Abstract
Sentiment analysis on microblogging platforms like Twitter is often hindered by informal language and the rapid evolution of slang. This paper demonstrates that word lengthening (e.g., "coooool") is not merely noise but a significant signal for subjectivity and sentiment. We propose an unsupervised pipeline to detect these lengthened forms and use them to automatically expand and adapt sentiment lexicons. Our findings show a strong correlation between the degree of lengthening and subjectivity, and our graph-based propagation methods successfully identify emerging slang and context-dependent sentiment.
Introduction
Traditional sentiment analysis relies on static lexicons that struggle with the creative and informal nature of social media. On platforms like Twitter, users frequently employ word lengthening as a textual proxy for prosody—the rhythm and emphasis used in spoken language. This study investigates how these non-canonical forms can be leveraged to automatically identify subjective content and expand sentiment dictionaries without the need for expensive human annotation.
Methodology
The researchers developed a multi-step pipeline to utilize word lengthening as a proxy for emphasis:
- Detection and Canonicalization: Words are reduced to a "condensed form" by collapsing repeated characters. The most frequent variation within a group is identified as the canonical form (e.g., "nice" for "niiice"). This strategy was found to be over 97% accurate in identifying the correct base word.
- Subjectivity Analysis: The relationship between word variations and subjectivity was analyzed by cross-referencing Twitter data with the OpinionFinder lexicon.
- Lexicon Expansion via Graph Propagation: A co-occurrence graph was built from a dataset of 500,000 tweets, with edges representing words appearing within a 3-word window of each other. Sentiment scores were propagated from seed words (e.g., "good", "bad") using two algorithms: the "Reviews" method (broad diffusion) and the "Web" method (strongest path diffusion).
Results and Discussion
The Subjectivity Trend
A key insight of this research is the clear correlation between word lengthening and subjectivity. As shown in the data, the more distinct lengthened variations a word has, the more likely it is to be a sentiment-bearing word. For instance, words with 10 or more variations are significantly more likely to be subjective than those with only two.
Lexicon Expansion and Slang Discovery
The system successfully identified modern slang and domain-specific terms that traditional lexicons often miss. The "Web" algorithm (based on Velikovich et al.) proved more robust in the noisy Twitter environment, as it focuses on the single strongest connection to a known sentiment word rather than allowing sentiment to "bleed" to all neighbors.
The algorithms correctly identified terms like "sus" (suspicious) and "smh" (shaking my head) as negative, and "win" as positive. The system also captured temporal sentiment; for example, "Canes" was identified as negative due to a specific sports loss during the sampling period. However, some misclassifications occurred, such as "holy" being flagged as negative due to its frequent co-occurrence with expletives.
Conclusion
Word lengthening is a pervasive feature of microblogs, appearing in approximately one out of every six tweets. This study proves that such linguistic variations are powerful indicators of sentiment. Future work will explore quantifying the strength of emotion based on the degree of lengthening and integrating these signals into real-time sentiment classifiers that account for emoticons, punctuation, and multilingual patterns.
Summary updated