TL;DR
This paper introduces the first Universal Dependencies (UD) treebank for Coptic, a crucial but under-resourced ancient Egyptian language. By adapting UD guidelines and leveraging existing Coptic resources, the authors have created a valuable dataset for computational linguistics research, enabling new avenues for Coptic language processing and typological studies.
What Problem Does This Paper Address?
Coptic, the final stage of the Ancient Egyptian language, holds immense historical and cultural significance. Despite this, digital resources for its study, particularly syntactically annotated data, have been scarce. Existing Coptic corpora often use specialized tagging schemes that are not easily interoperable with broader linguistic research. This paper addresses the need for a Coptic treebank that adheres to the Universal Dependencies (UD) framework, a widely adopted standard for cross-linguistic syntactic annotation. This allows Coptic to be integrated into comparative linguistic studies and benefit from advancements in NLP tools developed for other UD treebanks.
Methods
The authors constructed the Coptic UD Treebank by adapting the UD annotation scheme to the specific challenges of Coptic grammar and existing annotation conventions. Key methodological aspects include:
- Leveraging Existing Resources: The project utilized data from the Coptic Scriptorium (CS) project, which provided pre-tokenized and POS-tagged data. This significantly reduced annotation effort but also imposed constraints on segmentation and tagging choices.
- Adapting UD Guidelines: The paper details the process of mapping Coptic-specific grammatical phenomena to UD relations. This involved addressing challenges like:
- Bound Groups: Coptic often spells multi-word units together (e.g., clitic possessors with nouns). The authors adopted a multi-token approach in CoNLL-U format to represent these, using hyphens for morpheme separation and the
SpaceAfter=Nofeature. - Fusional Morphology: Coptic exhibits complex fusional morphology, where multiple grammatical functions are merged into single morphemes. The authors retained CS tokenization for fused auxiliaries and pronouns, using portmanteau tags in the MISC field and mapping them to appropriate UD relations (e.g.,
PRONfor fused auxiliaries with pronouns). - Ambiguous Passives: Coptic has ambiguous passive constructions, often using a 3rd person plural verb. The authors opted not to distinguish these morphologically, annotating them uniformly with regular dependent
nsubjandauxchildren (active syntax). - Non-Adjacent Appositions: UD guidelines typically require appositions to be adjacent. Coptic's structure, with particles often intervening, necessitated a more flexible interpretation of adjacency constraints for the
apposrelation. - Differential Object Marking: Coptic exhibits variations in how direct objects are marked depending on tense/aspect. The authors analyzed both constructions as
objto facilitate extraction of object arguments and identification of transitive verbs. - Inter-Annotator Agreement: To evaluate the annotation scheme, the authors conducted an experiment with three pairs of annotators (undergraduate students and expert annotators). They used Arborator as the annotation interface and measured agreement using Label Scores (LS) and Unlabeled Attachment Scores (UAS), both with and without punctuation.
Datasets
The primary dataset used for constructing the Coptic UD Treebank was sourced from the Coptic Scriptorium (CS) project. This data was chosen for its free availability under a Creative Commons license and its existing tokenization and POS tagging. The selection of texts aimed to cover a range of genres (hagiography, biblical narrative, epistles, sermons, letters, martyrdom) and include both translated biblical material and original Coptic works. The corpus ultimately encompasses over 20,000 tokens.
Key Insights and Ideas
- UD is Adaptable to Ancient Languages: The paper demonstrates that the UD framework, while designed with modern languages in mind, can be successfully adapted to ancient and morphologically complex languages like Coptic.
- Balancing Linguistic Fidelity and Practicality: The authors navigated a trade-off between strict linguistic analysis and the practical need to leverage existing tools and datasets. Decisions were often guided by Manning's Law (favoring unambiguous, deterministic, and easily convertible analyses).
- The Importance of Context for Ambiguity: Coptic's grammatical features, such as ambiguous passives and the placement of particles, highlight the challenges of applying rigid syntactic rules and the need for careful consideration of context.
- Inter-Annotator Agreement as a Measure of Scheme Robustness: The inter-annotator agreement experiments provided valuable feedback on the clarity and robustness of the adapted UD guidelines for Coptic. The improvement observed from pre- to post-adjudication rounds indicates that annotator experience and discussion significantly enhance accuracy.
- Coptic's Unique Syntactic Features: The paper sheds light on Coptic's distinct syntactic characteristics, such as its frequent use of bound groups, complex verbal morphology, and specific ways of marking objects and appositions, which differ from many Indo-European languages.
Key Results
- The paper successfully created the first dependency treebank for Coptic and the first for the Egyptian language family within the UD framework.
- The treebank currently contains over 20,000 tokens.
- Inter-annotator agreement scores were generally high, especially for expert annotators. Expert annotators achieved a Label Score (LS) of 0.92 (without punctuation) and 0.93 (with punctuation), considered very good agreement.
- Student annotators showed significant improvement after adjudication, with Group B achieving an LS of 0.81 and Group A scoring 0.88 post-adjudication.
- Unlabeled Attachment Scores (UAS) were also strong, with expert annotators reaching 95.8% (without punctuation) and 96.0% (with punctuation).
Limitations and Future Work
- Ambiguity of Passives: The decision not to distinguish morphologically between active and passive constructions due to ambiguity means that the treebank may not fully capture the nuances of Coptic passivization.
- Non-Native Annotators: The authors acknowledge that the use of non-native speakers as annotators, while common in UD projects, can affect inter-annotator agreement compared to native speakers.
- Future Work: The authors plan to expand the treebank to a size that allows for training robust parsers and evaluating parsing results on Coptic in future shared tasks. They also express a desire to discuss challenging issues, such as adjacency constraints for appositions and the encoding of pronominal/clausal double object constructions, with the broader UD community. The paper also notes that while the current treebank focuses on Sahidic, they intend to approach other Coptic dialects in the future.
Summary updated