Understanding how texts are produced is crucial not only for the development of theoretical models and writing strategies, but also for practical applications. However, the writing process itself--- including intermediate versions, copy-paste actions, input from co-authors or LLMs---remains invisible in the final text. This study addresses this gap by visualizing fine-grained keystroke logging data to capture both the product (final text) and the process (writer's actions) at sentence and text level. We design and implement custom JavaScript visualizations of linguistically processed keystroke logging data. Our pilot study examines data from nine students writing under identical conditions; we analyze temporal, spatial, and structural aspects of writing. The results reveal diverse, nonlinear writing strategies and suggest that individualized process visualizations can inform both document engineering and writing analytics. The novel visualization types we present demonstrate how process and product can be meaningfully integrated.
Non-robustness of image classifiers to subtle, adversarial perturbations is a well-known failure mode. Defenses against such attacks are typically evaluated by measuring the error rate on perturbed versions of the natural test set, quantifying the worst-case performance within a specified perturbation budget. However, these evaluations often isolate specific perturbation types, underestimating the adaptability of real-world adversaries who can modify or compose attacks in unforeseen ways. In this work, we show that models considered robust to strong attacks, such as AutoAttack, can be compromised by a simple modification of the weaker FGSM attack, where the adversarial perturbation is slightly transformed prior to being added to the input. Despite the attack's simplicity, robust models that perform well against standard FGSM become vulnerable to this variant. These findings suggest that current defenses may generalize poorly beyond their assumed threat models and can achieve inflated robustness scores under narrowly defined evaluation settings.
The perception of the readability of biomedical texts varies depending on the reader's profile, a disparity further amplified by the intrinsic complexity of these documents and the unequal distribution of health literacy within the population. Although 72% of Internet users consult medical information online, a significant proportion have difficulty understanding it. To ensure that texts are accessible to a diverse audience, it is essential to assess readability. However, conventional readability formulas, designed for general texts, do not take this diversity into account, underlining the need to adapt evaluation tools to the specific needs of biomedical texts and the heterogeneity of readers. To address this gap, we propose a novel readability assessment method tailored to three distinct audiences: expert adults, non-expert adults, and children. Our approach is built upon a structured, bilingual biomedical corpus of 20,008 documents (8,854 in French, 11,154 in English), compiled from multiple sources to ensure diversity in both content and audience. Specifically, the French corpus combines texts from Cochrane and Wikipedia/Vikidia, both of which are subsets of the CLEAR corpus, while the English corpus merges documents from the Cochrane Library, Plaba, and Science Journal for Kids. For each original expert-level text, domain specialists produced simplified variants calibrated specifically to the comprehension abilities of non-expert adults or children. Every document is therefore explicitly labeled by its target audience. Leveraging this resource, we trained a diverse suite of classifiers, from classical approaches (e.g., XGBoost, SVM) to classifiers built upon language models (e.g., BERT, CamemBERT, BioBERT, DrBERT). We then designed a hybrid architecture "BioReadNet" that integrates transformer embeddings with expert-driven linguistic features, achieving a macro-averaged F1 score of 0.987.
Image binarization is fundamental for document image processing. The performance of binarization algorithms depends on several factors that range from the quality of the digitalization devices to the intrinsic features of the document itself and the kind and intensity of the noises present in the image. This assessment on binarizing photographed documents evaluated the quality, time, space, and performance of five new algorithms and ninety-eight "classical" algorithms. The test data set is composed of laser and deskjet printed documents, photographed using six widely used mobile devices with the strobe flash on, off, and in auto modes under three different angles and places of capture.
This demo introduces Doenba Edit, a user-friendly, AI-powered platform developed by Librum Technologies, Inc., designed for seamless editing and typesetting of academic writing within a word processor-like interface. It supports the entire academic writing workflow, from outlining and idea development to drafting, revising, typesetting, and cross-referencing, assisted by integrated AI tools at every stage. This offers a comprehensive solution for enhancing both the quality and efficiency of producing scholarly work.
The availability of annotated corpora on coreference is a requirement for linguistics and NLP. This presupposes the availability of suitable annotation environments. Yet most of annotations tools for coreference are based on annotation models with limited expressiveness. We present here the OPERA annotation tool, which is based on an extended model for coreference annotation that, in addition of allowing work on referring expressions in the text (a widespread feature), enables the relations between entities to be annotated and characterized.
Recognition of historical documents is still an active research field due to the relatively low recognition accuracy achieved when processing old fonts or low-quality images. In this work, we investigate the use of Large Language Models (LLMs) for the correction of the OCR for old Greek documents. We examine two different old Greek datasets, one machine printed and one typewritten, using a Deep Network based OCR together with several known and easy-to-use LLMs for the correction of the result. Additionally, we synthetically produce erroneous texts and change the LLM prompts in order to further study the behavior of LLMs for correcting old Greek noisy text. Experimental results highlight the potential of LLMs for OCR correction of old Greek documents especially for the cases that the recognition results are relatively poor.
Mathematical formulas introduce complications to the standard approaches used in information retrieval. By studying how traditional (sparse) search systems perform in matching queries to documents, we hope to gain insights into which features in the formulas and in the accompanying natural language text signal likely relevance. In this paper, we focus on query rewriting for the ARQMath benchmarks recently developed as part of CLEF, the Conference and Labs of the Evaluation Forum. In particular, we improve mathematical community question answering applications by using responses from a large language model (LLM) to reformulate queries. Beyond simply replacing the query by the LLM response or concatenating the response to the query, we explore whether improvements accrue from the LLM selecting a subset of the query terms, augmenting the query with additional terms, or re-weighting the query terms. We also examine whether such query reformulation is equally advantageous for math features extracted from formulas and for keyword terms. As a final step, we use reciprocal rank fusion (RRF) to combine several component approaches in order to improve ranking results. In two experiments involving real-world mathematical questions, we show that combining four strategies for term selection, term augmentation, and term re-weighting improves nDCG'@1000 by 5%, MAP'@1000 by 7%, and P'@10 by more than 9% over using the question as given.
Advances in generative AI allow to create synthetic semi-structured documents with ease. Proprietary large language models (LLM) can now generate realistic receipts (a type of semi-structured data) with a prompt; however, their usefulness for training document understanding models is limited without accurate annotations. This work presents a framework that uses open-weight LLMs to create fullyannotated receipts. The framework can be extended or modified, and includes a step for self-assessment. We use the open dataset of receipts, SROIE, to test the usefulness of the generated receipts, and show that mixing both datasets can improve information extraction up to 32.9% for specific fields of the SROIE dataset.
With the exponential growth in the volume and complexity of structured and semi-structured documents, organizations face increasing challenges in extracting meaningful features for downstream analytics. This tutorial introduces a cutting-edge approach to automatic feature extraction using Large Language Models (LLMs), integrating generative AI with document engineering workflows. Participants will explore how LLMs can extract actionable insights from varied document types such as invoices, medical notes, legal filings, and audit reports. We demonstrate the integration of LLMs with machine learning models in end-to-end pipelines involving document parsing, feature extraction, and analytics. Real-world scenarios covered include business intelligence and globally significant domains such as: (1) detecting corruption from audit trails, (2) analyzing pregnancy risks from clinical narratives, (3) uncovering food adulteration in lab reports, (4) investigating accidents through unstructured reports, (5) monitoring war or geopolitical crises via leaked or public documents. This tutorial illustrates how global challenges---often rooted in complex documentation---can be addressed by transforming unstructured data into impactful solutions with social, economic, and ethical relevance.
We present the Di2Win Document Intelligence Platform (DIP). This modular AI-driven pipeline transforms raw document images --- captured by scanners or mobile phones --- into structured data and business actions in a single pass. The system comprises five loosely-coupled micro-services: (1) image-quality verification using a contrast-invariant model that flags blur, skew, and illumination issues above 100 ms per page; (2) document classification via a Transformer-base model with layout embeddings, delivering top-k types with calibrated confidence; (3) information extraction through i) Dilbert, a multimodal Token-Layout-Language model fine-tuned on weakly-labeled forms or ii) Delfos, a Large Language Model Mixture of Experts fine-tuned with well-defined prompts; (4) DataDrift, a powerful rules engine to avoid inconsistent outputs concerning the business process; and (5) process automation orchestrated by a Business Process Model Notation (BPMN) plus a Robot Process Automation (RPA) engine that routes results to databases, APIs, or human-review queues. All AI components are orchestrated through a messaging service to control the information flow, and the application exposes REST/gRPC endpoints to communicate with outside consumers. This enables the hot-swapping of models without downstream code changes by plugging a new message consumer into the messaging system. This also provides horizontal scalability since to increase the application throughput, we only need to add new AI engine consumers to the messaging system. Deployed in banking, insurance, and healthcare, the Di2Win DIP has processed more than 30 million pages, reducing average handling time by 79% and re-keying errors by 86 %, speeding up the workflows up to ten times. Our DocEng demonstration allows attendees to upload documents, observe live quality and confidence dashboards, and edit extracted fields with immediate feedback to the active-learning loop.
This study explores the use of Vision Large Language Models (VLLMs) for identifying items in complex graphical documents. In particular, we focus on looking for furniture objects (e.g. beds, tables, and chairs) and structural items (doors and windows) in floorplan images. We evaluate one object detection model (YOLO) and state-of-the-art VLLMs on two datasets featuring diverse floorplan layouts and symbols. The experiments with VLLMs are performed with a zero-shot setting, meaning the models are tested without any training or fine-tuning, as well as with a few-shot approach, where examples of items to be found in the image are given to the models in the prompt. The results highlight the strengths and limitations of VLLMs in recognizing architectural elements, providing guidance for future research in the use multimodal vision-language models for graphics recognition.
RVL-CDIP and Tobacco3482 are commonly used document classification benchmarks, but recent work on explainability has revealed that ID codes stamped on the documents in these datasets may be used by machine learning models to learn shortcuts on the classification task. In this paper, we present an in-depth investigation into the influence and impact of these ID codes on model performance. We annotate ID codes in documents from RVL-CDIP and Tobacco3482 and find that shallow learning models can achieve classification accuracy scores of roughly 40% on RVL-CDIP and 60% on Tobacco3482 using only features derived from the ID codes. We also find that a state-of-the-art document classifier sees a performance drop of 11 accuracy points on RVL-CDIP when ID codes are removed from the data. Finally, we train an ID code detection model in order to remove ID codes from RVL-CDIP and Tobacco3482 and make this data publicly available.
We discuss the use of hierarchical clustering to identify similar specimens in a large malware corpus. Clustering of any kind requires the use of a distance function, and evaluation of clustering algorithms requires criteria that involve some sort of ground truth. We use Jaccard Distance as the ground truth, and we compare the results of clustering when using MinHash and SuperMinHash, both of which are approximations of Jaccard, while supposedly being faster. This work therefore is a study of this tradeoff between speed and clustering quality.
Paragraph justification is based primarily on shrinking or stretching the interword blanks. While the blanks on a line are all scaled by the same amout, the amount in question varies from line to line. The quality of a paragraph's typographic color largely depends on the aforementioned variation being as small as possible. Yet, TEX'S paragraph justification algorithm addresses this problem in a rather coarse fashion. In this paper, we propose a refinement to the algorithm allowing to improve the situation without disturbing the general behavior of the algorithm too much, and without the need for manual intervention. We analyze the impact of our refinement on a large number of experiments through several statistical estimators. We also exhibit a number of typographical traits relatedto whitespace distribution that we believe may contribute to our perception of homogeneousness.
