PERSONAGE: Psychologically Informed Adaptation in Conversational AI

TL;DR

Researchers have developed PERSONAGE, an NLG engine that doesn't just "talk"—it expresses a personality. By mapping 67 psychological parameters to a standard generation pipeline, the system can project traits like Extraversion or Agreeableness with high precision. This is a leap from template-based systems to true stylistic adaptation.

Background: Why Personality Matters in UX

Humans are hardwired to adapt their speech to their partners—a process known as entrainment. Studies show that users find dialogue systems more competent and intelligent when the system’s personality matches their own (the similarity-attraction effect). However, most bots today are "personality-neutral" or rely on rigid, hand-crafted templates. PERSONAGE addresses this by building a generative bridge between psycholinguistics and computational linguistics.

The "Big Five" Framework

The system is grounded in the Big Five Model (OCEAN):

• Openness to Experience: Intellectual vs. Simple.
• Conscientiousness: Disciplined vs. Disorganized.
• Extraversion: Sociable vs. Reserved.
• Agreeableness: Friendly vs. Uncooperative.
• Neuroticism (Emotional Stability): Anxious vs. Calm.

Methodology: The 67-Parameter Pipeline

The genius of PERSONAGE lies in its integration into the standard NLG pipeline. It doesn't just swap words; it changes the logic of the utterance.

1. Content Planning

The system decides what to say. For instance, an extravert persona will have a higher Verbosity setting, including more restaurant attributes and repetitions. A disagreeable persona might trigger a Competence Mitigation parameter, saying something like, "Everybody knows that..." to belittle the user's request.

2. Sentence Planning (Aggregation & Pragmatics)

This layer controls how ideas are joined. An introvert might use "Period" aggregation to create short, formal sentences, while a neurotic persona might have high Stuttering and Filled Pause (e.g., "err", "mmm") markers to simulate anxiety.

3. Lexical Choice

The final step involves selecting the specific lexeme. Using resources like WordNet and VerbOcean, the system adjusts word length and frequency.

Experimental Battleground: The Restaurant Domain

The authors tested PERSONAGE by generating recommendations for NYC restaurants.

Results & Key Findings

The evaluation results were striking:

• Human Alignment: Human judges correctly identified the intended personality in 85% of cases.
• The Extravert Edge: Extraversion was the easiest to project (91.3% accuracy), largely due to explicit markers like exclamation points and high word counts.
• Naturalness vs. Accuracy: Interestingly, "neurotic" and "unconscientious" outputs were rated as less natural. This suggests that humans find extreme negative traits in AI slightly jarring or "robotic."

Critical Insight: Beyond Content

The study proves that Politeness and Style are not just "flavor text"—they are encoded in the syntax and rhetorical structure. For example, using a Tag Question (e.g., "...isn't it?") can modulate perceived Agreeableness, while Subject Implicitness (e.g., "The food is great" vs "The restaurant has great food") signals Extraversion.

Conclusion & Future Work

PERSONAGE represents a significant milestone in User-Adaptive Systems. While the current implementation uses rule-based logic, it provides the "psychological feature set" that future Large Language Model (LLM) fine-tuning can adopt. The next frontier? Real-time adaptation where the system senses your personality through your text and pivots its own persona to match or complement yours.

Field: Natural Language Generation / User Modeling Source: Springer Science+Business Media B.V. 2010