Back

Attention-based LSTM for Aspect-level Sentiment Classification

Yequan WangMinlie HuangXiaoyan ZhuLi Zhao

article2016EMNLP2,206 citations

Proposes an attention-based LSTM architecture that integrates aspect embeddings to dynamically focus on relevant context words, setting a new performance benchmark for fine-grained sentiment classification.

Listen

Standard sentiment analysis typically evaluates the overall tone of an entire sentence, which often misses critical nuances when customer feedback contains mixed opinions about different features. In real-world customer reviews, a single sentence frequently expresses opposite sentiments toward different topics, such as praising food quality while criticizing slow service. Effectively identifying these fine-grained opinions is crucial for organizations seeking actionable insights into specific operational strengths and weaknesses.

The article sets out to develop and evaluate a neural network approach that accurately classifies sentiment polarity at the aspect level by directly incorporating target aspect information into sentence modeling. Specifically, the authors aimed to demonstrate that using an attention mechanism enables the model to dynamically focus on the specific words in a sentence that correspond to a given aspect.

To accomplish this, the authors designed specialized neural network architectures based on Long Short-Term Memory networks (a type of machine learning model designed for sequential text data). They introduced mathematical representations called aspect embeddings and developed an attention mechanism to assign importance weights to different words based on the target aspect. The models were trained and evaluated on the standard SemEval 2014 benchmark dataset, which consists of real-world customer reviews from the restaurant and laptop domains labeled across three sentiment categories: positive, negative, and neutral.

The key findings demonstrate that explicitly connecting aspect data with attention mechanisms significantly improves classification performance over traditional baseline approaches. First, the proposed attention-based model with aspect embeddings (ATAE-LSTM) achieved the highest overall accuracy on restaurant aspect-level classification at 84.0% for three-class prediction, outperforming the standard baseline model's 82.0%. Second, on restaurant aspect-term classification, the model reached 77.2% accuracy, outperforming the standard baseline's 74.3%. Third, on the laptop review dataset, the model achieved 68.7% accuracy for three-way classification and 87.6% for binary classification, similarly surpassing conventional methods. Finally, qualitative visual analysis confirmed that the attention mechanism successfully shifts focus to relevant phrases—such as identifying delivery speed when evaluating service—while remaining robust against complex sentence structures and misleading negation words.

These findings imply that fine-grained opinion mining can be automated with higher reliability without relying on labor-intensive manual feature engineering or fragile syntax-parsing trees. For organizations, adopting aspect-aware attention models reduces the operational risk of misinterpreting customer feedback, allowing product and service teams to accurately identify specific drivers of customer satisfaction and dissatisfaction.

Based on these results, technical leaders should consider adopting aspect-embedding and attention-based architectures for customer feedback analytics rather than basic sentence-level sentiment classifiers. Moving forward, the article suggests extending the framework to model and process multiple aspects simultaneously rather than inputting each aspect separately, which could further improve computational efficiency and performance.

Confidence in these findings is supported by solid benchmark evaluation against established baselines. However, decision-makers should note that the current approach processes different aspects independently rather than concurrently, and the scope of testing was restricted to consumer electronics and restaurant reviews. Organizations applying this architecture to broader industry domains should conduct domain-specific pilot testing to validate performance before full deployment.

MicrosoftTsinghua Universityattention mechanismclassificationlong short term memory neural networksaspect-level sentiment classificationattention-based LSTMtarget-dependent sentiment classification
Wang et al (2016).pdf
  • Paper: A Structured Self-attentive Sentence Embedding, Zhouhan Lin et al. (2017). This study extends aspect-level models by introducing structured self-attentive sentence embeddings to capture multiple semantic components without relying on external task-specific context.
  • Paper: Graph Attention Networks, Petar Veličković et al. (2018). This work generalizes the attention paradigm beyond sequence models to graph-structured data by enabling nodes to dynamically weight and aggregate neighbor representations.
Cover for Attention-based LSTM for Aspect-level Sentiment Classification

Abstract

Aspect-level sentiment classification is a fine-grained task in sentiment analysis. Since it provides more complete and in-depth results, aspect-level sentiment analysis has received much attention these years. In this paper, we reveal that the sentiment polarity of a sentence is not only determined by the content but is also highly related to the concerned aspect. For instance, “The appetizers are ok, but the service is slow.”, for aspect taste, the polarity is positive while for service, the polarity is negative. Therefore, it is worthwhile to explore the connection between an aspect and the content of a sentence. To this end, we propose an Attention-based Long Short-Term Memory Network for aspect-level sentiment classification. The attention mechanism can concentrate on different parts of a sentence when different aspects are taken as input. We experiment on the SemEval 2014 dataset and results show that our model achieves state-of-the-art performance on aspect-level sentiment classification.

Topics

attention mechanismclassificationlong short term memory neural networkssentiment classificationpolarity classificationsemantic orientationrandom access memoryclassification models

Keywords

aspect-level sentiment classificationattention-based LSTMtarget-dependent sentiment classificationaspect term polarity classificationaspect embeddingsneural attention

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 2.1 Sentiment Classification at Aspect-level
  • 2.2 Sentiment Classification with Neural Networks
  • 3 Attention-based LSTM with Aspect Embedding
  • 3.1 Long Short-term Memory (LSTM)
  • 3.2 LSTM with Aspect Embedding (AE-LSTM)
  • 3.3 Attention-based LSTM (AT-LSTM)
  • 3.4 Attention-based LSTM with Aspect Embedding (ATAE-LSTM)
  • 3.5 Model Training
  • 4 Experiment
  • 4.1 Dataset
  • 4.2 Task Definition
  • 4.3 Comparison with baseline methods
  • 4.4 Qualitative Analysis
  • 4.5 Case Study
  • 5 Conclusion and Future Work
  • Acknowledgments
  • References

Knowls

  1. Knowl 1 — Attention-based LSTM with Aspect Embedding (ATAE-LSTM) Architecture

    model/method

    The Attention-based LSTM with Aspect Embedding (ATAE-LSTM) is a neural network model designed for aspect-level sentiment classification. It incorporates aspect information at two distinct stages: the LSTM input sequence and the attention weight computation.

    Let a sentence of length NN have word vectors [w1,w2,…,wN][w_1, w_2, \dots, w_N] where wt∈Rdww_t \in \mathbb{R}^{d_w}, and let va∈Rdav_a \in \mathbb{R}^{d_a} be the embedding vector for the target aspect. At each time step tt, the input to the LSTM cell is formed by concatenating the word embedding with the aspect embedding: xt=[wtva]∈Rdw+dax_t = \begin{bmatrix} w_t \\ v_a \end{bmatrix} \in \mathbb{R}^{d_w + d_a}

    The standard LSTM produces hidden state vectors H=[h1,h2,…,hN]∈Rd×NH = [h_1, h_2, \dots, h_N] \in \mathbb{R}^{d \times N}, where each ht∈Rdh_t \in \mathbb{R}^d carries both word and aspect semantics. An aspect-to-sentence attention mechanism then computes an attention weight vector α∈RN\alpha \in \mathbb{R}^N: M=tanh⁡([WhHWvva⊗eN])M = \tanh\left(\begin{bmatrix} W_h H \\ W_v v_a \otimes e_N \end{bmatrix}\right) α=softmax(wTM)\alpha = \text{softmax}(w^T M) r=HαTr = H \alpha^T where Wh∈Rd×dW_h \in \mathbb{R}^{d \times d}, Wv∈Rda×daW_v \in \mathbb{R}^{d_a \times d_a}, and w∈Rd+daw \in \mathbb{R}^{d + d_a} are trainable projection parameters, eN∈RNe_N \in \mathbb{R}^N is a column vector of ones, and va⊗eN=[va,va,…,va]∈Rda×Nv_a \otimes e_N = [v_a, v_a, \dots, v_a] \in \mathbb{R}^{d_a \times N} denotes repeating the aspect vector NN times. The vector r∈Rdr \in \mathbb{R}^d is the aspect-weighted sentence representation.

    The final sentence representation h∗∈Rdh^* \in \mathbb{R}^d combines the weighted representation rr and the final LSTM hidden state hNh_N: h∗=tanh⁡(Wpr+WxhN)h^* = \tanh(W_p r + W_x h_N) where Wp,Wx∈Rd×dW_p, W_x \in \mathbb{R}^{d \times d} are learnable projection matrices.

    The class prediction distribution y∈R∣C∣y \in \mathbb{R}^{|C|} over sentiment classes C={positive,negative,neutral}C = \{\text{positive}, \text{negative}, \text{neutral}\} is computed via a linear transformation followed by softmax: y=softmax(Wsh∗+bs)y = \text{softmax}(W_s h^* + b_s) where Ws∈R∣C∣×dW_s \in \mathbb{R}^{|C| \times d} and bs∈R∣C∣b_s \in \mathbb{R}^{|C|}.

  2. Knowl 2 — Attention-based LSTM (AT-LSTM) Architecture

    model/method

    The Attention-based LSTM (AT-LSTM) captures the key parts of a sentence corresponding to a given aspect by conditioning attention weights on the aspect embedding, without modifying the standard LSTM input.

    Given an input sentence of length NN with word embeddings [w1,w2,…,wN][w_1, w_2, \dots, w_N] (wt∈Rdw_t \in \mathbb{R}^d), the standard LSTM processes only the word embeddings xt=wtx_t = w_t to produce a hidden state matrix H=[h1,h2,…,hN]∈Rd×NH = [h_1, h_2, \dots, h_N] \in \mathbb{R}^{d \times N}. Given aspect embedding va∈Rdav_a \in \mathbb{R}^{d_a}, the attention matrix M∈R(d+da)×NM \in \mathbb{R}^{(d+d_a) \times N}, attention weight vector α∈RN\alpha \in \mathbb{R}^N, and weighted summary vector r∈Rdr \in \mathbb{R}^d are computed as: M=tanh⁡([WhHWvva⊗eN])M = \tanh\left(\begin{bmatrix} W_h H \\ W_v v_a \otimes e_N \end{bmatrix}\right) α=softmax(wTM)\alpha = \text{softmax}(w^T M) r=HαTr = H \alpha^T where Wh∈Rd×dW_h \in \mathbb{R}^{d \times d}, Wv∈Rda×daW_v \in \mathbb{R}^{d_a \times d_a}, and w∈Rd+daw \in \mathbb{R}^{d + d_a} are projection parameters, and eN∈RNe_N \in \mathbb{R}^N is a vector of ones.

    The final sentence representation is given by: h∗=tanh⁡(Wpr+WxhN)h^* = \tanh(W_p r + W_x h_N) where Wp,Wx∈Rd×dW_p, W_x \in \mathbb{R}^{d \times d}. Sentiment classification is performed using y=softmax(Wsh∗+bs)y = \text{softmax}(W_s h^* + b_s) with Ws∈R∣C∣×dW_s \in \mathbb{R}^{|C| \times d} and bs∈R∣C∣b_s \in \mathbb{R}^{|C|}.

  3. Knowl 3 — Aspect Embedding Learning

    model/method

    To represent predefined coarse-grained aspects (such as food, service, price, ambience, and anecdotes/miscellaneous) or fine-grained aspect terms in aspect-level sentiment classification, each unique aspect ii is assigned a continuous, low-dimensional aspect embedding vector vai∈Rdav_{a_i} \in \mathbb{R}^{d_a}.

    All aspect embeddings together form a parameter matrix A∈Rda×∣A∣A \in \mathbb{R}^{d_a \times |A|}, where ∣A∣|A| is the total number of distinct aspects. Unlike target-dependent methods that represent an aspect target simply by averaging the pretrained word embeddings of the words in the target phrase, aspect embeddings are dedicated parameter vectors learned end-to-end along with the model parameters via backpropagation.

  4. Knowl 4 — Model Training and Regularized Cross-Entropy Objective

    model/method

    The aspect-level sentiment classification models (AT-LSTM, ATAE-LSTM, and AE-LSTM) are trained end-to-end by minimizing the L2L_2-regularized multiclass cross-entropy loss: L(θ)=−∑i∑jyijlog⁡y^ij+λ∥θ∥2\mathcal{L}(\theta) = -\sum_{i} \sum_{j} y_i^j \log \hat{y}_i^j + \lambda \|\theta\|^2 where ii indexes training sentences, jj indexes sentiment classes (with ∣C∣=3|C| = 3 for positive, negative, and neutral), yij∈{0,1}y_i^j \in \{0, 1\} is the ground-truth binary indicator for class jj, y^ij\hat{y}_i^j is the predicted probability for class jj, λ\lambda is the L2L_2-regularization coefficient, and θ\theta is the set of all trainable parameters including LSTM gate matrices/biases, attention projection parameters, aspect embeddings, and word embeddings.

    Optimization is conducted using the AdaGrad algorithm, which adapts individual parameter learning rates based on historical gradient frequencies.

  5. Knowl 5 — Experimental Hyperparameter and Initialization Setup

    experimental setup

    Experiments on the SemEval 2014 Task 4 benchmark datasets (restaurant and laptop reviews) use the following implementation configurations:

    • Word Embeddings: Initialized with 300-dimensional GloVe vectors pre-trained on an 840-billion-token corpus. Out-of-vocabulary words (approximately 5% of total vocabulary) are initialized uniformly at random from U(−ϵ,ϵ)\mathcal{U}(-\epsilon, \epsilon) with ϵ=0.01\epsilon = 0.01.
    • Dimensions: Word vector dimension dw=300d_w = 300, aspect embedding dimension da=300d_a = 300, and LSTM hidden layer dimension d=300d = 300.
    • Optimization: AdaGrad optimizer with an initial learning rate of 0.010.01, mini-batch size of 2525, momentum of 0.90.9, and L2L_2-regularization weight λ=0.001\lambda = 0.001.
    • Parameter Initialization: All non-word embedding parameters are randomly sampled from U(−ϵ,ϵ)\mathcal{U}(-\epsilon, \epsilon) with ϵ=0.01\epsilon = 0.01.

  6. Knowl 6 — SemEval 2014 Task 4 Restaurant Aspect Distribution

    data/table

    The restaurant domain dataset from SemEval 2014 Task 4 contains labeled instances across five pre-identified coarse aspect categories: Food (Fo.), Price (Pr.), Service (Se.), Ambience (Am.), and Anecdotes/miscellaneous (An.). The distribution across sentiment classes is:

    Aspect Positive Negative Neutral
    Train Test Train Test Train Test
    Food 867 302 209 69 90 31
    Price 179 51 115 28 10 1
    Service 324 101 218 63 20 3
    Ambience 263 76 98 21 23 8
    Anecdotes/misc. 546 127 199 41 357 51
    Total 2179 657 839 222 500 94

    The total training set contains 3,518 aspect-sentiment pairs (2,179 positive, 839 negative, 500 neutral), and the test set contains 973 aspect-sentiment pairs (657 positive, 222 negative, 94 neutral).

  7. Knowl 7 — Aspect-Level Polarity Classification Performance on Restaurant Reviews

    data/table

    Aspect-level sentiment classification accuracy on the SemEval 2014 Task 4 restaurant dataset evaluates 3-class prediction (positive, negative, neutral) and binary prediction (ignoring neutral instances):

    Model Three-way (%) Pos./Neg. (%)
    LSTM 82.0 88.3
    TD-LSTM 82.6 89.1
    TC-LSTM 81.9 89.2
    AE-LSTM 82.5 88.9
    AT-LSTM 83.1 89.6
    ATAE-LSTM 84.0 89.9

    ATAE-LSTM attains the highest accuracy on both three-way (84.0%) and binary (89.9%) classification tasks. Standard LSTM achieves the lowest performance (82.0% three-way) because it cannot tailor sentence representations to the queried aspect. TC-LSTM underperforms TD-LSTM on three-way accuracy (81.9% vs. 82.6%) due to averaging word embeddings for target phrases.

  8. Knowl 8 — Aspect-Term-Level Polarity Classification Performance

    data/table

    Aspect-term-level polarity classification evaluates predicting the sentiment polarity of specific aspect terms occurring in sentences for both the Restaurant and Laptop subsets of SemEval 2014 Task 4:

    Model Restaurants Laptops
    Three-way (%) Pos./Neg. (%) Three-way (%) Pos./Neg. (%)
    LSTM 74.3 – 66.5 –
    TD-LSTM 75.6 – 68.1 –
    AE-LSTM 76.6 89.6 68.9 87.4
    ATAE-LSTM 77.2 90.9 68.7 87.6

    On the restaurant dataset, ATAE-LSTM achieves the best three-way accuracy of 77.2% and binary accuracy of 90.9%, outperforming LSTM (74.3%) and TD-LSTM (75.6%). On the laptop dataset, AE-LSTM attains the best three-way accuracy (68.9%) while ATAE-LSTM attains the best binary accuracy (87.6%).

  9. Knowl 9 — Dynamic Attention Focusing and Linguistic Robustness of Aspect Attention

    empirical result

    Qualitative analysis and attention visualization show that the aspect-to-sentence attention mechanism provides three key capabilities:

    1. Dynamic Aspect-Dependent Focus: When queried with different aspects on the same sentence, the attention vector α\alpha shifts to words semantically tied to the aspect (e.g., attending to multi-word phrases like 'fastest delivery times' when the aspect is service).
    2. Multi-Keyword Identification: The attention distribution assigns high weights across multiple dispersed sentiment words (e.g., 'tasteless' and 'too sweet' for the aspect food).
    3. Robustness to Negation and Syntactic Complexity: The model correctly resolves polarity in non-privative negation constructions (e.g., 'not just its superb cuisine, but also for its friendly owners and staff' correctly predicts positive sentiment for both food and service) and accurately processes long, syntactically complex sentences without relying on syntactic parsers that are prone to parsing errors.

  10. Knowl 10 — Limitation: Separate Evaluation of Multiple Sentence Aspects

    limitation

    The proposed attention-based LSTM architectures (AT-LSTM and ATAE-LSTM) process aspects independently. When a single review sentence mentions multiple aspects simultaneously (e.g., 'The appetizers are ok, but the service is slow'), each aspect must be fed to the model in a separate inference pass rather than being modeled and resolved jointly within a single forward pass.

Coverage note — None was omitted; all key contributions including the AE-LSTM, AT-LSTM, and ATAE-LSTM models, aspect embeddings, loss function, experimental results on aspect-level and aspect-term-level classification, attention analysis, and limitations are fully covered.

References

  1. 1.Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473.
  2. 2.Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, James Bergstra, Ian Goodfellow, Arnaud Bergeron, Nicolas Bouchard, David Warde-Farley, and Yoshua Bengio. 2012. Theano: new features and speed improvements. arXiv preprint arXiv:1211.5590.
  3. 3.Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. 2012. Large scale distributed deep networks. In Advances in Neural Information Processing Systems, pages 1223–1231.
  4. 4.Wankun Deng, Yongbo Wang, Zexian Liu, Han Cheng, and Yu Xue. 2014. Hemi: a toolkit for illustrating heatmaps. PloS one, 9(11):e111988.
  5. 5.Li Dong, Furu Wei, Chuanqi Tan, Duyu Tang, Ming Zhou, and Ke Xu. 2014. Adaptive recursive neural network for target-dependent twitter sentiment classification. In ACL (2), pages 49–54.
  6. 6.John Duchi, Elad Hazan, and Yoram Singer. 2011. Adaptive subgradient methods for online learning and stochastic optimization. The Journal of Machine Learning Research, 12:2121–2159.
  7. 7.David Golub and Xiaodong He. 2016. Character-level question answering with attention. arXiv preprint arXiv:1604.00727.
  8. 8.Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems, pages 1684–1692.
  9. 9.Sepp Hochreiter and Jurgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780.
  10. 10.Nobuhiro Kaji and Masaru Kitsuregawa. 2007. Building lexicon for sentiment analysis from massive collection of html documents. In EMNLP-CoNLL, pages 1075–1083.
  11. 11.Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami, and Chris Dyer. 2016. Neural architectures for named entity recognition. arXiv preprint arXiv:1603.01360.
  12. 12.Bing Liu. 2012. Sentiment analysis and opinion mining. Synthesis lectures on human language technologies, 5(1):1–167.
  13. 13.Tomas Mikolov, Martin Karafiat, Lukas Burget, Jan Cernocky, and Sanjeev Khudanpur. 2010. Recurrent neural network based language model. In INTERSPEECH, volume 2, page 3.
  14. 14.Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems, pages 3111–3119.
  15. 15.Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. 2014. Recurrent models of visual attention. In Advances in Neural Information Processing Systems, pages 2204–2212.
  16. 16.Saif M Mohammad, Svetlana Kiritchenko, and Xiaodan Zhu. 2013. Nrc-canada: Building the state-of-the-art in sentiment analysis of tweets. arXiv preprint arXiv:1308.6242.
  17. 17.Tony Mullen and Nigel Collier. 2004. Sentiment analysis using support vector machines with diverse information sources. In EMNLP, volume 4, pages 412–418.
  18. 18.Tetsuya Nasukawa and Jeonghee Yi. 2003. Sentiment analysis: Capturing favorability using natural language processing. In Proceedings of the 2nd international conference on Knowledge capture, pages 70–77. ACM.
  19. 19.Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. Proceedings of the Empiricial Methods in Natural Language Processing (EMNLP 2014), 12:1532–1543.
  20. 20.Veronica Perez-Rosas, Carmen Banea, and Rada Mihalcea. 2012. Learning sentiment lexicons in spanish. In LREC, volume 12, page 73.
  21. 21.Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th international workshop on semantic evaluation (SemEval 2014), pages 27–35.
  22. 22.Qiao Qian, Bo Tian, Minlie Huang, Yang Liu, Xuan Zhu, and Xiaoyan Zhu. 2015. Learning tag embeddings and tag-specific composition functions in recursive neural network. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, volume 1, pages 1365–1374.
  23. 23.Delip Rao and Deepak Ravichandran. 2009. Semi-supervised polarity lexicon induction. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics, pages 675–682. Association for Computational Linguistics.
  24. 24.Tim Rocktaschel, Edward Grefenstette, Karl Moritz Hermann, Tomás Kočiskỳ, and Phil Blunsom. 2015. Reasoning about entailment with neural attention. arXiv preprint arXiv:1509.06664.
  25. 25.Alexander M Rush, Sumit Chopra, and Jason Weston. 2015. A neural attention model for abstractive sentence summarization. arXiv preprint arXiv:1509.00685.
  26. 26.Richard Socher, Jeffrey Pennington, Eric H Huang, Andrew Y Ng, and Christopher D Manning. 2011. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 151–161. Association for Computational Linguistics.
  27. 27.Richard Socher, Alex Perelygin, Jean Y Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In EMNLP, volume 1631, page 1642. Citeseer.
  28. 28.Kai Sheng Tai, Richard Socher, and Christopher D Manning. 2015. Improved semantic representations from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075.
  29. 29.Duyu Tang, Bing Qin, Xiaocheng Feng, and Ting Liu. 2015a. Target-dependent sentiment classification with long short term memory. arXiv preprint arXiv:1512.01100.
  30. 30.Duyu Tang, Bing Qin, and Ting Liu. 2015b. Document modeling with gated recurrent neural network for sentiment classification. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1422–1432.
  31. 31.Wenpeng Yin, Hinrich Schutze, Bing Xiang, and Bowen Zhou. 2015. Abcnn: Attention-based convolutional neural network for modeling sentence pairs. arXiv preprint arXiv:1512.05193.

Citation

MLA
Wang, Y., et al. “Attention-based LSTM for Aspect-level Sentiment Classification”. Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, 2016, pp. 606–15, https://doi.org/10.18653/v1/D16-1058.
APA
Wang, Y., Huang, M., Zhu, X., & Zhao, L. (2016). Attention-based LSTM for Aspect-level Sentiment Classification. Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, 606–615. https://doi.org/10.18653/v1/D16-1058
Chicago
Wang, Y., M. Huang, X. Zhu, and L. Zhao. 2016. “Attention-based LSTM for Aspect-level Sentiment Classification”. Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, 606–15. https://doi.org/10.18653/v1/D16-1058.
Harvard
Wang, Y. et al. (2016) “Attention-based LSTM for Aspect-level Sentiment Classification”, Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pp. 606–615. Available at: https://doi.org/10.18653/v1/D16-1058.
Vancouver
1. Wang Y, Huang M, Zhu X, Zhao L (2016) Attention-based LSTM for Aspect-level Sentiment Classification. In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pp 606–615

BibTeX

@inproceedings{wang-etal-2016-attention,
    title = "Attention-based {LSTM} for Aspect-level Sentiment Classification",
    author = "Wang, Yequan  and
      Huang, Minlie  and
      Zhu, Xiaoyan  and
      Zhao, Li",
    editor = "Su, Jian  and
      Duh, Kevin  and
      Carreras, Xavier",
    booktitle = "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2016",
    address = "Austin, Texas",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/D16-1058/",
    doi = "10.18653/v1/D16-1058",
    pages = "606--615"
}
Metadata:ACL Anthology

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/