Attribution scores indicate the importance of different input parts and can, thus, explain model behaviour. Currently, prompt-based models are gaining popularity, i.a., due to their easier adaptability in low-resource settings. However, the quality of attribution scores extracted from prompt-based models has not been investigated yet. In this work, we address this topic by analyzing attribution scores extracted from prompt-based models w.r.t. plausibility and faithfulness and comparing them with attribution scores extracted from fine-tuned models and large language models. In contrast to previous work, we introduce training size as another dimension into the analysis. We find that using the prompting paradigm (with either encoder-based or decoder-based models) yields more plausible explanations than fine-tuning the models in low-resource settings and Shapley Value Sampling consistently outperforms attention and Integrated Gradients in terms of leading to more plausible and faithful explanations.

Annotators' sociodemographic backgrounds (i.e., the individual compositions of their gender, age, educational background, etc.) have a strong impact on their decisions when working on subjective NLP tasks, such as toxic language detection.Often, heterogeneous backgrounds result in high disagreements.To model this variation, recent work has explored sociodemographic prompting, a technique, which steers the output of prompt-based models towards answers that humans with specific sociodemographic profiles would give.However, the available NLP literature disagrees on the efficacy of this technique -it remains unclear for which tasks and scenarios it can help, and the role of the individual factors in sociodemographic prompting is still unexplored.We address this research gap by presenting the largest and most comprehensive study of sociodemographic prompting today.We analyze its influence on model sensitivity, performance and robustness across seven datasets and six instruction-tuned model families.We show that sociodemographic information affects model predictions and can be beneficial for improving zero-shot learning in subjective NLP tasks.However, its outcomes largely vary for different model types, sizes, and datasets, and are subject to large variance with regards to prompt formulations.Most importantly, our results show that sociodemographic prompting should be used with care for sensitive applications, such as toxicity annotation or when studying LLM alignment.1

Sheng Lu, Hendrik Schuff, Iryna Gurevych. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Abstract Many research topics in natural language processing (NLP), such as explanation generation, dialog modeling, or machine translation, require evaluation that goes beyond standard metrics like accuracy or F1score toward a more human-centered approach. Therefore, understanding how to design user studies becomes increasingly important. However, few comprehensive resources exist on planning, conducting, and evaluating user studies for NLP, making it hard to get started for researchers without prior experience in the field of human evaluation. In this paper, we summarize the most important aspects of user studies and their design and evaluation, providing direct links to NLP tasks and NLP-specific challenges where appropriate. We (i) outline general study design, ethical considerations, and factors to consider for crowdsourcing, (ii) discuss the particularities of user studies in NLP, and provide starting points to select questionnaires, experimental designs, and evaluation methods that are tailored to the specific NLP tasks. Additionally, we offer examples with accompanying statistical evaluation code, to bridge the gap between theoretical guidelines and practical applications.

Explainable question answering systems predict an answer together with an explanation showing why the answer has been selected. The goal is to enable users to assess the correctness of the system and understand its reasoning process. However, we show that current models and evaluation settings have shortcomings regarding the coupling of answer and explanation which might cause serious issues in user experience. As a remedy, we propose a hierarchical model and a new regularization term to strengthen the answer-explanation coupling as well as two evaluation scores to quantify the coupling. We conduct experiments on the HOTPOTQA benchmark data set and perform a user study. The user study shows that our models increase the ability of the users to judge the correctness of the system and that scores like F1 are not enough to estimate the usefulness of a model in a practical setting with human users. Our scores are better aligned with user experience, making them promising candidates for model selection.

Head-mounted augmented reality (AR) enables embodied in situ drawing in three dimensions (3D). We explore 3D drawing interactions based on uninstrumented, unencumbered (bare) hands that preserve the user's ability to freely navigate and interact with the physical environment. We derive three alternative interaction techniques supporting bare-handed drawing in AR from the literature and by analysing several envisaged use cases. The three interaction techniques are evaluated in a controlled user study examining three distinct drawing tasks: planar drawing, path description, and 3D object reconstruction. The results indicate that continuous freehand drawing supports faster line creation than the control point based alternatives, although with reduced accuracy. User preferences for the different techniques are mixed and vary considerably between the different tasks, highlighting the value of diverse and flexible interactions. The combined effectiveness of these three drawing techniques is illustrated in an example application of 3D AR drawing.

Mastering fine motor tasks, such as playing the guitar, takes years of time-consuming practice. Commonly, expensive guidance by experts is essential for adjusting the training program to the student's proficiency. In our work, we showcase the suitability of Electromyography to detect fine-grained hand and finger postures in an exemplary guitar tutor scenario. We present EMGuitar, an interactive guitar tutoring system, that assists students by reporting on play correctness and adjusts playback tempi automatically. We report person-dependent classification utilizing a ring of electrodes around the forearm with an F1 score of up to 0.89 on recorded calibration data. Furthermore, our system was received well by neither diminishing ease of use nor being disruptive for the participants. Based on the received comments, we identified the need for detailed play accuracy feedback down to individual chords, for which we suggest an adapted visualization and an algorithmic approach.

Interactive systems, such as online search interfaces, require appropriate input if they are to produce accurate information. Without this, they can be inaccurate if the user is uncertain about the keywords. Current systems do not have the means to detect uncertainty, which may lead to a negative user experience. We explore physiological and behavioral measurements as tools to implicitly detect users' uncertainty, and provide a method to integrate input variability in interactive systems. We conducted a laboratory study where participants answered questions of varying difficulty, recording physiological data via a key logger, an eye tracker, and a heart rate sensor. Our results show that participants spent significantly more time on difficult questions and looked longer at their answers before submitting them. Based on our results, we provide initial insights on how data from physiological sensors and logged user behavior can be utilized to enrich interactive systems and evaluate a user's uncertainty level.

Although more and more data is collected automatically, many interfaces still require manual input. When we, for example, enter our daily calorie intake or calculate our ecological footprint, we often have to guess the weight of the food or what distance we have covered with our car. In this paper, we propose a solution to overcome the problem of forcing users to enter a single value when they are unsure about the actual input. On the basis of a slider, we designed four input controls which allow the input of uncertain data in the form of probability distribution functions. To evaluate our input controls, we conducted two studies collecting subjective and objective feedback. Based on the evaluation, we derived implications for their usage. We additionally provide an open-source toolkit with the evaluated input controls that can be included in web applications and customized for different contexts and tasks.