Can AI Extract Antecedent Factors of Human Trust in AI? An Application of Information Extraction for Scientific Literature in Behavioural and Computer Sciences
Melanie McGrath, Harrison Bailey, Necva Bölücü, Xiang Dai, Sarvnaz Karimi, Cecile Paris
TL;DR
This work addresses how to extract antecedents of human trust in AI from scientific literature by constructing the Trust in AI dataset through expert-guided information extraction. It defines the problem as a two-task pipeline: Named Entity Recognition to identify factors and applications, and Relation Extraction to link factors to trust, using the formal dataset $D = \{S_i, P_i, L_i, R_i\}_{i=1}^N$ with relation labels such as unspecified, null, positive, negative, and interaction. The authors compare supervised models (e.g., RoBERTa-based NER and RE systems) with LLM-based, zero-/few-shot approaches, finding that supervised methods outperform prompt-based LLMs for this domain, underscoring the need for an annotated resource. They implement a rigorous annotation workflow, including LLM guidance and cross-annotator agreement, and provide a foundation for downstream decision-making and synthesis across Trust in AI research. Future work includes extending the dataset with entity resolution to link synonymous mentions and enhance cohesion, as well as discussing ethical and practical considerations of dataset sharing.
Abstract
Information extraction from the scientific literature is one of the main techniques to transform unstructured knowledge hidden in the text into structured data which can then be used for decision-making in down-stream tasks. One such area is Trust in AI, where factors contributing to human trust in artificial intelligence applications are studied. The relationships of these factors with human trust in such applications are complex. We hence explore this space from the lens of information extraction where, with the input of domain experts, we carefully design annotation guidelines, create the first annotated English dataset in this domain, investigate an LLM-guided annotation, and benchmark it with state-of-the-art methods using large language models in named entity and relation extraction. Our results indicate that this problem requires supervised learning which may not be currently feasible with prompt-based LLMs.