Towards Semantic Interoperability: Parallel Corpora as Linked Data Incorporating Named Entity Linking

Ranka Stanković, Milica Ikonić Nešić, Olja Perisic, Mihailo Škorić, Olivera KitanovićCOLING 2024

TL;DR

This paper presents a method for transforming parallel corpora into Linked Data using the NLP Interchange Format (NIF). This approach enhances semantic interoperability, allowing for richer cross-lingual analysis and integration with existing knowledge bases like Wikidata, ultimately benefiting digital humanities research and language technology development.

What Problem Does This Paper Address?

The paper tackles the challenge of making parallel corpora, which are crucial for multilingual studies and language technology, more accessible and interoperable. While parallel corpora exist, they are often siloed and difficult to integrate with other linguistic resources or knowledge graphs. The authors aim to bridge this gap by representing parallel corpora as Linked Data, enabling semantic enrichment and facilitating cross-lingual research, particularly for under-resourced languages. A specific challenge highlighted is the complexity of Serbian morphology (e.g., noun declensions) compared to Italian, making direct cross-lingual alignment and named entity recognition more intricate.

Methods

The core methodology involves several key steps:

  • NLP Interchange Format (NIF): The authors leverage NIF, a community standard for linguistic annotation, to represent textual data as RDF graphs. NIF's ability to handle string-based URIs and accommodate multilingual annotations is central to their approach.
  • Named Entity Recognition (NER) and Linking (NEL): Standard NER models were applied to identify persons, locations, and organizations in both Italian and Serbian texts. Crucially, these identified entities were then linked to Wikidata, providing standardized identifiers and enriching the data with external knowledge.
  • Data Conversion to RDF: The annotated parallel corpus was converted into RDF triples using the rdflib Python library. This involved creating classes for monolingual and bilingual corpora, sentences, words, and named entities, mapping them to NIF and other relevant ontologies (like OLIA, DBpedia, and Wikidata).
  • Cross-Lingual Alignment: For linking translation equivalents at the sentence level, the SKOS closeMatch property was employed, indicating that sentences are sufficiently similar to be used interchangeably.

Datasets

  • Italian-Serbian Parallel Corpus (It-Sr-NER): This corpus, comprising 10,000 aligned segments (sentences) from ten different novels, served as the primary dataset. For the case study, a subset of 1,000 aligned sentences was used.
  • Wikidata: Used as the target knowledge base for named entity linking.
  • Existing NLP Models: The paper mentions the use of pre-trained models like spaCy for Italian and the Jerteh-355-tesla model for Serbian NER.

Key Insights and Ideas

  • Semantic Interoperability through Linked Data: The paper strongly advocates for Linked Data principles as a way to achieve true semantic interoperability for linguistic resources, moving beyond simple data exchange to shared, well-defined meanings.
  • NIF as a Unifying Standard: NIF proves to be a flexible and powerful format for representing diverse linguistic annotations, including those from parallel corpora, making them compatible with the Linked Data ecosystem.
  • Bridging Digital Humanities and NLP: The work demonstrates a practical approach to integrating NLP technologies (NER, NEL) with digital humanities research by making complex linguistic data (parallel corpora) accessible and queryable through Linked Data.
  • Wikidata as a Central Hub: Linking named entities to Wikidata provides a crucial layer of standardization and enrichment, allowing for cross-lingual comparisons and deeper contextual understanding.

Key Results

  • The paper successfully transformed a subset of the Italian-Serbian parallel corpus into NIF-based RDF, creating a linked data representation.
  • Named entities (persons, locations, organizations) were identified and linked to Wikidata, enriching the corpus with external knowledge.
  • SPARQL queries were demonstrated to retrieve aligned sentences and named entities from the linked data representation, showcasing the queryability and interoperability of the data.
  • The authors highlight the benefits for multilingual research, translation studies, and cross-lingual information retrieval.

Limitations and Future Work

  • Verbosity of Transformation: The authors acknowledge that NIF, while powerful, can lead to a high degree of verbosity, increasing file sizes. They provide Table 3 to illustrate the size differences between TSV and NIF files, emphasizing the need for careful planning regarding scalability.
  • Documentation of NIF: The paper notes that the latest version of the NIF Ontology (2.1) lacks full documentation, which is why version 2.0 was used.
  • Future Work: The paper mentions plans for further research, including NEL model training and the publication of all 10,000 aligned segments in NIF.

Summary updated

BibTeX
@inproceedings{stankovic2024towards,
  author = {Ranka Stanković and Milica Ikonić Nešić and Olja Perisic and Mihailo Škorić and Olivera Kitanović},
  title = {{Towards Semantic Interoperability: Parallel Corpora as Linked Data Incorporating Named Entity Linking}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
}
Figure