Abstract
This research presents a specialized Natural Language Processing (NLP) pipeline designed to detect expressions of "loss" (grief) and "aggression" (threats) in the tweets of gang-involved youth in Chicago. By combining social work expertise with domain-adapted machine learning, the team built tools that outperform standard NLP models on highly non-standard dialects, providing a potential early-warning system for community violence intervention.
The Problem: The "Digital Street"
In urban environments like Chicago, social media—often referred to as the "digital street"—has become a primary arena for gang taunting and conflict escalation. However, standard NLP tools are notoriously ineffective at interpreting the language used in these communities.
The dialect, which includes African American Vernacular English (AAVE), gang-specific slang, and unique orthography, renders off-the-shelf tools largely useless. For example, a standard Stanford Part-of-Speech (POS) tagger achieves only 26% accuracy on this data. The authors aim to bridge this gap to help social workers identify when a youth’s grief over a fallen friend might be transitioning into a desire for violent retribution.
Methodology
The researchers employed an interdisciplinary approach, blending qualitative social work with quantitative computational linguistics:
- Deep Reads & Interpretation: Social workers collaborated with neighborhood youth to perform "deep reads" of tweets, providing the cultural context necessary to label data accurately (e.g., identifying that "smoke" can mean "to kill").
- Domain-Adapted POS Tagging: They built a POS tagger specifically for this sub-language, using domain adaptation techniques to leverage existing Twitter datasets while training on their manually annotated gang-related corpus.
- Glossary Extraction: Using machine translation alignment techniques, they created a "phrasebook" mapping gang slang to Standard American English.
- Emotion Detection: They utilized the Dictionary of Affect in Language (DAL), extending it with WordNet and a custom glossary to assign emotional scores to non-standard words and emojis.
- Cascading Classifiers: To find specific emotions in a sea of general tweets, they implemented a cascading SVM architecture. The first stage identifies if a tweet expresses either loss or aggression; the second stage then classifies the specific sentiment.
Datasets
The study centers on a newly created corpus of 820 tweets. The data was sourced from the public Twitter history of a high-profile female gang member in Chicago and her immediate communication network. This dataset captures a high-stakes, real-world environment of escalating conflict.
Key Insights
- The Failure of General Resources: The authors found that 56.9% of the words in their corpus missing from WordNet actually had incorrect definitions in Wiktionary, highlighting how general slang dictionaries miss lethal nuances.
- Aggression follows Loss: A critical qualitative finding was that aggressive posts were frequently preceded by posts reflecting loss or grief. This suggests that detecting "loss" is vital for violence prevention.
- Character N-grams are Essential: In their POS tagging ablation study, character n-grams were the most vital feature, providing robustness against high spelling variation and out-of-vocabulary words.
- Emojis as Language: Emojis were significant emotional markers. The team developed a custom "Emoji Lexicon" to translate these symbols into text descriptions for emotion scoring.
Key Results
The specialized tools significantly outperformed general-purpose baselines, particularly in POS tagging accuracy on the specific test set.
- POS Tagging: Their custom tagger reached 81.5% accuracy, compared to the 26% of the Stanford tagger and 78% of a standard Twitter tagger.
- Classification: The Cascading Classifier (CC) achieved an F-measure of 62.3% for aggression and 63.6% for loss, improving over unigram baselines by 13.7 and 5.8 points, respectively.
- High Recall: The system was tuned for high recall (over 90% for the CC model), ensuring that very few potentially violent escalations would be missed by human interveners.
Limitations and Future Work
The primary limitation is the scale and specificity of the dataset (820 tweets from a single network). It remains unclear how well these models generalize to other cities or gang cultures. Future work aims to expand the corpus and investigate the real-time relationship between Twitter expressions and physical-world violence to better support community outreach groups.
Summary updated