With the development of Semantic Web technologies in digital humanities, more and more users with little knowledge about these technologies need to interact with them. This paper presents Harold, a system for accessing the cultural heritage corpus without having to know any particular computer language, such as SPARQL. Harold is a conversational system with which a historian can interact using a user-friendly query interface together with navigation access to explore the corpus. Harold organizes the documents into a hierarchical structure using formal concept analysis (FCA) to provide a synthetic way to navigate the documents. The user may interact with the hierarchy concepts to focus on relevant documents and remove irrelevant ones. In addition, an ontology management interface is provided to assist the user in managing concepts related to their research problem. This ontology can be used by the retrieval process to better structure hierarchical access to the documents. Moreover, the concepts built by FCA provide interesting information that can guide the user to a new retrieval step to find more relevant documents related to their research problem.
This paper presents Harold, a conversational system that helps to explore textual corpora, such as the correspondence of Henri Poincaré. Currently developed in collaboration with the Henri Poincaré Archives, the system enables interactive exploration of the digitized and annotated correspondence using the Archives’ SPARQL endpoint. The Henri Poincaré Archives, a center for the history and philosophy of science, holds parts of the personal papers of various mathematicians and computer scientists. This includes more than 2, 000 letters exchanged by Henri Poincaré, now digitized, manually transcribed and enriched with metadata by the historians of the Archives. A key figure in mathematics, physics and philosophy of science, the Henri Poincaré (1854-1912) correspondence ofers valuable insight into his scientific activities and exchanges with fellow researchers of his time. Although the SPARQL endpoint supports advanced queries into the knowledge graph and ontologies developed at the Archives, it remains a challenge for non-experts, hence the need for an accessible interface like Harold. Even if Harold is originally developed for the Henri Poincaré Archives, it will be tested on other corpora as researchers working on other digital humanities corpora have expressed interest in this tool. This would allow us to test Harold on new ontologies and knowledge graphs with diferent vocabularies.
To simplify access to a corpus without requiring SPARQL queries, the Harold system ofers a more
user-friendly interface. This work addresses three key aspects: (1) Historical research questions are often
vague and cannot be answered with a single query, for example “What has Henri Poincaré exchanged
about mathematics for physics?”. Harold supports an iterative interactive exploration using FCA [
The paper is organized as follows. Section 2 presents the related work: solutions allowing nonspecialists to interact with ontologies, as well as the use of FCA in document access. Section 3 describes preliminaries: the letters of the Henri Poincaré correspondence, the general ontology used to annotate the letters, and the FCA theory. Section 4 presents the Harold system: the NLP extraction process to enrich the letter annotations by their content and shows the main functionalities of the Harold system illustrated by a running example. Section 5 describes the ontology management tool of the system.
Numerous works address the problem of interrogating knowledge base, especially represented in RDF
or accessible through a SPARQL endpoint. A classification of approaches that address this problem,
from the most informal (natural language) to the most formal (SPARQL), can be found in [
As query patterns are represented in an ontology, the transformation from a question in natural
language to a SPARQL query also uses SPARQL for mapping some part of the text of the question with
knowledge base resources. Other works propose to use specific user interfaces to query knowledge
bases without directly writing a SPARQL query. For example, Sparklis proposes an interface that assists
the user in constructing the SPARQL query [
Other systems propose to use a visual query language. In Nitelight, for example, the query takes
the form of a graph and the user has to construct the graph corresponding to the query [
Another point addressed in this work relies on a synthetic access instead of an individual letter access,
and the fact that a research problem required many steps and therefore interaction with the search
system. The work presented in this paper is inspired by the CreChainDo system [
Finally, the ontology building process presented in this paper is inspired by some of the many
approaches that can be found in the literature (see [
The Henri Poincaré correspondence corpus includes more than 2000 letters exchanged with more than 300 correspondents in five languages (mainly French with sporadic use of English, German, Sami, and Swedish), written between his entry to the École polytechnique in 1873 and his death in 1912. These documents have been digitized, manually transcribed, and semantically annotated with general metadata and contextual footnotes, forming the corpus’s critical apparatus, through over 30 years of historical work.
The letters are available in three formats: thematic printed volumes [
Each letter is described by semantic annotations using RDF triples, as illustrated in Fig. 1. The triples use properties of the Henri Poincaré ontology (ahpo prefix), such as the sender ( ahpo:sentBy), the recipient (ahpo:sentTo), the language (ahpo:language), the date (ahpo:writingDate), and the written place (ahpo:writtenAt).
ahpo:writtenAt ahpo:sentBy ahpo:language Letter_73 d c : t i t l e ... ahpo:writingDate ahpo:sentTo
FCA is a mathematical approach to data analysis based on lattice theory. A formal context is a triple
= (, , ), where is a set of individuals (called objects), is a set of properties (called attributes)
and is the relation on × stating that an object is described by a property [
Furthermore, the set of all formal concepts of the context = (, , ) is partially ordered
by extent inclusion, also called specialization (denoted by ≤ ) between concepts. ℒ = ⟨, ≤ ⟩ is a
complete lattice, called the concept lattice. The lattice ℒ can be drawn as a Hasse diagram where the
nodes are concepts and the arrows are specialization links. Fig. 2 illustrates a binary context (a) and its
corresponding lattice (b). The top concept contains all the letters; its intent is empty because there is no
common property shared by all the letters. In contrast, the bottom concept is defined by the set of all
properties, and its extent is empty because none of the letters is described by all the properties. Several
algorithms have been proposed for the construction of concept lattices; see [
The Harold system helps historians explore the Henri Poincaré correspondence by retrieving letters, analyzing their content, and acquiring knowledge. An NLP step for term identification in full texts enables the exploitation of both the letters’ content and their critical apparatus. Section 4.1 details the annotation process, while Section 4.2 introduces a running example, followed by descriptions and illustrations of Harold’s functionalities.
Historians analyzing the Henri Poincaré correspondence require analysis of both the full text and the critical apparatus, not just basic metadata. To enable this, terms were extracted from each letter and its critical apparatus using both a preexisting list of key terms and automatic noun phrase extraction. Scientific terms, including those of institutes such as Nobel committees, are primarily identified through noun phrases, while terms from Henri Poincaré ’s social environment, such as polytechnic slang, contain only one word. This list, curated by historians, is crucial as many terms have context-dependent meanings (e.g., "rat" refers to being late, not the animal). In addition, a list of cultural and scientific works cited in the letters is also maintained. All the terms extracted are kept in their original language, with spelling errors and abbreviations, the historians having the possibility to indicate all the variations of a given term.
As a running example, consider studying the impact of Henri Poincaré’s work on physics research. The historian can use Harold to search for letters exchanged with physicists Henri Poincaré knew or containing keywords related to mathematical physics (e.g., general terms: mathematics, physics, or more precise concepts: diferential equation , three-body problem). Solving this research problem requires multiple searches and content analysis. Based on the initial results, the historian can refine the searches with more specific terms or exclude irrelevant results. As they explore the letters, some knowledge can emerge within the search scope, particularly the vocabulary (i.e. terms) and concepts. Organizing these concepts in an ontology can enhance and refine the search process.
Section 4.3 presents the search form used to explore the corpus. In Section 4.4, the way the results are displayed is explained. Section 4.5 illustrates diferent types of interactions with Harold and shows the benefit of an interactive and interactive exploration process.
Harold allows to search, in the Henri Poincaré correspondence, letters that have some properties using a specific search form. This form also allows us to indicate which properties must be taken into account by the FCA using checkboxes. For example, a historian can indicate that they are interested in knowing who exchanged on a given subject (i.e., Sender and Recipient properties) and what terms appear in both the letters and their critical apparatus (Containing properties). The content of this form is transformed into a SPARQL query to search for letters that have the desired properties. This query simply filters all letters with respect to the given criteria and retrieves the properties chosen by the user. Finding all relevant letters in a single attempt is rare, requiring iterative interaction between the user and the system. Harold dynamically manages the set of letters used by FCA, which evolves by adding relevant letters through new queries or removing those irrelevant to the current research problem. In the interface, Starting a new search creates a new set, while Complete the current search adds new letters to the existing set.
Instead of simply listing the retrieved letters, Harold applies FCA to group them based on the selected properties, creating a synthetic view that highlights the shared characteristics (see Section 3.2). The resulting lattice is hierarchically displayed using depth-first search, as shown in Fig. 3. The sign µ expands the groups to show subgroups, while · collapses them. Each subgroup only displays properties not inherited from broader groups. For example, in Fig. 3, the 18 letters sent to Physics Nobel Committee also contain physique mathématique. The group size, set to 5, is adjustable by the user.
In this illustration, it can be seen, from the 39 letters containing physique mathématique, that 18 of these letters have been sent to the Nobel Physics Committee. Some potentially interesting terms for the research problem also appear: dérivées partielles de la physique mathématique (in English, partial derivatives of mathematical physics), propagation de la chaleur (in English, heat propagation), and problème de Dirichlet (in English, Dirichlet problem).
To find all letters related to the topic “Mathematics for physics”, a search for physique mathématique alone is insuficient, as many letters do not contain this exact phrase; Harold’s interactive features allow the user to refine searches, add relevant letters, or exclude irrelevant ones, thereby improving FCA results. As shown in Fig. 4, several actions are available for each letter group (here the 18 letters sent to the Physics Nobel Committee), the first three for letter analysis, then one for ontology management (see Section 5), and the last one for accessing the letters themselves. • Add the letters sent to the Physics Nobel Committee: This functionality aims to retrieve additional relevant letters. For example, the user might want to explore the letters sent to the Physics Nobel Committee to investigate their connection to mathematics for physics. Triggering this action adds these letters to the analysis set, similar to entering Physics Nobel Committee into the recipient ifeld and clicking Complete the current search. • Do not use “sent to the Physics Nobel Committee” in the result: This functionality excludes certain properties of FCA grouping to refine hierarchical results by removing irrelevant letter groups. For example, in the running case, it is applied to the sent to Physics Nobel Committee group, which is relevant to the research, but could also be used to exclude irrelevant groups such as sent by or sent to Henri Poincaré. • Remove letters sent to the Physics Nobel Committee: This functionality eliminates letters not interesting at all for the research problem. For the running example, removing the letters sent to this institute is not adequate but sometimes letters which are not helpful for a given research problem appear and this is this functionality that allows removing them from the analysis. • Add the concept “Physics Nobel Committee” to the ontology: This functionality helps the user to acquire knowledge of the content of the letter. The term Physics Nobel Committee is added to the set of unlabeled concepts of the ontology management interface, waiting to be placed in the ontology hierarchy, in this case under the concept institut (see Fig. 5).
The next section details the ontology management and its impact on the results.
The ontology interface allows for the manual creation of an ontology. It currently manages only atomic concepts and their relations. As shown in Fig. 5, it has two main parts: a Set of unclassified concepts , containing candidate concepts added by the user or retrieved from the Add to the ontology functionality, and a Set of classified concepts , organized hierarchically. For instance, Physics Nobel Committee is in the unclassified set, while the classified set includes mathématiques with 4 children in the hierarchy, physique with 3 children, and institut with 3 children. The hierarchy reflects specific/generic relationships between letters. This means that if fonction is a child of mathématiques in the hierarchy, it has to be interpreted as “a letter annotated by fonction will also be annotated by mathématiques.” Formally, using description logic notation, this means that the following formula is added to the ontology: Letter ⊓ ∃isAnnotatedBy.{"fonction"} ⊑ Letter ⊓ ∃isAnnotatedBy.{"mathématiques"} Hierarchy management, including moving concepts, is done by drag-and-drop. Here Physics Nobel Committee can be moved under institut (in English institute).
This hierarchy is used to improve the FCA process by enriching the binary context with new letter properties. For example, if letter l12 has the property fonction then it has also the property mathématiques. Thus, the historian obtains a more structured result than without this hierarchy. For example, letters containing propagation de la chaleur or théories électrodynamiques are linked to physique, creating a group of letters about physics (see Fig. 6). Similarly, a group of 40 letters related to both mathématiques and physique is formed, even if they do not contain these exact terms but include related concepts. Furthermore, using the ontology in the search process ensures that entering physique retrieves letters associated with the physique part of the ontology.
This paper presented Harold, an iterative and interactive system for exploring the Henri Poincaré correspondence without requiring knowledge of semantic web technologies. The system enables analysis of both metadata and full text, using FCA to group letters by shared properties. Ontology management, integrated with the exploration system, enhances knowledge discovery by organizing properties hierarchically, which improves both letter organization and retrieval.
The principles used to explore the Henri Poincaré correspondence are applicable to other corpora or data types. An ongoing application of Harold focuses on analyzing computer science publications, particularly in areas such as semantic web, ontology building, and digital humanities, with metadata including authors, publication venue, year, keywords, and for textual data, titles, and abstracts. Future work includes evaluating the system by comparing the results of two historians working on the same research problem, one using Harold and the other not, focusing on coverage, letter retrieval, and new knowledge acquired. Another focus is improving ontology management, including (1) incorporating various relationships between concepts, (2) refining the handling of letter properties (e.g., exchange with being more general than sent to), and (3) managing concept definitions (e.g., defining Henri Poincaré’s study years as between 1873 and 1879).
As Harold is still under development, a proper evaluation has to be performed. It would be necessary to compare how two Henri Poincaré specialists conduct research on a given topic with or without Harold. It should be done in a limited time frame and with the same research question for both historians. The set of letters selected by each researcher would then be analyzed, as only the historian using Harold would have built an ontology. Feedback from these researchers on Harold would also be insightful. Providing public access to an instance of Harold using the Henri Poincaré correspondence or other corpora could allow more people to test the system in the future.
The author(s) have not employed any Generative AI tools.