Machine Explanations and Human Understanding

Explanations are hypothesized to improve human understanding of machine learning models and achieve a variety of desirable outcomes, ranging from model debugging to enhancing human decision making. However, empirical studies have found mixed and even negative results. An open question, therefore, is under what conditions explanations can improve human understanding and in what way. Using adapted causal diagrams, we provide a formal characterization of the interplay between machine explanations and human understanding, and show how human intuitions play a central role in enabling human understanding. Specifically, we identify three core concepts of interest that cover all existing quantitative measures of understanding in the context of human-AI decision making: task decision boundary, model decision boundary, and model error. Our key result is that without assumptions about task-specific intuitions, explanations may potentially improve human understanding of model decision boundary, but they cannot improve human understanding of task decision boundary or model error. To achieve complementary human-AI performance, we articulate possible ways on how explanations need to work with human intuitions. For instance, human intuitions about the relevance of features (e.g., education is more important than age in predicting a person's income) can be critical in detecting model error. We validate the importance of human intuitions in shaping the outcome of machine explanations with empirical human-subject studies. Overall, our work provides a general framework along with actionable implications for future algorithmic development and empirical experiments of machine explanations.

Rethinking Explainability as a Dialogue: A Practitioner's Perspective

As practitioners increasingly deploy machine learning models in critical domains such as health care, finance, and policy, it becomes vital to ensure that domain experts function effectively alongside these models. Explainability is one way to bridge the gap between human decision-makers and machine learning models. However, most of the existing work on explainability focuses on one-off, static explanations like feature importances or rule lists. These sorts of explanations may not be sufficient for many use cases that require dynamic, continuous discovery from stakeholders. In the literature, few works ask decision-makers about the utility of existing explanations and other desiderata they would like to see in an explanation going forward. In this work, we address this gap and carry out a study where we interview doctors, healthcare professionals, and policymakers about their needs and desires for explanations. Our study indicates that decision-makers would strongly prefer interactive explanations in the form of natural language dialogues. Domain experts wish to treat machine learning models as "another colleague", i.e., one who can be held accountable by asking why they made a particular decision through expressive and accessible natural language interactions. Considering these needs, we outline a set of five principles researchers should follow when designing interactive explanations as a starting place for future work. Further, we show why natural language dialogues satisfy these principles and are a desirable way to build interactive explanations. Next, we provide a design of a dialogue system for explainability and discuss the risks, trade-offs, and research opportunities of building these systems. Overall, we hope our work serves as a starting place for researchers and engineers to design interactive explainability systems.

Towards a Science of Human-AI Decision Making: A Survey of Empirical Studies

As AI systems demonstrate increasingly strong predictive performance, their adoption has grown in numerous domains. However, in high-stakes domains such as criminal justice and healthcare, full automation is often not desirable due to safety, ethical, and legal concerns, yet fully manual approaches can be inaccurate and time consuming. As a result, there is growing interest in the research community to augment human decision making with AI assistance. Besides developing AI technologies for this purpose, the emerging field of human-AI decision making must embrace empirical approaches to form a foundational understanding of how humans interact and work with AI to make decisions. To invite and help structure research efforts towards a science of understanding and improving human-AI decision making, we survey recent literature of empirical human-subject studies on this topic. We summarize the study design choices made in over 100 papers in three important aspects: (1) decision tasks, (2) AI models and AI assistance elements, and (3) evaluation metrics. For each aspect, we summarize current trends, discuss gaps in current practices of the field, and make a list of recommendations for future research. Our survey highlights the need to develop common frameworks to account for the design and research spaces of human-AI decision making, so that researchers can make rigorous choices in study design, and the research community can build on each other's work and produce generalizable scientific knowledge. We also hope this survey will serve as a bridge for HCI and AI communities to work together to mutually shape the empirical science and computational technologies for human-AI decision making.

What to Learn, and How: Toward Effective Learning from Rationales

Learning from rationales seeks to augment model training with human-provided rationales (i.e., a subset of input tokens) that justify those labels. While intuitive, this idea has proven elusive in practice. We make two observations about human rationales via empirical analyses: 1) maximizing predicted rationale accuracy is not necessarily the optimal objective for improving model performance; 2) human rationales vary in whether they provide sufficient information for the model to exploit for prediction, and we can use this variance to assess a dataset's potential improvement from learning from rationales. Building on these insights, we propose loss functions and learning strategies, and evaluate their effectiveness on three datasets with human rationales. Our results demonstrate consistent improvements over baselines in both label performance and rationale performance, including a 3% accuracy improvement on MultiRC. Our work highlights the importance of understanding properties of human explanations and exploiting them accordingly in model training.

Investigating the Effect of Natural Language Explanations on Out-of-Distribution Generalization in Few-shot NLI

Although neural models have shown strong performance in datasets such as SNLI, they lack the ability to generalize out-of-distribution (OOD). In this work, we formulate a few-shot learning setup and examine the effects of natural language explanations on OOD generalization. We leverage the templates in the HANS dataset and construct templated natural language explanations for each template. Although generated explanations show competitive BLEU scores against groundtruth explanations, they fail to improve prediction performance. We further show that generated explanations often hallucinate information and miss key elements that indicate the label.

Decision-Focused Summarization

Relevance in summarization is typically defined based on textual information alone, without incorporating insights about a particular decision. As a result, to support risk analysis of pancreatic cancer, summaries of medical notes may include irrelevant information such as a knee injury. We propose a novel problem, decision-focused summarization, where the goal is to summarize relevant information for a decision. We leverage a predictive model that makes the decision based on the full text to provide valuable insights on how a decision can be inferred from text. To build a summary, we then select representative sentences that lead to similar model decisions as using the full text while accounting for textual non-redundancy. To evaluate our method (DecSum), we build a testbed where the task is to summarize the first ten reviews of a restaurant in support of predicting its future rating on Yelp. DecSum substantially outperforms text-only summarization methods and model-based explanation methods in decision faithfulness and representativeness. We further demonstrate that DecSum is the only method that enables humans to outperform random chance in predicting which restaurant will be better rated in the future.

On Positivity Bias in Negative Reviews

Prior work has revealed that positive words occur more frequently than negative words in human expressions, which is typically attributed to positivity bias, a tendency for people to report positive views of reality. But what about the language used in negative reviews? Consistent with prior work, we show that English negative reviews tend to contain more positive words than negative words, using a variety of datasets. We reconcile this observation with prior findings on the pragmatics of negation, and show that negations are commonly associated with positive words in negative reviews. Furthermore, in negative reviews, the majority of sentences with positive words express negative opinions based on sentiment classifiers, indicating some form of negation.

On the Diversity and Limits of Human Explanations

A growing effort in NLP aims to build datasets of human explanations. However, the term explanation encompasses a broad range of notions, each with different properties and ramifications. Our goal is to provide an overview of diverse types of explanations and human limitations, and discuss implications for collecting and using explanations in NLP. Inspired by prior work in psychology and cognitive sciences, we group existing human explanations in NLP into three categories: proximal mechanism, evidence, and procedure. These three types differ in nature and have implications for the resultant explanations. For instance, procedure is not considered explanations in psychology and connects with a rich body of work on learning from instructions. The diversity of explanations is further evidenced by proxy questions that are needed for annotators to interpret and answer open-ended why questions. Finally, explanations may require different, often deeper, understandings than predictions, which casts doubt on whether humans can provide useful explanations in some tasks.

Using Transformers to Provide Teachers with Personalized Feedback on their Classroom Discourse: The TalkMoves Application

TalkMoves is an innovative application designed to support K-12 mathematics teachers to reflect on, and continuously improve their instructional practices. This application combines state-of-the-art natural language processing capabilities with automated speech recognition to automatically analyze classroom recordings and provide teachers with personalized feedback on their use of specific types of discourse aimed at broadening and deepening classroom conversations about mathematics. These specific discourse strategies are referred to as "talk moves" within the mathematics education community and prior research has documented the ways in which systematic use of these discourse strategies can positively impact student engagement and learning. In this article, we describe the TalkMoves application's cloud-based infrastructure for managing and processing classroom recordings, and its interface for providing teachers with feedback on their use of talk moves during individual teaching episodes. We present the series of model architectures we developed, and the studies we conducted, to develop our best-performing, transformer-based model (F1 = 79.3%). We also discuss several technical challenges that need to be addressed when working with real-world speech and language data from noisy K-12 classrooms.

Understanding the Effect of Out-of-distribution Examples and Interactive Explanations on Human-AI Decision Making

Although AI holds promise for improving human decision making in societally critical domains, it remains an open question how human-AI teams can reliably outperform AI alone and human alone in challenging prediction tasks (also known as complementary performance). We explore two directions to understand the gaps in achieving complementary performance. First, we argue that the typical experimental setup limits the potential of human-AI teams. To account for lower AI performance out-of-distribution than in-distribution because of distribution shift, we design experiments with different distribution types and investigate human performance for both in-distribution and out-of-distribution examples. Second, we develop novel interfaces to support interactive explanations so that humans can actively engage with AI assistance. Using in-person user study and large-scale randomized experiments across three tasks, we demonstrate a clear difference between in-distribution and out-of-distribution, and observe mixed results for interactive explanations: while interactive explanations improve human perception of AI assistance's usefulness, they may magnify human biases and lead to limited performance improvement. Overall, our work points out critical challenges and future directions towards complementary performance.