Adaptor Grammars for the Linguist: Word Segmentation Experiments for Very Low-Resource Languages

Pierre Godard, Laurent Besacier, François Yvon, Martine Adda-Decker, Gilles Adda, Hélène Maynard, Annie RiallandEMNLP 2018

TL;DR

This paper explores using Adaptor Grammars to automatically segment words in very low-resource languages, a crucial step for linguistic documentation. By integrating linguist expertise into grammar rules, they significantly improve word segmentation performance, demonstrating a powerful framework for collaboration between linguists and computer scientists.

What Problem Does This Paper Address?

The world is facing a rapid loss of linguistic diversity, with many languages at risk of extinction. Computational Language Documentation (CLD) aims to leverage speech and language technologies to help preserve and document these endangered languages. A fundamental challenge in this process is word segmentation, especially for languages with limited existing digital resources (very low-resource languages). Unlike languages with clear spaces between words, many languages present a continuous stream of sounds or characters, making it difficult to identify word boundaries. Traditional methods often struggle with the scarcity of annotated data and the unique linguistic characteristics of these languages. This paper addresses this by proposing a method to improve unsupervised word discovery for such languages, specifically by incorporating linguistic knowledge into a flexible grammar framework.

Methods

The core of this work lies in the application and extension of Adaptor Grammars (AGs). AGs are a probabilistic grammar framework that extends Context-Free Grammars (CFGs) by allowing non-terminals to be "adapted," meaning they can be treated as whole units rather than just decomposable structures. This adaptability is key to handling linguistic variability.

The authors propose a systematic methodology for exploring a large space of possible grammars by distinguishing four levels of description:

  • Sentence Level: Models different hierarchies of word collocations.
  • Word Level: Defines how words are structured, from simple character sequences to more complex morphological structures (e.g., prefixes, stems, suffixes).
  • Syllable Level: Models phonotactic constraints within syllables.
  • Character Level: Captures relationships between characters, including digraphs and trigraphs.

They experiment with 162 different grammar configurations by combining various hypotheses at each level. A crucial aspect is the ability to specialize generic grammars with language-specific knowledge provided by linguists. This includes incorporating information about phonological constraints, morphological rules, and even lists of prefixes. The training is unsupervised, using sampling techniques (MCMC), and the evaluation focuses on metrics like token and type F-scores, as well as exact-match sentence segmentation.

Datasets

The experiments were conducted on corpora for two Northwestern Bantu languages:

  • Mboshi: Collected in Congo-Brazzaville.
  • Myene: Collected in Gabon (specifically the Orungu variant).

These corpora were collected using a mobile app designed for fieldwork and contain manual transcriptions with linguist-annotated word segmentations. Basic statistics for these corpora are provided in Table 1, showing the number of sentences, tokens, and types, along with average token length.

Key Insights and Ideas

  • Linguistic Expertise is Crucial for Low-Resource Languages: Integrating linguist knowledge, even in the form of simple rules or lists, significantly boosts word segmentation performance. This highlights the value of collaboration between computational linguists and field linguists.
  • Adaptor Grammars as a Flexible Framework: AGs provide a powerful and flexible way to represent and explore different linguistic hypotheses. Their ability to adapt non-terminals makes them well-suited for modeling the complex and often irregular structures found in low-resource languages.
  • Hierarchical Grammar Structure Matters: Modeling word collocations at the sentence level (A2, A3 hypotheses) consistently improves results, suggesting that understanding how words group together is important for segmentation.
  • Morphotactics are Key for Word Structure: Explicitly modeling morphotactics (the rules governing the sequence of morphemes within a word) at the word level (e.g., B2, B3, B4, B5, B6 hypotheses) leads to substantial improvements over simpler models.
  • Character-Level Features Can Be Language-Specific: Incorporating digraphs and trigraphs (D1+ variant) provided a benefit for Mboshi, likely due to its specific phonological inventory, but not for Myene, demonstrating the need for language-specific adaptations even at the character level.
  • Segmentation Performance Varies by Language: Myene generally shows superior segmentation performance, possibly due to its shorter average sentence length and thus a higher proportion of already known word boundaries, making segmentation a harder task in Mboshi.

Key Results

The paper reports significant improvements in word segmentation performance compared to a strong Bayesian baseline. The best performing grammars achieved around 30% higher token F-score than the baseline. For example, the best grammar for Mboshi (A3-B3-C3-D1+) achieved a 64.78% token F-measure, while the best for Myene (A3-B6+-C2-D1) reached 72.62%. This improvement is attributed to the systematic exploration of grammar configurations and the incorporation of language-specific knowledge. The study also demonstrated that improved grammatical descriptions correlate with better automatic analysis.

Limitations and Future Work

  • Limited Scope of Sentence-Level Structures: The paper acknowledges that exploring more realistic syntactic structures at the sentence level was left for future work.
  • Syllable Level Impact: The impact of syllable-level variants was found to be less significant for word segmentation compared to sentence and word levels, suggesting that while phonotactics are modeled, their contribution might be overshadowed by higher-level structures.
  • Character-Level Limitations: The character-level rules (D1+) showed language-specific benefits, indicating that a universal character-level hypothesis might not be optimal.
  • Future Work: The authors suggest further exploring the usability of output parses in speeding up the documentation process. They also propose analyzing the usability of tonal information for word segmentation, which was explored in a previous work. The potential for AGs to help linguists discover linguistic insights beyond just segmentation is also a promising avenue.

Summary updated

BibTeX
@inproceedings{godard2018adaptor,
  author = {Pierre Godard and Laurent Besacier and François Yvon and Martine Adda-Decker and Gilles Adda and Hélène Maynard and Annie Rialland},
  title = {{Adaptor Grammars for the Linguist: Word Segmentation Experiments for Very Low-Resource Languages}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure