[Paper Review] Product Function Need Recognition via Semi-supervised Attention Network

Hu Xu, Sihong Xie+2 more|arXiv (Cornell University)|Dec 6, 2017

Advanced Text Analysis Techniques23 references3 citations

TL;DR

This paper proposes a semi-supervised attention network (SAN) to recognize product function needs from customer questions in e-commerce QA forums. By leveraging unlabeled questions as contextual side information through an attention mechanism, SAN improves sequence labeling performance, achieving state-of-the-art F1-score of 0.776 on a newly annotated, dense QA corpus for product functionality extraction.

ABSTRACT

Functionality is of utmost importance to customers when they purchase products. However, it is unclear to customers whether a product can really satisfy their needs on functions. Further, missing functions may be intentionally hidden by the manufacturers or the sellers. As a result, a customer needs to spend a fair amount of time before purchasing or just purchase the product on his/her own risk. In this paper, we first identify a novel QA corpus that is dense on product functionality information \footnote{The annotated corpus can be found at \url{https://www.cs.uic.edu/~hxu/}.}. We then design a neural network called Semi-supervised Attention Network (SAN) to discover product functions from questions. This model leverages unlabeled data as contextual information to perform semi-supervised sequence labeling. We conduct experiments to show that the extracted function have both high coverage and accuracy, compared with a wide spectrum of baselines.

Motivation & Objective

• Address the challenge of identifying customer-asked product functions from unstructured QA data in e-commerce platforms.
• Overcome the limitations of sparse labeled data and out-of-vocabulary words in function need recognition by utilizing unlabeled data.
• Develop a novel deep learning framework that integrates attention mechanisms with semi-supervised learning for sequence labeling.
• Create a high-quality, dense QA corpus annotated for product functionality to support future research in function need recognition.
• Improve the coverage and accuracy of function extraction compared to supervised baselines by incorporating contextual information from unlabeled questions.

Proposed method

• Propose a Semi-supervised Attention Network (SAN) that combines a supervised Bi-LSTM sequence labeling model with an attention mechanism over unlabeled questions.
• Use attention to dynamically weigh the relevance of words in unlabeled questions when labeling a given labeled question, treating unlabeled data as side information.
• Concatenate labeled and unlabeled questions into a single input sequence, with a maximum length of 40 tokens per example.
• Train the model end-to-end using cross-entropy loss on labeled data while using attention to guide representation learning from unlabeled data.
• Fine-tune word embeddings for rare or unseen words using information from unlabeled data, improving generalization beyond the labeled vocabulary.
• Employ a two-layer Bi-LSTM on labeled questions and a single-layer Bi-LSTM on unlabeled questions to learn hierarchical representations.

Experimental results

Research questions

• RQ1 Can a semi-supervised deep learning model effectively recognize product function needs from customer questions in e-commerce QA forums?
• RQ2 How does leveraging unlabeled questions as contextual side information improve function recognition performance compared to purely supervised methods?
• RQ3 To what extent can attention mechanisms enhance representation learning in a semi-supervised sequence labeling setting for function need recognition?
• RQ4 Does the proposed SAN model outperform strong baselines such as CRF, S-BLSTM, and variants of SAN in terms of precision, recall, and F1-score?
• RQ5 Can the model generalize to unseen function expressions not present in the labeled training data by utilizing contextual cues from unlabeled data?

Key findings

• The proposed SAN model achieves an F1-score of 0.776, outperforming all baselines including CRF (0.692), S-BLSTM (0.749), and SAN(-)BLSTM2 (0.759).
• SAN improves recall by 5.8% over S-BLSTM (0.721 vs. 0.673), demonstrating better coverage of function needs, especially those involving rare or unseen words.
• The model maintains high precision (0.839) while significantly improving recall, indicating effective use of unlabeled data without sacrificing accuracy.
• The ablation study confirms that the second Bi-LSTM layer and the attention mechanism over unlabeled data are both critical for performance gains.
• The use of unlabeled data enables effective tuning of word embeddings for out-of-vocabulary words, enhancing model generalization.
• The newly annotated QA corpus, publicly available at https://www.cs.uic.edu/~hxu/, is dense in functionality information and serves as a valuable benchmark for future research in function need recognition.