Measuring Information in Text Explanations
Zining Zhu, Frank Rudzicz

TL;DR
This paper introduces an information-theoretic framework to evaluate text explanations in explainable AI, unifying different methods by quantifying relevance and informativeness through communication channels.
Contribution
It proposes a novel information-theoretic approach to assess text explanations, providing standardized metrics and insights into explanation mechanisms.
Findings
NLEs balance input and target information transmission.
Rationales do not show a trade-off between information types.
Information scores reveal underlying explanation differences.
Abstract
Text-based explanation is a particularly promising approach in explainable AI, but the evaluation of text explanations is method-dependent. We argue that placing the explanations on an information-theoretic framework could unify the evaluations of two popular text explanation methods: rationale and natural language explanations (NLE). This framework considers the post-hoc text pipeline as a series of communication channels, which we refer to as ``explanation channels''. We quantify the information flow through these channels, thereby facilitating the assessment of explanation characteristics. We set up tools for quantifying two information scores: relevance and informativeness. We illustrate what our proposed information scores measure by comparing them against some traditional evaluation metrics. Our information-theoretic scores reveal some unique observations about the underlying…
Peer Reviews
Decision·Submitted to ICLR 2024
Text-based explanation is a nascent field for which this paper offers a novel attempt. The paper shows the feasibility of applying mutual information measures, noting that currently it is "still unknown whether these tools can be used to examine information scores." The paper's demonstration of the feasibility of using various approximations to mutual information in this case is novel. The use in practice of these measures for evaluation is a worthwhile contribution. The entire field of natura
The conclusions are modest, and give limited insight. Despite this the approach has promise, as it plows new ground in an area where there is limited success today.
- the paper is original - we found only a paper with some similar concepts used in a different context - the authors tackle a relevant problem and propose a framework to evaluate the informativeness of text explanations. Furthermore, the framework contemplates a high degree of automation, making it feasible to deploy in production settings. - the authors propose measuring two key aspects of the explanations: relevance and predictive informativeness. - we consider the paper to be of good quali
- the structure of how experiments and results are reported can be improved. In particular, it would be helpful if the authors list the experiments performed, listing rationale behind the experiment, the procedure, aims, metrics, and other aspects of relevance.
* The way of forming the text explanation within the information theory framework is interesting. * The experiments have included three embedding models: RoBERTa, OpenAI, Cohere, to test the generalizability of the proposed relevance and informativeness scores, which highly depends on the used embeddings. * The experiments have considered two often-seen types: the rationale, which is defined to include the tokenwise explanation, and the NLE, which is defined to be explaining the input-target rel
* The writing can be improved: The goal of this paper can be more clear. I was lost in the middle of the paper and wondered (1) which parts are used for evaluation and (2) what are they truly evaluating for? I could only realize the goal after reading through the whole paper and gave it a guess. * This paper claims to unify the evaluation of rationale and natural language explanations. Unification would make me expect the proposed evaluation method can evaluate them in a good standing. However,
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsExplainable Artificial Intelligence (XAI) · Topic Modeling · Scientific Computing and Data Management
