answer selection?
• RQ4 Does the proposed model outperform existing state-of-the-art systems on standard benchmarks like TrecQA Raw?
• RQ5 Can a self-contained neural architecture achieve superior performance while maintaining efficiency and reducing dependency on external NLP components?

Key findings

• The proposed model achieves a 1.4% improvement in Mean Average Precision (MAP) over the previous state-of-the-art on the TrecQA Raw benchmark.
• The model records a 1.1% increase in Mean Reciprocal Rank (MRR), demonstrating better ranking quality for correct answers.
• The model outperforms existing methods without using any external tools such as syntactic parsers or POS taggers.
• The integration of attention mechanisms significantly improves the model's ability to focus on key interrogative words and relevant content.
• Wide convolutions enhance the model's sensitivity to critical terms, contributing to improved performance on factoid question answering.
• The ablation study confirms that both the attention mechanism and pairwise ranking are essential components for achieving optimal performance.