Empowering Discovery,
Enhancing Knowledge

In this paper, we argue that the way we have been training and evaluating ML models has largely forgotten the fact that they are applied in an organization or societal context as they provide value to people. We show that with this perspective we fundamentally change how we evaluate and select machine learning models.

As powerful and complex language models are being released to the public, understanding their behaviour is more important than ever. Although Explainable Artificial Intelligence (XAI) approaches have been widely applied to NLP models, the explanations they provide may still be complex to understand for human interpreters as these may not be aligned with the reasoning process they apply in language-based tasks. Furthermore, such a misalignment is also present in most XAI datasets as they are not structured to reflect such a fundamental property. Striving to bridge the gap between model and human reasoning, we propose ad hoc formalizations to structure and detail the thought process applied by human interpreters when performing a set of NLP tasks of interest. Hence, we define rationale mappings, ie, representations that organize humans’ analytical reasoning steps when identifying and associating the essential parts of the texts involved in a language-based task leading to its output. These are organized in tree structures referred to as rationale trees and characterized for each task to enhance their expressiveness. Furthermore, we describe their data collection and storage process. We argue these structures would result in a better alignment between model and human reasoning, hence improving models’ explanations, while still being suited for standard explainability processes.

In many practical applications, machine learning models are embedded into a pipeline involving a human actor that decides whether to trust the machine prediction or take a default route (eg, classify the example herself). Selective classifiers have the option to abstain from making a prediction on an example they do not feel confident about. Recently, the notion of the value of a machine learning model has been introduced as a way to jointly consider the benefit of a correct prediction, the cost of an error, and that of abstaining. In this paper, we study how active learning of selective classifiers is affected by the focus on value. We show that the performance of the state-of-the-art active learning strategies drops significantly when we evaluate them based on value rather than accuracy. Finally, we propose a novel value-aware active learning strategy that outperforms the state-of-the-art ones when the cost of incorrect …

Computer-implemented method of extracting knowledge — TU Delft Research Portal Skip to main navigation Skip to search Skip to main content TU Delft Research Portal Home TU Delft Research Portal Logo Help & FAQ Home Research units Researchers Research output Datasets Projects Press/Media Prizes Activities Search by expertise, name or affiliation Computer-implemented method of extracting knowledge AMA Balayn (Inventor), G. He (Inventor), Jie Yang (Inventor), Ujwal Gadiraju (Inventor), Andrea Hu (Inventor) Organisation & Governance Web Information Systems Research output: Patent Overview Original language English IPC G06N Priority date 12/04/22 Publication status Published – 2023 Bibliographical note Patent: OCT-22-002 Applicant: TU Delft Cite this APA Author BIBTEX Harvard Standard RIS Vancouver Powered by Pure, Scopus & Elsevier Fingerprint Engine™ All content on this site: …