TL;DR
This paper introduces extensions to the Universal Dependencies (UD) annotation scheme to better capture the nuances of both modern and historical German. By creating the first UD treebank for Middle High German (MHG) and adapting existing UD resources, the authors provide a more robust framework for analyzing German syntax across time, revealing interesting diachronic shifts in linguistic structures.
What Problem Does This Paper Address?
Analyzing historical languages presents unique challenges for linguistic annotation. Existing Universal Dependencies (UD) schemes, while powerful, often lack the granularity to capture specific syntactic phenomena that evolve over time. For Middle High German (MHG), a period crucial for understanding the development of modern German, there was a significant lack of reliably annotated treebanks. This gap made it difficult to conduct detailed syntactic analyses and compare historical language use with its modern counterpart. The authors aim to bridge this gap by extending the UD scheme and creating a new, annotated corpus for MHG.
Methods
The core of this work involves extending the existing UD annotation scheme for German. The authors identified areas where the standard UD scheme was insufficient for accurately representing German syntax, particularly concerning the distinction between arguments and modifiers, and various types of clausal complements. They propose new subtypes for existing UD labels (e.g., obl:arg for prepositional objects, xcomp:subj for infinitival subjects) to provide finer-grained distinctions. These extensions are designed to be backward compatible with the general UD framework. The annotation process was carried out using the INCEPTION tool, and inter-annotator agreement was measured using Krippendorff's alpha and standard dependency parsing metrics like UAS and LAS.
Datasets
The study utilizes two primary datasets:
- Modern German (ModG): A subset of the German GSD treebank, specifically the development set, which was manually annotated in the UD style.
- Middle High German (MHG): A newly annotated corpus based on the Reference Corpus of Middle High German (ReM). This corpus consists of four verse texts and comprises approximately 29,000 tokens. This is highlighted as the first UD treebank for Middle High German.
Key Insights and Ideas
- Diachronic Syntactic Shifts: The extensions reveal significant differences in the distribution of syntactic labels between modern and historical German, offering insights into how language structures evolve.
- Need for Granularity: The paper demonstrates that a more detailed annotation scheme is crucial for capturing the specificities of historical languages and for robust cross-temporal linguistic analysis.
- UD's Adaptability: The work showcases the flexibility of the Universal Dependencies framework, which allows for language-specific extensions and subtypes to accommodate diverse linguistic phenomena.
- MHG Specifics: The study highlights unique MHG constructions, such as the partitive construction and the use of "floating quantifiers," which require specialized annotation.
Key Results
- High Inter-Annotator Agreement: The authors achieved high inter-annotator agreement scores, with alpha values ranging from 0.85 for MHG to over 0.96 for ModG, indicating the reliability of their extended annotation scheme.
- First MHG UD Treebank: The creation of a UD treebank for Middle High German, containing nearly 29,000 tokens, is a significant contribution to the field of historical linguistics and computational linguistics.
- Distributional Differences: Statistical analysis revealed notable differences in the frequency of certain dependency relations between MHG and ModG. For instance, subjects are more frequent in MHG, while prepositional phrases (case) and determiners (det) are more prevalent in ModG.
- Subtype Representation: The study found that while unmodified labels (like
xcomp,advmod) are dominant in both datasets, specific subtypes are strongly represented in one language over the other, reflecting historical language changes. For example,nmod:part(partitive) is frequent in MHG but rare in ModG.
Limitations and Future Work
- Linguistic Test Limitations: The authors acknowledge that applying linguistic tests to historical data can be challenging, which sometimes leads to ambiguity in annotation, particularly for distinguishing between locative and directional elements.
- Data Size Discrepancy: Direct comparison of corpus statistics between MHG and ModG is somewhat limited due to the significant difference in dataset sizes and text types (religious texts for MHG vs. news/reviews for ModG).
- Future Work: The authors plan to annotate more MHG data from different genres, develop automatic parsers for their extended scheme, and further investigate the evolution of syntactic constructions. They also make their annotated data publicly available under a free license.
Summary updated