Jeffrey P. Bigham

Prakhar Gupta, Cathy Jiao, Yi-Ting Yeh et al. · cmu

Instruction tuning is an emergent paradigm in NLP wherein natural language instructions are leveraged with language models to induce zero-shot performance on unseen tasks. Instructions have been shown to enable good performance on unseen tasks and datasets in both large and small language models. Dialogue is an especially interesting area to explore instruction tuning because dialogue systems perform multiple kinds of tasks related to language (e.g., natural language understanding and generation, domain-specific interaction), yet instruction tuning has not been systematically explored for dialogue-related tasks. We introduce InstructDial, an instruction tuning framework for dialogue, which consists of a repository of 48 diverse dialogue tasks in a unified text-to-text format created from 59 openly available dialogue datasets. Next, we explore cross-task generalization ability on models tuned on InstructDial across diverse dialogue tasks. Our analysis reveals that InstructDial enables good zero-shot performance on unseen datasets and tasks such as dialogue evaluation and intent detection, and even better performance in a few-shot setting. To ensure that models adhere to instructions, we introduce novel meta-tasks. We establish benchmark zero-shot and few-shot performance of models trained using the proposed framework on multiple dialogue tasks.

Prakhar Gupta, Harsh Jhamtani, Jeffrey P. Bigham · cmu, microsoft-research

Target-guided response generation enables dialogue systems to smoothly transition a conversation from a dialogue context toward a target sentence. Such control is useful for designing dialogue systems that direct a conversation toward specific goals, such as creating non-obtrusive recommendations or introducing new topics in the conversation. In this paper, we introduce a new technique for target-guided response generation, which first finds a bridging path of commonsense knowledge concepts between the source and the target, and then uses the identified bridging path to generate transition responses. Additionally, we propose techniques to re-purpose existing dialogue datasets for target-guided generation. Experiments reveal that the proposed techniques outperform various baselines on this task. Finally, we observe that the existing automated metrics for this task correlate poorly with human judgement ratings. We propose a novel evaluation metric that we demonstrate is more reliable for target-guided response evaluation. Our work generally enables dialogue system designers to exercise more control over the conversations that their systems produce.

Michelle S. Lam, Fred Hohman, Dominik Moritz et al. · apple-ml, cmu

AI policy sets boundaries on acceptable behavior for AI models, but this is challenging in the context of large language models (LLMs): how do you ensure coverage over a vast behavior space? We introduce policy maps, an approach to AI policy design inspired by the practice of physical mapmaking. Instead of aiming for full coverage, policy maps aid effective navigation through intentional design choices about which aspects to capture and which to abstract away. With Policy Projector, an interactive tool for designing LLM policy maps, an AI practitioner can survey the landscape of model input-output pairs, define custom regions (e.g., "violence"), and navigate these regions with if-then policy rules that can act on LLM outputs (e.g., if output contains "violence" and "graphic details," then rewrite without "graphic details"). Policy Projector supports interactive policy authoring using LLM classification and steering and a map visualization reflecting the AI practitioner's work. In an evaluation with 12 AI safety experts, our system helps policy designers craft policies around problematic model behaviors such as incorrect gender assumptions and handling of immediate physical safety threats.

Kundan Krishna, Saurabh Garg, Jeffrey P. Bigham et al.

For most natural language processing tasks, the dominant practice is to finetune large pretrained transformer models (e.g., BERT) using smaller downstream datasets. Despite the success of this approach, it remains unclear to what extent these gains are attributable to the massive background corpora employed for pretraining versus to the pretraining objectives themselves. This paper introduces a large-scale study of self-pretraining, where the same (downstream) training data is used for both pretraining and finetuning. In experiments addressing both ELECTRA and RoBERTa models and 10 distinct downstream classification datasets, we observe that self-pretraining rivals standard pretraining on the BookWiki corpus (despite using around $10\times$--$500\times$ less data), outperforming the latter on $7$ and $5$ datasets, respectively. Surprisingly, these task-specific pretrained models often perform well on other tasks, including the GLUE benchmark. Besides classification tasks, self-pretraining also provides benefits on structured output prediction tasks such as span based question answering and commonsense inference, often providing more than $50\%$ of the performance boosts provided by pretraining on the BookWiki corpus. Our results hint that in many scenarios, performance gains attributable to pretraining are driven primarily by the pretraining objective itself and are not always attributable to the use of external pretraining data in massive amounts. These findings are especially relevant in light of concerns about intellectual property and offensive content in web-scale pretraining data.

Jessica Huynh, Ting-Rui Chiang, Jeffrey Bigham et al.

Dialog system developers need high-quality data to train, fine-tune and assess their systems. They often use crowdsourcing for this since it provides large quantities of data from many workers. However, the data may not be of sufficiently good quality. This can be due to the way that the requester presents a task and how they interact with the workers. This paper introduces DialCrowd 2.0 to help requesters obtain higher quality data by, for example, presenting tasks more clearly and facilitating effective communication with workers. DialCrowd 2.0 guides developers in creating improved Human Intelligence Tasks (HITs) and is directly applicable to the workflows used currently by developers and researchers.

Colin Lea, Dianna Yee, Jaya Narain et al.

Many consumer speech recognition systems are not tuned for people with speech disabilities, resulting in poor recognition and user experience, especially for severe speech differences. Recent studies have emphasized interest in personalized speech models from people with atypical speech patterns. We propose a query-by-example-based personalized phrase recognition system that is trained using small amounts of speech, is language agnostic, does not assume a traditional pronunciation lexicon, and generalizes well across speech difference severities. On an internal dataset collected from 32 people with dysarthria, this approach works regardless of severity and shows a 60% improvement in recall relative to a commercial speech recognition system. On the public EasyCall dataset of dysarthric speech, our approach improves accuracy by 30.5%. Performance degrades as the number of phrases increases, but consistently outperforms ASR systems when trained with 50 unique phrases.

Jessica Huynh, Alfredo Gomez, Athiya Deviyani et al.

Autoraters, also referred to as LLM-as-judges, are increasingly used for evaluation and automated content moderation. However, there is limited statistical analysis of how modifications in a rubric presented to both humans and autoraters affect their score agreement. Rubrics that ask for an overall or \emph{holistic} judgment - for example, rating the ``quality'' of an essay - may be inconsistently interpreted due to the complexity or subjectivity of the criteria. Conversely, rubrics can ask for \emph{analytic} judgments, which decompose assessment criteria - for example, ``quality'' into ``fluency'' and ``organization''. While these rubrics can be edited to improve the individual accuracy of both human and automated scoring, this approach may result in disagreement between the two scores, or with the associated holistic judgment. Designing and deploying reliable autoraters requires understanding not just the relationship between human and autorater annotations but how that relationship changes as holistic or analytic judgments are elicited. The results indicate that rubric edits providing representative examples and additional context, and reducing positional bias in the rubric increased human-autorater agreement, while higher rubric complexity and conservative aggregation methods tended to decrease it. The findings from the automatic essay scoring and instruction-following evaluation domains suggest that practitioners should carefully analyze domain- and rubric-specific performance to move towards higher human-autorater agreement.

Yi-Hao Peng, Samarth Das, Jeffrey P. Bigham et al.

Generative user interfaces (UIs) create new opportunities to adapt interfaces to individual users on demand, but personalization remains difficult because desirable UI properties are subjective, hard to articulate, and costly to infer from sparse feedback. We study this problem through a new dataset in which 20 trained designers each provide pairwise judgments over the same 600 generated UIs, enabling direct analysis of preference divergence. We find substantial disagreement across designers (average kappa = 0.25), and written rationales reveal that even when designers appeal to similar concepts such as hierarchy or cleanliness, designers differ in how they define, prioritize, and apply those concepts. Motivated by these findings, we develop a sample-efficient personalization method that represents a new user in terms of prior designers rather than a fixed rubric of design concepts. In a technical evaluation, our preference model outperforms both a pretrained UI evaluator and a larger multimodal model, and scales better with additional feedback. When used to personalize generation, it also produces interfaces preferred by 12 new designers over baseline approaches, including direct user prompting. Our findings suggest that lightweight preference elicitation can serve as a practical foundation for personalized generative UI systems.

Tom Gunter, Zirui Wang, Chong Wang et al.

We present foundation language models developed to power Apple Intelligence features, including a ~3 billion parameter model designed to run efficiently on devices and a large server-based language model designed for Private Cloud Compute. These models are designed to perform a wide range of tasks efficiently, accurately, and responsibly. This report describes the model architecture, the data used to train the model, the training process, how the models are optimized for inference, and the evaluation results. We highlight our focus on Responsible AI and how the principles are applied throughout the model development.

Yi-Hao Peng, Faria Huq, Yue Jiang et al.

Enabling machines to understand structured visuals like slides and user interfaces is essential for making them accessible to people with disabilities. However, achieving such understanding computationally has required manual data collection and annotation, which is time-consuming and labor-intensive. To overcome this challenge, we present a method to generate synthetic, structured visuals with target labels using code generation. Our method allows people to create datasets with built-in labels and train models with a small number of human-annotated examples. We demonstrate performance improvements in three tasks for understanding slides and UIs: recognizing visual elements, describing visual content, and classifying visual content types.

Atieh Taheri, Hamza El Alaoui, Patrick Carrington et al.

Ableist microaggressions remain pervasive in everyday interactions, yet interventions to help people recognize them are limited. We present an experiment testing how AI-mediated dialogue influences recognition of ableism. 160 participants completed a pre-test, intervention, and a post-test across four conditions: AI nudges toward bias (Bias-Directed), inclusion (Neutral-Directed), unguided dialogue (Self-Directed), and a text-only non-dialogue (Reading). Participants rated scenarios on standardness of social experience and emotional impact; those in dialogue-based conditions also provided qualitative reflections. Quantitative results showed dialogue-based conditions produced stronger recognition than Reading, though trajectories diverged: biased nudges improved differentiation of bias from neutrality but increased overall negativity. Inclusive or no nudges remained more balanced, while Reading participants showed weaker gains and even declines. Qualitative findings revealed biased nudges were often rejected, while inclusive nudges were adopted as scaffolding. We contribute a validated vignette corpus, an AI-mediated intervention platform, and design implications highlighting trade-offs conversational systems face when integrating bias-related nudges.

Faria Huq, Zora Zhiruo Wang, Zhanqiu Guo et al.

Despite rapid progress in autonomous web agents, human involvement remains essential for shaping preferences and correcting agent behavior as tasks unfold. However, current agentic systems lack a principled understanding of when and why humans intervene, often proceeding autonomously past critical decision points or requesting unnecessary confirmation. In this work, we introduce the task of modeling human intervention to support collaborative web task execution. We collect CowCorpus, a dataset of 400 real-user web navigation trajectories containing over 4,200 interleaved human and agent actions. We identify four distinct patterns of user interaction with agents -- hands-off supervision, hands-on oversight, collaborative task-solving, and full user takeover. Leveraging these insights, we train language models (LMs) to anticipate when users are likely to intervene based on their interaction styles, yielding a 61.4-63.4% improvement in intervention prediction accuracy over base LMs. Finally, we deploy these intervention-aware models in live web navigation agents and evaluate them in a user study, finding a 26.5% increase in user-rated agent usefulness. Together, our results show structured modeling of human intervention leads to more adaptive, collaborative agents.

Jason Wu, Eldon Schoop, Alan Leung et al.

Large language models (LLMs) struggle to consistently generate UI code that compiles and produces visually relevant designs. Existing approaches to improve generation rely on expensive human feedback or distilling a proprietary model. In this paper, we explore the use of automated feedback (compilers and multi-modal models) to guide LLMs to generate high-quality UI code. Our method starts with an existing LLM and iteratively produces improved models by self-generating a large synthetic dataset using an original model, applying automated tools to aggressively filter, score, and de-duplicate the data into a refined higher quality dataset. The original LLM is improved by finetuning on this refined dataset. We applied our approach to several open-source LLMs and compared the resulting performance to baseline models with both automated metrics and human preferences. Our evaluation shows the resulting models outperform all other downloadable baselines and approach the performance of larger proprietary models.

Jason Wu, Yi-Hao Peng, Amanda Li et al.

User interface (UI) design is a difficult yet important task for ensuring the usability, accessibility, and aesthetic qualities of applications. In our paper, we develop a machine-learned model, UIClip, for assessing the design quality and visual relevance of a UI given its screenshot and natural language description. To train UIClip, we used a combination of automated crawling, synthetic augmentation, and human ratings to construct a large-scale dataset of UIs, collated by description and ranked by design quality. Through training on the dataset, UIClip implicitly learns properties of good and bad designs by i) assigning a numerical score that represents a UI design's relevance and quality and ii) providing design suggestions. In an evaluation that compared the outputs of UIClip and other baselines to UIs rated by 12 human designers, we found that UIClip achieved the highest agreement with ground-truth rankings. Finally, we present three example applications that demonstrate how UIClip can facilitate downstream applications that rely on instantaneous assessment of UI design quality: i) UI code generation, ii) UI design tips generation, and iii) quality-aware UI example search.

Faria Huq, Zora Zhiruo Wang, Frank F. Xu et al. · cmu

While much work on web agents emphasizes the promise of autonomously performing tasks on behalf of users, in reality, agents often fall short on complex tasks in real-world contexts and modeling user preference. This presents an opportunity for humans to collaborate with the agent and leverage the agent's capabilities effectively. We propose CowPilot, a framework supporting autonomous as well as human-agent collaborative web navigation, and evaluation across task success and task efficiency. CowPilot reduces the number of steps humans need to perform by allowing agents to propose next steps, while users are able to pause, reject, or take alternative actions. During execution, users can interleave their actions with the agent by overriding suggestions or resuming agent control when needed. We conducted case studies on five common websites and found that the human-agent collaborative mode achieves the highest success rate of 95% while requiring humans to perform only 15.2% of the total steps. Even with human interventions during task execution, the agent successfully drives up to half of task success on its own. CowPilot can serve as a useful tool for data collection and agent evaluation across websites, which we believe will enable research in how users and agents can work together. Video demonstrations are available at https://oaishi.github.io/cowpilot.html

Kundan Krishna, Sanjana Ramprasad, Prakhar Gupta et al. · cmu

LLMs can generate factually incorrect statements even when provided access to reference documents. Such errors can be dangerous in high-stakes applications (e.g., document-grounded QA for healthcare or finance). We present GenAudit -- a tool intended to assist fact-checking LLM responses for document-grounded tasks. GenAudit suggests edits to the LLM response by revising or removing claims that are not supported by the reference document, and also presents evidence from the reference for facts that do appear to have support. We train models to execute these tasks, and design an interactive interface to present suggested edits and evidence to users. Comprehensive evaluation by human raters shows that GenAudit can detect errors in 8 different LLM outputs when summarizing documents from diverse domains. User studies demonstrate that using GenAudit can substantially improve the performance of humans at finding errors in LLM-generated summaries. We release our tool (GenAudit) and fact-checking model for public use.

Yasmine Kotturi, Angel Anderson, Glenn Ford et al.

Generative AI platforms and features are permeating many aspects of work. Entrepreneurs from lean economies in particular are well positioned to outsource tasks to generative AI given limited resources. In this paper, we work to address a growing disparity in use of these technologies by building on a four-year partnership with a local entrepreneurial hub dedicated to equity in tech and entrepreneurship. Together, we co-designed an interactive workshops series aimed to onboard local entrepreneurs to generative AI platforms. Alongside four community-driven and iterative workshops with entrepreneurs across five months, we conducted interviews with 15 local entrepreneurs and community providers. We detail the importance of communal and supportive exposure to generative AI tools for local entrepreneurs, scaffolding actionable use (and supporting non-use), demystifying generative AI technologies by emphasizing entrepreneurial power, while simultaneously deconstructing the veneer of simplicity to address the many operational skills needed for successful application.

Faria Huq, Jeffrey P. Bigham, Nikolas Martelaro

Large language models (LLMs) that have been trained on a corpus that includes large amount of code exhibit a remarkable ability to understand HTML code. As web interfaces are primarily constructed using HTML, we design an in-depth study to see how LLMs can be used to retrieve and locate important elements for a user given query (i.e. task description) in a web interface. In contrast with prior works, which primarily focused on autonomous web navigation, we decompose the problem as an even atomic operation - Can LLMs identify the important information in the web page for a user given query? This decomposition enables us to scrutinize the current capabilities of LLMs and uncover the opportunities and challenges they present. Our empirical experiments show that while LLMs exhibit a reasonable level of performance in retrieving important UI elements, there is still a substantial room for improvement. We hope our investigation will inspire follow-up works in overcoming the current challenges in this domain.

Fred Hohman, Chaoqun Wang, Jinmook Lee et al. · apple-ml, cmu

On-device machine learning (ML) moves computation from the cloud to personal devices, protecting user privacy and enabling intelligent user experiences. However, fitting models on devices with limited resources presents a major technical challenge: practitioners need to optimize models and balance hardware metrics such as model size, latency, and power. To help practitioners create efficient ML models, we designed and developed Talaria: a model visualization and optimization system. Talaria enables practitioners to compile models to hardware, interactively visualize model statistics, and simulate optimizations to test the impact on inference metrics. Since its internal deployment two years ago, we have evaluated Talaria using three methodologies: (1) a log analysis highlighting its growth of 800+ practitioners submitting 3,600+ models; (2) a usability survey with 26 users assessing the utility of 20 Talaria features; and (3) a qualitative interview with the 7 most active users about their experience using Talaria.

Yi-Hao Peng, Dingzeyu Li, Jeffrey P. Bigham et al.

User interface (UI) agents promise to make inaccessible or complex UIs easier to access for blind and low-vision (BLV) users. However, current UI agents typically perform tasks end-to-end without involving users in critical choices or making them aware of important contextual information, thus reducing user agency. For example, in our field study, a BLV participant asked to buy the cheapest available sparkling water, and the agent automatically chose one from several equally priced options, without mentioning alternative products with different flavors or better ratings. To address this problem, we introduce Morae, a UI agent that automatically identifies decision points during task execution and pauses so that users can make choices. Morae uses large multimodal models to interpret user queries alongside UI code and screenshots, and prompt users for clarification when there is a choice to be made. In a study over real-world web tasks with BLV participants, Morae helped users complete more tasks and select options that better matched their preferences, as compared to baseline agents, including OpenAI Operator. More broadly, this work exemplifies a mixed-initiative approach in which users benefit from the automation of UI agents while being able to express their preferences.

Yi-Hao Peng, Jeffrey P. Bigham, Jason Wu

Generative models, such as large language models and text-to-image diffusion models, are increasingly used to create visual designs like user interfaces (UIs) and presentation slides. Finetuning and benchmarking these generative models have often relied on datasets of human-annotated design preferences. Yet, due to the subjective and highly personalized nature of visual design, preference varies widely among individuals. In this paper, we study this problem by introducing DesignPref, a dataset of 12k pairwise comparisons of UI design generation annotated by 20 professional designers with multi-level preference ratings. We found that among trained designers, substantial levels of disagreement exist (Krippendorff's alpha = 0.25 for binary preferences). Natural language rationales provided by these designers indicate that disagreements stem from differing perceptions of various design aspect importance and individual preferences. With DesignPref, we demonstrate that traditional majority-voting methods for training aggregated judge models often do not accurately reflect individual preferences. To address this challenge, we investigate multiple personalization strategies, particularly fine-tuning or incorporating designer-specific annotations into RAG pipelines. Our results show that personalized models consistently outperform aggregated baseline models in predicting individual designers' preferences, even when using 20 times fewer examples. Our work provides the first dataset to study personalized visual design evaluation and support future research into modeling individual design taste.

Shruti Palaskar, Leon Gatys, Mona Abdelrahman et al.

Safety evaluation of multimodal foundation models often treats vision and language inputs separately, missing risks from joint interpretation where benign content becomes harmful in combination. Existing approaches also fail to distinguish clearly unsafe content from borderline cases, leading to problematic over-blocking or under-refusal of genuinely harmful content. We present Vision Language Safety Understanding (VLSU), a comprehensive framework to systematically evaluate multimodal safety through fine-grained severity classification and combinatorial analysis across 17 distinct safety patterns. Using a multi-stage pipeline with real-world images and human annotation, we construct a large-scale benchmark of 8,187 samples spanning 15 harm categories. Our evaluation of eleven state-of-the-art models reveals systematic joint understanding failures: while models achieve 90%-plus accuracy on clear unimodal safety signals, performance degrades substantially to 20-55% when joint image-text reasoning is required to determine the safety label. Most critically, 34% of errors in joint image-text safety classification occur despite correct classification of the individual modalities, further demonstrating absent compositional reasoning capabilities. Additionally, we find that models struggle to balance refusing unsafe content while still responding to borderline cases that deserve engagement. For example, we find that instruction framing can reduce the over-blocking rate on borderline content from 62.4% to 10.4% in Gemini-1.5, but only at the cost of under-refusing on unsafe content with refusal rate dropping from 90.8% to 53.9%. Overall, our framework exposes weaknesses in joint image-text understanding and alignment gaps in current models, and provides a critical test bed to enable the next milestones in research on robust vision-language safety.

Hua Shen, Tiffany Knearem, Reshmi Ghosh et al.

Recent advances in general-purpose AI underscore the urgent need to align AI systems with human goals and values. Yet, the lack of a clear, shared understanding of what constitutes "alignment" limits meaningful progress and cross-disciplinary collaboration. In this position paper, we argue that the research community should explicitly define and critically reflect on "alignment" to account for the bidirectional and dynamic relationship between humans and AI. Through a systematic review of over 400 papers spanning HCI, NLP, ML, and more, we examine how alignment is currently defined and operationalized. Building on this analysis, we introduce the Bidirectional Human-AI Alignment framework, which not only incorporates traditional efforts to align AI with human values but also introduces the critical, underexplored dimension of aligning humans with AI -- supporting cognitive, behavioral, and societal adaptation to rapidly advancing AI technologies. Our findings reveal significant gaps in current literature, especially in long-term interaction design, human value modeling, and mutual understanding. We conclude with three central challenges and actionable recommendations to guide future research toward more nuanced, reciprocal, and human-AI alignment approaches.

Kundan Krishna, Prakhar Gupta, Sanjana Ramprasad et al.

While the NLP community has produced numerous summarization benchmarks, none provide the rich annotations required to simultaneously address many important problems related to control and reliability. We introduce a Wikipedia-derived benchmark, complemented by a rich set of crowd-sourced annotations, that supports $8$ interrelated tasks: (i) extractive summarization; (ii) abstractive summarization; (iii) topic-based summarization; (iv) compressing selected sentences into a one-line summary; (v) surfacing evidence for a summary sentence; (vi) predicting the factual accuracy of a summary sentence; (vii) identifying unsubstantiated spans in a summary sentence; (viii) correcting factual errors in summaries. We compare various methods on this benchmark and discover that on multiple tasks, moderately-sized fine-tuned models consistently outperform much larger few-shot prompted language models. For factuality-related tasks, we also evaluate existing heuristics to create training data and find that training on them results in worse performance than training on $20\times$ less human-labeled data. Our articles draw from $6$ domains, facilitating cross-domain analysis. On some tasks, the amount of training data matters more than the domain where it comes from, while for other tasks training specifically on data from the target domain, even if limited, is more beneficial.

Colin Lea, Zifang Huang, Dhruv Jain et al.

Voice assistants have become an essential tool for people with various disabilities because they enable complex phone- or tablet-based interactions without the need for fine-grained motor control, such as with touchscreens. However, these systems are not tuned for the unique characteristics of individuals with speech disorders, including many of those who have a motor-speech disorder, are deaf or hard of hearing, have a severe stutter, or are minimally verbal. We introduce an alternative voice-based input system which relies on sound event detection using fifteen nonverbal mouth sounds like "pop," "click," or "eh." This system was designed to work regardless of ones' speech abilities and allows full access to existing technology. In this paper, we describe the design of a dataset, model considerations for real-world deployment, and efforts towards model personalization. Our fully-supervised model achieves segment-level precision and recall of 88.6% and 88.4% on an internal dataset of 710 adults, while achieving 0.31 false positives per hour on aggressors such as speech. Five-shot personalization enables satisfactory performance in 84.5% of cases where the generic model fails.

Jessica Huynh, Jeffrey Bigham, Maxine Eskenazi

Crowdsourcing requesters on Amazon Mechanical Turk (AMT) have raised questions about the reliability of the workers. The AMT workforce is very diverse and it is not possible to make blanket assumptions about them as a group. Some requesters now reject work en mass when they do not get the results they expect. This has the effect of giving each worker (good or bad) a lower Human Intelligence Task (HIT) approval score, which is unfair to the good workers. It also has the effect of giving the requester a bad reputation on the workers' forums. Some of the issues causing the mass rejections stem from the requesters not taking the time to create a well-formed task with complete instructions and/or not paying a fair wage. To explore this assumption, this paper describes a study that looks at the crowdsourcing HITs on AMT that were available over a given span of time and records information about those HITs. This study also records information from a crowdsourcing forum on the worker perspective on both those HITs and on their corresponding requesters. Results reveal issues in worker payment and presentation issues such as missing instructions or HITs that are not doable.

Kundan Krishna, Jeffrey Bigham, Zachary C. Lipton

Pretraining techniques leveraging enormous datasets have driven recent advances in text summarization. While folk explanations suggest that knowledge transfer accounts for pretraining's benefits, little is known about why it works or what makes a pretraining task or dataset suitable. In this paper, we challenge the knowledge transfer story, showing that pretraining on documents consisting of character n-grams selected at random, we can nearly match the performance of models pretrained on real corpora. This work holds the promise of eliminating upstream corpora, which may alleviate some concerns over offensive language, bias, and copyright issues. To see whether the small residual benefit of using real data could be accounted for by the structure of the pretraining task, we design several tasks motivated by a qualitative study of summarization corpora. However, these tasks confer no appreciable benefit, leaving open the possibility of a small role for knowledge transfer.

Prakhar Gupta, Yulia Tsvetkov, Jeffrey P. Bigham

Open-domain neural dialogue models have achieved high performance in response ranking and evaluation tasks. These tasks are formulated as a binary classification of responses given in a dialogue context, and models generally learn to make predictions based on context-response content similarity. However, over-reliance on content similarity makes the models less sensitive to the presence of inconsistencies, incorrect time expressions and other factors important for response appropriateness and coherence. We propose approaches for automatically creating adversarial negative training data to help ranking and evaluation models learn features beyond content similarity. We propose mask-and-fill and keyword-guided approaches that generate negative examples for training more robust dialogue systems. These generated adversarial responses have high content similarity with the contexts but are either incoherent, inappropriate or not fluent. Our approaches are fully data-driven and can be easily incorporated in existing models and datasets. Experiments on classification, ranking and evaluation tasks across multiple datasets demonstrate that our approaches outperform strong baselines in providing informative negative examples for training dialogue systems.

Jason Wu, Gabriel Reyes, Sam C. White et al.

Numerous accessibility features have been developed and included in consumer operating systems to provide people with a variety of disabilities additional ways to access computing devices. Unfortunately, many users, especially older adults who are more likely to experience ability changes, are not aware of these features or do not know which combination to use. In this paper, we first quantify this problem via a survey with 100 participants, demonstrating that very few people are aware of built-in accessibility features on their phones. These observations led us to investigate accessibility recommendation as a way to increase awareness and adoption. We developed four prototype recommenders that span different accessibility categories, which we used to collect insights from 20 older adults. Our work demonstrates the need to increase awareness of existing accessibility features on mobile devices, and shows that automated recommendation could help people find beneficial accessibility features.

Yi-Hao Peng, JiWoong Jang, Jeffrey P. Bigham et al.

Presenters commonly use slides as visual aids for informative talks. When presenters fail to verbally describe the content on their slides, blind and visually impaired audience members lose access to necessary content, making the presentation difficult to follow. Our analysis of 90 presentation videos revealed that 72% of 610 visual elements (e.g., images, text) were insufficiently described. To help presenters create accessible presentations, we introduce Presentation A11y, a system that provides real-time and post-presentation accessibility feedback. Our system analyzes visual elements on the slide and the transcript of the verbal presentation to provide element-level feedback on what visual content needs to be further described or even removed. Presenters using our system with their own slide-based presentations described more of the content on their slides, and identified 3.26 times more accessibility problems to fix after the talk than when using a traditional slide-based presentation interface. Integrating accessibility feedback into content creation tools will improve the accessibility of informational content for all.

Colin Lea, Vikramjit Mitra, Aparna Joshi et al.

The ability to automatically detect stuttering events in speech could help speech pathologists track an individual's fluency over time or help improve speech recognition systems for people with atypical speech patterns. Despite increasing interest in this area, existing public datasets are too small to build generalizable dysfluency detection systems and lack sufficient annotations. In this work, we introduce Stuttering Events in Podcasts (SEP-28k), a dataset containing over 28k clips labeled with five event types including blocks, prolongations, sound repetitions, word repetitions, and interjections. Audio comes from public podcasts largely consisting of people who stutter interviewing other people who stutter. We benchmark a set of acoustic models on SEP-28k and the public FluencyBank dataset and highlight how simply increasing the amount of training data improves relative detection performance by 28\% and 24\% F1 on each. Annotations from over 32k clips across both datasets will be publicly released.

Xiaoyi Zhang, Lilian de Greef, Amanda Swearngin et al.

Many accessibility features available on mobile platforms require applications (apps) to provide complete and accurate metadata describing user interface (UI) components. Unfortunately, many apps do not provide sufficient metadata for accessibility features to work as expected. In this paper, we explore inferring accessibility metadata for mobile apps from their pixels, as the visual interfaces often best reflect an app's full functionality. We trained a robust, fast, memory-efficient, on-device model to detect UI elements using a dataset of 77,637 screens (from 4,068 iPhone apps) that we collected and annotated. To further improve UI detections and add semantic information, we introduced heuristics (e.g., UI grouping and ordering) and additional models (e.g., recognize UI content, state, interactivity). We built Screen Recognition to generate accessibility metadata to augment iOS VoiceOver. In a study with 9 screen reader users, we validated that our approach improves the accessibility of existing mobile apps, enabling even previously inaccessible apps to be used.

Claudia Flores-Saviaga, Yuwen Li, Benjamin V. Hanrahan et al.

Crowd work has the potential of helping the financial recovery of regions traditionally plagued by a lack of economic opportunities, e.g., rural areas. However, we currently have limited information about the challenges facing crowd work-ers from rural and super rural areas as they struggle to make a living through crowd work sites. This paper examines the challenges and advantages of rural and super rural AmazonMechanical Turk (MTurk) crowd workers and contrasts them with those of workers from urban areas. Based on a survey of421 crowd workers from differing geographic regions in theU.S., we identified how across regions, people struggled with being onboarded into crowd work. We uncovered that despite the inequalities and barriers, rural workers tended to be striving more in micro-tasking than their urban counterparts. We also identified cultural traits, relating to time dimension and individualism, that offer us an insight into crowd workers and the necessary qualities for them to succeed on gig platforms. We finish by providing design implications based on our findings to create more inclusive crowd work platforms and tools

Jonathan Dinu, Jeffrey Bigham, J. Zico Kolter

As machine learning and algorithmic decision making systems are increasingly being leveraged in high-stakes human-in-the-loop settings, there is a pressing need to understand the rationale of their predictions. Researchers have responded to this need with explainable AI (XAI), but often proclaim interpretability axiomatically without evaluation. When these systems are evaluated, they are often tested through offline simulations with proxy metrics of interpretability (such as model complexity). We empirically evaluate the veracity of three common interpretability assumptions through a large scale human-subjects experiment with a simple "placebo explanation" control. We find that feature attribution explanations provide marginal utility in our task for a human decision maker and in certain cases result in worse decisions due to cognitive and contextual confounders. This result challenges the assumed universal benefit of applying these methods and we hope this work will underscore the importance of human evaluation in XAI research. Supplemental materials -- including anonymized data from the experiment, code to replicate the study, an interactive demo of the experiment, and the models used in the analysis -- can be found at: https://doi.pizza/challenging-xai.

Jaylin Herskovitz, Jason Wu, Samuel White et al.

Augmented Reality (AR) technology creates new immersive experiences in entertainment, games, education, retail, and social media. AR content is often primarily visual and it is challenging to enable access to it non-visually due to the mix of virtual and real-world content. In this paper, we identify common constituent tasks in AR by analyzing existing mobile AR applications for iOS, and characterize the design space of tasks that require accessible alternatives. For each of the major task categories, we create prototype accessible alternatives that we evaluate in a study with 10 blind participants to explore their perceptions of accessible AR. Our study demonstrates that these prototypes make AR possible to use for blind users and reveals a number of insights to move forward. We believe our work sets forth not only exemplars for developers to create accessible AR applications, but also a roadmap for future research to make AR comprehensively accessible.

Amy Pavel, Gabriel Reyes, Jeffrey P. Bigham

Audio descriptions make videos accessible to those who cannot see them by describing visual content in audio. Producing audio descriptions is challenging due to the synchronous nature of the audio description that must fit into gaps of other video content. An experienced audio description author will produce content that fits narration necessary to understand, enjoy, or experience the video content into the time available. This can be especially tricky for novices to do well. In this paper, we introduce a tool, Rescribe, that helps authors create and refine their audio descriptions. Using Rescribe, authors first create a draft of all the content they would like to include in the audio description. Rescribe then uses a dynamic programming approach to optimize between the length of the audio description, available automatic shortening approaches, and source track lengthening approaches. Authors can iteratively visualize and refine the audio descriptions produced by Rescribe, working in concert with the tool. We evaluate the effectiveness of Rescribe through interviews with blind and visually impaired audio description users who preferred Rescribe-edited descriptions to extended descriptions. In addition, we invite novice users to create audio descriptions with Rescribe and another tool, finding that users produce audio descriptions with fewer placement errors using Rescribe.

Prakhar Gupta, Jeffrey P. Bigham, Yulia Tsvetkov et al.

Dialogue systems pretrained with large language models generate locally coherent responses, but lack the fine-grained control over responses necessary to achieve specific goals. A promising method to control response generation is exemplar-based generation, in which models edit exemplar responses that are retrieved from training data, or hand-written to strategically address discourse-level goals, to fit new dialogue contexts. But, current exemplar-based approaches often excessively copy words from the exemplar responses, leading to incoherent replies. We present an Exemplar-based Dialogue Generation model, EDGE, that uses the semantic frames present in exemplar responses to guide generation. We show that controlling dialogue generation based on the semantic frames of exemplars, rather than words in the exemplar itself, improves the coherence of generated responses, while preserving semantic meaning and conversation goals present in exemplar responses.

Kundan Krishna, Amy Pavel, Benjamin Schloss et al.

Despite diverse efforts to mine various modalities of medical data, the conversations between physicians and patients at the time of care remain an untapped source of insights. In this paper, we leverage this data to extract structured information that might assist physicians with post-visit documentation in electronic health records, potentially lightening the clerical burden. In this exploratory study, we describe a new dataset consisting of conversation transcripts, post-visit summaries, corresponding supporting evidence (in the transcript), and structured labels. We focus on the tasks of recognizing relevant diagnoses and abnormalities in the review of organ systems (RoS). One methodological challenge is that the conversations are long (around 1500 words), making it difficult for modern deep-learning models to use them as input. To address this challenge, we extract noteworthy utterances---parts of the conversation likely to be cited as evidence supporting some summary sentence. We find that by first filtering for (predicted) noteworthy utterances, we can significantly boost predictive performance for recognizing both diagnoses and RoS abnormalities.

Kundan Krishna, Sopan Khosla, Jeffrey P. Bigham et al.

Following each patient visit, physicians draft long semi-structured clinical summaries called SOAP notes. While invaluable to clinicians and researchers, creating digital SOAP notes is burdensome, contributing to physician burnout. In this paper, we introduce the first complete pipelines to leverage deep summarization models to generate these notes based on transcripts of conversations between physicians and patients. After exploring a spectrum of methods across the extractive-abstractive spectrum, we propose Cluster2Sent, an algorithm that (i) extracts important utterances relevant to each summary section; (ii) clusters together related utterances; and then (iii) generates one summary sentence per cluster. Cluster2Sent outperforms its purely abstractive counterpart by 8 ROUGE-1 points, and produces significantly more factual and coherent sentences as assessed by expert human evaluators. For reproducibility, we demonstrate similar benefits on the publicly available AMI dataset. Our results speak to the benefits of structuring summaries into sections and annotating supporting evidence when constructing summarization corpora.

Prakhar Gupta, Shikib Mehri, Tiancheng Zhao et al.

The aim of this paper is to mitigate the shortcomings of automatic evaluation of open-domain dialog systems through multi-reference evaluation. Existing metrics have been shown to correlate poorly with human judgement, particularly in open-domain dialog. One alternative is to collect human annotations for evaluation, which can be expensive and time consuming. To demonstrate the effectiveness of multi-reference evaluation, we augment the test set of DailyDialog with multiple references. A series of experiments show that the use of multiple references results in improved correlation between several automatic metrics and human judgement for both the quality and the diversity of system output.

Luz Rello, Ricardo Baeza-Yates, Abdullah Ali et al.

Dyslexia is a specific learning disorder related to school failure. Detection is both crucial and challenging, especially in languages with transparent orthographies, such as Spanish. To make detecting dyslexia easier, we designed an online gamified test and a predictive machine learning model. In a study with more than 3,600 participants, our model correctly detected over 80% of the participants with dyslexia. To check the robustness of the method we tested our method using a new data set with over 1,300 participants with age customized tests in a different environment -- a tablet instead of a desktop computer -- reaching a recall of over 72% for the class with dyslexia for children 9 years old or older. Our work shows that dyslexia can be screened using a machine learning approach. An online screening tool based on our methods has already been used by more than 200,000 people.

Susumu Saito, Chun-Wei Chiang, Saiph Savage et al.

Workers in crowd markets struggle to earn a living. One reason for this is that it is difficult for workers to accurately gauge the hourly wages of microtasks, and they consequently end up performing labor with little pay. In general, workers are provided with little information about tasks, and are left to rely on noisy signals, such as textual description of the task or rating of the requester. This study explores various computational methods for predicting the working times (and thus hourly wages) required for tasks based on data collected from other workers completing crowd work. We provide the following contributions. (i) A data collection method for gathering real-world training data on crowd-work tasks and the times required for workers to complete them; (ii) TurkScanner: a machine learning approach that predicts the necessary working time to complete a task (and can thus implicitly provide the expected hourly wage). We collected 9,155 data records using a web browser extension installed by 84 Amazon Mechanical Turk workers, and explored the challenge of accurately recording working times both automatically and by asking workers. TurkScanner was created using ~150 derived features, and was able to predict the hourly wages of 69.6% of all the tested microtasks within a 75% error. Directions for future research include observing the effects of tools on people's working practices, adapting this approach to a requester tool for better price setting, and predicting other elements of work (e.g., the acceptance likelihood and worker task preferences.)

Danna Gurari, Qing Li, Abigale J. Stangl et al.

The study of algorithms to automatically answer visual questions currently is motivated by visual question answering (VQA) datasets constructed in artificial VQA settings. We propose VizWiz, the first goal-oriented VQA dataset arising from a natural VQA setting. VizWiz consists of over 31,000 visual questions originating from blind people who each took a picture using a mobile phone and recorded a spoken question about it, together with 10 crowdsourced answers per visual question. VizWiz differs from the many existing VQA datasets because (1) images are captured by blind photographers and so are often poor quality, (2) questions are spoken and so are more conversational, and (3) often visual questions cannot be answered. Evaluation of modern algorithms for answering visual questions and deciding if a visual question is answerable reveals that VizWiz is a challenging dataset. We introduce this dataset to encourage a larger community to develop more generalized algorithms that can assist blind people.

Kotaro Hara, Abi Adams, Kristy Milland et al.

A growing number of people are working as part of on-line crowd work, which has been characterized by its low wages; yet, we know little about wage distribution and causes of low/high earnings. We recorded 2,676 workers performing 3.8 million tasks on Amazon Mechanical Turk. Our task-level analysis revealed that workers earned a median hourly wage of only ~\$2/h, and only 4% earned more than \$7.25/h. The average requester pays more than \$11/h, although lower-paying requesters post much more work. Our wage calculations are influenced by how unpaid work is included in our wage calculations, e.g., time spent searching for tasks, working on tasks that are rejected, and working on tasks that are ultimately not submitted. We further explore the characteristics of tasks and working patterns that yield higher hourly wages. Our analysis informs future platform design and worker tools to create a more positive future for crowd work.

Ting-Hao Kenneth Huang, Walter S. Lasecki, Amos Azaria et al.

Intelligent conversational assistants, such as Apple's Siri, Microsoft's Cortana, and Amazon's Echo, have quickly become a part of our digital life. However, these assistants have major limitations, which prevents users from conversing with them as they would with human dialog partners. This limits our ability to observe how users really want to interact with the underlying system. To address this problem, we developed a crowd-powered conversational assistant, Chorus, and deployed it to see how users and workers would interact together when mediated by the system. Chorus sophisticatedly converses with end users over time by recruiting workers on demand, which in turn decide what might be the best response for each user sentence. Up to the first month of our deployment, 59 users have held conversations with Chorus during 320 conversational sessions. In this paper, we present an account of Chorus' deployment, with a focus on four challenges: (i) identifying when conversations are over, (ii) malicious users and workers, (iii) on-demand recruiting, and (iv) settings in which consensus is not enough. Our observations could assist the deployment of crowd-powered conversation systems and crowd-powered systems in general.

Ting-Hao 'Kenneth' Huang, Yun-Nung Chen, Jeffrey P. Bigham

Output-agreement mechanisms such as ESP Game have been widely used in human computation to obtain reliable human-generated labels. In this paper, we argue that a "time-limited" output-agreement mechanism can be used to create a fast and robust crowd-powered component in interactive systems, particularly dialogue systems, to extract key information from user utterances on the fly. Our experiments on Amazon Mechanical Turk using the Airline Travel Information System (ATIS) dataset showed that the proposed approach achieves high-quality results with an average response time shorter than 9 seconds.

Michael Nebeling, Anhong Guo, Kyle Murray et al.

In this paper we introduce a paradigm for completing complex tasks from wearable devices by leveraging crowdsourcing, and demonstrate its validity for academic writing. We explore this paradigm using a collaborative authoring system, called WearWrite, which is designed to enable authors and crowd workers to work together using an Android smartwatch and Google Docs to produce academic papers, including this one. WearWrite allows expert authors who do not have access to large devices to contribute bits of expertise and big picture direction from their watch, while freeing them of the obligation of integrating their contributions into the overall document. Crowd workers on desktop computers actually write the document. We used this approach to write several simple papers, and found it was effective at producing reasonable drafts. However, the workers often needed more structure and the authors more context. WearWrite addresses these issues by focusing workers on specific tasks and providing select context to authors on the watch. We demonstrate the system's feasibility by writing this paper using it.

Walter S. Lasecki, Christopher M. Homan, Jeffrey P. Bigham

Crowdsourcing can solve problems that current fully automated systems cannot. Its effectiveness depends on the reliability, accuracy, and speed of the crowd workers that drive it. These objectives are frequently at odds with one another. For instance, how much time should workers be given to discover and propose new solutions versus deliberate over those currently proposed? How do we determine if discovering a new answer is appropriate at all? And how do we manage workers who lack the expertise or attention needed to provide useful input to a given task? We present a mechanism that uses distinct payoffs for three possible worker actions---propose,vote, or abstain---to provide workers with the necessary incentives to guarantee an effective (or even optimal) balance between searching for new answers, assessing those currently available, and, when they have insufficient expertise or insight for the task at hand, abstaining. We provide a novel game theoretic analysis for this mechanism and test it experimentally on an image---labeling problem and show that it allows a system to reliably control the balance betweendiscovering new answers and converging to existing ones.

Walter S. Lasecki, Samuel C. White, Kyle I. Murray et al.

Crowd algorithms often assume workers are inexperienced and thus fail to adapt as workers in the crowd learn a task. These assumptions fundamentally limit the types of tasks that systems based on such algorithms can handle. This paper explores how the crowd learns and remembers over time in the context of human computation, and how more realistic assumptions of worker experience may be used when designing new systems. We first demonstrate that the crowd can recall information over time and discuss possible implications of crowd memory in the design of crowd algorithms. We then explore crowd learning during a continuous control task. Recent systems are able to disguise dynamic groups of workers as crowd agents to support continuous tasks, but have not yet considered how such agents are able to learn over time. We show, using a real-time gaming setting, that crowd agents can learn over time, and `remember' by passing strategies from one generation of workers to the next, despite high turnover rates in the workers comprising them. We conclude with a discussion of future research directions for crowd memory and learning.