This paper, "ASGM-KG: Unveiling the Dynamics of Alluvial Gold Mining Through Knowledge Graphs," represents the inaugural effort to systematically apply a Knowledge Graph (KG) to the study of artisanal and small-scale gold mining (ASGM). This research addresses a significant gap by constructing the first known dataset and knowledge graph dedicated to ASGM, highlighting its environmental and health impacts, particularly mercury pollution in tropical forests. This pioneering work is crucial, as ASGM is a major contributor to environmental degradation yet has been understudied in this structured and accessible format.

Resource Description Framework (RDF): At the core of ASGM-KG is the Resource Description Framework (RDF), a standard model for data interchange on the Web. RDF uses a triple-based structure to link data points, making it an essential tool for developing knowledge graphs. This framework facilitates the integration of diverse data sources into a unified, queryable structure, enabling complex relationships and dependencies within the ASGM domain to be mapped and analyzed systematically.

Our contribution through ASGM-KG is pivotal, as it introduces the first dataset that encapsulates validated and structured information about ASGM impacts, which can guide policymakers and researchers. The methodology leverages advanced Natural Language Processing powered by large language models for RDF triple extraction from varied texts, followed by a meticulous validation via our novel Data Assessment Semantics (DAS) framework. Experimentally, ASGM-KG has proven effective, aligning with verified data sources and enabling robust querying and data interpretation. The significance of our work lies in its application to environmental monitoring and policy-making, potentially serving as a blueprint for addressing similar challenges in other sectors.

The code has been written in python programming language and put into the code directory. Before running the code, please install the requirements.txt file using the following command

After installing all the required packages, ensure that the following configurations -

1. Take the session of Microsoft Copilot (pro if you have) and create a cookies.json file in the code directory.
2. Get API-key of page rank from open-page-rank website.
3. If you want to extract the information from the PDFs itself, please see the TABLE-2 of the paper and the corresponding listed urls and download the pdfs and save them in the data/pdf directory.

In the code folder, the rdf_extraction.py (as shown by Figure 2 in the paper) file contains the code of extracting the RDF format from the PDF and save the extracted information in csv format to the specified folder by the user (if not provided, the default directory is Output).

The RDF_extraction function takes three arguments - pdf_path, cookies_path, and output_filename. pdf_path, and cookies_path are required however if output_filename is not provided the default output_filename will be assigned to Output.

The DAS_validation.py is responsible for validating the extracted RDF in realtime as shown in Figure 3 in the paper using DAS_validation which takes four arguments, entity1, relation, entity2 and cookies_path. Here entity1,relation, and entity2 refer to the Subject, Predicate and Object properties of the RDF.

You need to run only the main.py file in the code directory to regenerate the experimental results using the following command

python3 code/main.py

In the main function, we are doing the following things -

1. Getting the pdfs from the data/pdf directory.
2. Setting up the file path of cookies.json file.
3. Extracting the pdfs using RDF_extraction function.
4. creating the main dataframe of extracted information from step 3 using create_rdf_dataset function.
5. Finally, we are validating each of the extracted RDF triples using our DAS Validation framework by calling the DAS_Validation function.

A sample output from the RDF extraction looks like below:

This is an example of extraction from the following text -

"The Amazon region is a unique environmental icon. Spanning more than a third of the South American continent, it contains the greatest share of biodiversity in the world. Further, it is home to more than 34 million human inhabitants, including some 3 million indigenous peoples. It is an invaluable source of water, food, shelter, medicines, and culture to these people of diverse origins, stretching back thousands of years. However, today the Amazon is under threat. Artisanal and small- scale gold mining is a prominent source of livelihoods and income in the region. These informal, unregulated operations make heavy use of mercury in the gold purification process, which is then released into the water and air. This mercury is carried far from the mines, irreversibly contaminating plants and animals throughout the region."

A sample output of our DAS Validation framework is following:

As shown in the table, the output of DAS framework contains the RDF triplet (Subject, Predicate and Object), followed by the Factual Validation based on the Logical Reasoning of LLM model applied to the top K most web references according to the Page-Rank's relevancy and coherence score sorted. For demonstration purposes and limited resource, we took the top most reference and context K as 3 which can be increased according to the available resources. Our final resultant knowledge graph on ASGM (ASGM-KG) is validated by this approach and only the factually correct information is mostly considered to be included in the graph.

As mentioned in the paper, DAS validation framework has been applied to the CoDEx: A Comprehensive Knowledge Graph Completion Benchmark small dataset (CODEX-S) that contains 1838 positive and negative triplet based on Wikidata and Wikipedia data. While applying the DAS validation framework, this framework not only outperforms the existing model on classifying the triplet mentioned in the CoDEx paper, but it also successfully indentifies the mislabeled triplet. The following table shows the results of DAS validation framework on CoDEx-S dataset -

This Table shows a summary of our framework while applying on the Codex-S test dataset describing the result of the triples from positive and negative triples indicated by ground truth (GT) column, the prediction (PR) refers the output of the framework and actual ground truth (AGT) actual label of the ground truth by cross-checking the open-source knowledge and expert opinion. The Web Reference and Logical Reasoning columns show the context of the top 3 Web references and the explanation behind the prediction of the framework respectively. Whenever the GT = AGT and PR = GT then it implies that our framework correctly labels the triple and if PR ≠ GT but PR = AGT this implies the ground truth (GT) is wrong labeled.

Following two figures illustrate the performance of our framework on without (left) and with (right) ground truth checking.