PaperNotes丨Personalized Microblog Sentiment Classification via Adversarial Cross-lingual Multi-task

Personalized Microblog Sentiment Classification via Adversarial Cross-lingual Multi-task Learning

---

0.Information

Year : 2018
Authors : Weichao Wang1, Shi Feng1, Wei Gao2, Daling Wang1, Yifei Zhang1
Conference : EMNLP 2018

Wang, Weichao, et al. “Personalized microblog sentiment classification via adversarial cross-lingual multi-task learning.” Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.

---

1.Summary

This paper propose an ACNN model and an ACNN-AMT model. ACNN model can capturing user’s individuality on microblogs(Sina,Twtter) based on attention-mechanism. ACNN-AMT alleviating the data insufficiency problem in personalized sentiment classification.

---

2.Research Objective(s)

Propose a novel user-attention-based Convolutional Neural Network(CNN) model with adversarial cross-lingual learning framework.

---

3.Background / Problem Statement

1. Most of the existing microblog sentiment classification algorithms largely ignore latent personal distinctions among microblog users.
2. The sentiment expression in posts can be affected by user’s language habit, personal character, opinionbias and so on, and even the same sentence can deliver quite different sentiment polarities depending on user’s underlying context.
3. By cross-lingual learning from same(or similar) user’s posts, we can extract the common language-independent features to enhance the representations for both monolingual and cross-lingual sentiment classfication.

---

4.Method(s)

Model

ACNN

ACNN-AMT(user-attention-based CNN model with adversarial muti-task learning)

---

Steps

1. Incorporate user attention into CNN model to capture user’s language-specific individuality from the posts.
2. Propose an adversarial muti-task learning framework to co-train cross-lingual sentiment features with the help of user’s consistent individuality.

---

Basic

1. CNN
2. Attention Mechanism
3. Adversarial muti-task learning.
4. NSC-UPA¹

---

5.Evaluation

Dataset

1. The first dataset was crawed using built-in-house crawler from Sina Weibo.
2. Extract the Twiter homepage links for crawling the Twitter dataset via Twitter API.
3. https://github.com/githubfordata/data

---

Compared with baseline

Measure:

• Accuracy

---

Visualization

---

6.Conclusion

1. Propose an ACNN model for capturing user’s individuality on microblogs.
2. Propose an ACNN-AMT model for alleviating the data insufficiency problem in personalized sentiment classification. With the help of consistency of user individuality as bridge across two languages, we can learn language-specific features and language-independent feature to enrich the user post representation.

---

7.Notes

1. Microblogs sentiment classification.
2. Cross-language dataset. For one user, use his Sina microblog post and Twitter posts.

---

8.Code

None

---

1. Chen, Huimin, et al. “Neural sentiment classification with user and product attention.” Proceedings of the 2016 conference on empirical methods in natural language processing. 2016. ↩︎