Retrieval-Augmented Generation (RAG) allows overcoming the limited knowledge of LLMs by extending the input with external information. As a consequence, the contextual inputs to the model become much longer slowing down decoding time affecting the time a user has to wait for an answer. We address this challenge by presenting COCOM, an effective context compression method, reducing long contexts to only a handful of Context Embeddings, speeding up the generation time by a large margin. Our method allows for different compression rates, trading off decoding time for answer quality. Compared to earlier methods, COCOM allows for handling multiple contexts more effectively, significantly reducing decoding time for long inputs. Our method demonstrates an inference speed-up of up to 5.69 times while achieving higher performance compared to existing efficient context compression methods

The late interaction paradigm introduced with ColBERT stands out in the neural Information Retrieval space, offering a compelling effectiveness-efficiency trade-off across many benchmarks. Efficient late interaction retrieval is based on an optimized multi-step strategy, where an approximate search first identifies a set of candidate documents to re-rank exactly. In this work, we introduce SPLATE, a simple and lightweight adaptation of the ColBERTv2 model which learns an "MLM adapter'', mapping its frozen token embeddings to a sparse vocabulary space with a partially learned SPLADE module. This allows us to perform the candidate generation step in late interaction pipelines with traditional sparse retrieval techniques, making it particularly appealing for running ColBERT in CPU environments. Our SPLATE ColBERTv2 pipeline achieves the same effectiveness as the PLAID ColBERTv2 engine by re-ranking 50 documents that can be retrieved under 10ms.

Middle training methods aim to bridge the gap between the Masked Language Model (MLM) pre-training and the final finetuning for retrieval. Recent models such as CoCondenser, RetroMAE, and LexMAE argue that the MLM task is not sufficient enough to pre-train a transformer network for retrieval and hence propose various tasks to do so. Intrigued by those novel methods, we noticed that all these models used different finetuning protocols, making it hard to assess the benefits of middle training. We propose in this paper a benchmark of CoCondenser, RetroMAE, and LexMAE, under the same finetuning conditions. We compare both dense and sparse approaches under various finetuning protocols and middle training on different collections (MS MARCO, Wikipedia). We use additional middle training baselines, such as a standard MLM finetuning on the retrieval collection, optionally augmented by a CLS predicting the passage term frequency. For the sparse approach, our study reveals that there is almost no statistical difference between those methods: the more effective the finetuning procedure is, the less difference there is between those models. For the dense approach, RetroMAE using MS MARCO as middle-training collection shows excellent results in almost all the settings. Finally, we show that middle training on the retrieval collection, thus adapting the language model to it, is a critical factor. Overall, a better experimental setup should be adopted to evaluate middle training methods.

Sparse neural retrievers, such as DeepImpact, uniCOIL and SPLADE, have been introduced recently as an efficient and effective way to perform retrieval with inverted indexes. They aim to learn term importance and, in some cases, document expansions, to provide a more effective document ranking compared to traditional bag-of-words retrieval models such as BM25. However, these sparse neural retrievers have been shown to increase the computational costs and latency of query processing compared to their classical counterparts. To mitigate this, we apply a well-known family of techniques for boosting the efficiency of query processing over inverted indexes: static pruning. We experiment with three static pruning strategies, namely document-centric, term-centric and agnostic pruning, and we assess, over diverse datasets, that these techniques still work with sparse neural retrievers. In particular, static pruning achieves 2x speedup with negligible effectiveness loss (≤ 2% drop) and, depending on the use case, even 4x speedup with minimal impact on the effectiveness (≤ 8% drop). Moreover, we show that neural rerankers are robust to candidates from statically pruned indexes.

We consider the Document Understanding problem of segmenting tables in rows. We propose a method that first enumerates virtual row separator candidates and then select the correct ones thanks to a classification task, solved using supervised structured machine learning. Interestingly, the task is the joint-classification of virtual separators and real text lines. We describe and tested several alternative candidate generation methods and report the results of our experiment for each, on two different types of registry books from the 19th century.

The cTDaR competition aims at benchmarking state-of-the-art table detection (TRACK A) and table recognition (TRACK B) methods. In particular, we wish to investigate and compare general methods that can reliably and robustly identify the table regions within a document image on the one hand, and the table structure on the other hand. Due to the presence of hand-drawn tables and handwritten text, the methods must be robust against various noise conditions, interfering annotations, and variations of the tables. Two new challenging datasets were created to test the behaviour of state-of-the-art table detection and recognition systems on real world data. One dataset consists of modern documents, while the other consists of archival documents with presence of hand-drawn tables and handwritten text. The evaluation scheme is adapted from the ICDAR 2013 Table competition. We received results of Track A from 11 teams and results of Track B from 2 teams. Results for Track A are very good for the top participants. The winner and his runner-up are very close while using very different approaches. Track B was more challenging and only one participant was able to produce good results.

Recent advances in document understanding, especially text recognition, provide new opportunities to address the page segmentation problem. In this paper, we propose a method to groups text lines into semantic objects. We model a page as a graph where nodes represent text lines and the edges their geometric relations. The logical segmentation task then refers to identify all text lines belonging to some logical sub-division of the page. We model this task as categorizing edges as relevant or not to build the targeted sub-division (sub-graph). This edge categorization is performed using structured machine learning algorithms (graph Conditional Random Field and Edge Convolutional Network). We use a connected components-based approach following the edge classification for aggregating the nodes. This simple approach shows very robust results for various layout and various page sub-division. We experiment on table segmentation into multiple sub-divisions (rows, columns, and cells) and minutes segmentation into resolutions. Our sub-division and page-layout oblivious approach shows near-par performance as compared to task dedicated approaches and even outperforms them in certain setups.

We present in this paper experiments on Table Recognition in hand-written register books. We first explain how the problem of row and column detection is modelled, and then compare two Machine Learning approaches (Conditional Random Field and Graph Convolutional Network) for detecting these table elements. Evaluation was conducted on death records provided by the Archives of the Diocese of Passau. With an F-1 score of 89, both methods provide a quality which allows for Information Extraction. Software and dataset are open source/data.

We present a method for extracting structured elements of information, called structured data (sdata), from ocr'ed pages. The method first analyzes the layout of the page, building several concurrent layout structures. Then a tagging step is performed in order to tag textual elements based on their content. Combining the layout structures and the tagged elements, layout models for representing the structured data are inferred for the current page. These models are used to correct or tag some elements missed by the tagging step. The final set of structured data is extracted. An evaluation is presented.

After two participations to the INEX competition in the Structure Extraction task, which consists in building navigation tools for digitised books by constructing hyperlinked table of contents from OCR text and layout information, we present in this paper some reflections about this competition regarding its dataset, and its evaluation measure. We point out some issues, and propose some recommendations for improving the groundtruth and the measures.

In this paper we are interested in the problem of book splitting or more generally of indexing the logical parts of a document. This involves determining the boundaries of these parts as well as their label. We report here on the combined use of generic methods published in previous papers. We discuss the effect of combining several methods, also from a quality assurance perspective. Our experiments ground on real case studies of technical documents, such as books of specifications.

We present in this paper a method for document layout analysis based on identifying the function of document elements (what they do). This approach is orthogonal and complementary to the traditional view based on the form of document elements (how they are constructed). One key advantage of such functional knowledge is that the functions of some document elements are very stable from document to document and over time. Relying on the stability of such functions, the method is not impacted by layout variability, a key issue in logical document analysis and is thus very robust and versatile. The method starts the recognition process by using functional knowledge and uses in a second step formal knowledge as a source of feedback in order to correct some errors. This allows the method to adapt to specific documents by using formal specificities.

In this paper, we present a method for structuring a document according to the information present in its Table of Contents. The detection of the ToC as well as the determination of the parts it refers to in the document body rely on a series of generic properties characterizing any ToC, while its hierarchization is achieved using clustering techniques. We also report on the robustness and performance of the method before discussing it, in light of related work.