TL;DR
This paper introduces GLAUx, a large-scale, diachronic corpus of Ancient Greek texts annotated with NLP methods. It tackles the significant methodological challenges of creating such a resource, from encoding textual variants to handling linguistic ambiguity, aiming to unlock new avenues for studying language evolution.
What Problem Does This Paper Address?
Studying the evolution of languages over long periods is crucial for understanding linguistic change. Ancient Greek, with its exceptionally long and well-documented history, is an ideal candidate for such research. However, existing digital resources, while extensive, often lack the detailed linguistic annotations necessary for advanced computational analysis, or they are not publicly accessible. Furthermore, creating a comprehensive corpus of historical texts presents unique challenges:
- Textual Variants: Ancient texts are often transmitted through multiple manuscript versions, each with variations that need to be accounted for.
- Extralinguistic Variation: The sheer diversity of text types (literary, papyrological, epigraphical), genres, dialects, and historical periods within Ancient Greek creates significant challenges for automated analysis.
- Linguistic Ambiguity: Natural language is inherently ambiguous, and historical texts can exhibit ambiguities that are particularly difficult for NLP tools to resolve, especially when these ambiguities are linked to linguistic change.
The GLAUx project aims to address these issues by building a large, diverse, and richly annotated corpus of Ancient Greek, making it accessible for computational linguistic research.
Methods
The GLAUx project employs a range of NLP techniques to automatically annotate its corpus. Key methods include:
- Morphological and Part-of-Speech (POS) Tagging: The RFTagger, an HMM-based tagger, is used for this task. It leverages decision trees for contextual probabilities and suffix features for lexical probabilities. A lexicon generated by the Morpheus morphological analysis tool is used to constrain the tagger's output, improving accuracy.
- Lemmatization: The Lemming tool, a log-linear model, is used for lemmatization. It utilizes edit trees, lemma, POS, morphology, and dictionary features. The accuracy is further enhanced by constraining Lemming's output with lemmas recognized by Morpheus.
- Syntactic Annotation: The corpus is annotated with dependency information using the Stanford Graph-Based Dependency Parser, a biaffine neural (LSTM) parser. The annotation follows the AGDT guidelines, which are based on the Prague Dependency Treebanks. The authors plan to transition to the Universal Dependencies (UD) standard in the future.
- Semantic Role Annotation: A custom annotation standard and training data were developed for semantic role labeling, as existing resources were deemed too fine-grained or outdated. The semantic role labeler is based on a Random Forest classifier.
A significant aspect of their approach is the development of rule-based systems to handle specific challenges, such as generating "original" morphological information when editors have standardized spelling or morphology.
Datasets
The GLAUx corpus draws from several sources:
- Literary Texts: Primarily from the Open Greek and Latin project and various web sources.
- Papyrological Texts: Collected from the Integrating Digital Papyrology project.
- Epigraphical Texts: Intended to be added from the Packard Humanities Institute.
The paper also mentions the use of manually annotated Greek treebanks (PROIEL, AGDT, Gorman, Pedalion) as training data for their NLP models. These treebanks, particularly the Pedalion treebanks, were curated to include a variety of text types less represented by major projects, aiming to improve the robustness of the models.
Key Insights and Ideas
- The Importance of "Bridging Contexts": The paper highlights how ambiguous contexts, where a linguistic form could be interpreted in multiple ways (e.g., a clause being either complement or purpose), are crucial sites for linguistic change. This insight is fundamental to understanding how languages evolve.
- The Trade-off Between Standardization and Authenticity: For historical texts, editors often standardize spelling and morphology. While this aids readability, it can obscure the original linguistic usage. GLAUx attempts to preserve both original and standardized forms, and develops methods to infer original morphology from standardized texts, offering a richer analytical perspective.
- Genre as a Stronger Predictor of Lexical Meaning than Time: The authors cite research suggesting that text genre has a more significant impact on lexical semantics than chronological period, implying that models trained on one genre might perform poorly on another.
- The Need for Diverse Training Data: To combat the "out-of-domain" problem in NLP for historical texts, the GLAUx project emphasizes the importance of including a wide variety of text types, including less common ones like mathematical texts and private letters, in their training data.
Key Results
The paper reports high accuracies for its NLP annotations:
- POS/Morphological Tagging: Accuracy ranges from 0.908 (philosophical treatises) to 0.961 (biblical texts), with an average of 0.945.
- Lemmatization: Overall accuracy is 0.980, with variations across genres (e.g., 0.965 for comedies to 0.975 for epic poetry).
- Syntactic Parsing: Labeled attachment scores (LAS) range from 0.751 (philosophical/scientific prose) to 0.881 (biblical texts).
- Semantic Role Labeling: Accuracy ranges from 0.687 (poetic texts) to 0.838 (religious texts), with a relatively small training set.
Limitations and Future Work
- Text Preservation and Critical Apparatus: The GLAUx corpus often lacks critical apparatus from original editions, meaning users should be aware that the underlying data may not always be perfect and might include editorial alterations.
- Extralinguistic Variation Challenges: Despite efforts, the inherent diversity of Ancient Greek remains a challenge for NLP models. The authors acknowledge that accuracy drops when analyzing out-of-domain data.
- Syntactic Annotation Format: While the AGDT format was used for historical reasons, the authors plan to migrate to the Universal Dependencies (UD) standard for better interoperability.
- Semantic Annotation Granularity: The developed semantic role set is less granular than some pedagogical frameworks, but was chosen for its suitability for automatic prediction.
Future work includes:
- Expanding the corpus with epigraphical data and extending the literary data to the 8th century AD.
- Improving NLP models, potentially by exploring ELECTRA transformer models.
- Adding a derivational annotation layer.
- Transitioning to the Universal Dependencies (UD) syntactic annotation format.
- Developing detailed documentation and a user interface for querying the corpus.
- Collaborating with other digital humanities projects to enhance accessibility and functionality.
Summary updated