AIKoGAM: An AI-driven Knowledge Graph of the Antiquities Market: Toward Automatised Methods to Identify Illicit Trafficking Networks

1 Introduction

The antiquities market, both legal and illegal, is a widely acknowledged phenomenon, prompting significant attention from national governments and international bodies aiming to regulate it. This commitment is reflected in dedicated funding programs aimed at comprehending and addressing the issue.

As emphasised by Graham et al. (2023), there is a growing trend towards digital and technological solutions, evidenced by European Commission funding for several initiatives. These include RITHMS (2023), a Multiplex SNA-based platform, and projects employing 3D photogrammetry, AI for border control (CORDIS 2022a), chemical marking, miniature devices, blockchain against illicit activities (CORDIS 2022b), and technologies for the identification, traceability, and provenance research of cultural goods (CORDIS 2022c; Patias & Georgiadis 2023).

Parallelly, the debate surrounding the antiquities ownership, market and cultural rights persists (Campfens 2020; Cuno 2012; Hixenbaugh 2019), with efforts to involve art market representatives in decision-making processes.

Despite claims by collectors, dealers, and auction houses of adhering to ethical practices, UNESCO (2022b) data indicates a shortfall in meeting standards for diligently vetted objects with solid ownership histories. The survey revealed a significant lack of due-diligence policies (15%), with the 33% neglecting provenance research for lower-priced items, and the 54% not giving attention to objects originating from conflict zones, leading to a market with a wide range of unlawfully sourced artefacts.

The 2022 ‘Operation Pandora VII’, executed by Law Enforcement Agencies (LEAs), underscores this, with 8,495 online checks leading to the confiscation of 4,017 illicit goods (EUROPOL 2023; Giovanelli 2023). However, as Brodie et al. (2022: 117) argue, ‘if the reported arrests and seizures are only the tip of a much larger dark criminal iceberg, then the indications are that the illicit trade in cultural objects is very poorly controlled’. Nonetheless, it is worth noting that most of the seizures and identifications were made possible through crosschecks against LEAs databases.

While databases of stolen goods exist (CECOJI-CNRS and EC-DGHOME 2019: 82–89; Oosterman 2019), their effectiveness is limited when dealing with illegally excavated items lacking prior documentation (Rush & Millington 2015; EC-DGEAC et al. 2019: 177; Roodt 2016: 221). These databases prove inadequate in such cases, with a few exceptions in which LEAs have managed to seize written and photographic evidence of the unlawful origin of cultural goods. Noteworthy instances include the integration of the ‘Medici dossier’ within the Italian Carabinieri’s LEONARDO database (Watson & Todeschini 2007; Gill & Tsirogiannis 2016; CC-TPC 2023b; Pellegrini 2023) and the seizure of Almagi`a’s Archives by the New York County District Attorney (Bogdanos & Iyer 2021; Giovanelli 2021b). In such scenarios, the monitoring of marketplaces by both LEAs (CC-TPC 2023a) and archaeologists/experts (Gill & Tsirogiannis 2016; Tsirogiannis 2021) has proven instrumental in successfully identifying looted cultural goods and eventually enabling their recovery.

Significant gaps persist in understanding the antiquities trade, particularly the ‘demand’ side (Brodie et al. 2022). While we concur with Brodie et al. (2022: 121) that many cultural goods appearing on the ‘nearly infinite number of storefronts for cultural objects’ being declared as looted stem from old illegal excavations, these marketplaces harbour valuable data yet to be fully utilised. Galleries, auction houses, and dealers catalogues and websites form extensive archives within a big data framework (Boyd & Crawford 2011), offering an overview, albeit partial (Brodie et al. 2022; Yates 2022), of the trade’s current state and historical evolution.

Provenance statements, critical in investigations and in understanding trade dynamics, are increasingly scrutinized for their quality and verifiability (Albertson 2020a; Brodie 2015; Davis 2020; Fabiani & Marrone 2021; Fletcher 2020; Gill & Tsirogiannis 2016; Giovanelli 2018; Levine 2009; Mackenzie & Yates 2020; Tsirogiannis 2015a; Tsirogiannis 2015b; Tsirogiannis 2016; Tsirogiannis 2021). Provenance is crucial for unravelling ‘an object’s trajectory as it changes legal status while crossing international borders’ (Gerstenblith 2020), and the examination of the ‘provenance given by the art dealer’ (UNESCO 2021) is a critical step in initiating a recovery procedure, as potential evidence to support restitution claims (UNESCO 2022a).

Acknowledging the significance of grasping market dynamics, the pivotal role of provenance, the growing emphasis on digital progress, and the challenges and resolutions emerging from extensive data exploration in the humanities (Smith 2018), this paper proposes a semi-automated system combining Natural Language Processing (NLP), Machine Learning (ML), and Social Network Analysis (SNA) to construct a Knowledge Graph (KG) of the antiquities market. Our aim is to generate fresh insights into the antiquities market dynamics and its actors, incorporating Artificial Intelligence (AI) technologies to correlate the market with existing trafficking networks knowledge.

2 Background: Antiquities Trade as a Complex Network

3 Methodology

In line with current trends in computational research on the illicit trafficking of cultural property, our work focuses primarily on researching, exploiting, and refining various ML tools and algorithms. Our goal is to develop a semi-automated, interoperable, and accessible methodology that can benefit other research groups and LEAs alike.

This research initiative, named AIKoGAM (Giovanelli & Traviglia 2023), aims to establish a common ontology following the LOD and Findable, Accessible, Interoperable, and Reusable (FAIR) (Wilkinson et al. 2016) principles, and seeks to create an open data environment that fosters knowledge generation from existing resources and encourages data enrichment, sharing, and reuse.

3.1 Information extraction

3.1.1 Web harvesting

Online data accessibility for antiquities in the market has increased since Christie’s introduced digital access to bidding in 2006 (Christie’s 2012), further boosted by Sotheby’s partnership with eBay in 2014 (eBay Inc. Staff 2014). Consequently, auction houses, galleries, and dealers have increasingly adopted digital platforms to showcase their catalogues.

The Information Retrieval (IR) module in AIKoGAM starts with identifying and examining data sources, including selecting target websites, locating pertinent data linked to archaeological objects and provenance, comprehending the structure, quality, and nature of these data, scrutinising data source usage policies, assessing data availability, considering access restrictions, and formulating a scraping strategy.

For auctions, specific sales and corresponding lots can usually be located through a straightforward website architecture. This typically involves navigating from the ‘Home page’ to the ‘Past auctions page’ or the ‘Future auctions page’, leading to a ‘Sale page’ (Figure 1) and finally reaching the ‘Lot page’. For galleries, the website structure may vary, but it usually involves browsing through an ‘inventory page’ (Figure 2) displaying all the objects or categorised pages (Egyptian, Roman, Greek etc.) listing objects within each respective category.

Figure 1

Figure 2

Although these web pages appear straightforward and present the relevant information about the objects in a human-readable fashion (see Figure 3), underlying Hyper-Text Markup Language (HTML) reveals significant differences in structure among websites. Some adopt structured graph databases, making data retrieval easy for both users and software thanks to their contextual schemas. Conversely, many websites rely on HTML elements (such as headings, paragraphs, lists, etc.) combined with Cascading Style Sheets (CSS) selectors for content organisation. CSS selectors enable selective styling of HTML elements or groups of elements, enhancing the visual appeal and layout.

Figure 3

Certain web pages are well structured, adhering to a consistent layout and appearance. They use designated HTML elements and CSS classes to maintain this predictability. Structured web pages facilitate information retrieval because data is clearly organised in an anticipated manner: product listings are structured using ‘<div>’ elements for individual products, simplifying data extraction using CSS selectors.

Some others do not have a strict structure, resulting in a free-form layout. Retrieving information from such pages can be challenging as content is not neatly organised. Information may be embedded within paragraphs or other textual content, necessitating more complex extraction methods.

Sotheby’s utilises an Apollo GraphQL Database for systematic data organisation, allowing straightforward data access via the database Application Programming Interface (API), or through uncomplicated scraping methods. Christie’s has a different structure where certain information like object names, estimates, or image URLs are retrievable via API calls, while other details, such as provenance, literature, description, and exhibitions, require HTML element detection with specific CSS selectors.

To handle these differences, the ‘harvesting.py’ script in Python was developed, utilising Beautiful Soup 4 (Richardson 2019) and Requests (Chandra & Varanasi 2015) libraries for structured data collection. Initially, manual fine-tuning is required to maximise data collection automation and minimise manual cleaning and restructuring of the data afterwards.

3.1.2 Ontology mapping and Knowledge Graph database building

To maximise interoperability and applicability across various websites, the harvesting module does not directly produce a uniform database due to varying keys in the harvested dictionaries. We developed an ontology and a vocabulary to map harvested data to specific entities and relationships. This facilitates future AI automation and reduces the need for manual cleaning. The vocabulary ensures the mapping of different keys pointing to the same logical instance, which lays the foundations for the KG and future interoperability with other ontologies and LOD.

For AIKoGAM, we adapted existing ontologies for general entities and proposed a new ontology for provenance reasoning, capturing both the synchronic relationship of each owner to the objects and the diachronic dimension (Figure 4).

Figure 4

A single provenance statement is treated as an event involving the object, the seller, the buyer, and the subsisting market event (such as an auction sale or the presence in a specific market environment, inferred from statements like ‘Bought by the present owner in London antiquities market, 1980’). To build a Neo4j database instance with this schema, we developed the ‘mapping.py’ script. This script employs data-cleaning techniques and regular expressions and generates unique identifiers for each sale and lot utilising the Secure Hash Algorithm 512, which produces a fixed string of characters representing the data. Thanks to the algorithm’s collision resistance properties (USDOC-NIST 2015), we can guarantee uniqueness and adhere to data indexing constraints.

To enrich the provenance information of the objects, we have devised the ‘event_extractor.py’ script that leverages the Spacy library NER models (Montani et al. 2020) to split provenance statements into ‘events’. Each of these events ideally contains an actor, a location, and a time-span indication. Recognising linguistic diversity within our dataset, our ‘detect_language_batch’ function uses the LangID Python library (Heafield, Kshirsagar & Barona 2015), a high-speed language identification tool pre-trained across 97 languages. This function associates the language identifier with the corresponding SpaCy NER model, facilitating the categorisation and storage of the relevant SpaCy model variable within each statement. The script systematically processes each provenance statement, extracting NER entities based on the identified language. To address discrepancies in entity labels across distinct SpaCy models, we pass the recognised entities through a constant dictionary that maps entities to specific standardised labels, ensuring uniformity in the output. The script then stores within each dataset entry the various ‘events’ in which the object participated. This information is structured as shown in the following example: the provenance statement ‘Sotheby’s Hong Kong 04 August 2022 lot 1095’ is processed by the ‘detect language batch’ function into the dictionary:

{‘text’: ‘Sotheby’s Hong Kong 04 August 2022 lot 1095’, ‘detected language’: ‘en’,

‘spacy model’: ‘en_core_web_md}.

Subsequently, The ‘event_extractor.py’ script stores the event within the entry as:

{‘label’: ‘Sotheby’s Hong Kong 04 August 2022 lot 1095’, ‘ORG’: ‘Sotheby’s Hong Kong’,

‘DATE’: ‘04 August 2022’,

‘NUMBER’: ‘1095’}.

The ‘kg_construction.py’ script then builds into a desktop instance of a Neo4j (2023b) the enhanced KG by automatically extracting events from each object and constructing new nodes based on them. These event nodes are connected to the corresponding artworks through a ‘PARTICIPATED_TO_EVENT’ relationship.

Through CYPHER queries (Neo4j 2023a), the native query language for Neo4j, the named entities are extracted from the event nodes properties and generates new nodes as specified in the ontology schema depicted in Figure 4. The event node remains central, receiving relationships from the artworks and projecting or receiving new relationships with other relevant entities.

To evaluate the general SpaCy models accuracy, we checked the results of a dataset sample (976 items). Of the 2671 entities extracted from these 976 items, 94 were incorrectly identified, resulting in an overall accuracy of approximately 96.49% (considering a correct identification if a collection name or an organisation name is labelled under ‘PERSON’ tag and vice versa, given the general tendency of having personal names in organisation names and the high chance of error. Given that our ontology mapping considers both ‘PERSON’ entities and ‘ORG’ entities as ‘actor’ entities, we consider both identifications as correct).

Examining the error rates across different models reveals variations in performance (see Tables 1 and 2).

Table 1

Error Rates per Model.

MODEL	ERROR RATE (%)
en_core_web_md	3.35
zh_core_web_sm	6.17
de_core_news_sm	6.25
es_core_news_sm	6.67
fr_core_news_sm	11.46
it_core_news_sm	16.67
pl_core_news_sm	20.00
nl_core_news_sm	32.36
lt_core_news_sm	33.33
pt_core_news_sm	33.33
da_core_news_sm	66.67

Table 2

All Errors and Their Counts.

PREDICTED	TRUE	COUNT
ORG	OTHER	17
PERSON	OTHER	16
GPE	OTHER	11
GPE	PERSON	10
PERSON	GPE	4
ORG	GPE	4
FAC	PERSON	4
GPE	ORG	4
ORG	DATE	3
NORP	OTHER	3
NORP	PERSON	2
GPE	DATE	2
FAC	ORG	2
FAC	GPE	2
MONEY	OTHER	1
PRODUCT	ORG	1
PRODUCT	GPE	1
NORP	ORG	1
PRODUCT	OTHER	1
WORK OF ART	ORG	1
NORP	DATE	1
WORK OF ART	DATE	1
WORK OF ART	OTHER	1

The most frequent error involves entities better identifiable as ‘OTHER’ being misclassified as ‘ORG’, occurring 17 times. This pattern suggests a challenge where entities difficult to categorise are seen as organisations, indicating potential areas for model improvement.

Analysing misclassifications by entity type, ‘PERSON’ and ‘ORG categories are often misclassification of ‘OTHER’ entities, while ‘GPE’ misclassifies ‘PERSON’ and vice versa. This highlights challenges in accurately identifying individuals and geopolitical entities in specific contexts.

While the overall accuracy of the NER models is commendable, there are notable variations in performance across different models, with specific challenges in accurately classifying certain entity types. Understanding these nuances is crucial for refining and optimising the models to enhance their performance in real-world applications.

4 Preliminary Results

Through the transformative process, we evolved from an original database with 30,954 entries (30,422 objects, 529 sales, and 3 entities) to a KG database containing 111,979 nodes and 237,394 relationships. This enhanced KG offers a detailed representation of provenance data, enabling us to extract valuable insights.

Figure 5 summarises the node and relationship counts in the KG: 13,482 actors involved in selling or acquiring 30,422 antiquities dealt with by Christie’s, Sotheby’s, and Phoenix Ancient Art from 1999 to 2023.

Figure 5

After a manual examination of the data, duplicate occurrences in the ‘actors’ became apparent. We introduced the ‘similarity.py’ script with a ‘find_similar_nodes’ function, identifying similar ‘actor’ nodes. The function stores pairs of nodes with similar names along with their similarity score. The ‘ngram_index’ function generates the index for identifying similar strings by decomposing them into smaller units known as n-grams (Krátký et al. 2011). The function retrieves the name of each ‘actor’ node, normalises it, and prepares it for subsequent comparisons. Within the loop that iterates over all the nodes again, the function systematically compares the normalised names of the two nodes. With an exact match the function categorises the two nodes as similar and the shorter of the two names is added to the list of merged strings. When the normalised names do not align exactly, the function employs fuzzy string matching (SeatGeek 2023) to calculate the similarity score. If this score exceeds the 0.9 threshold, the node pair is appended to the list of similar nodes. For instances where the similarity score is a perfect match (score 1.00), a similar consolidation process to the exact match scenario is executed.

The ‘semi_auto_evaluation.py’ script addresses occurrences of nodes representing the same entities but perceived as different due to variations in naming conventions. The comparison is performed on single words composing the string of the node name, through the ‘process name’ function. The automated portion of the script identifies nodes with single words that match precisely and employs fuzzy matching (threshold 0.72) to calculate similarity scores. This accommodates cases where names are similar but not identical, addressing variations like ‘thomas’, ‘Thmas’, and ‘Tommaso’. Where exact fuzzy matches are not found, the script checks for length consistency to filter out minor discrepancies and considers a specific lesser similarity ratio (0.68 to 0.72). This is needed due to the heterogeneity of the data sources and languages, where substantial changes in name representation may occur. The script recognises potential matches by comparing trailing names, a mechanism useful for dealing with variations in naming conventions such as differences in titles, initials, and other suffixes, as well as residual contextual information. In cases where no duplicates are identified, the output is automatically generated. When matches are identified, the script prompts the user for manual confirmation based on the presented potential matches. This human intervention enables the mapping of nodes with strongly diverse representations. As a result, a dictionary is stored with the first encountered name as the key and the names to be merged as values. Subsequently, the head node is connected to the neighbours of the nodes to be merged, transferring their relevant edges and attributes, and then they are removed from the graph. This ensures the preservation of the graph structure and essential relationships during the clean-up.

The process identified 2,540 potentially similar nodes, with 791 (5.87% of the total actors) automatically merged. The ‘semi_auto_evaluation.py’ implementation facilitated the merging of an additional 907 actors (6.72% of the total actors), averaging 2.2 duplicates per node. This process resulted in a total of 1,698 nodes representing alternative spellings of the same entity. Nodes mislabelled or too general accounted for 523. These nodes were removed from further projections and re-labelled onto the Neo4j graph instance. In summary, from the initial 13,482 actors identified by SpaCy, our refined scripts and processes reduced their number to 11,479 single actors.

For SNA on the KG, we experimented with an artworks-actor bipartite projection (Newman 2003; Liew et al. 2023). The top centrality degrees (Freeman 1978) in Figure 6 show that aside from Christie’s and Sotheby’s there are other notable actors. A prominent figure is Stephen Junkunc, a well-known collector of Chinese art, whose collection is highly desirable and has gained prominence in the market (Sotheby’s 2019). Notably, in 2018 Sotheby’s withdrew the sale of a statue of Buddha suspected to be a looted relic from the Longmen Grottoes and coming from his collection (The Value 2018).

Figure 6

Noteworthy names already connected with cases of illicit trafficking of antiquities appear, such as Royal Athena (Galleries), N(icholas) Koutoulakis, and E(lie) Borowski, a piece of evidence that very well aligns with the finding of Fabiani and Marrone (2021), whose datasets cover sales up to 2015. These three actors have been significantly involved in numerous cases of cultural object trafficking, as documented in previous works (Chasing Aphrodite 2012; Gill 2019; Tsirogiannis 2021).

To gain a deeper understanding, we build an additional layer, generating a monopartite undirected weighted network of actors (see Figure 7). In this process, we excluded the nodes representing Christie’s and Sotheby’s to avoid bias in our analysis. This allowed us to focus on the relationships among other actors within the network. Removal of out-of-scale high-degree nodes does not lead to network fragmentation, as demonstrated in previous research (Nguyen et al. 2021).

Figure 7

We refactored the artworks connected to actors in our previous layer, transforming those nodes into undirected edges between actors (Ramasco & Morris 2006): an edge between actori and actorj has thus weight wij equal to the number of artworks both dealt.

The monopartite projection, as indicated by the metrics in Table 3, appears relatively large, characterised by significant complexity and interconnection. The Average Degree (AD) suggests moderately linked individuals, fostering a decentralised pattern of influence. Centrality measures reinforce this, indicating a lack of concentrated control. Substantial mutual connectivity among entities facilitates efficient communication, as reflected in the Average Closeness Centrality (AC). A relatively short Average Path Length (APL) and a high Average Clustering Coefficient (ACC) highlight tightly interconnected clusters with cohesive subgroups. Despite sparse overall connections (GED), the network has a substantial diameter (GD), with long-distance connections contributing to the overall complexity.

To better understand the intricacies of the network, we employ various centrality measures. Figure 8 showcases scatter plots and linear regression lines, offering a clinical exploration of the distribution of nodes of the network’s inherent structure. The network manifests a moderate closeness, emphasising closely linked nodes, especially at lower centrality levels (Figures 8b, 8d, 8f). Parsing the correlations between centrality measures provides deeper insights into the roles and relationships among network entities. The substantial positive correlation (Pearson’s r of 0.715) between degree and betweenness centralities (Figure 8a) underscores the pivotal roles of highly connected nodes. Nodes with elevated degrees are crucial intermediaries, fostering efficient communication pathways and facilitating the seamless flow of information across disparate sections of the network. This interconnectedness is supported by the moderate relationship (Pearson’s r of 0.268) between degree and eigenvector centrality (Figure 8c). Nodes with high-degree centrality are likely, although not necessarily, to exhibit high eigenvector centrality, solidifying their status as influential ‘hubs’ adept at efficiently connecting with other significant nodes.

Figure 8

The predictive relationship (z-score of 33.20) observed in betweenness centralities (Figure 9c), along with a low standard deviation and an average of 0.00074, indicates a consistent pattern with a modest prevalence of intermediary nodes, underscoring the significance of specific nodes playing crucial roles. Beyond being well-connected, these nodes also tend to forge connections with other influential nodes, solidifying their central position in the network.

Figure 9

In real-world networks, nodes with high betweenness centrality can act as ‘bottlenecks’ for information or resources, playing a crucial role in maintaining efficient communication and coordination (Goh, Kahng & Kim 2003). The minimal correlation observed between betweenness and eigenvector centrality (Figure 8e) suggests that nodes acting as intermediaries do not necessarily have a high influence on the broader network. Seemingly, the positive correlation (Pearson’s r of 0.271) seen in Figure 8d implies that nodes facilitating communication pathways have reasonably efficient access to the broader network. The interplay of these centrality measures, as revealed by variations in Pearson’s correlation coefficients across scatter plots, underscores the nuanced dynamics within the network. Positive correlations suggest a general trend of nodes with higher centrality in one measure exhibiting higher centrality in others. Importantly, these correlations extend beyond mere structural relationships, as evidenced by the scattered distribution of data points (Kanyou, Kouokam & Emvudu 2022).

Tables 5, 6, 7, and 8 provide a comprehensive overview of the top 50 nodes based on their centrality values. Additionally, Table 4 presents nodes that consistently appear between the top 50 centralities.

Revisiting previously discussed actors and introducing other well-known figures, Table 4 sheds light on the actors’ dynamics. Robin Symes, whose ‘role as a liaison for antiquities smuggling networks across the world spans multiple decades and jurisdictions’ (Akers 2023), emerges as a key player. His centrality scores of 0.300 (closeness), 0.024 (betweenness), and 0.011 (degree) affirm his influential role and position as a bridge or intermediary within the network. The efficient flow of information or transactions reflected by Symes’ central presence within the network very well aligns with the global influence he once had as a dealer (Frammolino & Flech 2005). Fortuna Fine Arts, with a degree of 0.011, betweenness of 0.017, and closeness of 0.284, is another actor, well-known due to its recurrent associations with illicit antiquities cases (Albertson 2020b), that stands out: while its degree centrality suggests a moderate level of direct connections, the betweenness centrality indicates a role in facilitating the flow of information between other nodes. The closeness centrality suggests a relatively efficient connection to other nodes in the network, emphasising its importance in the overall structure. Recent legal developments, including charges related to provenance forgery and impersonation of deceased collectors (Rogers, Vere-Hodge & Dimmek 2021), emphasise the nature of its engagements.

Similar to Fortuna Fine Arts’ in centrality scores is N(icolas) Koutoulakis: both entities exhibit a moderate degree centrality and a high closeness centrality, suggesting their involvement in connecting diverse entities, while in betweenness centrality N Koutoulakis has a significantly higher value, suggesting a more critical role in connecting different parts of the network and controlling the flow of information. Royal Athena (Galleries), highlighted for its high degree in the bipartite projection (see Figure 6) and its connection to illicit antiquities cases, maintains prominence with a substantial closeness centrality of 0.329, a high betweenness of 0.072 and one of the highest degree centralities (0.029). Royal Athena appears closely connected to other entities, facilitating efficient and direct interactions and playing a role of intermediary influencing the network dynamics. With one of the highest degree centralities, Royal Athena exhibits extensive direct connections, emphasizing its influential position. This suggests that Royal Athena is a central figure in terms of connecting with other entities and is involved with a notable number of direct relationships.

The eigenvector centrality offers unique insights into the structural dynamics of the network (Table 8). Unlike other centrality measures, eigenvector centrality places equal importance on the quality of connections, specifically those linked to other influential nodes. Among the top 50 nodes, a distinctive pattern emerges where individuals such as scholars (e.g., Salomon Reinach, Adolf Michaelis, and Gavin Hamilton) and entities not exclusively characterised as dealers but rather as figures intricately involved with artworks in other ways are featured. This observation reinforces the role of experts within the network, suggesting that their influence is not solely derived from transactional engagements but also from their connections with and knowledge of other influential nodes (Lawler 2005; Brodie 2009; Brodie 2011). Their significance extends beyond transactional networks, encompassing the broader and more nuanced web of interactions that shape the cultural heritage landscape.

5 Discussion and Future Work

The preliminary results of our work demonstrate the high potential of describing and investigating the antiquities market using AIKoGAM’s methodology. Further analysis of ego networks of specific actors, encompassing all the possible layers that the KG offers, might lead to new knowledge and to transaction patterns previously unnoticed, as demonstrated by Graham et al. (2023).

While the potential of this application is evident, there are still several issues we need to address. General models like the one provided by SpaCy fail to identify certain entities correctly, leading to erroneous entities and relationships, especially with non-English statements. For example, sometimes the adopted model recognises entities such as ‘Lyon’ as actors in our network, instead of as locations. One possible solution is to manually tag a large dataset of provenance statements for NER and Relationship Extraction (RE) and fine-tune transformers models, such as RoBERTa for RE and DeBERTa-CRF for NER due to their performance on short texts (Wang et al. 2021), in order to make them domain-specific.

Another approach, as demonstrated by Graham, Yates, and El-Roby (2023), involves developing prompts for LLMs. Within our specific work, where we often encounter juxtaposed short phrases lacking evident contextual clues, this approach has varied success rates. Some prompts adeptly extract entities and links, while others stumble when devoid of supplementary contextual data. A promising solution could involve the training or fine-tuning of a LLM using domain-specific literature.

As demonstrated, our research faced a challenge due to diverse raw data in terms of quality, language, and structure, specifically regarding inconsistencies in the representation of the names of individuals or organisations. With the developed scripts, we achieved significant improvements in data quality and accuracy by integrating automated scripts with manual inspection. This allowed us to detect and resolve duplicate occurrences, showcasing a deep understanding of the complexities arising from naming conventions and linguistic variations across various datasets. The application of advanced techniques along with flexibility in handling diverse spellings, played a crucial role in identifying nodes that likely represented the same ‘actor’ entity. Human-assisted fine-tuning enhanced the precision of the KG, while building a mapping dictionary for future automatising.

Looking ahead, the mapped dataset serves as a valuable resource for training and refining advanced NLP models, and opens avenues for more detailed and insightful analyses of the KG, enabling trend predictions based on enriched entity relationships.

The scripts can serve as templates for methodologies in similar projects dealing with diverse and heterogeneous data. They can be adapted to handle complex datasets where fully automated processes may encounter challenges. AIKoGAM modules are versatile and can be adjusted and expanded for similar projects, ensuring a balance between automated efficiency and human expertise.

The semi-automated data collection shows promise, but initial fine-tuning of each different website’s HTML structure remains necessary. Research is ongoing on the feasibility of applying AI language models to HTML parsing and unsupervised information extraction (Gur et al. 2023; Zhang et al. 2023). Ideally, galleries and auction houses should adhere to precise, structured, and shared open-data ontologies to enable the creation of a real KG of circulating artefacts. This is definitely a path that could be pursued at the political and decision-making level.

A promising avenue of exploration would involve merging other similar datasets, as the one discussed in Fabiani and Marrone (2021), which encompasses data from external auction catalogues and extend over a wider time frame. Lastly the KG would benefit from the implementation of Computer Vision (CV) and embedding of visual data (Eyharabide, Bekkouch & Constantin 2021). By combining textual and visual data, similarity calculations, clustering and identification of the same objects/highly similar objects described differently and/or with misaligned provenance information would be possible with less human inspection. This implementation would facilitate integration with external image-based databases (such as LEONARDO or INTERPOL Databases) as well.

5 Conclusions

AIKoGAM stands as a substantial stride in comprehending the antiquities market and combatting the illicit trafficking of cultural artefacts. Its innovation lies in the utilisation of different automated and semi-automated features to establish and analyse a domain-specific KG. While the initial outcomes offer promise, the genuine potential of this tool resides in its capacity to establish a framework for the subsequent implementation of advanced ML algorithms.

Adopting general-purpose NER models for automated information extraction, AIKoGAM has successfully collected data from various sources and built an interconnected KG, even though with some errors and inconsistencies that needed manual inspection. Further improvements in domain-specific models will be fundamental to refine the NER results and to accurately automatise the extraction of relationships.

A key aspect of AIKoGAM is the potential to foster cooperation and interoperability with related tools and initiatives. The depth and breadth of the KG, as well as the new knowledge that can be generated, would benefit from the integration of previous works described in Section 2, as well as other existing or future domain-specific databases. Standardised data formats and ontology mapping would enable seamless integration and cross-referencing of data from a variety of sources and external databases, and this is something that should involve market representatives.

The refined dataset becomes a valuable asset for a wide range of applications, from training machine learning models to constructing KGs, while the scripts contribute proven methodologies for addressing intricacies in naming conventions and linguistic variations.

In conclusion, AIKoGAM’s methodology provides a foundation for future work in combating illicit trafficking of cultural property. As fast-paced advancements are made in NER, data integration, CV, and LOD, AIKoGAM has the potential to become a tool for researchers and LEAs, fostering a more transparent and secure cultural heritage landscape. Continued research and collaboration will be key in refining the methodology, and enabling the development of advanced big data models that can analyse, predict, and prevent the illegal trade of cultural heritage.

Data Accessibility Statement

The scripts that support the findings of this study are openly available in Zenodo at https://doi.org/10.5281/zenodo.8184155.

Competing Interests

The authors have no competing interests to declare.