[Paper Review] Evaluating Human-Language Model Interaction

Mina Lee, Megha Srivastava+16 more|arXiv (Cornell University)|Dec 19, 2022

Topic ModelingComputer Science39 citations

TL;DR

HALIE is a framework to evaluate interactive human-LM interactions across targets, perspectives, and criteria, showing that better non-interactive LM performance does not always yield better interactive outcomes across five tasks.

ABSTRACT

Many real-world applications of language models (LMs), such as writing assistance and code autocomplete, involve human-LM interaction. However, most benchmarks are non-interactive in that a model produces output without human involvement. To evaluate human-LM interaction, we develop a new framework, Human-AI Language-based Interaction Evaluation (HALIE), that defines the components of interactive systems and dimensions to consider when designing evaluation metrics. Compared to standard, non-interactive evaluation, HALIE captures (i) the interactive process, not only the final output; (ii) the first-person subjective experience, not just a third-party assessment; and (iii) notions of preference beyond quality (e.g., enjoyment and ownership). We then design five tasks to cover different forms of interaction: social dialogue, question answering, crossword puzzles, summarization, and metaphor generation. With four state-of-the-art LMs (three variants of OpenAI's GPT-3 and AI21 Labs' Jurassic-1), we find that better non-interactive performance does not always translate to better human-LM interaction. In particular, we highlight three cases where the results from non-interactive and interactive metrics diverge and underscore the importance of human-LM interaction for LM evaluation.

Motivation & Objective

• Define a multidimensional framework (HALIE) for evaluating human-LM interaction beyond final outputs.
• Model the interaction as a system with states, actions, and prompts to capture the interactive process.
• Develop metrics across three dimensions (targets, perspectives, criteria) to assess traces of interaction.
• Instantiate HALIE on five tasks to study how LMs perform in interactive settings.
• Provide guidelines and release data to support future interactive LM evaluation.

Proposed method

• Define a system as LM + UI + system logic that constructs prompts and invokes the LM.
• Represent interaction as a trace of state-action pairs generated by user-system prompts and LM responses.
• Propose HALIE’s three evaluation dimensions: targets (process + output), perspectives (first-person user vs third-party), and criteria (quality vs preference).
• Design five interactive tasks (social dialogue, question answering, crossword puzzles, summarization, metaphor generation) and implement task-specific interactive systems.
• Evaluate four LMs (TextDavinci, TextBabbage, Davinci, Jumbo) to compare interactive performance against non-interactive performance.
• Analyze interaction traces to identify cases where non-interactive superiority does not translate to better interactive performance.

Experimental results

Research questions

• RQ1 RQ1: Does optimizing for non-interactive performance translate to better interactive performance in real-use interactive settings?
• RQ2 RQ2: How do first-person (user) evaluations compare to third-party evaluations in interactive LM interaction?
• RQ3 RQ3: Do quality metrics align with user preferences in interactive use, or can preferences diverge from quality?
• RQ4 How do different tasks (from social dialogue to metaphor generation) reveal divergences between interactive and non-interactive evaluation?

Key findings

• Better non-interactive performance does not always translate to better interactive performance across tasks.
• For QA, some models with lower non-interactive accuracy can outperform others in interactive roles in certain domains.
• First-person user evaluations can diverge from third-party annotations on outputs like summarization.
• Users can perceive models as more helpful than they actually improve task performance (e.g., crossword puzzles).
• Instruction-tuned models often score higher on many quality metrics but may lose specificity.
• In dialogue, users favored TextDavinci for many metrics, yet preferred Davinci for specificity, highlighting preference-driven model selection.