TL;DR
This paper introduces Generalized Glossing Guidelines (GGG), a new YAML-based notation for interlinear glossed text (IGT) that explicitly handles complex, non-concatenative morphological processes like infixation, reduplication, and tonal overwriting. GGG aims to be both human-readable and machine-parseable, offering a more comprehensive and consistent way to annotate morphology compared to existing schemes.
What Problem Does This Paper Address?
Linguistic annotation, particularly for interlinear glossed text (IGT), is crucial for both human understanding and computational processing of language. Traditional IGT conventions, like the Leipzig Glossing Rules (LGR), are well-suited for concatenative morphology (where morphemes are added sequentially, like prefixes and suffixes). However, many languages exhibit non-concatenative morphology, where changes occur within a word (e.g., vowel changes, tone shifts, or insertions of material in the middle of a root). Existing IGT formats struggle to represent these processes explicitly, often leading to ambiguity or requiring ad-hoc solutions. This makes it difficult for both linguists to communicate analyses and for computational models to accurately capture the full complexity of a language's morphology. The authors' work on the wav2gloss project highlighted this gap when dealing with languages like Mixtec, Nahuatl, and Totonac, which heavily rely on non-concatenative processes.
Methods
The core of GGG is an "Item-and-Process" (IP) framework, which contrasts with the more common "Item-and-Arrangement" (IA) approach. Instead of just listing morphemes and their order, GGG explicitly models morphological processes that transform morphemes. The notation is structured in a YAML format with four key fields for each segment of text:
- Surface Representation (sr): Shows the actual spoken or written form, with curly brackets
{}indicating where non-concatenative processes have applied. - Lexical Representation (lx): This is the most crucial field. It explicitly details the morphological changes. Non-concatenative processes are represented using bracket-and-arrow notation, e.g.,
{A>B}signifies that stringAis replaced by stringB. This field aims to capture the output of morphological processes before phonological rules apply. - Gloss (gl): Associates linguistic meanings or grammatical features with the morphemes and processes in the
srandlxfields. - Translation (tr): An idiomatic translation of the segment.
GGG builds upon existing conventions like LGR but formalizes the representation of non-concatenative alternations, making them as explicit as affixation. It aims to be human-readable and editable with standard text tools, differentiating it from more complex XML-based formats like Xigt.
Datasets
The paper uses examples from several languages to illustrate GGG, including:
- Yoloxóchitl Mixtec: Used extensively to demonstrate tonal overwriting and other non-concatenative processes.
- German: To show apophony (umlaut).
- Motu: To illustrate reduplication and infixation.
- Tagalog: For infixation.
- Ulwa: For possessive marking via infixes.
- Latin: For infixation.
- Mangap-Mbula: For reduplication.
- Balsas Nahuatl: For complex reduplication.
- Pima: For infixing reduplication.
- Murle: For subtractive morphology.
- Arabic and Hebrew: For transfixation.
- Irish and Totonac: For apophony.
The paper does not introduce a new benchmark dataset but rather proposes a new annotation format that can be applied to existing linguistic data.
Key Insights and Ideas
- Unified Representation for Non-Concatenative Morphology: The most significant insight is the development of a single, consistent notation (
{A>B}inlx,{A}insr) to handle a wide range of non-concatenative processes. This contrasts with older systems that might have had separate, ad-hoc notations for each type of process. - Explicitly Modeling Processes: GGG treats morphological processes (like vowel change or tone shift) as first-class citizens, on par with morphemes (like prefixes or suffixes). This explicit modeling is key to its descriptive power.
- Separation of Morphology and Phonology: The
lxfield represents the output of morphological processes, while thesrfield shows the result after phonological rules have applied. This clear distinction is crucial for understanding how languages work. - Human- and Machine-Readability: The YAML format and the explicit notation are designed to be easily understood by humans using text editors and also parsed by computational tools, bridging the gap between linguistic analysis and NLP applications.
- Descriptive vs. Theoretical: GGG is presented as a descriptive tool, aiming to accurately represent observed linguistic phenomena without necessarily imposing a deep theoretical framework. This makes it more broadly applicable.
Key Results
The paper demonstrates the linguistic adequacy of GGG by showing how it can represent various non-concatenative morphological processes across different languages. A comparison table (Table 3) shows GGG achieving a score of 11, compared to LGR (6.5) and Xigt (7) in terms of representing different morphological processes. This suggests GGG is more comprehensive in its coverage of these complex phenomena. The authors also highlight that GGG's explicitness allows for easier discovery of morphological patterns, as one can directly see the relationship between forms and processes without needing to infer them from complex rules.
Limitations and Future Work
- Complexity of Annotation: The authors acknowledge that achieving full GGG annotation requires a thorough linguistic analysis of a language's morphology, which can be time-consuming.
- Parsing Overhead: Because the
lx,sr, andglfields must be aligned (same number of tokens), context-free parsing is not directly applicable, adding some overhead to parsing tools. - Future Work: The authors express a desire to develop tools that can facilitate the annotation process, potentially reducing the human intervention needed to go from raw data to interlinear annotations. They also hope that GGG will be widely adopted and useful for other researchers.
Summary updated