Natural language database interfaces broaden data access, yet they remain brittle under input ambiguity. Standard approaches often collapse uncertainty into a single query, offering little support for mismatches between user intent and system interpretation. We reframe this challenge through pragmatic inference: while users economize expressions, systems operate on priors over the action space that may not align with the users’. In this view, pragmatic repair—incremental clarification through minimal interaction—is a natural strategy for resolving underspecification. We present PleaSQLarify, which operationalizes pragmatic repair by structuring interaction around interpretable decision variables that enable efficient clarification1. A visual interface2 complements this by surfacing the action space for exploration, requesting user disambiguation, and making belief updates traceable across turns. In a study with twelve participants, PleaSQLarify helped users recognize alternative interpretations and efficiently resolve ambiguity. Our findings highlight pragmatic repair as a design principle that fosters effective user control in natural language interfaces.

Abstract Large language models (LLMs) represent words through contextual word embeddings encoding different language properties like semantics and syntax. Understanding these properties is crucial, especially for researchers investigating language model capabilities, employing embeddings for tasks related to text similarity, or evaluating the reasons behind token importance as measured through attribution methods. Applications for embedding exploration frequently involve dimensionality reduction techniques, which reduce high‐dimensional vectors to two dimensions used as coordinates in a scatterplot. This data transformation step introduces uncertainty that can be propagated to the visual representation and influence users' interpretation of the data. To communicate such uncertainties, we present LayerFlow – a visual analytics workspace that displays embeddings in an interlinked projection design and communicates the transformation, representation, and interpretation uncertainty. In particular, to hint at potential data distortions and uncertainties, the workspace includes several visual components, such as convex hulls showing 2D and HD clusters, data point pairwise distances, cluster summaries, and projection quality metrics. We show the usability of the presented workspace through replication and expert case studies that highlight the need to communicate uncertainty through multiple visual components and different data perspectives.

Treemaps are a powerful tool for representing hierarchical data in a space-efficient manner and are used in various domains, including network security or software development. However, interpreting the topology encoded by nested rectangles can be challenging, particularly compared to tree-structured representations like node-link diagrams or icicle plots. To address this challenge, we introduce TreEducation, a visual education platform designed to improve the visualization literacy skills required for reading treemaps among non-expert users. TreEducation is an online application that combines visualizations, interactions, and gamification elements to facilitate understanding of eight different treemap layout algorithms and enhance students' learning process. We evaluated TreEducation in a classroom setting and a controlled environment. Our results indicate a significant knowledge gain of students training exclusively with TreEducation and the usefulness of competition as a social gamification element included in our competitive quiz.

Humans struggle to perceive and interpret high-dimensional data. Therefore, high-dimensional data are often projected into two dimensions for visualization. Many applications benefit from complex nonlinear dimensionality reduction techniques, but the effects of individual high-dimensional features are hard to explain in the two-dimensional space. Most visualization solutions use multiple two-dimensional plots, each showing the effect of one high-dimensional feature in two dimensions; this approach creates a need for a visual inspection of k plots for a k-dimensional input space. Our solution, Feature Clock, provides a novel approach that reduces the need to inspect these k plots to grasp the influence of original features on the data structure depicted in two dimensions. Feature Clock enhances the explainability and compactness of visualizations of embedded data and is available in an open-source Python library1.

To study the role of perceptual cues on categorization and decision making, participants are typically tested in (perception) experiments with a fixed set of randomized or pseudo-randomized trials. In linguistics and psycholinguistics, for instance, studies often investigate the relative weighting of different cues for a linguistic contrast (e.g., intonation vs. word order). For categorization beyond the segmental level (e.g., /p/ vs. /b/), it is important to establish that results generalise to different words or sentences, which necessitates the use of a range of different items. This may limit the number of conditions (cues and cue combinations) that can be sensibly tested in the same experiment. We show that Active Learning (AL) systems provide a solution: Since stimulus selection is informed by the system's learning mechanism (presenting obvious conditions less often than uncertain conditions), they allow for efficient testing of numerous conditions and different items in the same experiment. In this paper, we compared two weighting approaches (probability-based vs. regression-based) to model the outcome of three simulated scenarios with three binary factors each. Results show that valid results (i.e., little error between predicted values and the actual responses at the end of the experiment) are obtained after about half of the trials of an original psycholinguistic experiment we replicated. For simulations with interactions between factors, the regression-based approach performed better. Our findings bear implications for the application of AL in psycholinguistic research (extraction of cue weights, inferential statistics, and a stopping criterion during an on-going experiment), which we will discuss.

Large language models (LLMs) are widely deployed in various downstream tasks, e.g., auto-completion, aided writing, or chat-based text generation. However, the considered output candidates of the underlying search algorithm are under-explored and under-explained. We tackle this shortcoming by proposing a tree-in-the-loop approach, where a visual representation of the beam search tree is the central component for analyzing, explaining, and adapting the generated outputs. To support these tasks, we present generAItor, a visual analytics technique, augmenting the central beam search tree with various task-specific widgets, providing targeted visualizations and interaction possibilities. Our approach allows interactions on multiple levels and offers an iterative pipeline that encompasses generating, exploring, and comparing output candidates, as well as fine-tuning the model based on adapted data. Our case study shows that our tool generates new insights in gender bias analysis beyond state-of-the-art template-based methods. Additionally, we demonstrate the applicability of our approach in a qualitative user study. Finally, we quantitatively evaluate the adaptability of the model to few samples, as occurring in text-generation use cases.

Language models are widely used for different Natural Language Processing tasks while suffering from a lack of personalization. Personalization can be achieved by, e.g., fine-tuning the model on training data that is created by the user (e.g., social media posts). Previous work shows that the acquisition of such data can be challenging. Instead of adapting the model's parameters, we thus suggest selecting a model that matches the user's mental model of different thematic concepts in language. In this article, we attempt to capture such individual language understanding of users. In this process, two challenges have to be considered. First, we need to counteract disengagement since the task of communicating one's language understanding typically encompasses repetitive and time-consuming actions. Second, we need to enable users to externalize their mental models in different contexts, considering that language use changes depending on the environment. In this article, we integrate methods of gamification into a visual analytics (VA) workflow to engage users in sharing their knowledge within various contexts. In particular, we contribute the design of a gameful VA playground called Concept Universe. During the four-phased game, the users build personalized concept descriptions by explaining given concept names through representative keywords. Based on their performance, the system reacts with constant visual, verbal, and auditory feedback. We evaluate the system in a user study with six participants, showing that users are engaged and provide more specific input when facing a virtual opponent. We use the generated concepts to make personalized language model suggestions.

Neural language models are widely used; however, their model parameters often need to be adapted to the specific domains and tasks of an application, which is time- and resource-consuming. Thus, adapters have recently been introduced as a lightweight alternative for model adaptation. They consist of a small set of task-specific parameters with a reduced training time and simple parameter composition. The simplicity of adapter training and composition comes along with new challenges, such as maintaining an overview of adapter properties and effectively comparing their produced embedding spaces. To help developers overcome these challenges, we provide a twofold contribution. First, in close collaboration with NLP researchers, we conducted a requirement analysis for an approach supporting adapter evaluation and detected, among others, the need for both intrinsic (i.e., embedding similarity-based) and extrinsic (i.e., prediction-based) explanation methods. Second, motivated by the gathered requirements, we designed a flexible visual analytics workspace that enables the comparison of adapter properties. In this paper, we discuss several design iterations and alternatives for interactive, comparative visual explanation methods. Our comparative visualizations show the differences in the adapted embedding vectors and prediction outcomes for diverse human-interpretable concepts (e.g., person names, human qualities). We evaluate our workspace through case studies and show that, for instance, an adapter trained on the language debiasing task according to context-0 (decontextualized) embeddings introduces a new type of bias where words (even gender-independent words such as countries) become more similar to female- than male pronouns. We demonstrate that these are artifacts of context-0 embeddings, and the adapter effectively eliminates the gender information from the contextualized word representations.

With the success of contextualized language models, much research explores what these models really learn and in which cases they still fail. Most of this work focuses on specific NLP tasks and on the learning outcome. Little research has attempted to decouple the models’ weaknesses from specific tasks and focus on the embeddings per se and their mode of learning. In this paper, we take up this research opportunity: based on theoretical linguistic insights, we explore whether the semantic constraints of function words are learned and how the surrounding context impacts their embeddings. We create suitable datasets, provide new insights into the inner workings of LMs vis-a-vis function words and implement an assisting visual web interface for qualitative analysis.

Abstract Language models, such as BERT, construct multiple, contextualized embeddings for each word occurrence in a corpus. Understanding how the contextualization propagates through the model's layers is crucial for deciding which layers to use for a specific analysis task. Currently, most embedding spaces are explained by probing classifiers; however, some findings remain inconclusive. In this paper, we present LMFingerprints, a novel scoring‐based technique for the explanation of contextualized word embeddings. We introduce two categories of scoring functions, which measure (1) the degree of contextualization, i.e., the layerwise changes in the embedding vectors, and (2) the type of contextualization, i.e., the captured context information. We integrate these scores into an interactive explanation workspace. By combining visual and verbal elements, we provide an overview of contextualization in six popular transformer‐based language models. We evaluate hypotheses from the domain of computational linguistics, and our results not only confirm findings from related work but also reveal new aspects about the information captured in the embedding spaces. For instance, we show that while numbers are poorly contextualized, stopwords have an unexpected high contextualization in the models' upper layers, where their neighborhoods shift from similar functionality tokens to tokens that contribute to the meaning of the surrounding sentences.

We present VisInReport, a visual analytics tool that supports the manual analysis of discourse transcripts and generates reports based on user interaction. As an integral part of scholarly work in the social sciences and humanities, discourse analysis involves an aggregation of characteristics identified in the text, which, in turn, involves a prior identification of regions of particular interest. Manual data evaluation requires extensive effort, which can be a barrier to effective analysis. Our system addresses this challenge by augmenting the users' analysis with a set of automatically generated visualization layers. These layers enable the detection and exploration of relevant parts of the discussion supporting several tasks, such as topic modeling or question categorization. The system summarizes the extracted events visually and verbally, generating a content-rich insight into the data and the analysis process. During each analysis session, VisInReport builds a shareable report containing a curated selection of interactions and annotations generated by the analyst. We evaluate our approach on real-world datasets through a qualitative study with domain experts from political science, computer science, and linguistics. The results highlight the benefit of integrating the analysis and reporting processes through a visual analytics system, which supports the communication of results among collaborating researchers.

Rita Sevastjanova, Aikaterini-Lida Kalouli, Christin Beck, Hanna Schäfer, Mennatallah El-Assady. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Abstract Communication consists of both meta‐information as well as content. Currently, the automated analysis of such data often focuses either on the network aspects via social network analysis or on the content, utilizing methods from text‐mining. However, the first category of approaches does not leverage the rich content information, while the latter ignores the conversation environment and the temporal evolution, as evident in the meta‐information. In contradiction to communication research, which stresses the importance of a holistic approach, both aspects are rarely applied simultaneously, and consequently, their combination has not yet received enough attention in automated analysis systems. In this work, we aim to address this challenge by discussing the difficulties and design decisions of such a path as well as contribute CommAID, a blueprint for a holistic strategy to communication analysis. It features an integrated visual analytics design to analyze communication networks through dynamics modeling, semantic pattern retrieval, and a user‐adaptable and problem‐specific machine learning‐based retrieval system. An interactive multi‐level matrix‐based visualization facilitates a focused analysis of both network and content using inline visuals supporting cross‐checks and reducing context switches. We evaluate our approach in both a case study and through formative evaluation with eight law enforcement experts using a real‐world communication corpus. Results show that our solution surpasses existing techniques in terms of integration level and applicability. With this contribution, we aim to pave the path for a more holistic approach to communication analysis.

Linguistic insight in the form of high-level relationships and rules in text builds the basis of our understanding of language. However, the data-driven generation of such structures often lacks labeled resources that can be used as training data for supervised machine learning. The creation of such ground-truth data is a time-consuming process that often requires domain expertise to resolve text ambiguities and characterize linguistic phenomena. Furthermore, the creation and refinement of machine learning models is often challenging for linguists as the models are often complex, in-transparent, and difficult to understand. To tackle these challenges, we present a visual analytics technique for interactive data labeling that applies concepts from gamification and explainable Artificial Intelligence (XAI) to support complex classification tasks. The visual-interactive labeling interface promotes the creation of effective training data. Visual explanations of learned rules unveil the decisions of the machine learning model and support iterative and interactive optimization. The gamification-inspired design guides the user through the labeling process and provides feedback on the model performance. As an instance of the proposed technique, we present QuestionComb , a workspace tailored to the task of question classification (i.e., in information-seeking vs. non-information-seeking questions). Our evaluation studies confirm that gamification concepts are beneficial to engage users through continuous feedback, offering an effective visual analytics technique when combined with active learning and XAI.

Abstract We present ThreadReconstructor, a visual analytics approach for detecting and analyzing the implicit conversational structure of discussions, e.g., in political debates and forums. Our work is motivated by the need to reveal and understand single threads in massive online conversations and verbatim text transcripts. We combine supervised and unsupervised machine learning models to generate a basic structure that is enriched by user‐defined queries and rule‐based heuristics. Depending on the data and tasks, users can modify and create various reconstruction models that are presented and compared in the visualization interface. Our tool enables the exploration of the generated threaded structures and the analysis of the untangled reply‐chains, comparing different models and their agreement. To understand the inner‐workings of the models, we visualize their decision spaces, including all considered candidate relations. In addition to a quantitative evaluation, we report qualitative feedback from an expert user study with four forum moderators and one machine learning expert, showing the effectiveness of our approach.

Topic modeling algorithms are widely used to analyze the thematic composition of text corpora but remain difficult to interpret and adjust. Addressing these limitations, we present a modular visual analytics framework, tackling the understandability and adaptability of topic models through a user-driven reinforcement learning process which does not require a deep understanding of the underlying topic modeling algorithms. Given a document corpus, our approach initializes two algorithm configurations based on a parameter space analysis that enhances document separability. We abstract the model complexity in an interactive visual workspace for exploring the automatic matching results of two models, investigating topic summaries, analyzing parameter distributions, and reviewing documents. The main contribution of our work is an iterative decision-making technique in which users provide a document-based relevance feedback that allows the framework to converge to a user-endorsed topic distribution. We also report feedback from a two-stage study which shows that our technique results in topic model quality improvements on two independent measures.

Abstract We present NEREx, an interactive visual analytics approach for the exploratory analysis of verbatim conversational transcripts. By revealing different perspectives on multi‐party conversations, NEREx gives an entry point for the analysis through high‐level overviews and provides mechanisms to form and verify hypotheses through linked detail‐views. Using a tailored named‐entity extraction, we abstract important entities into ten categories and extract their relations with a distance‐restricted entity‐relationship model. This model complies with the often ungrammatical structure of verbatim transcripts, relating two entities if they are present in the same sentence within a small distance window. Our tool enables the exploratory analysis of multi‐party conversations using several linked views that reveal thematic and temporal structures in the text. In addition to distant‐reading, we integrated close‐reading views for a text‐level investigation process. Beyond the exploratory and temporal analysis of conversations, NEREx helps users generate and validate hypotheses and perform comparative analyses of multiple conversations. We demonstrate the applicability of our approach on real‐world data from the 2016 U.S. Presidential Debates through a qualitative study with three domain experts from political science.