TL;DR
This paper tackles the challenge of automatically segmenting words in Inuinnaqtun, a low-resource, polysynthetic Indigenous language facing extinction. By adapting an unsupervised "Adaptor Grammars" approach, they achieve promising results, paving the way for better NLP tools to help preserve such languages.
What Problem Does This Paper Address?
Natural Language Processing (NLP) has made great strides with well-resourced languages. However, Indigenous polysynthetic languages, like Inuinnaqtun, present significant hurdles. These languages are characterized by extremely complex word structures, where a single word can convey the meaning of an entire sentence by combining a root word with numerous prefixes and suffixes. This morphological complexity, coupled with a scarcity of linguistic resources and tools, makes traditional NLP tasks like morphological analysis and machine translation very difficult. Furthermore, Inuinnaqtun is critically endangered, with fewer than two generations remaining, making the development of NLP tools for its preservation an urgent task.
Methods
The core of this research is the adaptation of an unsupervised morphological segmentation approach based on Adaptor Grammars (AG). AGs are a probabilistic framework that can learn grammatical structures from unsegmented text. The authors leverage a specific AG model that views word structure as a sequence of a word base, followed by optional lexical suffixes, and finally grammatical ending suffixes.
They build upon existing AG work by:
- Scholar-seeded learning: Incorporating linguistic knowledge by collecting a list of known affixes from dictionaries and websites. This helps guide the grammar learning process.
- Adaptation for Inuinnaqtun: Tailoring the AG framework to the specific structure of Inuinnaqtun, which differs from some previously studied polysynthetic languages in how it sequences prefixes, stems, and suffixes.
The process involves defining the grammar rules (non-terminals, terminals, production rules) and collecting a large unsegmented word list to discover and learn morphological patterns.
Datasets
- Training Data: A lexicon of over 50,000 unsegmented Inuinnaqtun words, with lengths ranging from three to thirty letters. This lexicon was compiled from various sources, including the Nunavut government website, open-source dictionaries, and grammar books.
- Experimental Corpus: A manually curated corpus containing 190 word bases and 571 affixes.
- Golden Testing Set: A manually crafted set of 1,055 unique segmented words for evaluation.
Key Insights and Ideas
- Adaptor Grammars are effective for low-resource polysynthetic languages: The study demonstrates that AGs, even with limited data and a scholar-seeded approach, can achieve significant progress in segmenting complex words in a language like Inuinnaqtun.
- Linguistic knowledge significantly boosts performance: The "scholar-seeded" AG model, which incorporates pre-identified affixes, outperforms the standard AG model and the baseline Morfessor, highlighting the value of even partial linguistic input.
- Inuinnaqtun's complexity poses challenges: The tendency for over-segmentation in both AG models suggests that the intricate and sometimes irregular morphological rules of Inuinnaqtun require further refinement of the models.
- Preservation through NLP: The work underscores the critical role of NLP in revitalizing and preserving endangered Indigenous languages, not just for linguistic research but also for transmitting ancestral knowledge and culture.
Key Results
The AG-based approach showed promising results on the Inuinnaqtun test set:
- AG-Standard Model: Achieved a precision of 50.76%, recall of 80.30%, and an F1 score of 62.20%. This is an improvement over the baseline Morfessor (58.87% F1).
- AG-Scholar Model: Significantly outperformed both the baseline and the standard AG model, achieving a precision of 71.06%, recall of 82.83%, and an F1 score of 76.49%. This represents a substantial gain, particularly in precision.
Qualitatively, the AG-Scholar model demonstrated a better ability to identify common lexical and grammatical suffixes.
Limitations and Future Work
- Over-segmentation: The authors note that both AG models tend to over-segment complex morphemes, likely due to the inherent irregularities and morphophonological phenomena in Inuinnaqtun.
- Future Work: The researchers plan to develop more efficient unsupervised morphological segmentation methods. They also aim to extend their research to other endangered Indigenous languages and dialects, with potential applications in Machine Translation and Information Retrieval.
Summary updated