Feelings Are Contagious: Scaling Sentiment Lexicons via Games and Graphs

Executive Summary

TL;DR: Building a comprehensive sentiment lexicon is no longer just a choice between slow manual annotation and noisy automated scraping. This research presents a "best of both worlds" strategy: using a competitive online game (Emot) to gather high-fidelity human associations and a propagation algorithm (Botemot) to spread those "feelings" across a 43-million-relation lexical network.

Background Positioning: This work sits at the intersection of Human Computation (GWAPs) and Graph-based Semi-Supervised Learning. It transforms the JeuxDeMots (JDM) network from a standard semantic graph into a living, "emotional" knowledge base.

Problem & Motivation: The Bottleneck of Human Emotion

Why is sentiment analysis still hard? Most lexicons are:

1. Too Small: Limited to a few thousand words.
2. Too Rigid: Constrained to basic polarities (Positive/Negative) or small sets like "The Big Six" emotions.
3. Noisy: Crowdsourcing via paid tasks (AMT) often lacks the intrinsic motivation found in games, leading to "trolling" or low-effort responses.

The authors' core insight is transitivity. If a "tumor" leads to "death," and "death" evokes "fear," then the sentiment of fear should logically propagate to the concept of a tumor. Feelings are not isolated; they are contagious neighbors in our mental lexicon.

Methodology: The "Contagion" Architecture

The system operates in a dual-loop known as Never Ended Learning.

1. High-Precision Seeding (The Emot Game)

Players are presented with words and must associate them with feelings. The "competitive" aspect ensures quality: players want their answers to match others' to gain points. This creates a validated "nucleus" of data.

2. Massive Expansion (Botemot Algorithm)

Botemot treats the lexical network as a substrate for infection. It identifies a target word ($T$) and looks at its neighbors linked by six specific "contagious" relations:

• Synonyms & Hypernyms: Identity and hierarchy.
• Characteristics & Consequences: Attributes and causal outcomes.

The algorithm calculates a weighted union of sentiments from these neighbors, filtered by a tolerance factor ($\alpha$) and a polarity filter (ensuring a negative word doesn't accidentally inherit a positive joy sentiment due to a weak link).

Note: The JDM graph provides the relational structure; Botemot provides the emotional flow.

Experiments & Results: Proving the Contagion

The authors evaluated the algorithm against 350,000 existing validated relations and conducted manual audits of new proposals.

• Performance Metrics: The weighted precision, recall, and F1-score all hovered around 0.95. This suggests the algorithm perfectly mirrors human intuition in the vast majority of cases.
• The Power of Weights: By assigning negative weights to "false" associations (e.g., Fly is NOT an agent of Ostrich), the system creates Inhibitors. These inhibitors stop the "contagion" of incorrect sentiments, preventing the propagation of noise.
• Comparison with Emolex: When tested against the established Emolex (translated to French), Botemot achieved 92% total inclusion. Remarkably, where Botemot differed, it was often richer—providing nuanced synonyms (e.g., "anguish" vs. just "fear") that the original Emolex lacked.

Figure: The high F1-score validates that sentiment can be reliably inferred from semantic neighbors.

Critical Analysis & Conclusion

Takeaway

The success of Botemot proves that lexical networks are not just semantic; they are affective. By leveraging the existing topology of language, we can scale emotional intelligence in AI without requiring humans to label every single word in the dictionary.

Limitations

• Dominant Sense Bias: In polysemous words (words with multiple meanings), the "loudest" emotional meaning tends to drown out others. For example, the tragic sense of a word might contaminate its neutral technical usage.
• Language Specificity: While propagation is universal, the underlying JDM network is specifically tuned for French.

Future Outlook

This framework is a blueprint for self-enriching networks. As the GWAP continues to feed the nucleus, the "contagion" will reach deeper into the long tail of rare words, eventually creating a sentiment resource with the breadth of a full encyclopedia and the precision of a human expert.