Abstract
Semantic lexicon induction is critical for understanding domain-specific language, yet traditional methods struggle with the informal and evolving nature of social media. This paper introduces RelWBound, a system designed to extract high-quality semantic lexicons from Twitter. By replacing rigid linguistic parsers with a combination of semantic pattern filtering and a novel Term Boundary Score (TBS), the authors effectively identify both single-word and multi-word terms. Experimental results show that RelWBound achieves up to 92% accuracy, significantly outperforming state-of-the-art baselines and capturing a wealth of terms not found in standard databases like WordNet.
The Challenge of Social Media NLP
Traditional Natural Language Processing (NLP) relies heavily on semantic lexicons—structured lists of "is-a" relationships (e.g., "pizza is a FOOD")—to interpret text. However, hand-built resources like WordNet cannot keep pace with the rapid evolution of language on platforms like Twitter. Social media presents two primary hurdles:
- Informal Language: Slang ("bball"), abbreviations ("cod"), and unconventional grammar frequently cause standard part-of-speech (POS) taggers and parsers to fail.
- Multi-word Terms: Meaningful entities are often phrases rather than single words. While "nugget" is generic, "chicken nugget" is a specific food item. Existing methods often struggle to identify the correct boundaries of these phrases, frequently extracting noisy fragments.
The RelWBound Methodology
The authors propose RelWBound (Related Words + Boundary detection), a four-stage framework that operates without fragile linguistic tools:
- Context Pattern Selection: Starting with 10 seed terms per category, the system identifies N-gram patterns (e.g., "I ate ()"). These are filtered using semantically related words generated by the DISCO tool to ensure the patterns are relevant to the target category.
- Candidate Discovery: The system searches a massive Twitter corpus for terms fitting these filtered patterns, ranking the top 2,000 candidates.
- Term Boundary Detection: To solve the phrase-length problem, the authors introduce the Term Boundary Score (TBS). By calculating the harmonic mean of a term's frequency in leftmost, rightmost, and middle context positions, the system identifies "complete" terms. Fragments (e.g., "sandwich today with") are penalized because they rarely appear in all three positions.
- Candidate Ranking: Final candidates are ranked based on the ratio of reliable patterns that extract them versus general, noisy patterns.
Experimental Evaluation
The study utilized a corpus of 114 million tweets and focused on three categories: Food & Drinks, Games & Sports, and Vehicles. RelWBound was compared against established baselines, including BASILISK and DISCO-based similarity measures. Human annotators evaluated the top 100 learned terms for both category correctness and proper boundary identification.
Key Results and Insights
RelWBound demonstrated a clear advantage over traditional methods, particularly in handling the noise of Twitter data:
- Superior Accuracy: In the "Food & Drinks" category, RelWBound achieved 92% accuracy for the top 100 terms. It also reached 83% for Sports and 57% for Vehicles, outperforming all baselines by a wide margin.
- The Necessity of Boundary Detection: The "RelW" version of the algorithm (without TBS) performed significantly worse, confirming that boundary detection is essential for filtering out phrase fragments.
- Beyond WordNet: A significant portion of the correct terms found (up to 48% in Sports) were absent from WordNet. This highlights the system's ability to capture modern slang, brand names (e.g., "Nutella," "Candy Crush"), and niche multi-word expressions.
Conclusion and Future Work
The success of RelWBound suggests that complex syntactic parsers are not strictly necessary for high-quality lexicon induction on social media. Simple N-gram patterns, when augmented with semantic filtering and robust boundary detection, are surprisingly effective. Future work may explore iterative "bootstrapping" to further expand these lexicons and address the ambiguity challenges found in categories like "Vehicles," where brand names often appear in non-vehicular contexts.
Summary updated