Abstract
The Irish Universal Dependencies Treebank (IUDT) serves as a critical resource for Irish Natural Language Processing. However, standard universal annotations often overlook language-specific morphological complexities. This study details the integration of rich morphological features, including initial mutations and dialectal variations, into the IUDT. Using MaltParser and MaltOptimizer, we demonstrate that incorporating these fine-grained linguistic details significantly enhances automated dependency parsing performance, particularly for low-resource languages where morphology provides essential structural cues.
Introduction
Natural Language Processing (NLP) relies heavily on treebanks—collections of sentences annotated with syntactic structure—to train automated systems. The Universal Dependencies (UD) project seeks to standardize these annotations globally to facilitate cross-linguistic research. However, Irish presents unique challenges as a morphologically rich language featuring complex inflections and "initial mutations" (where the start of a word changes based on the preceding word).
Previous versions of the Irish UD treebank lacked these nuances, treating the language as if it were morphologically simpler than it is. This paper addresses this gap by mapping complex Irish-specific morphological data into the UD format and evaluating whether this additional information improves the accuracy of automated dependency parsers.
Methodology
The researchers employed a multi-step approach to enrich the treebank. First, they mapped the output of an existing Finite State Morphological Analyser for Irish to the UD feature inventory. Because the standard UD set was insufficient to capture the full breadth of Irish grammar, they introduced several language-specific features, such as Form (to capture lenition and eclipsis), Dialect, and NounType.
The primary dataset, IUDT version 1.4, consists of 1,020 trees (approximately 23,684 tokens). Parsing experiments were conducted using MaltParser, a transition-based dependency parser. To ensure performance gains were attributable to data quality rather than software configuration, MaltOptimizer was used to automatically determine the best algorithmic settings for the Irish dataset.
Results and Analysis
The inclusion of morphological features led to a clear quantitative improvement in parsing accuracy, measured by Labelled Attachment Score (LAS) and Unlabelled Attachment Score (UAS). The baseline IUDT (without features) achieved a LAS of 68.27% on the test set. With the addition of morphological features and algorithmic optimization, the LAS rose to 72.18%.
Analysis reveals that morphology acts as a structural map in Irish. For example, the Voice=Auto feature (marking autonomous verbs) helped the parser correctly identify direct objects, which are often confused with subjects in Irish. Notably, a version of the treebank using only universal features performed worse than the version including Irish-specific features (69.26% vs 70.63% LAS). This highlights that for low-resource languages, local linguistic nuances are not just "noise"—they are critical data points that compensate for the information loss inherent in universal standardization.
Conclusion
This work demonstrates that integrating language-specific morphological features significantly boosts the performance of dependency parsers for Irish. While the IUDT remains relatively small, these results provide a roadmap for improving NLP tools for other morphologically rich, low-resource languages. Future research will focus on expanding the treebank size and testing these features with neural dependency parsing architectures.
Summary updated