Towards Language Technology Evaluation Grounded in Human Judgment
Abstract:
Evaluation of language technology is a crucial research problem because it directly impacts the development, deployment, and trustworthiness of AI-driven systems that interact through human language. Despite progress in benchmark-driven evaluation, there is growing evidence that performance on static datasets often fails to reflect how systems are perceived and judged by humans in real-world use. The emergence of large generative models with increasingly diverse and powerful capabilities has amplified this ongoing challenge in evaluation, raising uncertainty about whether their stronger technical capabilities translate into better human experience. This potential mismatch highlights the limitations of previous evaluation approaches based on fixed tasks and model-centric metrics, creating an urgent need for evaluation frameworks that are more closely grounded in empirical human judgment and real interactions.
In this thesis, we examine how the evaluation of language technology can be grounded in human judgment by identifying where existing benchmarks fall short and designing evaluation methodologies that better reflect how humans assess language systems. Our work focuses on the use of empirical human signals such as human annotations, preference judgments, and feedback during interactions, which reflect actual human needs or values. Grounding evaluation in these signals leads to a more faithful evaluation of models in settings where social appropriateness, interaction experience, and user needs and perceptions are central.
Specifically, we introduce a social intelligence data infrastructure that supports evaluation grounded in human judgment through a comprehensive taxonomy of social intelligence in AI systems and a data library of 480 natural language processing (NLP) datasets. This infrastructure clarifies how socially intelligent behaviors have been benchmarked in prior work, and it provides a structured foundation for analyzing evaluation practices and identifying gaps in how language models are assessed in socially grounded contexts. Building on insights from this work, we conduct an interactive evaluation of multiple Large Audio Models by collecting large-scale human preference data from real users. Through analyses of preference rankings and qualitative feedback, we empirically demonstrate a substantial gap between performance on existing speech benchmarks and human judgments during interaction.
On top of identifying human-relevant evaluation dimensions in actual interactions, we also introduce scalable evaluation methodologies that enhance alignment with human judgment. We develop a collaborative human-in-the-loop data annotation process that allocates labeling jobs based on model uncertainty, enabling data creation that is better aligned with human values at the same cost level across various NLP and computational social science datasets. In addition, we propose the DnA-Eval framework which decomposes complex evaluative judgments into interpretable sub-criteria and aggregates them systematically. The framework results in model-based evaluations that align more closely with labels derived from human judgment.