TL;DR
This paper introduces a novel, coarse-grained tag set for annotating modality in Oceanic languages, which often express modality through verb moods rather than auxiliaries. The authors demonstrate high inter-annotator agreement, suggesting their approach is efficient and adaptable for similar low-resource languages.
What Problem Does This Paper Address?
Documenting and analyzing the grammatical structures of low-resource languages, particularly those in Oceania, presents unique challenges. While many European languages express modality (e.g., possibility, necessity) using modal auxiliaries (like "might" or "must") or adverbs, Oceanic languages often rely on verb moods, such as realis and irrealis. This difference makes applying existing modality tagging frameworks, which are often built around lexical expressions, difficult and potentially inaccurate for these languages. The authors aim to develop and validate a more suitable method for capturing modality in these mood-prominent Oceanic languages.
Methods
The core of this paper is the development and application of a new tag set for annotating modality. Instead of focusing on specific modal words, the authors tag entire clauses. Their approach is inspired by a branching-times framework, which categorizes modality into three domains: factual (actual present and past), counterfactual (situations that did not occur), and possible (future events or quantification over actual/counterfactual domains).
The annotation process involved:
- Corpus Selection: Identifying comparable narrative and explanatory texts across five Oceanic languages (Daakaka, Dalkalaen, Mavea, Nafsan, and Saliba-Logea).
- Annotation Units: Segmenting texts into clauses, which served as the units for TAM (tense, aspect, modality) annotation.
- Manual Tagging: Two annotators manually tagged each clause for modality using the developed tag set.
- Inter-Annotator Consistency Analysis: Comparing the annotations to identify and resolve disagreements, calculating agreement scores (Krippendorff's alpha) for each category.
Datasets
The study utilized corpora from five Oceanic languages:
- Daakaka
- Dalkalaen
- Mavea
- Nafsan
- Saliba-Logea
These corpora are described as being richly annotated with morpheme-by-morpheme glosses, part-of-speech tags, and English translations. The authors imported these corpora into the ANNIS platform for searching and analysis. A subset of 26 texts, totaling 1953 clauses, was selected for TAM annotation.
Key Insights and Ideas
- Modality in Oceanic Languages: The paper highlights that modality in many Oceanic languages is primarily expressed through verb moods (realis/irrealis) rather than dedicated modal auxiliaries or adverbs, necessitating a different annotation approach.
- Coarse-Grained Tagging Efficiency: A coarse-grained tag set distinguishing between factual, counterfactual, and possible modalities proved more efficient and yielded higher inter-annotator agreement compared to more fine-grained distinctions.
- Branching-Times Framework: The theoretical underpinning of the tag set, based on a branching-times framework, allows for a more nuanced capture of modality than a simple realis/irrealis distinction, without getting bogged down in the often elusive distinctions between modal flavors (like epistemic vs. deontic).
- Clause-Level Annotation: Tagging entire clauses, rather than just specific words, is crucial for languages where modality is embedded in the verbal complex or grammatical structure.
Key Results
The analysis of inter-annotator consistency showed that the tag set works particularly well in the mood category, achieving a Krippendorff's alpha score of 0.85, which is considered acceptable.
- The factual tag demonstrated the highest consistency within the mood category, with only 3.17% inconsistencies across 1481 clauses.
- The possible tag had the highest inter-annotator inconsistency (12.7%), often stemming from mismatches with temporal references.
- The event category showed the lowest inter-annotator agreement (alpha = 0.79), while polarity had the highest (alpha = 0.91).
Limitations and Future Work
- Limitations:
- The authors acknowledge that the tag set for temporal and aspectual reference might need to be more fine-grained for languages with complex tense or aspect systems.
- The "possible" tag, while encompassing future and quantification, might still be broad and could benefit from further refinement depending on specific language needs.
- The study focused on a subset of texts, and further exploration across larger portions of the corpora is implied.
- Future Work:
- The authors suggest that their tag set, due to its efficiency and stability, could be a useful alternative for other languages and purposes, especially for mood-prominent languages.
- They also mention carrying out field work to elicit modal-temporal contexts that were rarely attested in the corpora, indicating ongoing research in this area.
Summary updated