Ensuring that generated text is accurately attributed to its underlying sources is critical for the transparency, trustworthiness, and verifiability of large language model (LLM) outputs. In this work, we conduct a comparative study of post-hoc context-attribution methods, focusing on the use of cross-encoders—both frozen and fine-tuned—as well as proprietary and open-source LLMs in low-annotation settings. We explore strategies for leveraging frozen LLMs for context-attribution without fine-tuning, and we develop techniques to optimize cross-encoder performance for semantic alignment between generated text and source material. Our evaluation spans four datasets: ASQA, ELI5, TREC-RAG, and a proprietary legal corpus, and includes both answer-level and sentence-level attribution tasks. Additionally, we investigate the impact of training small cross-encoders on synthetic data to assess their scalability and deployment potential in resource-constrained environments. Our results demonstrate that cross-encoders prove to be valid alternatives to LLMs for post-generation answer-level context-attribution. Moreover, after proper hyperparameter tuning, the same model can achieve performance comparable to proprietary LLM performance for sentence- and answer-level context-attribution. Finally, trained solely on synthetic data, small cross-encoders’ performance can be further improved while offering a scalable and cost-effective solution.

After the impressive diffusion of social media and microblogging websites of the last few years, the identification of users having the capability of influencing other users' choices is an important research topic because of the opportunities it can offer to many business companies. Most of the existing approaches, however, detect influencers by relying on centrality measures computed on networks that connect users having different types of inter-relationships. In this paper, we propose a method capable of finding influential users by exploiting the contents of the messages posted by them to express opinions on items, by modeling these contents with a three-layer network. Layers represent users, items, and keywords, along with intra-layer interactions among the actors of the same layer. Inter-layer connections are triples (u, i, k) expressing the information that a user u comments on an item i by using a keyword k. By exploiting multilinear algebra, we present a method capable of extracting the most active users stating their point of view about dominant items tagged with dominant keywords. We conduct a series of experiments on different realworld datasets collected from Twitter and Yelp Social Networks about different topics. Experimental results show the ability of our approach to find influential users that are both authoritative in the user network and very active in posting opinions about the topic of interest.

This tutorial is targeted at academics and practitioners, both within and outside of the Document Engineering community, who are confronted with table processing tasks such as information extraction and conversion, or have an interest in the topic, and wish to deepen their understanding of the state-of-the-art in this field.

If smartly utilized, Big Data locked in unstructured sources, such as PDF documents, can yield unprecedented insights in solving tough business issues, optimizing business processes and improving customer relations. The challenge addressed in this paper is to unlock the value held in data plunged in unstructured document. We describe how a contextual workflow based approach is used to address, in a semantic and flexible way, various problems arising in processing data contained into documents. We present the MANTRA Smart Data Platform, which enables to turn Big Data into Smart Data by means of contextual workflows composed by smart-cloud applications (APPs for short). Among the others, the MANTRA Language APP executes MANTRA rules that are able to extract and annotate information contained in heterogeneous sources (raw text, PDF, HTML or other presentation-oriented document format). Such rules exploit syntactic and semantic expressions, visual and spatial features, and natural language capabilities. Real cases of applications are showing that the proposed approach is able to process a large amount of heterogeneous input documents, as well as extract and consolidate the information of interest.

The amount of data in our world has been exploding. Hence, integrating, managing and analyzing large amounts of data – i.e. big data - will become a key issue for businesses for better operating and competing in today’s markets. We strongly believe that big data are really useful if used in a smart way. So we introduce the concept of Smart Data that is web and enterprise structured and unstructured big data with explicit and implicit semantics that leverages context to understand intent for better driving business processes and for better and more informed decisions making. This paper proposes a language able to give a representation of big data based on ontologies and a system that implements an approach capable to satisfy the increasing need for efficiency and scalability in semantic data management. The proposed MANTRA Language allows for: (i) representing the semantics of data by knowledge representation constructs; (ii) acquiring big data from disparate heterogeneous sources (e.g. data bases, documents), integrating and managing them; (iii) reasoning and querying with big data. The syntax of the proposed language is partially derived from logic programming, but the semantic is completely revised. The novelty of the language we propose is that a class can be thought of as a flexible collection of structurally heterogeneous individuals that have different properties (schema-less). The language also allows executing efficient querying and reasoning for revealing implicit knowledge. These have been achieved by using a triple-based data persistency model and a scalable No-SQL storage system.

Table understanding is a well studied problem in document analysis, and many academic and commercial approaches have been developed to recognize tables in several document formats, including plain text, scanned page images and born-digital, object-based formats such as PDF. Despite the abundance of these techniques, an objective comparison of their performance is still missing. The Table Competition held in the context of ICDAR 2013 is our first attempt at objectively evaluating these techniques against each other in a standardized way, across several input formats. The competition independently addresses three problems: (i) table location, (ii) table structure recognition, and (iii) these two tasks combined. We received results from seven academic systems, which we have also compared against four commercial products. This paper presents our findings.

This paper presents a methodology for the evaluation of table understanding algorithms for PDF documents. The evaluation takes into account three major tasks: table detection, table structure recognition and functional analysis. We provide a general and flexible output model for each task along with corresponding evaluation metrics and methods. We also present a methodology for collecting and ground-truthing PDF documents based on consensus-reaching principles and provide a publicly available ground-truthed dataset.

Deep Web pages convey very relevant information for different application domains like e-government, e-commerce, social networking. For this reason there is a constant high interest in efficiently, effectively and automatically extracting data from Deep Web data sources. In this paper we present SILA, a novel Spatial Instance Learning Approach, that allows for extracting data records from Deep Web pages by exploiting both the spatial arrangement and the presentation features of data items/fields produced by layout engines of Web browsers in visualizing Deep Web pages on the screen. SILA is independent from the internal HTML encodings of Web pages, and allows for recognizing data records in pages having multiple data regions in which data items are arranged by many different presentation layouts. Experimental results show that SILA has very high precision and recall and that it works much better than MDR and ViNTs approaches.

In last years the huge relevance of accessing and acquiring information made available by Web pages and business documents has grown much further. Thus, wrapping information from documents in HTML and PDF formats is receiving increasing interest. In this paper we present a textual query language, named ViQueL, that allows for querying information in both Web and PDF documents on the base of its spatial arrangement. The proposed language is founded on spatial grammars, i.e. context free grammars extended by spatial constructs. The main feature of ViQueL is that it make possible to identify and extract relevant information from HTML and PDF documents on the base of their visual appearance by using easy-to-write queries. Despite a considerable expressive power, combined complexity of ViQueL is in P-Time. Moreover, experiments show that ViQueL is reasonably efficient for real life extraction tasks.

Querying data from presentation formats like HTML, for purposes such as information extraction, requires the consideration of tree structures as well as the consideration of spatial relationships between laid out elements. The underlying rationale is that frequently the rendering of tree structures is very involved and undergoing more frequent updates than the resulting layout structure. Therefore, in this paper, we present Spatial XPath (SXPath), an extension of XPath 1.0 that allows for inclusion of spatial navigation primitives into the language resulting in conceptually simpler queries on Web documents. The SXPath language is based on a combination of a spatial algebra with formal descriptions of XPath navigation, and maintains polynomial time combined complexity. Practical experiments demonstrate the usability of SXPath.

This paper presents PDF-TREX, an heuristic approach for table recognition and extraction from PDF documents.The heuristics starts from an initial set of basic content elements and aligns and groups them, in bottom-up way by considering only their spatial features, in order to identify tabular arrangements of information. The scope of the approach is to recognize tables contained in PDF documents as a 2-dimensional grid on a Cartesian plane and extract them as a set of cells equipped by 2-dimensional coordinates. Experiments, carried out on a dataset composed of tables contained in documents coming from different domains, shows that the approach is well performing in recognizing table cells.The approach aims at improving PDF document annotation and information extraction by providing an output that can be further processed for understanding table and document contents.

Information extraction is of paramount importance in several real world applications in the areas of business intelligence, competitive and military intelligence. Although several sophisticated and indeed complex approaches were proposed, they are still limited in many aspects. In this paper the novel ontology-based system named XONTO, that allows the semantic extraction of information from PDF unstructured documents, is presented. The XONTO system is founded on the idea of self-describing ontologies in which objects and classes can be equipped by a set of rules named descriptors. These rules represent patterns that allow to automatically recognize and extract ontology objects contained in PDF documents also when information is arranged in tabular form. This way a self-describing ontology expresses the semantic of the information to extract and the rules that, in turn, populate itself. In the paper XONTO system behaviors and structure are sketched by means of a running example.