Notes on Villarroel Ordenes et al. (2017) – Consumer Sentiment Analysis in Social Media

Main Topic or Phenomenon

This paper addresses the limitations of traditional sentiment analysis in consumer research, which typically relies on simple positive/negative word counts. The authors argue that consumers use a much richer array of explicit and implicit linguistic features to express sentiment in online reviews and social media, and that understanding these nuances is crucial for predicting consumer behavior and sales outcomes.

Theoretical Construct

The paper builds on Speech Act Theory (SAT) as the primary theoretical framework. SAT focuses on how language functions to communicate intent beyond just the literal meaning of words.

Key definitions:

• Explicit sentiment expressions: Direct emotional language with varying activation levels (e.g., “good” vs. “awesome”) that can be boosted (“very good”) or attenuated (“kind of good”)
• Implicit sentiment expressions: Language that conveys sentiment without using emotion words, including:
  • Directive acts: Recommendations to others (“You should stay here”)
  • Commissive acts: Commitments to future action (“I will come back”)
  • Assertive acts: Statements of fact (“We got a discount”)
• Discourse patterns: How sentiment develops across sentences, including incoherence (variability in sentiment) and trends (increasing/decreasing positivity)

Key Findings

1. Activation levels matter: High-activation positive words (“awesome”) have significantly stronger effects on overall sentiment than low-activation words (“nice”), with the effect roughly doubling the probability of a higher star rating.

2. Implicit expressions have asymmetric effects: Directive and commissive speech acts have stronger impacts on overall sentiment than assertive acts, with directives showing particularly strong effects.

3. Discourse patterns influence sentiment: Sentiment incoherence across sentences is associated with more negative overall ratings. Counterintuitively, positive trends (increasing positivity toward the end) are associated with more negative overall sentiment, while negative trends correlate with more positive sentiment.

4. Predictive superiority: The nuanced sentiment model better predicts sales performance than traditional valence-based approaches.

5. Context dependency: Effects vary by product type (books vs. hotels) and platform (reviews vs. social media).

Boundary Conditions and Moderators

1. Product/Service Context:

   • Commissive language more frequent in hotel reviews than book reviews (repeat patronage likelihood)
   • Negative high-activation words had stronger effects in hotels but weaker effects in books (possibly due to genre descriptions)

2. Platform Differences:

   • Social media posts are much shorter, forcing more explicit sentiment expression
   • Effects of implicit expressions less pronounced on Twitter/Facebook due to length constraints

3. Review Length:

   • Discourse patterns only meaningful for reviews with 3+ sentences
   • Longer reviews associated with more negative sentiment in hotels

4. First-person Pronoun Usage:

   • Positive effect on sentiment in hotel reviews (shared experiences with “we”)
   • Negative effect in book reviews (individual “I” experiences)

Building on Previous Work

The paper extends sentiment analysis research by moving beyond the “bag of words” approach that dominated prior work. While previous studies (Pang & Lee, Das & Chen, Berger & Milkman) focused on simple positive/negative classifications, this work introduces:

• Theoretical grounding: Uses SAT to provide theoretical justification for different types of sentiment expression
• Granular analysis: Distinguishes activation levels within valence categories
• Implicit recognition: Identifies sentiment conveyed without emotion words
• Dynamic patterns: Examines how sentiment develops across message sequences

The paper challenges the assumption that all positive (or negative) words have equal impact and shows that context and linguistic structure matter significantly.

Major Theoretical Contribution

The paper’s primary theoretical contribution is demonstrating that consumer sentiment is a multidimensional construct that requires analysis at multiple linguistic levels. It provides a theoretically grounded framework for understanding how consumers actually express sentiment in digital environments, moving beyond simplistic word counting to capture the full richness of consumer expression. This contributes to a more nuanced understanding of digital consumer behavior and the psychology of online expression.

Major Managerial Implication

Managers should move beyond simple sentiment monitoring tools that only count positive and negative words. The research suggests that:

• Activation level monitoring: Pay special attention to high-activation positive language as it’s more predictive of actual satisfaction
• Implicit sentiment tracking: Monitor recommendations and future commitment language, not just direct emotional expressions
• Discourse pattern analysis: Incoherent sentiment patterns may indicate ambivalent customers who need targeted intervention
• Context-specific approaches: Tailor sentiment analysis approaches to specific product categories and platforms

Unexplored Theoretical Factors

Several potential moderators and factors were not explored:

1. Cultural factors: How sentiment expression varies across different cultural contexts and languages
2. Individual differences: Personality traits (extraversion, neuroticism) that might influence sentiment expression patterns
3. Temporal dynamics: How sentiment expression patterns change over a customer’s relationship lifecycle with a brand
4. Social influence: How exposure to others’ reviews influences one’s own sentiment expression patterns
5. Emotional regulation strategies: How individual differences in emotion regulation affect the linguistic patterns used
6. Brand relationship strength: How existing brand loyalty moderates the relationship between linguistic patterns and sentiment
7. Purchase involvement: How high vs. low involvement purchases influence sentiment expression complexity
8. Demographic factors: Age, education, and digital nativity effects on sentiment expression sophistication

Reference

Villarroel Ordenes, Francisco, Stephan Ludwig, Ko de Ruyter, Dhruv Grewal, and Martin Wetzels (2017), “Unveiling What Is Written in the Stars: Analyzing Explicit, Implicit, and Discourse Patterns of Sentiment in Social Media,” Journal of Consumer Research, 43 (6), 875–94.