TL;DR
This paper introduces a novel method for analyzing the emotional sentiment of Chinese comments from news portals and forums. By expanding an existing Chinese emotion lexicon (DUTIR) with network-specific language and emojis, they develop a quantitative approach to measure positive and negative emotions, demonstrating improved accuracy over baseline methods on a custom annotated dataset.
What Problem Does This Paper Address?
Understanding public opinion and sentiment is crucial, especially in the digital age where online comments offer a rich source of data. While emotion analysis has been explored for English text, particularly on social media platforms like Twitter, there's a gap in robust methods for analyzing Chinese comments from news portals and forums. These platforms often contain unique linguistic expressions and informal language that require specialized lexicons and analysis techniques. The authors aim to fill this gap by developing a method to accurately quantify positive and negative emotions in Chinese online discourse.
Methods
The core of the paper's contribution lies in its methodology for emotion analysis.
- Emotion Lexicon Expansion: The authors start with the DUTIR lexicon, which contains over 27,000 Chinese emotional words, idioms, and proverbs. They expand this lexicon in two key ways:
- PMI and Synonym Lexicon: They use Point Mutual Information (PMI) to identify candidate words from Sina news, Baidu post bar, and Tianya forum comments that are semantically related to existing DUTIR words. These candidates are then assigned the emotional category and value of their closest synonym in DUTIR.
- Co-occurrence with Emojis: Recognizing that emojis often enhance emotional tone, they analyze the co-occurrence of DUTIR words with emojis. The top 100 emojis are then assigned the emotional category and value of the DUTIR word they most frequently co-occur with.
- This expanded lexicon is named DUTIRII.
- Emotion Value Calculation: For each comment, after Chinese word segmentation, they calculate:
- Word-level Polarity: Each word
wijis assigned a positivep(wij)or negativen(wij)emotion value based on its polarity (positive for appreciation/happiness, negative for anger/disgust/fear/sadness) and its intensity value from DUTIRII. Surprise is excluded as it's not distinctly positive or negative. - Comment-level Emotion: The total positive
p(ci)and negativen(ci)emotion values for a commentciare the sum of its words' respective emotion values. - Average Emotion: To account for comment length and avoid bias, they define average positive
pa(ci)and negativena(ci)emotion values. - Overall Positive Emotion Percentage (P(C)): Finally, they compute the overall positive emotion percentage for a set of comments
Cusing the formula:P(C) = Σpa(ci) / (Σpa(ci) + Σna(ci)).
Datasets
The paper utilizes and introduces datasets for their research:
- Sina News, Baidu Post Bar, and Tianya Forum Comments: These are the primary sources for expanding the emotion lexicon and for large-scale analysis. The paper mentions a massive dataset of over 16 million comments from Sina news, over 800,000 from Baidu post bar, and over 7 million from Tianya forum.
- Manual Annotation Dataset: To evaluate their method, the authors constructed a new dataset. They collected 10,000 comments from Sina news, Baidu post bar, and Tianya forum, which were then annotated by 5 graduate students. Each comment received a positive emotion score from 0 to 10. Spam comments were filtered out. This dataset contains 8,614 annotated comments across "Positive," "Negative," and "Objective" categories.
Key Insights and Ideas
- Network Language Matters: The expansion of the DUTIR lexicon to include network-specific language and emojis is crucial for accurately capturing the sentiment expressed in online comments. Standard lexicons might miss nuances.
- Emoji as Sentiment Indicators: The co-occurrence analysis with emojis highlights their role as strong indicators of emotional tone, effectively bridging the gap between text and user sentiment.
- Platform Rules Influence Sentiment: A significant finding is that comments on forums tend to be more positive than those on news portals. This is attributed to the more structured and rule-bound nature of forums, which may encourage more positive interactions.
- News Category Sentiment: Different news categories elicit distinct emotional responses. Entertainment and sports news comments are generally more positive, likely due to fan engagement, while international and social news comments tend to be more negative, possibly reflecting reporting on disputes or unfortunate events.
Key Results
- Improved Accuracy: The proposed method, using the expanded DUTIRII lexicon, achieved significant improvements in Mean Squared Error (MSE) and Mean Absolute Error (MAE) compared to a baseline method that did not account for word intensity or used the original DUTIR lexicon.
- Low MAE: The paper reports a Mean Absolute Error (MAE) of only 2.2% for their method on the manual annotation dataset, indicating high accuracy.
- Large-Scale Applicability: The method is deemed suitable for large-scale text analysis due to its efficiency and accuracy, especially when word combinations are not considered.
- Forum Positivity: Comments on forums (Tianya and Baidu) showed higher positive emotion percentages (P(C) values) compared to Sina news comments. Baidu post bar, with stricter rules, exhibited even more positive comments.
Limitations and Future Work
- Word Combination Ignored: The authors acknowledge that their method does not consider the emotional impact of word combinations or sentence structure, which could lead to more nuanced sentiment analysis.
- Surprise Emotion: The surprise emotion is not classified as positive or negative, which might limit the completeness of the emotional analysis in certain contexts.
- Future Work: The paper suggests that future work could explore the impact of word combinations and further refine the lexicon and analysis techniques for more comprehensive emotion detection.
Summary updated