Linking Text Corpora to Lexicographical Resources Using Wikibase and OntoLex- Lemon, Artificial Intelligence Conference

Linking Text Corpora to Lexicographical Resources Using Wikibase and OntoLex-Lemon Keywords: Linked Data, Wikibase, OntoLex-Lemon, Serbian, srpELTeC Topic: Knowledge representation, reasoning, and planning; semantic networks; knowledge representation techniques This abstract showcases a novel approach to linking textual data from a corpus to lexicographical resources using Wikibase as a collaborative platform. Specifically, we present a data model and an initial use case focused on the integration of a Serbian literary corpus (SrpELTeC) in NIF (Natural Language Interchange Format) with a Serbian dictionary following the OntoLex-Lemon lexicographical framework. Our work aims to combine linguistic annotations, including morphosyntactic, semantic, and philological information, with lexicon entries, facilitating a broader linguistic knowledge graph. Wikibase, a versatile extension of MediaWiki, serves as the infrastructure behind Wikidata, one of the largest crowdsourced, queriable knowledge graphs to date (Vrandečić & Krötzsch, 2014). Our proposed model aims to explore the synergy between textual corpora and lexical data. Our proposed use case shows how a Serbian text corpus can be linked to a Serbian dictionary, specifically leveraging the OntoLex-Lemon model, which has been increasingly utilized for linguistic linked data. The SrpELTeC corpus is enriched with linguistic annotations based on the NIF Ontology (Hellmann et al., 2013), including part of speech tags, lemmas, and named entity recognition (NER) results. For this, we employed the BEaST tagger for Serbian (Stanković et al. 2020, Stanković et. al. 2022), generating annotations such as lemma and part of speech categories mapped to the OLiA (Ontology for Linguistic Annotation) lexical categories. This mapping allows for the identification of candidate lexicon entries within the SrpMD Serbian dictionary (Krstev, 2008; Stanković et al., 2018), which is structured using the OntoLex-Lemon model (McCrae et al., 2017). The smooth integration of this model into a Wikibase instance (Lindemann et al., 2023) provides an efficient means of linking corpus tokens to their corresponding dictionary entries. In addition to the morphosyntactic and lexical layers, our approach also incorporates named entity recognition. The SrpNER tool (Krstev et al., 2014; Nešić et al., 2022) was used to classify entities into seven predefined categories: PERS (persons), ORG (organizations), LOC (locations), ROLE, EVENT, DEMO, and WORK. These categories were then mapped to OLiA, Wikidata, and DBpedia entities, enabling more comprehensive knowledge graph integration. Our data model aims to link corpus tokens to several key components: 1) Lexeme Nodes (a corpus token is linked to a lexeme node; each lexeme has a unique Part of Speech and a canonical citation form as lemma), 2) Lexical Forms (inflected forms with a description of their grammatical features), and, eventually also to Lexical Senses, i.e., to instances of the three OntoLex core classes. Integration of word senses and named entities directly into the Wiki framework allows for seamless alignment between corpus tokens and structured knowledge graphs such as Wikidata, creating a rich, interconnected resource that can be queried and expanded collaboratively. The presentation will provide a visual representation of the data model as graph visualization, along with textual explanations of each component. Additionally, sample SPARQL queries will illustrate how tokens are annotated, linked to dictionary entries, and associated with Wikidata entities. We aim to demonstrate how a representation according to the model we propose not only facilitates the creation of a multilingual linguistic knowledge graph but also serves as a proof of concept for linking lexical resources across diverse languages and source formats. As for possible applications of such an interlinked resource, for example, AI systems may be fine-tuned using rich corpus annotations. Lexicographical data, like word senses, lemmas, and grammatical features, help AI systems disambiguate meanings and understand context. This enhances Natural Language Understanding tasks such as text summarization, question answering, and semantic search. Trusted lexical knowledge can contribute to the reliability to algorithm-based tools, especially relevant in the legal or the medical domain.