Source-linked AI summary
DOER: Dual Cross-Shared RNN for Aspect Term-Polarity Co-Extraction
Huaishao Luo, Tianrui Li, Bing Liu, Junbo Zhang
TL;DR
Aspect term extraction and aspect sentiment classification are often handled separately, though practical applications need both. DOER treats them as sequence-labeling tasks and jointly generates aspect-polarity pairs; it outperforms baselines on three benchmark datasets.
Problem
Aspect term extraction and sentiment classification are often solved separately, although practical applications need both and sentences may contain variable numbers of pairs with variable-length aspects.
Source: §1 · pp. 1–2Method
DOER uses dual recurrent networks for aspect extraction and sentiment classification, with a cross-shared unit modeling their interaction and auxiliary tasks supporting labeling.
Source: §1 · p. 2Results
DOER improves over the best pipelined model by 3.88% on SL, 5.24% on SR, and 2.63% on ST in aspect term-polarity pair F1.
Source: §3.5 · p. 7Takeaways & Limitations
The authors present DOER as an end-to-end unified solution that generates aspect terms and polarities together instead of building two separate models.
Source: §1 · p. 2
Abstract
from arXiv · showhide
This paper focuses on two related subtasks of aspect-based sentiment analysis, namely aspect term extraction and aspect sentiment classification, which we call aspect term-polarity co-extraction. The former task is to extract aspects of a product or service from an opinion document, and the latter is to identify the polarity expressed in the document about these extracted aspects. Most existing algorithms address them as two separate tasks and solve them one by one, or only perform one task, which can be complicated for real applications. In this paper, we treat these two tasks as two sequence labeling problems and propose a novel Dual crOss-sharEd RNN framework (DOER) to generate all aspect term-polarity pairs of the input sentence simultaneously. Specifically, DOER involves a dual recurrent neural network to extract the respective representation of each task, and a cross-shared unit to consider the relationship between them. Experimental results demonstrate that the proposed framework outperforms state-of-the-art baselines on three benchmark datasets.
1 Introduction
Aspect term extraction and aspect sentiment classification are often handled separately, although practical applications need both tasks solved together. DOER addresses this challenge by jointly labeling aspect terms and polarities while modeling interaction between the tasks.
1 Introduction
- The two subtasks are rarely solved together, and pipeline approaches require separate models and can propagate aspect-extraction errors into polarity classification.ATE extracts aspects receiving opinions; ASC identifies the polarity expressed on those extracted aspects.Source: §1 · p. 1
- DOER treats ATE and ASC as separate sequence-labeling tasks, rather than combining their boundaries and polarities into collapsed tags.The introduction describes separate tag sets and representations as a way to keep aspect-term meaning distinct from the words expressing polarity.Source: §1 · p. 2
- Sentences can contain arbitrary numbers of aspect-polarity pairs, and aspect terms can span any number of words.These properties make the co-extraction task difficult to solve.Source: §1 · p. 2
- DOER uses a dual RNN and cross-shared unit to generate aspect term-polarity pairs end to end, with auxiliary tasks and ReGU supporting the task representations.The cross-shared unit models interaction between ATE and ASC; the auxiliary tasks enhance their labeling, and ReGU adds gated, skip-like input transfer.Source: §1 · p. 2
- DOER also introduces two auxiliary tasks to enhance ATE and ASC labeling and ReGU to improve feature extraction.ReGU uses a gate to transfer input to output, like a skip connection, supporting deeper training and more useful features.Source: §1 · p. 2
2 Methodology
DOER formulates aspect term extraction and aspect sentiment classification as two sequence-labeling tasks. Separate recurrent branches model each task, while cross-sharing and auxiliary objectives connect and enhance their representations.
2.1 Problem Statement
- DOER assigns aspect-boundary tags and polarity tags to each word, with separate tag sets for the two sequence-labeling tasks.ATE uses B, I, and O; ASC uses PO, NT, NG, CF, and O.Source: §2.1 · p. 3
2.2 Model Overview
- Two stacked bidirectional ReGU branches generate separate ATE and ASC representations, with gates controlling information flow through the recurrent states.The forget gate controls flow across timestamps, while the residual gate controls flow from the previous layer.Source: §2.2 · pp. 3–4
- The cross-shared unit uses tensor-based composition and attention to let ATE and ASC representations exchange task-related information.Its scores measure correlations between aspect-term and polarity representations, and row-based softmax forms the cross-task enhancements.Source: §2.2 · p. 4
- Auxiliary objectives predict average aspect length and word-level sentiment labels, while CRF inference produces the final ATE and ASC tag sequences.The joint loss combines both task losses with the auxiliary losses and a regularization parameter.Source: §2.2 · pp. 4–5, §2.3 · p. 5
3 Experiments
The experiments evaluate DOER on three benchmark datasets using exact-match F1 for aspect term-polarity pairs. DOER improves over pipeline and collapsed baselines, while ablations and analyses examine its components and cross-task interactions.
3.1 Datasets
- SL covers laptop reviews, SR restaurant reviews, and ST English tweets; SL and SR results average 10 runs, while ST uses ten-fold cross-validation.Evaluation uses F1 based on exact matching of the aspect term and its polarity.Source: §3.1 · pp. 5–6
3.2 Word Embeddings
- Domain-specific embeddings use Amazon reviews for SL and ST and Yelp reviews for SR, alongside general-purpose GloVe embeddings.The Amazon corpus contains 142.8M reviews and the Yelp corpus 2.2M restaurant reviews.Source: §3.2 · p. 6
3.3 Settings
- DOER uses Adam with learning rate 0.001, batch size 16, dropout 0.5, and gradient-norm clipping at 5.K is set to 5, and word embeddings remain fixed during training.Source: §3.3 · p. 6
3.4 Baseline Methods
- The comparison includes pipeline, joint, and collapsed systems, plus neural sequence-labeling baselines and DOER component variants.The baselines include DE-CNN+TNet, E2E-TBSA, S-BiLSTM, and S-BiReGU.Source: §3.4 · pp. 6–7
3.5 Results and Analysis
- 3.88%, 5.24%, and 2.63% over DE-CNN+TNet on SL, SR, and ST, and 2.36%, 2.87%, and 2.24% over E2E-TBSA.The improvements over both baselines are statistically significant (p < 0.05).Source: Table 2 · p. 7, §3.5 · pp. 7–8
- Ablations report gains from ReGU, CSU, aspect-length enhancement, and sentiment-lexicon enhancement, with the full DOER combining these components.The authors attribute DOER’s performance mainly to enhanced features from the two auxiliary tasks and interaction between the ATE and ASC routes.Source: §3.5 · p. 8
- DOER achieves new state-of-the-art aspect term extraction scores against DE-CNN, and outperforms DOER∗, its variant without ASC.The authors interpret the difference from DOER∗ as indicating that ATE–ASC interaction can improve ATE performance.Source: Table 3 · p. 8, §3.5 · p. 8
- Case examples show CSU avoiding error words; auxiliary tasks improve sentiment representations and address misjudgments on long aspect terms.On SL, K=5 performs best in the 1–10 sweep; CSU attention differs across ATE and ASC, with OS attending to itself for ATE and “great” for ASC.Source: §3.5 · p. 8, Figure 5 · p. 8, Figure 6 · p. 8
4 Related Work
Related work covers aspect term extraction and aspect sentiment classification, which are commonly studied separately. DOER instead treats them as parallel sequence-labeling tasks solved simultaneously, distinguishing its joint approach from collapsed models.
4 Related Work
- Prior ATE methods include unsupervised mining, rules, and topic models, while supervised ATE is usually sequence labeling; ASC work uses context relations, CNNs, attention, and memory networks.The cited approaches generally target either ATE or ASC.Source: §4 · p. 9
- Earlier joint and collapsed approaches addressed related sequence-labeling problems, while other systems map sentences to predefined aspects before classifying polarity.The authors argue predefined aspect terms may miss important aspects in reviews.Source: §4 · p. 9
- DOER solves ATE and ASC simultaneously as parallel sequence-labeling tasks, unlike Li et al. (2019)’s collapsed approach to the same objective.The paper also distinguishes its objective and model from joint named-entity-and-sentiment work.Source: §4 · p. 9
5 Conclusion
DOER addresses aspect term-polarity co-extraction with joint sequence labeling that models interaction between separate extraction and classification routes. Auxiliary tasks support sentiment representation and long-aspect handling; experiments on three benchmark datasets show it outperforms baselines.
5 Conclusion
- DOER significantly outperforms baselines on aspect term-polarity co-extraction across three benchmark datasets.The conclusion reports this as evidence of the framework's effectiveness.Source: §5 · p. 9
- DOER uses joint sequence labeling to model interaction between separate aspect term extraction and sentiment classification routes.Source: §5 · p. 9
- Two auxiliary tasks enhance sentiment representation and alleviate difficulty with long aspect terms.Source: §5 · p. 9