We introduce SmolDocling, an ultra-compact vision-language model targeting end-to-end document conversion. Our model comprehensively processes entire pages by generating DocTags, a new universal markup format that captures all page elements in their full context with location. Unlike existing approaches that rely on large foundational models, or ensemble solutions that rely on handcrafted pipelines of multiple specialized models, SmolDocling offers an end-to-end conversion for accurately capturing content, structure and spatial location of document elements in a 256M parameters vision-language model. SmolDocling exhibits robust performance in correctly reproducing document features such as code listings, tables, equations, charts, lists, and more across a diverse range of document types including business documents, academic papers, technical reports, patents, and forms -- significantly extending beyond the commonly observed focus on scientific papers. Additionally, we contribute novel publicly sourced datasets for charts, tables, equations, and code recognition. Experimental results demonstrate that SmolDocling competes with other Vision Language Models that are up to 27 times larger in size, while reducing computational requirements substantially. The model is currently available, datasets will be publicly available soon.

We present Deep Search DocQA. This application enables information extraction from documents via a question-answering conversational assistant. The system integrates several technologies from different AI disciplines consisting of document conversion to machine-readable format (via computer vision), finding relevant data (via natural language processing), and formulating an eloquent response (via large language models). Users can explore over 10,000 Environmental, Social, and Governance (ESG) disclosure reports from over 2000 corporations. The Deep Search platform can be accessed at: https://ds4sd.github.io.

Document understanding is a key business process in the data-driven economy since documents are central to knowledge discovery and business insights. Converting documents into a machine-processable format is a particular challenge here due to their huge variability in formats and complex structure. Accordingly, many algorithms and machine-learning methods emerged to solve particular tasks such as Optical Character Recognition (OCR), layout analysis, table-structure recovery, figure understanding, etc. We observe the adoption of such methods in document understanding solutions offered by all major cloud providers. Yet, publications outlining how such services are designed and optimized to scale in the cloud are scarce. In this paper, we focus on the case of document conversion to illustrate the particular challenges of scaling a complex data processing pipeline with a strong reliance on machine-learning methods on cloud infrastructure. Our key objective is to achieve high scalability and responsiveness for different workload profiles in a well-defined resource budget. We outline the requirements, design, and implementation choices of our document conversion service and reflect on the challenges we faced. Evidence for the scaling behavior and resource efficiency is provided for two alternative workload distribution strategies and deployment configurations. Our best-performing method achieves sustained throughput of over one million PDF pages per hour on 3072 CPU cores across 192 nodes.

Accurate document layout analysis is a key requirement for high-quality PDF document conversion. With the recent availability of public, large ground-truth datasets such as PubLayNet and DocBank, deep-learning models have proven to be very effective at layout detection and segmentation. While these datasets are of adequate size to train such models, they severely lack in layout variability since they are sourced from scientific article repositories such as PubMed and arXiv only. Consequently, the accuracy of the layout segmentation drops significantly when these models are applied on more challenging and diverse layouts. In this paper, we presentDocLayNet, a new, publicly available, document-layout annotation dataset in COCO format. It contains 80863 manually annotated pages from diverse data sources to represent a wide variability in layouts. For each PDF page, the layout annotations provide labelled bounding-boxes with a choice of 11 distinct classes. DocLayNet also provides a subset of double- and triple-annotated pages to determine the inter-annotator agreement. In multiple experiments, we provide baseline accuracy scores (in mAP) for a set of popular object detection models. We also demonstrate that these models fall approximately 10% behind the inter-annotator agreement. Furthermore, we provide evidence that DocLayNet is of sufficient size. Lastly, we compare models trained on PubLayNet, DocBank and DocLayNet, showing that layout predictions of the DocLayNet-trained models are more robust and thus the preferred choice for general-purpose document-layout analysis.

Foundational Models (FMs) have demonstrated unprecedented capabilities including zero-shot learning, high fidelity data synthesis, and out of domain generalization. However, the parameter capacity of FMs is still limited, leading to poor out-of-the-box performance of FMs on many expert tasks (e.g. retrieval of car manuals technical illustrations from language queries), data for which is either unseen or belonging to a long-tail part of the data distribution of the huge datasets used for FM pre-training. This underlines the necessity to explicitly evaluate and finetune FMs on such expert tasks, arguably ones that appear the most in practical real-world applications. In this paper, we propose a first of its kind FETA benchmark built around the task of teaching FMs to understand technical documentation, via learning to match their graphical illustrations to corresponding language descriptions. Our FETA benchmark focuses on text-to-image and image-to-text retrieval in public car manuals and sales catalogue brochures. FETA is equipped with a procedure for completely automatic annotation extraction (code would be released upon acceptance), allowing easy extension of FETA to more documentation types and application domains in the future. Our automatic annotation leads to an automated performance metric shown to be consistent with metrics computed on human-curated annotations (also released). We provide multiple baselines and analysis of popular FMs on FETA leading to several interesting findings that we believe would be very valuable to the FM community, paving the way towards real-world application of FMs for many practical expert tasks currently being `overlooked' by standard benchmarks focusing on common objects.

The number of published PDF documents in both the academic and commercial world has increased exponentially in recent decades. There is a growing need to make their rich content discoverable to information retrieval tools. Achieving high-quality semantic searches demands that a document's structural components such as title, section headers, paragraphs, (nested) lists, tables and figures (including their captions) are properly identified. Unfortunately, the PDF format is known to not conserve such structural information because it simply represents a document as a stream of low-level printing commands, in which one or more characters are placed in a bounding box with a particular styling. In this paper, we present a novel approach to document structure recovery in PDF using recurrent neural networks to process the low-level PDF data representation directly, instead of relying on a visual re-interpretation of the rendered PDF page, as has been proposed in previous literature. We demonstrate how a sequence of PDF printing commands can be used as input into a neural network and how the network can learn to classify each printing command according to its structural function in the page. This approach has three advantages: First, it can distinguish among more fine-grained labels (typically 10-20 labels as opposed to 1-5 with visual methods), which results in a more accurate and detailed document structure resolution. Second, it can take into account the text flow across pages more naturally compared to visual methods because it can concatenate the printing commands of sequential pages. Last, our proposed method needs less memory and it is computationally less expensive than visual methods. This allows us to deploy such models in production environments at a much lower cost. Through extensive architectural search in combination with advanced feature engineering, we were able to implement a model that yields a weighted average F1 score of 97% across 17 distinct structural labels. The best model we achieved is currently served in production environments on our Corpus Conversion Service (CCS), which was presented at KDD18. This model enhances the capabilities of CCS significantly, as it eliminates the need for human annotated label ground-truth for every unseen document layout. This proved particularly useful when applied to a huge corpus of PDF articles related to COVID-19.

Over the past few decades, the amount of scientific articles and technical literature has increased exponentially in size. Consequently, there is a great need for systems that can ingest these documents at scale and make the contained knowledge discoverable. Unfortunately, both the format of these documents (e.g. the PDF format or bitmap images) as well as the presentation of the data (e.g. complex tables) make the extraction of qualitative and quantitive data extremely challenging. In this paper, we present a modular, cloud-based platform to ingest documents at scale. This platform, called the Corpus Conversion Service (CCS), implements a pipeline which allows users to parse and annotate documents (i.e. collect ground-truth), train machine-learning classification algorithms and ultimately convert any type of PDF or bitmap-documents to a structured content representation format. We will show that each of the modules is scalable due to an asynchronous microservice architecture and can therefore handle massive amounts of documents. Furthermore, we will show that our capability to gather groundtruth is accelerated by machine-learning algorithms by at least one order of magnitude. This allows us to both gather large amounts of ground-truth in very little time and obtain very good precision/recall metrics in the range of 99% with regard to content conversion to structured output. The CCS platform is currently deployed on IBM internal infrastructure and serving more than 250 active users for knowledge-engineering project engagements.