Benchmarks that reflect the diversity and complexity of real-world documents are essential for accurately evaluating Automatic Text Recognition (ATR) systems, especially Vision-Large Language Models (vLLMs). Although recent models demonstrate impressive performance, they are often evaluated on datasets containing modern, printed texts mostly written in English, which limits their relevance to many practical applications. Therefore, selecting a model for a specific use case requires evaluating it on data that matches the target documents. This highlights the importance of representative benchmarks for real-world applications. In this paper, we introduce METATR (v1.0), a multilingual, evolving benchmark designed to evaluate ATR models across a wide range of documents, facilitating meaningful model comparison and selection. The benchmark was designed to maximize diversity by including documents from various public collections. These documents cover 29 languages and include texts with multiple scripts and layouts. Beyond the dataset itself, METATR defines a standardized prompting and normalization methodology and establishes a dynamic evaluation framework. This approach is intended to produce reproducible results while remaining extensible over time. We evaluated a wide range of state-of-the-art systems, including open-source models, such as Churro and PeroOCR, as well as closed-source models, such as Gemini 3 and Claude. Results are reported across various dimensions, including performance at the dataset and language levels, robustness to handwritten documents, and computational efficiency. Our findings show that, although proprietary models achieve the most consistent performance, substantial variability persists across scripts and layouts. Overall, METATR provides a multidimensional, practitioner-oriented framework for assessing multilingual ATR in real-world conditions and tracking progress as the field evolves.

Information Extraction processes in handwritten documents tend to rely on obtaining an automatic transcription and performing Named Entity Recognition (NER) over such transcription. For this reason, in publicly available datasets, the performance of the systems is usually evaluated with metrics particular to each dataset. Moreover, most of the metrics employed are sensitive to reading order errors. Therefore, they do not reflect the expected final application of the system and introduce biases in more complex documents. In this paper, we propose and publicly release a set of reading order independent metrics tailored to Information Extraction evaluation in handwritten documents. In our experimentation, we perform an in-depth analysis of the behavior of the metrics to recommend what we consider to be the minimal set of metrics to evaluate a task correctly.

The PRET19 project aims to carry out a comprehensive analysis of the nineteenth-century loan registers of several Parisian university libraries. These historical documents provide invaluable insights into the circulation of books and the intellectual engagement of the academic community during a transformative period. By reconstructing the relationships and trends between borrowers, the project offers a unique perspective on the intellectual landscape of Parisian universities. In this first phase, the registers come from three different university libraries and exhibit a great diversity in layout, handwriting and content, which poses significant challenges for data processing. To address these challenges, we developed a document processing workflow that effectively combines automatic handwritten text recognition (HTR) with manual processing and validation. In addition, we provide a detailed description of the database, which is designed to comprehensively model the information extracted from the registries, ensuring that the data is structured in a way that adequately responds to anticipated research requests. Once completed, the processed data will be from winter 2024 accessible via a dedicated website, providing a comprehensive digital resource for historical research.

PyLaia is one of the most popular open-source software for Automatic Text Recognition (ATR), delivering strong performance in terms of speed and accuracy. In this paper, we outline our recent contributions to the PyLaia library, focusing on the incorporation of reliable confidence scores and the integration of statistical language modeling during decoding. Our implementation provides an easy way to combine PyLaia with n-grams language models at different levels. One of the highlights of this work is that language models are completely auto-tuned: they can be built and used easily without any expert knowledge, and without requiring any additional data. To demonstrate the significance of our contribution, we evaluate PyLaia’s performance on twelve datasets, both with and without language modelling. The results show that decoding with small language models improves the Word Error Rate by 13 https://gitlab.teklia.com/atr/pylaia ), and twelve open-source models are released on Hugging Face ( https://huggingface.co/collections/Teklia/pylaia-65f16e9ae0aa03690e9e9f80 ).

This study evaluates the recently proposed Document Attention Network (DAN) for extracting key-value information from Uruguayan birth certificates, handwritten in Spanish. We investigate two annotation strategies for automatically transcribing handwritten documents, fine-tuning DAN with minimal training data and annotation effort. Experiments were conducted on two datasets containing the same images (201 scans of birth certificates written by more than 15 different writers) but with different annotation methods. Our findings indicate that normalized annotation is more effective for fields that can be standardized, such as dates and places of birth, whereas diplomatic annotation performs much better for fields containing names and surnames, which can not be standardized.

This paper presents a complete processing workflow for extracting information from French census lists from 1836 to 1936. These lists contain information about individuals living in France and their households. We aim at extracting all the information contained in these tables using automatic handwritten table recognition. At the end of the Socface project, in which our work is taking place, the extracted information will be redistributed to the departmental archives, and the nominative lists will be freely available to the public, allowing anyone to browse hundreds of millions of records. The extracted data will be used by demographers to analyze social change over time, significantly improving our understanding of French economic and social structures. For this project, we developed a complete processing workflow: large-scale data collection from French departmental archives, collaborative annotation of documents, training of handwritten table text and structure recognition models, and mass processing of millions of images. We present the tools we have developed to easily collect and process millions of pages. We also show that it is possible to process such a wide variety of tables with a single table recognition model that uses the image of the entire page to recognize information about individuals, categorize them and automatically group them into households. The entire process has been successfully used to process the documents of a departmental archive, representing more than 450,000 images.

In recent advances in automatic text recognition (ATR), deep neural networks have demonstrated the ability to implicitly capture language statistics, potentially reducing the need for traditional language models. This study directly addresses whether explicit language models, specifically n-gram models, still contribute to the performance of state-of-the-art deep learning architectures in the field of handwriting recognition. We evaluate two prominent neural network architectures, PyLaia [23] and DAN [8], with and without the integration of explicit n-gram language models. Our experiments on three datasets - IAM [19], RIMES [11], and NorHand v2 [2] - at both line and page level, investigate optimal parameters for n-gram models, including their order, weight, smoothing methods and tokenization level. The results show that incorporating character or subword n-gram models significantly improves the performance of ATR models on all datasets, challenging the notion that deep learning models alone are sufficient for optimal performance. In particular, the combination of DAN with a character language model outperforms current benchmarks, confirming the value of hybrid approaches in modern document analysis systems.

This paper presents Callico, a web-based open source platform designed to simplify the annotation process in document recognition projects. The move towards data-centric AI in machine learning and deep learning underscores the importance of high-quality data, and the need for specialised tools that increase the efficiency and effectiveness of generating such data. For document image annotation, Callico offers dual-display annotation for digitised documents, enabling simultaneous visualisation and annotation of scanned images and text. This capability is critical for OCR and HTR model training, document layout analysis, named entity recognition, form-based key value annotation or hierarchical structure annotation with element grouping. The platform supports collaborative annotation with versatile features backed by a commitment to open source development, high-quality code standards and easy deployment via Docker. Illustrative use cases - including the transcription of the Belfort municipal registers, the indexing of French World War II prisoners for the ICRC, and the extraction of personal information from the Socface project's census lists - demonstrate Callico's applicability and utility.

Deep neural networks are becoming increasingly powerful and large and always require more labelled data to be trained. However, since annotating data is time-consuming, it is now necessary to develop systems that show good performance while learning on a limited amount of data. These data must be correctly chosen to obtain models that are still efficient. For this, the systems must be able to determine which data should be annotated to achieve the best results. In this paper, we propose four estimators to estimate the confidence of object detection predictions. The first two are based on Monte Carlo dropout, the third one on descriptive statistics and the last one on the detector posterior probabilities. In the active learning framework, the three first estimators show a significant improvement in performance for the detection of document physical pages and text lines compared to a random selection of images. We also show that the proposed estimator based on descriptive statistics can replace MC dropout, reducing the computational cost without compromising the performances.

In this paper, we address the challenge of document image analysis for historical index table documents with handwritten records. Demographic studies can gain insight from the use of automatic document analysis in such documents through the study of population movements. To evaluate the efficacy of automatic layout analysis tools, we release the PARES dataset [6], which contains 250 labeled index table images originating from French archives. Also, we run state-of-the-art algorithms (U-FCN, R-CNN and Transformers) in order to detect the lines within index tables, a common prerequisite for handwritten text recognition (HTR). Our results indicate that text line extraction works well with the U-FCN model, while also indicating that Transformer architectures show promise for accurate text line detection in such historical documents with great efficiency. This is a encouraging step towards a Transformer-based architecture for both layout and content detection. This process and dataset represent a first step to automatically analyze handwritten and historical index tables. In addition to this paper and the PARES [6] dataset of historical index tables of 250 images, we release segmentation masks, the code we used to train and test the models, and the models themselves.

Recent advancements in Deep Learning-based Handwritten Text Recognition (HTR) have led to models with remarkable performance on both modern and historical manuscripts in large benchmark datasets. Nonetheless, those models struggle to obtain the same performance when applied to manuscripts with peculiar characteristics, such as language, paper support, ink, and author handwriting. This issue is very relevant for valuable but small collections of documents preserved in historical archives, for which obtaining sufficient annotated training data is costly or, in some cases, unfeasible. To overcome this challenge, a possible solution is to pretrain HTR models on large datasets and then fine-tune them on small single-author collections. In this paper, we take into account large, real benchmark datasets and synthetic ones obtained with a styled Handwritten Text Generation model. Through extensive experimental analysis, also considering the amount of fine-tuning lines, we give a quantitative indication of the most relevant characteristics of such data for obtaining an HTR model able to effectively transcribe manuscripts in small collections with as little as five real fine-tuning lines.

We propose a Transformer-based approach for information extraction from digitized handwritten documents. Our approach combines, in a single model, the different steps that were so far performed by separate models: feature extraction, handwriting recognition and named entity recognition. We compare this integrated approach with traditional two-stage methods that perform handwriting recognition before named entity recognition, and present results at different levels: line, paragraph, and page. Our experiments show that attention-based models are especially interesting when applied on full pages, as they do not require any prior segmentation step. Finally, we show that they are able to learn from key-value annotations: a list of important words with their corresponding named entities. We compare our models to state-of-the-art methods on three public databases (IAM, ESPOSALLES, and POPP) and outperform previous performances on all three datasets.

We propose a new database for information extraction from historical handwritten documents. The corpus includes 5,393 finding aids from six different series, dating from the 18th-20th centuries. Finding aids are handwritten documents that contain metadata describing older archives. They are stored in the National Archives of France and are used by archivists to identify and find archival documents. Each document is annotated at page-level, and contains seven fields to retrieve. The localization of each field is not available in such a way that this dataset encourages research on segmentation-free systems for information extraction. We propose a model based on the Transformer architecture trained for end-to-end information extraction and provide three sets for training, validation and testing, to ensure fair comparison with future works. The database is freely accessible ( https://zenodo.org/record/7868059 ).

In this paper, we explore different ways of training a model for handwritten text recognition when multiple imperfect or noisy transcriptions are available. We consider various training configurations, such as selecting a single transcription, retaining all transcriptions, or computing an aggregated transcription from all available annotations. In addition, we evaluate the impact of quality-based data selection, where samples with low agreement are removed from the training set. Our experiments are carried out on municipal registers of the city of Belfort (France) written between 1790 and 1946. The results show that computing a consensus transcription or training on multiple transcriptions are good alternatives. However, selecting training samples based on the degree of agreement between annotators introduces a bias in the training data and does not improve the results. Our dataset is publicly available on Zenodo.

This paper presents a complete workflow designed for extracting information from Quebec handwritten parish registers. The acts in these documents contain individual and family information highly valuable for genetic, demographic and social studies of the Quebec population. From an image of parish records, our workflow is able to identify the acts and extract personal information. The workflow is divided into successive steps: page classification, text line detection, handwritten text recognition, named entity recognition and act detection and classification. For all these steps, different machine learning models are compared. Once the information is extracted, validation rules designed by experts are then applied to standardize the extracted information and ensure its consistency with the type of act (birth, marriage and death). This validation step is able to reject records that are considered invalid or merged. The full workflow has been used to process over two million pages of Quebec parish registers from the 19–20th centuries. On a sample comprising 65% of registers, 3.2 million acts were recognized. Verification of the birth and death acts from this sample shows that 74% of them are considered complete and valid. These records will be integrated into the BALSAC database and linked together to recreate family and genealogical relations at large scale.

We propose a new database for information extraction from historical handwritten documents. The corpus includes 5,393 finding aids from six different series, dating from the 18th-20th centuries. Finding aids are handwritten documents that contain metadata describing older archives. They are stored in the National Archives of France and are used by archivists to identify and find archival documents. Each document is annotated at page-level, and contains seven fields to retrieve. The localization of each field is not available in such a way that this dataset encourages research on segmentation-free systems for information extraction. We propose a model based on the Transformer architecture trained for end-to-end information extraction and provide three sets for training, validation and testing, to ensure fair comparison with future works. The database is freely accessible at https://zenodo.org/record/7868059.

Text line segmentation is one of the key steps in historical document understanding. It is challenging due to the variety of fonts, contents, writing styles and the quality of documents that have degraded through the years. In this paper, we address the limitations that currently prevent people from building line segmentation models with a high generalization capacity. We present a study conducted using three state-of-the-art systems Doc-UFCN, dhSegment and ARU-Net and show that it is possible to build generic models trained on a wide variety of historical document datasets that can correctly segment diverse unseen pages. This paper also highlights the importance of the annotations used during training: each existing dataset is annotated differently. We present a unification of the annotations and show its positive impact on the final text recognition results. In this end, we present a complete evaluation strategy using standard pixel-level metrics, object-level ones and introducing goal-oriented metrics.

In this paper, we propose an evaluation of several state-of-the-art open-source natural language processing (NLP) libraries for named entity recognition (NER) on handwritten historical documents: spaCy, Stanza and Flair. The comparison is carried out on three low-resource multilingual datasets of handwritten historical documents: HOME (a multilingual corpus of medieval charters), Balsac (a corpus of parish records from Quebec), and Esposalles (a corpus of marriage records in Catalan). We study the impact of the document recognition processes (text line detection and handwriting recognition) on the performance of the NER. We show that current off-the-shelf NER libraries yield state-of-the-art results, even on low-resource languages or multilingual documents using multilingual models. We show, in an end-to-end evaluation, that text line detection errors have a greater impact than handwriting recognition errors. Finally, we also report state-of-the-art results on the public Esposalles dataset.

Handwritten Text Recognition (HTR) is an open problem at the intersection of Computer Vision and Natural Language Processing. The main challenges, when dealing with historical manuscripts, are due to the preservation of the paper support, the variability of the handwriting – even of the same author over a wide time-span – and the scarcity of data from ancient, poorly represented languages. With the aim of fostering the research on this topic, in this paper we present the Ludovico Antonio Muratori (LAM) dataset, a large line-level HTR dataset of Italian ancient manuscripts edited by a single author over 60 years. The dataset comes in two configurations: a basic splitting and a date-based splitting which takes into account the age of the author. The first setting is intended to study HTR on ancient documents in Italian, while the second focuses on the ability of HTR systems to recognize text written by the same writer in time periods for which training data are not available. For both configurations, we analyze quantitative and qualitative characteristics, also with respect to other line-level HTR benchmarks, and present the recognition performance of state-of-the-art HTR architectures. The dataset is available for download at https://aimagelab.ing.unimore.it/go/lam.

In this paper, we introduce an open database of historical handwritten documents fully annotated in Norwegian, the first of its kind, allowing the development of handwritten text recognition models (HTR) in Norwegian. In order to evaluate the performance of state-of-the-art HTR models on this new base, we conducted a systematic survey of open-source HTR libraries published between 2019 and 2021, identified ten libraries and selected four of them to train HTR models. We trained twelve models in different configurations and compared their performance on both random and scripter-based data splitting. The best recognition results were obtained by the PyLaia and Kaldi libraries which have different and complementary characteristics, suggesting that they should be combined to further improve the results.