In this paper, the authors proposed a sentiment analysis tool- SentiMoji that leverages the emoticons in the text to identify sentiments in it. They evaluate the performance of SentiMoji using four existing datasets which are Jira issue comments, Stack Overflow questions answers, Code Review comments and Java API Library. SentiMoji is trained based on a small amount of labelled SE-related data as well as large scale emoji labelled data from both Twitter and Github. It is directly built upon DeepMoji and it follows two-stage approaches: 1) fine-tune DeepMoji using emoji-labelled texts from GitHub to incorporate technical knowledge. The fine-tuned model is still a representation model based on the emoji-prediction task and we call it DeepMoji-SE; 2) use DeepMoji-SE to obtain vector representations of the sentiment-labelled texts and then use these vectors as features to train the sentiment classifier. SentiMoji fine-tuned the DeepMoji using the chain-thaw approach. To evaluate the performance of SentiMoji, they compare the performance of this tool with existing SentiStrength, Senti4SD, SentiCR and SentiStrength-SE. They notice that SentiMoji outperformed all the existing sentiment analysis methods in SE on all the selected benchmark datasets. Furthermore, they also observe that the GitHub post does not make a key contribution while the combination of large scale Tweets and a small amount of labelled data can achieve satisfactory performance.
Contributions of The Paper
Proposed a sentiment analysis tool based on DeppMoji that can leverage the emoticons in the body of the text
Provide the replication package of their proposed tool for further research
Contrast the performance of SentiMoji with the existing sentiment analysis tools in SE
Qualitative analysis shows that the proposed sentiment analysis tool can obtain useful knowledge from the Tweets rather than from the Github Post
Publisher
ESEC/FSE'19
Link to The Paper
https://arxiv.org/pdf/1907.02202.pdf
Name of The Authors
Zhenpeng Chen, Yanbin CAo, Xuan Lu
Year of Publication
2019
Summary
In this paper, the authors proposed a sentiment analysis tool- SentiMoji that leverages the emoticons in the text to identify sentiments in it. They evaluate the performance of SentiMoji using four existing datasets which are Jira issue comments, Stack Overflow questions answers, Code Review comments and Java API Library. SentiMoji is trained based on a small amount of labelled SE-related data as well as large scale emoji labelled data from both Twitter and Github. It is directly built upon DeepMoji and it follows two-stage approaches: 1) fine-tune DeepMoji using emoji-labelled texts from GitHub to incorporate technical knowledge. The fine-tuned model is still a representation model based on the emoji-prediction task and we call it DeepMoji-SE; 2) use DeepMoji-SE to obtain vector representations of the sentiment-labelled texts and then use these vectors as features to train the sentiment classifier. SentiMoji fine-tuned the DeepMoji using the chain-thaw approach. To evaluate the performance of SentiMoji, they compare the performance of this tool with existing SentiStrength, Senti4SD, SentiCR and SentiStrength-SE. They notice that SentiMoji outperformed all the existing sentiment analysis methods in SE on all the selected benchmark datasets. Furthermore, they also observe that the GitHub post does not make a key contribution while the combination of large scale Tweets and a small amount of labelled data can achieve satisfactory performance.
Contributions of The Paper
Comments
No response