Abstract
This paper introduces a system designed for automatically generating personalized annotation tags to label Twitter user's interests and concerns. We applied TFIDF ranking and TextRank to extract keywords from Twitter messages to tag the user. The user tagging precision we obtained is comparable to the precision of keyword extraction from web pages for content-targeted advertising.
Introduction
Twitter serves as a massive repository of "everyday life logs," yet extracting meaningful insights from its informal, short-form content remains a significant challenge. Unlike structured documents like news articles, tweets are short, riddled with slang, emoticons, and typos, and often lack a coherent topical sequence.
This research addresses the User Tagging Problem: the automatic generation of a set of keywords (tags) that accurately reflect a user's personal interests and concerns based on their tweet history. Solving this is critical for commercial applications like personalized advertising and "who to follow" recommendation engines. Given the scarcity of human-labeled datasets for user interests, this study focuses on unsupervised methods.
Methodology
The proposed system architecture follows a two-stage pipeline: extensive preprocessing followed by keyword extraction.
Preprocessing
The preprocessing stage acts as a "noise filter" to prepare the informal text for analysis:
- Reply Removal: Filtering out @replies, as these often reflect the interests of the person being replied to rather than the user themselves.
- Slang/Abbreviation Mapping: Using the "NoSlang" dictionary to map abbreviations (e.g., "bff" to "best friend forever") to ensure the downstream Part-of-Speech (POS) tagger functions correctly.
- POS Filtering: Retaining only nouns and adjectives as candidate tags.
- Standard NLP: Applying Porter stemming and stopword removal.
Extraction Algorithms
Two primary algorithms were compared:
- TFIDF Ranking: Treating a user’s entire tweet history as a single document and comparing word frequencies against a broader Twitter corpus.
- TextRank: A graph-based model inspired by Google’s PageRank. Words are nodes, and edges are created if words co-occur within the same tweet. The "importance" of a word is determined by its connectivity to other important words.
Experimental Setup
The system was evaluated using a Twitter corpus of 11,376 users, each with 180–200 messages. An evaluation set of 156 randomly selected users was analyzed by three human judges who rated the accuracy of the generated tags against the users' full message histories.
Results and Discussion
The results indicate that preprocessing is paramount; replacing slang and abbreviations is essential for the POS tagger to recognize grammatical roles correctly.
As shown in the table above, TextRank outperformed TFIDF significantly for the primary tag (Top-1), achieving 67.3% precision compared to 59.6% for TFIDF. This suggests that the graph-based approach is superior at identifying the "core" theme of a user's profile. While TextRank was better for the most prominent tags, both methods converged to similar performance (around 58-59%) when looking at the top-10 tags.
Notably, despite the inherent noise of Twitter, these precision scores are comparable to results previously reported for keyword extraction from formal web pages for content-targeted advertising. The researchers also found that User Message Entropy (H)—a measure of vocabulary diversity—correlates strongly with tagging precision. Systems struggle to find meaningful tags for users with low entropy (repetitive or very limited unique words).
Conclusion and Future Work
This study demonstrates that personalized tags can be effectively extracted from messy social media data using unsupervised methods. However, several limitations remain for future exploration:
- Phrase Extraction: The current system extracts isolated words. Extracting multi-word phrases (e.g., "Google Wave") would provide richer context.
- Temporal Dynamics: Future work could incorporate temporal weighting to account for how interests change over time.
- Sentiment Analysis: Identifying what a user talks about is different from identifying how they feel. Incorporating sentiment would prevent tagging a user with a brand they mention only in a negative context.
Summary updated