From Pixels to Personas: Investigating and Modeling Self-Anthropomorphism in Human-Robot Dialogues

Yu Li, Devamanyu Hazarika, Di Jin, Julia Hirschberg, Yang Liu

TL;DR

This paper investigates "self-anthropomorphism" (SA) in AI dialogue systems – how bots express human-like traits like preferences and emotions. They introduce a novel dataset, PIX2PERSONA, and models to transition between SA and non-SA responses, aiming for more ethical and engaging AI interactions across different embodiments.

What Problem Does This Paper Address?

As AI becomes more sophisticated, it increasingly mimics human behavior, leading to self-anthropomorphism (SA) in its responses. While this can enhance user engagement, it also raises ethical concerns, such as misleading users about AI capabilities and fostering misplaced trust. Current dialogue datasets often lack clear distinctions between SA and non-SA responses, making it difficult to study and control this phenomenon. The paper aims to address this by systematically analyzing SA in existing datasets, developing methods to transition between SA and non-SA responses, and creating a new dataset to facilitate this research.

Methods

The authors employ a multi-pronged approach:

  • SA Classification: They use GPT-4, guided by specific linguistic criteria (embodiment, relation-seeking, self-expression, identity), to classify bot responses as either self-anthropomorphic (SA) or non-self-anthropomorphic (NSA). This classifier was validated with human annotators, achieving a Cohen's Kappa of 0.83.
  • Prevalence Analysis: They analyze the frequency of SA responses across various dialogue datasets, categorizing them by dialogue task type (open-domain, knowledge-grounded, conversational recommendation, task-oriented).
  • Linguistic Analysis: They use the LIWC tool to identify linguistic features that correlate with SA responses, such as first-person singular pronouns.
  • Transformation Models:
    • SA to NSA: They develop a prompt-based method using GPT-4 to transform SA responses into NSA ones, focusing on ethical considerations.
  • NSA to SA: They also develop a method to transform NSA responses into SA ones, using in-context learning with GPT-4 to encourage more human-like traits, particularly for task-oriented dialogues where NSA is common.
  • Dual-Capability Model: They fine-tune a Mistral-7B model to perform both SA and NSA transformations, aiming for a versatile model that can adapt to different AI embodiment needs.
  • Dataset Creation: They introduce PIX2PERSONA, a dataset containing approximately 10K dialogue turns from 15 diverse datasets, each enhanced with paired SA and NSA responses for original bot turns.

Datasets

The study utilizes and analyzes several existing dialogue datasets, including:

  • Open-Domain Dialogue: DailyDialog, PERSONACHAT, EMPATHETICDIALOGUES, PROSOCIALDIALOG, HH-RLHF, SODA, BlenderBot 3X.
  • Knowledge-Grounded Dialogue: Topical-Chat, Wizard of Wikipedia, Wizard of the Internet.
  • Conversational Recommendation: OPENDIALKG, DuRecDial 2.0.
  • Task-Oriented Dialogue: AirDialogue, MultiWOZ 2.2, MultiDoc2Dial.

The paper also introduces PIX2PERSONA, a novel dataset comprising about 10K dialogue turns from each of the 15 datasets mentioned above, totaling 143K dialogue turns. For each original bot response, PIX2PERSONA provides a paired SA and NSA response, serving as a resource for training and evaluating models that can dynamically adjust anthropomorphism levels.

Key Insights and Ideas

  • SA Varies by Task: Open-domain dialogues tend to have a higher prevalence of SA responses, suggesting bots in these contexts are more likely to exhibit human-like traits. Task-oriented dialogues, conversely, show fewer SA responses, likely due to their focus on objective information.
  • Ethical vs. Engaging: There's a trade-off between ethical AI (NSA responses, often involving disclaimers) and engaging AI (SA responses). The paper aims to bridge this by enabling transitions between the two.
  • Linguistic Markers of SA: The use of "I" statements and first-person singular pronouns strongly correlates with SA responses, while "you" and politeness markers are less common in SA.
  • Disclaimers in NSA: NSA responses, especially in open-domain dialogues, frequently use disclaimers to clarify AI limitations, which can sometimes reduce semantic similarity to the original response.
  • Model Capabilities: Language models like GPT-4 naturally lean towards generating NSA responses, aligning with ethical guidelines. However, transforming NSA to SA requires more nuanced prompting and fine-tuning.

Key Results

  • Classifier Performance: The GPT-4 based classifier achieved 81.92% F1 score for identifying SA responses.
  • Prevalence: PersonaChat and SODA showed the highest SA prevalence (81% and 69% respectively), while MultiWOZ 2.2 and MultiDoc2Dial had the lowest (5% and 0%).
  • Transformation Success:
    • The SA to NSA transformation model showed a significant reduction in SA classifications across datasets.
  • The NSA to SA transformation model successfully increased SA responses in task-oriented dialogues, overcoming GPT-4's natural tendency towards NSA.
  • The fine-tuned Mistral-7B model achieved high accuracy in transforming responses to both SA (up to 87.9%) and NSA (up to 98.3%) across various tasks.
  • Semantic Preservation: Sentence similarity scores and human evaluations indicated robust semantic preservation during transformations, ranging from 0.5 to 0.8 for sentence similarity. NSA transformations in task-oriented dialogues showed the highest semantic retention (93%).

Limitations and Future Work

  • Scope of SA Classification: The study uses a binary classification of SA, which might not fully capture the nuanced requirements of different AI embodiments (e.g., robots vs. chatbots). Future work could tailor SA levels to specific embodiments.
  • Classifier Ambiguity: The broad linguistic guidelines for the SA classifier can lead to some ambiguity, requiring more precise definitions for certain expressions.
  • Disclaimer Generation: The reliance on LLMs for generating disclaimers in NSA responses means that some necessary disclaimers might be missed if not present in the training data. Enhancing models' ability to detect and incorporate ethical disclaimers is crucial.
  • Ethical Dual Use: The dataset, with its SA responses, could be used to create more human-like AI assistants, raising ethical considerations about transparency and user education.
  • Future Directions: The authors suggest exploring optimal SA qualities for diverse AI embodiments, enhancing classifier precision, and ensuring comprehensive ethical disclaimer coverage. They also emphasize the importance of educating the community about ethical AI training.

Summary updated

Abstract

Self-anthropomorphism in robots manifests itself through their display of human-like characteristics in dialogue, such as expressing preferences and emotions. Our study systematically analyzes self-anthropomorphic expression within various dialogue datasets, outlining the contrasts between self-anthropomorphic and non-self-anthropomorphic responses in dialogue systems. We show significant differences in these two types of responses and propose transitioning from one type to the other. We also introduce Pix2Persona, a novel dataset aimed at developing ethical and engaging AI systems in various embodiments. This dataset preserves the original dialogues from existing corpora and enhances them with paired responses: self-anthropomorphic and non-self-anthropomorphic for each original bot response. Our work not only uncovers a new category of bot responses that were previously under-explored but also lays the groundwork for future studies about dynamically adjusting self-anthropomorphism levels in AI systems to align with ethical standards and user expectations.
BibTeX
@inproceedings{li2024from,
  author = {Yu Li and Devamanyu Hazarika and Di Jin and Julia Hirschberg and Yang Liu},
  title = {{From Pixels to Personas: Investigating and Modeling Self-Anthropomorphism in Human-Robot Dialogues}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.findings-emnlp.567},
}
Figure