[Paper Review] A Unified Model for Opinion Target Extraction and Target Sentiment Prediction

Xin Li, Lidong Bing+2 more|arXiv (Cornell University)|Nov 13, 2018

Sentiment Analysis and Opinion MiningComputer Science30 references85 citations

TL;DR

The paper presents an end-to-end unified TBSA model using two stacked LSTMs with boundary guidance, sentiment consistency, and opinion-enhanced target word detection to jointly extract targets and predict their sentiments. It achieves state-of-the-art results on benchmark datasets through an integrated tagging framework and auxiliary tasks.

ABSTRACT

Target-based sentiment analysis involves opinion target extraction and target sentiment classification. However, most of the existing works usually studied one of these two sub-tasks alone, which hinders their practical use. This paper aims to solve the complete task of target-based sentiment analysis in an end-to-end fashion, and presents a novel unified model which applies a unified tagging scheme. Our framework involves two stacked recurrent neural networks: The upper one predicts the unified tags to produce the final output results of the primary target-based sentiment analysis; The lower one performs an auxiliary target boundary prediction aiming at guiding the upper network to improve the performance of the primary task. To explore the inter-task dependency, we propose to explicitly model the constrained transitions from target boundaries to target sentiment polarities. We also propose to maintain the sentiment consistency within an opinion target via a gate mechanism which models the relation between the features for the current word and the previous word. We conduct extensive experiments on three benchmark datasets and our framework achieves consistently superior results.

Motivation & Objective

• Motivate the need for end-to-end Target-Based Sentiment Analysis (TBSA) that jointly extracts targets and predicts sentiments.
• Propose a unified tagging scheme and a two-stage RNN architecture to model inter-task dependencies.
• Introduce boundary-guided prediction, sentiment consistency, and opinion-enhanced target word detection to improve TBSA performance.
• Demonstrate that the integrated model outperforms state-of-the-art baselines on SemEval ABSA and Twitter datasets.
• Provide ablation studies to quantify contributions of each component.

Proposed method

• Two stacked LSTMs (LSTM^S for the complete TBSA task using unified tags; LSTM^T for auxiliary target boundary prediction).
• Boundary Guidance (BG): encode boundary-to-unified tag transitions via a transition matrix W^tr and map boundary scores to unified tags; combine with primary scores using a confidence-based proportion alpha_t.
• Sentiment Consistency (SC): gate-based mechanism to mix current and previous word features to maintain sentiment consistency inside an opinion target.
• Opinion-Enhanced (OE) Target Word Detection: auxiliary token-level classifier to detect potential target words using context window around opinion words; refines boundary information.
• Training combines three losses: L^S, L^T, L^O (word/token-level cross-entropy for TBSA, boundary, and target word detection respectively).
• Unified tagging scheme: Y^S = {B-POS, I-POS, E-POS, S-POS, B-NEG, I-NEG, E-NEG, S-NEG, B-NEU, I-NEU, E-NEU, S-NEU, O}.

Experimental results

Research questions

• RQ1 Can TBSA be effectively solved in an end-to-end manner using a unified tagging scheme without relying on separate target boundary predictions at inference time?
• RQ2 How can boundary information be exploited to guide a unified TBSA tagger without hard constraints that propagate errors?
• RQ3 Does enforcing sentiment consistency within a multi-word target improve prediction accuracy in TBSA?
• RQ4 Does incorporating an auxiliary target word detection task improve boundary quality and overall TBSA performance?
• RQ5 What is the impact of key hyperparameters (boundary guidance weight alpha_t, context window s) on TBSA performance?

Key findings

• The unified two-LSTM framework with BG, SC, and OE components achieves superior F1 scores across datasets compared to strong baselines and state-of-the-art methods.
• Base model + BG improves TBSA performance over the base model, indicating the value of boundary-guided transitions.
• Full model (Base + BG + SC + OE) yields the best results, demonstrating the complementary benefits of SC and OE for boundary-guided TBSA.
• Compared to HAST-TNet pipeline, the proposed unified model gains 2.6%, 2.4%, and 0.40% absolute F1 on D_L, D_R, and D_T respectively.
• Ablation studies show SC and OE are necessary to achieve state-of-the-art results, with SC mainly helping sentiment consistency and OE improving boundaryquality.
• Case analysis demonstrates the model’s ability to correctly predict sentiments within multi-word targets when boundary guidance and auxiliary signals are used.