Context

Document understanding is an essential task in various application areas such as data invoice extraction, subject review, medical prescription analysis, etc., and holds significant commercial potential. Several approaches are proposed in the literature, but datasets' availability and data privacy challenge them. Considering the problem of information extraction from documents, different aspects must be taken into account, such as (1) document classification, (2) text localization, (3) OCR (Optical Character Recognition), (4) table extraction, and (5) key information detection.

In this context, machine vision and, more precisely, deep learning models for image processing are attractive methods. In fact, several models for document analysis were developed for text box detection, text extraction, table extraction, etc. Different kinds of deep learning approaches, such as GNN, are used to tackle these tasks. On the other hand, the extracted text from documents can be represented using different embeddings based on recent NLP approaches such as Transformers. Also, understanding spatial relationships is critical for text document extraction results for some applications such as invoice analysis.

Thus, the aim is to capture the structural connections between keywords (invoice number, date, amounts) and the main value (the desired information). An effective approach requires a combination of visual (spatial) and textual information.

Novelty for this edition:

After the success of VINALDO 2023, in the second edition of the VINALDO workshop, we encourage the description of novel problems or applications for document analysis in the area of information retrieval that has emerged in recent years. On the other hand, we want to highlight a particular topic namely “Multi-view and Multimodal approaches”. In fact, the VINALDO workshop aims to combine visual and textual information for document analysis, in this context, multi-view and multimodal methods have really an important advantage in dealing with different types of data. Thus, we encourage works that combine machine vision and NLP through Multiview or/and multimodal approaches. Finally, we also encourage works that combine NLP and computer vision methods and develop new document datasets for novel applications.

The VINALDO workshop aims to bring together an area for experts from industry, science, and academia to exchange ideas and discuss ongoing research in graph representation learning for scanned document analysis.

Topics of interests

• Multi-view document representation

• Multi-view algorithms for document clustering

• Multimodal document classification

• Multimodal deep networks

• Multi-view models for document ranking

• Document structure and layout learning

• OCR based methods

• Semi-supervised methods for document analysis

• Dynamic graph analysis

• Information Retrieval and Extraction form documents

• Knowledge graph for semantic document analysis

• Semantic understanding of document content

• Entity and link prediction in graphs

• Merging ontologies with graph-based methods using NLP techniques

• Cleansing and image enhancement techniques for scanned document

• Font text recognition in a scanned document

• Table identification and extraction from scanned documents

• Handwriting detection and recognition in documents

• Signature detection and verification in documents

• Visual document structure understanding

• Visual Question Answering

• Invoice analysis

• Scanned documents classification

• Scanned documents summarization

• Scanned documents translation

• Graph-based approaches for a spatial component in a scanned document

• Graph representation learning for NLP

Submission

The workshop is open to original papers of theoretical or practical nature. Papers should be formatted according to LNCS instructions for authors. VINALDO 2023 will follow a double blind review process. Authors should not include their names and affiliations anywhere in the manuscript. Authors should also ensure that their identity is not revealed indirectly by citing their previous work in the third person, and omit acknowledgements until the camera-ready version. Papers have to be submitted via the workshop's EasyChair submission page.

We welcome the following types of contributions:

• Full research papers (12-15 pages): finished or consolidated R&D works, to be included in one of the Workshop topics

• Short papers (6-8 pages): ongoing works with relevant preliminary results, opened to discussion.

At least one author of each accepted paper must register for the workshop, in order to present the paper. For further instructions please refer to the ICDAR 2024 page.

Contact

Rim Hantach, Engie, France

Rafika Boutalbi, Aix Marseille University, LIS Lab