TL;DR
This paper introduces LiLa, a Linked Data-based Knowledge Base designed to harmonize diverse lemmatization strategies for Latin linguistic resources. By creating a central hub for lemmas, LiLa aims to make previously siloed resources like corpora, lexica, and NLP tools interoperable, enabling richer cross-resource querying and analysis.
What Problem Does This Paper Address?
The field of Digital Humanities and Computational Linguistics for ancient languages like Latin faces a significant challenge: a lack of interoperability between existing linguistic resources. While there's a growing wealth of digitized Latin texts and various tools for their analysis, these resources often employ different annotation schemas and conceptual models. This incompatibility, particularly in how words are lemmatized (reduced to their base or dictionary form), severely limits their potential for combined use. For highly inflected languages like Latin, where a single word can have numerous forms, harmonizing lemmatization is crucial for enabling researchers to effectively query and link diverse datasets, from historical texts to lexical databases and NLP tools.
Methods
The core of LiLa's approach is building a Knowledge Base using the Linked Data paradigm. This involves representing linguistic data as a network of interconnected resources, using Semantic Web standards like RDF and ontologies.
Key methodological aspects include:
- Lemma as a Pivotal Node: Lemmas are central to the LiLa Knowledge Base, acting as the primary link between different types of linguistic information (lexical, morphological, syntactic, etc.).
- Ontology-Driven Design: LiLa leverages existing ontologies like OLiA (for linguistic annotation) and Ontolex/Lemon (for lexical resources) to describe the data. This ensures compatibility with other Semantic Web initiatives.
- Harmonization Strategies: The paper details how LiLa tackles the complexities of Latin lemmatization:
- Canonical Forms: Identifying and collecting canonical forms (lemmas) from various sources.
- Hypolemmas: Introducing "hypolemmas" to harmonize strategies for participles, treating them as forms of a verbal paradigm that can be used as lemmas.
- Form Variants: Managing alternations in spelling, endings, and graphical variations as different written representations of the same lemma.
- Data Integration: Integrating existing resources like the Lemlat morphological analyzer, the Word Formation Latin (WFL) lexicon, and the PROIEL treebank into the Knowledge Base.
- SPARQL for Querying: Using SPARQL to query the RDF data stored in a triple store.
Datasets
The paper discusses the integration of several existing Latin linguistic resources:
- Lemlat: A Latin morphological analyzer, providing the fundamental list of Latin lemmas for the LiLa Knowledge Base.
- Word Formation Latin (WFL) Lexicon: Enhances the lexical basis with information on derivational morphology.
- PROIEL Treebank: A large, diverse annotated corpus of Latin texts (specifically the UD release 2.3 version) used for textual resource integration and evaluation.
- Lexical Resources: Mentions the inclusion of the Lewis-Short dictionary, Thesaurus Linguae Latinae, Neulateinische Wortliste, Latin Vallex, and Latin WordNet (LWN) as sources of lexical information.
- Onomasticon: Portions of the Forcellini dictionary's Onomasticon are being added.
No new benchmark datasets were introduced; rather, the paper focuses on integrating and harmonizing existing ones.
Key Insights and Ideas
- The Power of the Lemma: The lemma is identified as the crucial interface for connecting diverse linguistic resources, acting as a universal key for textual occurrences, lexical entries, and NLP tool outputs.
- Linked Data as a Solution: The Linked Data paradigm is presented as a robust framework for achieving interoperability, enabling the creation of a rich, interconnected web of linguistic data.
- Latin's Complexity as a Test Case: The highly inflected and diachronically diverse nature of Latin makes it an ideal, albeit challenging, testbed for developing and evaluating lemmatization harmonization strategies.
- The "Hypolemma" Concept: The introduction of "hypolemmas" is a novel approach to specifically address the challenge of harmonizing participle lemmatization, offering a more nuanced representation.
- Diagnosing Architecture Issues: Mismatches and multiple matches during the lemma matching process are valuable for identifying and diagnosing problems within the Knowledge Base architecture and the integrated resources.
Key Results
The paper demonstrates the feasibility of integrating diverse Latin resources into a Linked Data framework. A key result is the successful matching of PROIEL corpus tokens to LiLa's lemma collection:
- String Match: 162,998 tokens (81.52%) were unambiguously matched to a lemma via simple string comparison.
- PoS Disambiguation: An additional 6,262 tokens were matched using PoS tags for disambiguation.
- Combined Coverage: String matching and PoS-driven disambiguation together covered 84.64% of the PROIEL tokens.
- Onomasticon Matches: 7,252 proper names from PROIEL were matched to entries in the Forcellini dictionary.
The paper also highlights that unresolved mismatches (10,555 tokens) often point to issues with tokenization, lemmatization strategies in source resources (like Lemlat treating deadjectival adverbs as adjectives), or missing entities.
Limitations and Future Work
Limitations:
- Incomplete Coverage: The current lemma collection does not yet include all named entities from the Forcellini Onomasticon or the full du Cange glossary.
- Deadjectival Adverbs: Deadjectival adverbs are currently missing as lemmas in LiLa due to Lemlat's lemmatization strategy.
- Ambiguity: Despite disambiguation efforts, a significant number of tokens (11,865) remain associated with multiple lemmas, indicating inherent ambiguity or challenges in the current architecture.
- Out-of-Domain Performance: The accuracy of NLP tools applied to Latin texts can be affected by their "out-of-domain" nature, a challenge that LiLa aims to address.
Future Work:
- Efficient PoS Tagging/Lemmatization: Developing efficient strategies for automatic PoS tagging and lemmatization of Latin corpora that currently lack this annotation.
- Integrating More Lexical Resources: Incorporating the full Forcellini Onomasticon and the du Cange glossary into the lemma collection.
- Semantic Search: Including the Latin WordNet (LWN) to support semantic-based search capabilities.
- Testing NLP Tools: Evaluating the performance of existing NLP tools and models on out-of-domain Latin texts.
- Generating Deadjectival Adverbs: Systematically generating and including deadjectival adverbs as lemmas.
- Expanding PoS Tagset Alignment: Aligning other morphological features beyond Universal PoS tags.
Summary updated