For the past ten years, the history of fashion has attracted a growing public [Petrov 2019] [Stewart and Marcketti 2012]. But despite the interest shown in these collections, textile heritage remains under threat. This is particularly true of silk heritage, whose fragility poses major conservation problems and whose protection requires substantial investment. Few materials are of such historical, cultural and artistic importance; but a significant part of the silk-related collections are preserved by small institutions, lacking generally the means to promote their collections [Portalés et al 2018] [Gaitán et al 2019] [Pagán et al 2020]. Outside of the major events devoted to them, we also note the under-use of textile collections by visitors [Stewart and Marcketti 2012]. This is why the digitization of cultural heritage has become one of the crucial issues for its promotion [Eastermann 2015].

When users are interested in textile cultural heritage online, they are generally faced with two situations. The large generalist museums have the technical, financial and human resources to disseminate their data on the Web, but they do not systematically highlight their textile collections on their websites. Conversely, small museums do not have the same resources as their better-endowed counterparts [Claerr and Westeel 2010] to promote specialized collections online, which can be particularly rich. This is particularly the case for silk heritage, as the example of the Rhône-Alpes region clearly illustrates [Foron-Dauphin and Cano 2016] [Fournier et al 2016]. These institutions do describe their collections, but the digital data they produce is, most of the time, inaccessible online. Without a digital presence, these collections remain generally unknown, except to a few connoisseurs and specialists.

The Semantic Web offers particularly interesting prospects in this domain; it proposes to cultural heritage institutions to increase the visibility of their metadata on the Web by “freeing” it from catalogs; to aggregate this data with others and provide access to it in a federated way; to interconnect data with each other and, ultimately, to enrich this information more easily thanks to the links created between them. By encouraging the dissemination of structured data, Semantic Web technologies indeed facilitate the creation of links between dispersed historical sources [Meroño-Peñuela et al 2015]. The Semantic Web is not a new Web that replaces the current Web, but an extension of it [Berners-Lee, Hendler, and Lassila 2001], enabling navigation not between documents via hypertext links, but directly between data. This implies that all information is provided with a well-defined meaning, making it interpretable and reusable by data consumers, both human and machine, who will then be able to process it automatically and develop new applications and services [Bermès 2013]. Museums are particularly interested in the aggregation possibilities offered by Semantic Web technologies [Freire et al 2019], especially for heterogeneous metadata [Peroni, Tomasi, and Vitali 2013]. For example, the Europeana Fashion portal provides access to the collections of over thirty European institutions. However, it primarily aggregates data from major European museums only [Suls 2017], resulting in a notable exclusion of smaller museums. This approach neglects the promotion of more specialized collections, notably those focusing on silk. To address this deficiency, the H2020 SilkNow project[1] has developed ADASilk, an innovative exploratory search engine. This engine offers federated searches across diverse cultural heritage collections, with a specific emphasis on silk-related artifacts.

ADASilk is based on a knowledge graph that enables data from a variety of sources to be aggregated and presented online. A knowledge graph is a set of interconnected descriptions of entities - objects, people, events, concepts or situations - which can be enriched by a retrospective acquisition of information and integrated into ontologies [Gruber 1993]. This approach has proven particularly successful for other types of cultural heritage data: for example, bibliographic data describing musical works for the DOREMUS project [Achichi et al 2018], archival data for the ArchOnto project [Koch, Ribeiro, and Lopes 2020], or historical data for the WarSampo project [Hyvönen et al 2021]. The SilkNow Knowledge Graph represents the data contained in the catalogues of 21 cultural heritage institutions[2] , following a subset of the CIDOC Conceptual Reference Model or CIDOC-CRM. Collected online or provided directly by the institution concerned[3], the data describe objects that may be entirely made of silk, or incorporate other textile materials (natural fibers, metal threads), or be made of composite materials (e.g. wood in the case of furniture). The SilkNow project focused on objects produced or consumed in Europe between the mid-15th to the middle of the 19th century.

In this paper, we present the different steps that allowed us to build the SilkNow data model. We also describe the methodology we used to create a conceptual model extending the CIDOC CRM, and to integrate the annotations produced with machine learning methods. In a first part, we discuss the contributions of Semantic Web technologies for the protection and enrichment of silk heritage. In a second part, we explain the approach we adopted to develop the SilkNow data model. In the third part, we show how we used the Provenance Data Model or Prov DM to integrate the semantic annotations produced on these data thanks to the analysis of texts and images. We also propose to extend the CIDOC CRM with new classes and properties, so as to be able to express the complex semantics of the data describing the production process of silk fabrics.

The benefits of using this type of technology are particularly apparent when it comes to religious textiles. Religious textile collections are preserved in a wide variety of institutions. While collections of “ordinary” textiles are generally preserved in fashion museums, they can also be found in ethnological or decorative art museums. Religious textiles are preserved not only in national heritage institutions, but also in small and medium-sized ecclesiastical museums. In addition, churches, cathedrals and convents may use these fabrics as part of their liturgy and daily devotions. There are also collections preserved by companies that produced these fabrics: such is the case of the Garín archives - one of the SilkNow consortium's partners - which conserve very rich samples of religious fabrics representative of Spanish silk manufacturing in the 18th and 19th centuries [Alba et al 2020]. The integration of the inventories produced by these structures into the SilkNow Knowledge Graph enables them to be displayed via ADASilk, thus improving the accessibility and visibility of their collections.

In some cases, small heritage structures do not have a catalog or even an inventory, which makes it extremely difficult to determine the origin of the fabrics kept there. These fabrics, widely dispersed across Europe, are even incomplete[22]. As a result, information is fragmented and disconnected, making it difficult in many cases to correctly identify and date these artifacts, and hence to catalogue them. ADASilk enables smaller organizations, particularly church museums and religious institutions, to trace the links between the objects they conserve and other artifacts held in distant collections, and to integrate them into a renewed chronology. This can help create new synergies for ecclesiastical collections, cathedrals and parishes that are custodians of this heritage [Alba et al 2020].

Small parishes and diocesan museums often lack the resources — both financial and human — to properly conserve these works. This information is also indispensable when these establishments want to restore the fabrics they conserve, as they rarely have experts capable of carrying out the appropriate research. This results in a lack of appropriate methodology and restoration that should stem from a thorough knowledge of weaving techniques, materials, design and historical and artistic links in order to better preserve these assets [Alba et al 2019]. ADASilk enables them to compare the fabrics they conserve with those in other collections, and thus obtain the information they need to launch better-documented restoration campaigns. It can be very difficult to distinguish the origin and technique of certain fabrics without applying a technical analysis which, in some cases, leads to the destruction of part of the textile [Alba et al 2020]. Comparing fabrics held in other collections - and already catalogued - is therefore particularly useful when it comes to analysing and identifying old and rare fabrics without having to damage them.

In addition to integrating heterogeneous data into the knowlegde graph, the SilkNow ontology provides a structure for automatic image recognition, facilitating the discovery of similarities between several textiles preserved in various European institutions. Image recognition performed on images from Garín's collections and images from public museums [Dorozynski, Clermont, and Rottensteiner 2019] [Clermont et al 2020] enabled [Alba et al 2020] to highlight the great similarities between an 18th-century dalmatic preserved by the Centro de Documentación y Museo Textil (CDMT) in Terrassa (Spain) and Garín's 19th-century Francia model. A comparison of these two objects through their motifs reveals that Garín drew inspiration from Lyonnais motifs for its religious textiles. These images help us to understand the link between Valencian and French motifs. On the other hand, during the 19th century, silk industries turned to revivals, reembracing ancient motifs that may have appeared obsolete and reinterpreting them to suit the tastes of each period. This complicates the process of dating and attribution. Sometimes, textiles found in geographically distant locations exhibit strikingly similar features; and the same fabric design can be observed in both secular and ecclesiastical garments. These confusing scenarios prompt researchers to question whether these designs are original or, alternatively, imitations of other textiles.

Silk heritage is of crucial importance for understanding European history and culture. But some conservation institutions lack the means to promote it and make it accessible to the public. More generally, silk heritage suffers from a lack of visibility, especially online, which has become the preferred tool for searching and discovering new information. Semantic Web technologies, however, offer the possibility of disseminating cultural heritage data more easily on the Web by improving its discoverability, but also of enriching this data by linking it to other information.

The SilkNow project wished to exploit these potentialities to enhance the visibility of the European silk heritage, by developing the exploratory search engine ADASilk dedicated to these specialized collections. ADASilk is based on a knowledge graph, whose creation is based on the SilkNow ontology - a subset of the CIDOC CRM and CRMsci ontology. The richness of the metadata contained in the knowledge graph has also led us to develop the SilkNow extension, a conceptual model for the description of ancient silk textiles. We believe that this specialized extension represents a first step in the development of a domain ontology for textile cultural heritage. Data reusers indeed play a vital role in data modeling. They can suggest collaborative and shared models to ensure this data is interoperable and completely reusable in the context of the Semantic Web. We have thus conceived the SilkNow extension as a first step towards the creation and dissemination of a new compatible model compatible with the CIDOC CRM SIG. To achieve such recognition, it is essential to extend the SilkNow framework to create an ontology that encompasses not only the results of textile manufacture, but also the places and methods of production, the circumstances surrounding this production and the memories associated with it. In addition, we have placed special emphasis in this article on the methodology we have used to create this specialized extension, so as to enable other research projects and institutions to adopt the same approach. We believe that such an approach could be systematized when the aim is to promote and protect an endangered heritage using the Web.

By creating the ADASilk exploratory search engine, SilkNow has demonstrated the value of using Semantic Web technologies to aggregate heterogeneous heritage data, provide access to it and increase its visibility. It follows in the footsteps of other projects that also aim to promote fragile and endangered heritage and demonstrates the validity of this type of approach for large-scale processing of cultural heritage data. SilkNow also illustrates the important work that needs to be done on heritage data if the objectives of the Semantic Web are to be achieved. But to bring this kind of project to fruition, a strong institutional commitment is required. Not only from the point of view of producing open, high-quality data, which is the essential fuel for this type of project, but also from an organizational and technical point of view. Such a commitment is not accessible to small and medium-sized conservation institutions [Sebastián-Lozano et al 2023]. Indeed, such projects require significant investments, and therefore remain inaccessible to small institutions, which would be the first to benefit from them. Having demonstrated the value of these approaches, it is therefore necessary to consider the development of tools to access these technologies, and thus make them accessible to the greatest number of people.

The integration of data into the SilkNow Knowledge Graph could be considerably facilitated by the use of automatic classification tools, based on the exploitation of Linked Open Data. Case studies have demonstrated the value of developing such methods for creating links (or properties) between entities [Bianchini and Bargioni 2021]. But they have also shown the need for semantic data to be available from a high-quality standard and as complete as possible [Freire, Manguinhas, and Isaac 2020] [Bianchini and Bargioni 2021]; without this, crucial aspects of the data may be left out, and links between data may thus be missing. We must be aware, however, that the diversity of cataloguing practices - and sometimes errors in metadata - as well as the inherent complexity of historical data, often unstructured or a little structured [Meroño-Peñuela et al 2015], can be a brake on this kind of approach, both from the point of view of automatic classification and the development of a knowledge graph based on CIDOC CRM.