Language Explanations Research Articles

Explainable Artificial Intelligence: Evaluating the Objective and Subjective Impacts of xAI on Human-Agent Interaction

Intelligent agents must be able to communicate intentions and explain their decision-making processes to build trust, foster confidence, and improve human-agent team dynamics. Recognizing this need, academia and industry are rapidly proposing new ideas, methods, and frameworks to aid in the design of more explainable AI. Yet, there remains no standardized metric or experimental protocol for benchmarking new methods, leaving researchers to rely on their own intuition or ad hoc methods for assessing new concepts. In this work, we present the first comprehensive (n = 286) user study testing a wide range of approaches for explainable machine learning, including feature importance, probability scores, decision trees, counterfactual reasoning, natural language explanations, and case-based reasoning, as well as a baseline condition with no explanations. We provide the first large-scale empirical evidence of the effects of explainability on human-agent teaming. Our results will help to guide the future of explainability research by highlighting the benefits of counterfactual explanations and the shortcomings of confidence scores for explainability. We also propose a novel questionnaire to measure explainability with human participants, inspired by relevant prior work and correlated with human-agent teaming metrics.

Supervising Model Attention with Human Explanations for Robust Natural Language Inference

Natural Language Inference (NLI) models are known to learn from biases and artefacts within their training data, impacting how well they generalise to other unseen datasets. Existing de-biasing approaches focus on preventing the models from learning these biases, which can result in restrictive models and lower performance. We instead investigate teaching the model how a human would approach the NLI task, in order to learn features that will generalise better to previously unseen examples. Using natural language explanations, we supervise the model’s attention weights to encourage more attention to be paid to the words present in the explanations, significantly improving model performance. Our experiments show that the in-distribution improvements of this method are also accompanied by out-of-distribution improvements, with the supervised models learning from features that generalise better to other NLI datasets. Analysis of the model indicates that human explanations encourage increased attention on the important words, with more attention paid to words in the premise and less attention paid to punctuation and stopwords.

Nice Perfume. How Long Did You Marinate in It? Multimodal Sarcasm Explanation

Sarcasm is a pervading linguistic phenomenon and highly challenging to explain due to its subjectivity, lack of context and deeply-felt opinion. In the multimodal setup, sarcasm is conveyed through the incongruity between the text and visual entities. Although recent approaches deal with sarcasm as a classification problem, it is unclear why an online post is identified as sarcastic. Without proper explanation, end users may not be able to perceive the underlying sense of irony. In this paper, we propose a novel problem -- Multimodal Sarcasm Explanation (MuSE) -- given a multimodal sarcastic post containing an image and a caption, we aim to generate a natural language explanation to reveal the intended sarcasm. To this end, we develop MORE, a new dataset with explanation of 3510 sarcastic multimodal posts. Each explanation is a natural language (English) sentence describing the hidden irony. We benchmark MORE by employing a multimodal Transformer-based architecture. It incorporates a cross-modal attention in the Transformer's encoder which attends to the distinguishing features between the two modalities. Subsequently, a BART-based auto-regressive decoder is used as the generator. Empirical results demonstrate convincing results over various baselines (adopted for MuSE) across five evaluation metrics. We also conduct human evaluation on predictions and obtain Fleiss' Kappa score of 0.4 as a fair agreement among 25 evaluators.

Hybrid Autoregressive Inference for Scalable Multi-Hop Explanation Regeneration

Regenerating natural language explanations in the scientific domain has been proposed as a benchmark to evaluate complex multi-hop and explainable inference. In this context, large language models can achieve state-of-the-art performance when employed as cross-encoder architectures and fine-tuned on human-annotated explanations. However, while much attention has been devoted to the quality of the explanations, the problem of performing inference efficiently is largely under studied. Cross-encoders, in fact, are intrinsically not scalable, possessing limited applicability to real-world scenarios that require inference on massive facts banks. To enable complex multi-hop reasoning at scale, this paper focuses on bi-encoder architectures, investigating the problem of scientific explanation regeneration at the intersection of dense and sparse models. Specifically, we present SCAR (for Scalable Autoregressive Inference), a hybrid framework that iteratively combines a Transformer-based bi-encoder with a sparse model of explanatory power, designed to leverage explicit inference patterns in the explanations. Our experiments demonstrate that the hybrid framework significantly outperforms previous sparse models, achieving performance comparable with that of state-of-the-art cross-encoders while being approx 50 times faster and scalable to corpora of millions of facts. Further analyses on semantic drift and multi-hop question answering reveal that the proposed hybridisation boosts the quality of the most challenging explanations, contributing to improved performance on downstream inference tasks.

DeXAR

The sensor-based recognition of Activities of Daily Living (ADLs) in smart-home environments is an active research area, with relevant applications in healthcare and ambient assisted living. The application of Explainable Artificial Intelligence (XAI) to ADLs recognition has the potential of making this process trusted, transparent and understandable. The few works that investigated this problem considered only interpretable machine learning models. In this work, we propose DeXAR, a novel methodology to transform sensor data into semantic images to take advantage of XAI methods based on Convolutional Neural Networks (CNN). We apply different XAI approaches for deep learning and, from the resulting heat maps, we generate explanations in natural language. In order to identify the most effective XAI method, we performed extensive experiments on two different datasets, with both a common-knowledge and a user-based evaluation. The results of a user study show that the white-box XAI method based on prototypes is the most effective.

Teachers’ Verbal Lexical Explanation for Second Language Vocabulary Learning: A Meta‐Analysis

AbstractThis article reports a meta‐analysis of studies on the effects of teachers’ verbal lexical explanation (TVLE), as one type of lexical focus on form, on second language vocabulary learning. The dataset for this meta‐analysis included 14 studies, representing a total of 36 independent samples (N = 3,304). The results of this study reveal that TVLE resulted in more vocabulary gains than an absence of TVLE, at posttest and delayed posttest. In addition, explanation in the first language was more effective than explanation in the second language at posttest and delayed posttest. We further identified a range of potential moderators that may influence the effects of TVLE. In particular, we found that the overall effectiveness of TVLE was greater when the degree of vocabulary knowledge was measured through recall tests rather than through recognition tests in the long term.

Interpretable clinical time-series modeling with intelligent feature selection for early prediction of antimicrobial multidrug resistance

Electronic health records provide rich, heterogeneous data about the evolution of the patients’ health status. However, such data need to be processed carefully, with the aim of extracting meaningful information for clinical decision support. In this paper, we leverage interpretable (deep) learning and signal processing tools to deal with multivariate time-series data collected from the Intensive Care Unit (ICU) of the University Hospital of Fuenlabrada (Madrid, Spain). The presence of antimicrobial multidrug-resistant (AMR) bacteria is one of the greatest threats to the health system in general and to the ICUs in particular due to the critical health status of the patients therein. Thus, early identification of bacteria at the ICU and early prediction of their antibiotic resistance are key for the patients’ prognosis. While intelligent data-based processing and learning schemes can contribute to this early prediction, their acceptance and deployment in the ICUs require the automatic schemes to be not only accurate but also understandable by clinicians. Accordingly, we have designed trustworthy intelligent models for the early prediction of AMR based on the combination of meaningful feature selection with interpretable recurrent neural networks. These models were created using irregularly sampled clinical measurements, both considering the health status of the patient and the global ICU environment. We explored several strategies to cope with strongly imbalance data, since only a few ICU patients are infected by AMR bacteria. It is worth noting that our approach exhibits a good balance between performance and interpretability, especially when considering the difficulty of the classification task at hand. A multitude of factors are involved in the emergence of AMR (several of them not fully understood), and the records only contain a subset of them. In addition, the limited number of patients, the imbalance between classes, and the irregularity of the data render the problem harder to solve. Our models are also enriched with SHAP post-hoc interpretability and validated by clinicians who considered model understandability and trustworthiness of paramount concern for pragmatic purposes. Moreover, we use linguistic fuzzy systems to provide clinicians with explanations in natural language. Such explanations are automatically generated from a pool of interpretable rules that describe the interaction among the most relevant features identified by SHAP. Notice that clinicians were especially satisfied with new insights provided by our models. Such insights helped them to trust the automatic schemes and use them to make (better) decisions to mitigate AMR spreading in the ICU. All in all, this work paves the way towards more comprehensible time-series analysis in the context of early AMR prediction in ICUs and reduces the time of detection of infectious diseases, opening the door to better hospital care.

AI-GlobalEvents: A Software for analyzing, identifying and explaining global events with Artificial Intelligence

AI-GlobalEvents is a decision support dashboard for policy planners to conduct strategic threat assessments based on global news and events. It uses AI services and algorithms to automatically aggregate global news from online news portals, websites and social media for analyzing the events and identifying critical events requiring imminent attention. The software uses AI algorithms like Entity Detection, Sentiment Analysis, Anomaly detection and Regression to produce explanations in natural language. It is capable of presenting the result in a wide range of platforms including mobile phones (Android or iOS), tablets or desktop environments. AI-GlobalEvents is available at https://github.com/DrSufi/GlobalEvent.

Patient-Friendly Radiology Reporting-Implementation and Outcomes.

To evaluate patient use and experience with patient-centered radiology reports provided via a radiology-specific patient portal in an outpatient setting. Patient-centered interactive radiology reports were made available to all patients who had imaging examinations performed at any of seven outpatient imaging centers via a radiology-specific patient portal, using commercially available software. We recorded portal usage statistics including report access rate, report view time, and use of interactive diagrams and plain language explanations. We assessed patients' subjective report comprehension and experience via survey questions. Results were stratified by age. The majority of patients accessed their online radiology report (59%). For patients who used the interactive plain language definitions and diagrams, the average report viewing time was 5.8 ± 6.6 min compared with 1.8 ± 1.9 min for patients viewing only the raw text report (P < .005). The majority (84.7%) of patients reported that the definitions and diagrams helped them understand their report, and 86.7% of patients endorsed an overall positive experience viewing their report online. Our results suggest that patient-centered radiology reporting is feasible in the outpatient private practice setting and that dedicated patient-centered content has the potential to increase engagement across all age groups. Our experience suggests that patient-centered radiology reports may improve subjective patient comprehension and positively impact the patient experience.

Classification of Covid-19 chest X-ray images by means of an interpretable evolutionary rule-based approach.

In medical practice, all decisions, as for example the diagnosis based on the classification of images, must be made reliably and effectively. The possibility of having automatic tools helping doctors in performing these important decisions is highly welcome. Artificial Intelligence techniques, and in particular Deep Learning methods, have proven very effective on these tasks, with excellent performance in terms of classification accuracy. The problem with such methods is that they represent black boxes, so they do not provide users with an explanation of the reasons for their decisions. Confidence from medical experts in clinical decisions can increase if they receive from Artificial Intelligence tools interpretable output under the form of, e.g., explanations in natural language or visualized information. This way, the system outcome can be critically assessed by them, and they can evaluate the trustworthiness of the results. In this paper, we propose a new general-purpose method that relies on interpretability ideas. The approach is based on two successive steps, the former being a filtering scheme typically used in Content-Based Image Retrieval, whereas the latter is an evolutionary algorithm able to classify and, at the same time, automatically extract explicit knowledge under the form of a set of IF-THEN rules. This approach is tested on a set of chest X-ray images aiming at assessing the presence of COVID-19.

Explainable discrete Collaborative Filtering

Using hashing to learn the binary codes of users and items significantly improves the efficiency and reduces the space consumption of the recommender system. However, existing hashing-based recommender systems remain black boxes without any explainable outputs that illustrate why the system recommends the items. In this paper, we present a new end-to-end discrete recommendation framework based on the multi-task learning to simultaneously perform explainable and efficient recommendation. Toward this goal, an Explainable Discrete Collaborative Filtering (EDCF) method is proposed to preserve the user-item interaction features and semantic text features into binary hash codes by adaptively exploiting the correlations between the preference prediction task and the explanation generation task. At the online recommendation stage, EDCF makes efficient top-K recommendation by calculating the Hamming distances between the feature hash codes, and simultaneously generates natural language explanations for recommendation results through the explanation generation module. To obtain the hash codes directly from the end-to-end neural network, we introduce an attentive TextCNN and an Adaptive Tanh layer in the preference prediction task. For explanation generation, Long Short-Term Memory is employed to generate the explanations for recommendation results from the binary hash codes of user and item. Experiments demonstrate the superiority of the proposed method.

Human Language Explanation for a Decision Making Agent via Automated Rationale Generation

Advances in reinforcement learning (RL) algorithms have made them become increasingly capable in many tasks in recent years.However, the vastmajority of RL algorithms are not readily interpretable off the shelf. Moreover, the task of generating explanations in the form of human language has not been sufficiently addressed for these RL algorithms in previous works. Human language explanations have the advantages of being easy to understand, and they can help increase the satisfaction experienced by the end user while using the product. In this paper, we propose a method for generating explanations in the form of free-text human language to help the end user better understand the behaviors of RL agents. Our work involves generating explanations for both single actions and sequences of actions. We also create an open dataset as a baseline for future research. Our proposed method is evaluated in two simulated environments: Pong and the Minimalistic Gridworld Environment (MiniGgrid). The results demonstrate that our models are able to consistently generate accurate rationales, which are highly correlated with the expert rationales. Hence, this work offers a solution for bridging the gap of trust encountered when employing RL agents in virtual world or real-world applications.

ContrXT: Generating contrastive explanations from any text classifier

The need for explanations of ML systems is growing as new models outperform their predecessors while becoming more complex and less comprehensible for their end-users. Though several XAI methods have been proposed in recent years, not enough attention was paid to explaining how models change their behaviour in contrast with previous ones (e.g., due to retraining). In such cases, an XAI system should explain why the model changes its predictions concerning past outcomes. Capturing and understanding such differences is crucial, as the need for trust is key in any application to support human-AI decision-making processes.This is the idea of ContrXT, a novel approach that (i) traces the decision criteria of a black box text classifier by encoding the changes in the decision logic through Binary Decision Diagrams. Then (ii) it provides global, model-agnostic, Time-Contrastive (T-contrast) explanations in natural language, estimating why – and to what extent – the model has modified its behaviour over time. We implemented and evaluated ContrXT over several supervised ML models trained on a benchmark dataset and a real-life application, showing it is effective in catching majorly changed classes and in explaining their variation through a user study.The approach has been implemented, and it is available to the community both as a python package and through REST API, providing contrastive explanations as a service.

قراءات قرآنية أشكل توجيهها على بعض العلماء دراسة تحليلية لنماذج مختارة

This research aims to study examples of Qira'at that appear to be ambiguous to some scholars. The importance of this research lies in scrutinising the discourse between Qira’at and its explanation in the Arabic language. The two researchers followed the inductive approach to clarify models of Qira'at that appears ambiguous to them and the analytical approach to liberate the main problem and carry out guidance and weighting. The research concluded with some results, the most prominent of which are: The ambiguity in Qira'at is a relative matter that is due to a person; And whoever among the scholars think very deep and conduct a research about it will get a clear understanding of it, its meaning will be understandable, also its clear that some scholars’ objections to Qira'at were due to their jealousy of the Qur’an and their reverence for it.

Generating visual explanations with natural language

AbstractWe generate natural language explanations for a fine‐grained visual recognition task. Our explanations fulfill two criteria. First, explanations are class discriminative, meaning they mention attributes in an image which are important to identify a class. Second, explanations are image relevant, meaning they reflect the actual content of an image. Our system, composed of an explanation sampler and phrase‐critic model, generates class discriminative and image relevant explanations. In addition, we demonstrate that our explanations can help humans decide whether to accept or reject an AI decision.

Explaining the impact of source behaviour in evidential reasoning

Explanation abilities are required for data-driven models, where the high number of parameters may render its internal reasoning opaque to users. Despite the natural transparency brought by the graphical model structure of Bayesian networks, decisions trees or valuation networks, additional explanation abilities are still required due to both the complexity of the problem as well as the consequences of the decision to be taken. Threat assessment is an example of such a complex problem in which several sources with partially unknown behaviour provide information on distinct but related frames of discernment. In this paper, we propose a solution as an evidential network with explanation abilities to detect and investigate threat to maritime infrastructure. We propose a post-hoc explanation approach to an already transparent by design threat assessment model, combining feature relevance and natural language explanations with some visual support. To this end, we extend the sensitivity analysis method of generation of explanations for evidential reasoning to a multi-source model where sources can have several and disparate behaviours. Natural language explanations are generated on the basis of a series of sensitivity measures quantifying the impact of both direct reports and source models. We conclude on challenges to be addressed in future work.

Knowledge-based relational search in cultural heritage linked data

Abstract This article presents a new knowledge-based approach for finding interesting semantic relations between resources in a knowledge graph (KG). The idea is to characterize the notion of ‘interesting connection’ in terms of generic ontological explanation patterns that are applied to an underlying linked data repository to instantiate connections. In this way, (1) semantically uninteresting connections can be ruled out effectively and (2) natural language explanations about the connections can be created for the end-user. The idea has been implemented and tested based on a KG of biographical data extracted from the life stories of 13,144 prominent historical persons in Finland, enriched by data linking to collection databases of museums, libraries, and archives. The demonstrator is in use as part of the BiographySampo portal of interlinked biographies.

Interpretable Mamdani neuro-fuzzy model through context awareness and linguistic adaptation

Interpretable machine learning is trending as it aims to build a human-understandable decision process. There are two main types of machine learning systems: white-box and black-box models. White-box models are inherently interpretable but commonly suffer from under-fitting phenomena; on the other hand, black-box models perform quite well in a wide range of application domain problems, but their reasoning behind a decision is hard or even impossible to understand. In the soft-computing area, fuzzy inference systems are rule-based systems that use fuzzy reasoning, bringing human perception modeling and computing with word capability. These rule-based systems are designed either manually or automatically but are commonly optimized to fit better some phenomena’ data (in a supervised learning task). After the optimization process, the initial semantic meaning of fuzzy sets is modified (slightly, in the best cases), creating a gray-box model. The principal objective of the proposed methodology in this paper is to extract a high-quality rule in terms of comprehensibility, accuracy and fidelity. This is accomplished by using a fuzzy linguistic interpretable model from an optimized neuro-fuzzy model, considering the initial knowledge context with which it was built. A grammar-guided genetic algorithm is used as the optimization process to find the interpretable description of the model. A collection of 16 datasets for classification tasks were used to evaluate our proposal, obtaining an f1-score of 0.814 with 0.026 standard deviation in the optimized model; the obtained fidelity, in terms of similarity from the interpretable model to the optimized one, was 0.93 of mean with 0.018 standard deviation. Obtained results show that neuro-fuzzy systems could play an important role in interpretable machine learning, providing natural language explanations from previous knowledge.

Dissemination and Implementation Science Approaches for Occupational Safety and Health Research: Implications for Advancing Total Worker Health.

Total Worker Health® (TWH), an initiative of the U.S. National Institute for Occupational Safety and Health, is defined as policies, programs, and practices that integrate protection from work-related health and safety hazards by promoting efforts that advance worker well-being. Interventions that apply the TWH paradigm improve workplace health more rapidly than wellness programs alone. Evidence of the barriers and facilitators to the adoption, implementation, and long-term maintenance of TWH programs is limited. Dissemination and implementation (D&I) science, the study of methods and strategies for bridging the gap between public health research and practice, can help address these system-, setting-, and worker-level factors to increase the uptake, impact, and sustainment of TWH activities. The purpose of this paper is to draw upon a synthesis of existing D&I science literature to provide TWH researchers and practitioners with: (1) an overview of D&I science; (2) a plain language explanation of key concepts in D&I science; (3) a case study example of moving a TWH intervention down the research-to-practice pipeline; and (4) a discussion of future opportunities for conducting D&I science in complex and dynamic workplace settings to increase worker safety, health, and well-being.

A plea for proportionality: A reply to Yun-chien Chang and Xin Dai

Abstract After a survey of the proportionality principle and its discontents (irrationality, conceptual flaws, and bias), this article addresses the dual remedy proposed by proportionality critics, which consists in: (i) skipping the first to third prongs of the proportionality analysis (legitimate objective, suitability, and necessity) and (2) concentrating on balancing in the metric form of a cost-benefit analysis. I argue that discarding the three prongs, and thus abandoning the asymmetrical structure of the sequential proportionality analysis, would kill off the specific quality and function of rights which is to constitute a special protection that triggers the obligation to explain and justify interference. Although recent empirical studies have demonstrated that the proportionality test indeed does not deliver predictable legal outcomes, a simple cost-benefit analysis is worse. It would not secure more reliable outcomes either but only offers a sham-rationality. Assigning numbers to the goods and interests at stake cannot replace the work of a multidimensional assessment and argumentative explanation in natural language which is needed for making not only a sound, but also a transparent and thus contestable, decision. What legal life needs is a culture of justification, not a culture of calculation.