Sentic computing for social good: sentiment analysis on toxic comment
With the neural network revolutions and increased computational power, Artificial Intelligence has been applied in various fields for improving life, such as concept-level sentiment analysis. We focused on one of the sentiment analysis applications: toxic comments detection. These inappropriate mess...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project (FYP) |
| Language: | English |
| Published: |
Nanyang Technological University
2022
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/156503 |
| _version_ | 1811693808934453248 |
|---|---|
| author | Wang Jingtan |
| author2 | Erik Cambria |
| author_facet | Erik Cambria Wang Jingtan |
| author_sort | Wang Jingtan |
| collection | NTU |
| description | With the neural network revolutions and increased computational power, Artificial Intelligence has been applied in various fields for improving life, such as concept-level sentiment analysis. We focused on one of the sentiment analysis applications: toxic comments detection. These inappropriate messages, hiding in the massive data, result in verbal violence to the receiver. Therefore, we aimed to detect the toxicity of content given raw textual input, outputting whether toxic or not. We selected an open-source multilabel dataset with around 150k samples. Each sentence is assigned 6 categories of toxic behaviors. We intended to predict the belonging of a text in these 6 labels. To achieve this, we reviewed and experimented the state-of-art methods in this field, known as the pre-trained model. We then improved the models based on the issues we noticed during experiments: imbalanced multilabel. We reviewed various approaches discussed in papers and journals, such as external knowledge of minority labels, cost-sensitive metrics, and resampling. We then compared them for an effective way to address the imbalance. Note that due to resources constraint, we only sampled ten percent of original data for our experimentation. Overall, we discovered the best fitting pre-trained model, BERT, and improved it in the imbalanced multilabel classification by using focal loss and random oversampling. We hope the reviews, the experimentation, and the result can contribute to the toxic comment challenge. We also pointed out the limitation in this project: the lack of resources and some unexpected behaviors, as well as possible future directions: active learning and data-augmentation supported resampling. |
| first_indexed | 2024-10-01T06:57:34Z |
| format | Final Year Project (FYP) |
| id | ntu-10356/156503 |
| institution | Nanyang Technological University |
| language | English |
| last_indexed | 2024-10-01T06:57:34Z |
| publishDate | 2022 |
| publisher | Nanyang Technological University |
| record_format | dspace |
| spelling | ntu-10356/1565032022-04-18T08:49:51Z Sentic computing for social good: sentiment analysis on toxic comment Wang Jingtan Erik Cambria School of Computer Science and Engineering cambria@ntu.edu.sg Engineering::Computer science and engineering With the neural network revolutions and increased computational power, Artificial Intelligence has been applied in various fields for improving life, such as concept-level sentiment analysis. We focused on one of the sentiment analysis applications: toxic comments detection. These inappropriate messages, hiding in the massive data, result in verbal violence to the receiver. Therefore, we aimed to detect the toxicity of content given raw textual input, outputting whether toxic or not. We selected an open-source multilabel dataset with around 150k samples. Each sentence is assigned 6 categories of toxic behaviors. We intended to predict the belonging of a text in these 6 labels. To achieve this, we reviewed and experimented the state-of-art methods in this field, known as the pre-trained model. We then improved the models based on the issues we noticed during experiments: imbalanced multilabel. We reviewed various approaches discussed in papers and journals, such as external knowledge of minority labels, cost-sensitive metrics, and resampling. We then compared them for an effective way to address the imbalance. Note that due to resources constraint, we only sampled ten percent of original data for our experimentation. Overall, we discovered the best fitting pre-trained model, BERT, and improved it in the imbalanced multilabel classification by using focal loss and random oversampling. We hope the reviews, the experimentation, and the result can contribute to the toxic comment challenge. We also pointed out the limitation in this project: the lack of resources and some unexpected behaviors, as well as possible future directions: active learning and data-augmentation supported resampling. Bachelor of Engineering (Computer Science) 2022-04-18T08:49:50Z 2022-04-18T08:49:50Z 2022 Final Year Project (FYP) Wang Jingtan (2022). Sentic computing for social good: sentiment analysis on toxic comment. Final Year Project (FYP), Nanyang Technological University, Singapore. https://hdl.handle.net/10356/156503 https://hdl.handle.net/10356/156503 en SCSE21-0238 application/pdf Nanyang Technological University |
| spellingShingle | Engineering::Computer science and engineering Wang Jingtan Sentic computing for social good: sentiment analysis on toxic comment |
| title | Sentic computing for social good: sentiment analysis on toxic comment |
| title_full | Sentic computing for social good: sentiment analysis on toxic comment |
| title_fullStr | Sentic computing for social good: sentiment analysis on toxic comment |
| title_full_unstemmed | Sentic computing for social good: sentiment analysis on toxic comment |
| title_short | Sentic computing for social good: sentiment analysis on toxic comment |
| title_sort | sentic computing for social good sentiment analysis on toxic comment |
| topic | Engineering::Computer science and engineering |
| url | https://hdl.handle.net/10356/156503 |
| work_keys_str_mv | AT wangjingtan senticcomputingforsocialgoodsentimentanalysisontoxiccomment |