Phonological Features for Morphological Inflection

Adam Wiemerslage, Miikka Silfverberg, Mans HuldenEMNLP 2018

TL;DR

This paper explores using phonetic representations (IPA segments and distinctive features) for morphological inflection, a key NLP task. Surprisingly, phonetic inputs performed comparably to orthographic ones, suggesting that orthography already encodes much of the necessary phonological information for inflection.

What Problem Does This Paper Address?

Morphological inflection, the process of changing a word's form to indicate grammatical function (e.g., "walk" to "walks"), is crucial for Natural Language Processing (NLP) but challenging, especially for morphologically rich languages. Traditional approaches often rely on orthographic (written) representations. However, orthographies can be inconsistent and don't always directly reflect pronunciation, leading to potential difficulties for NLP models. This paper investigates whether using phonetic representations, which more closely mirror spoken language, can improve morphological inflection tasks and offer benefits like cross-lingual transfer learning.

Methods

The authors employ an encoder-decoder architecture with attention, a standard approach for sequence-to-sequence tasks like machine translation and morphological inflection. Their key innovation lies in experimenting with different input representations for the encoder:

  • Orthographic: Standard character-based input.
  • IPA Segments: International Phonetic Alphabet symbols representing sounds.
  • Phonological Distinctive Features: A more granular representation where each sound is described by a set of binary features (e.g., [+voice], [-nasal]).

They use the Python libraries Epitran for grapheme-to-phoneme (G2P) conversion to get IPA segments and PanPhon to convert IPA segments into distinctive feature vectors. The model then learns to map these input representations to the correct inflected output sequence.

Datasets

The experiments were conducted on a subset of data from the CoNLL-SIGMORPHON 2017 Shared Task on Universal Morphological Reinflection. This dataset covers 52 languages, and the authors selected 8 languages that were supported by their chosen G2P tools (Epitran and PanPhon) and exhibited typological diversity. The data was split into two settings: "Medium" (around 1,000 training examples) and "High" (around 10,000 training examples).

Key Insights and Ideas

  • Phonetic vs. Orthographic Performance: Counterintuitively, phonetic representations (both IPA and distinctive features) did not significantly outperform orthographic representations for morphological inflection. This suggests that orthographic systems, despite their inconsistencies, often implicitly encode enough phonological information for inflectional tasks.
  • Redundancy of Detailed Phonology: Providing detailed phonological representations like IPA segments appears to be largely redundant when compared to orthographic inputs, as the distributional properties of graphemes already capture much of the relevant information.
  • IPA vs. Distinctive Features: IPA segments and distinctive feature bundles yielded very similar accuracies. This implies that the distributional properties captured by IPA segments are already rich enough, and the added granularity of distinctive features doesn't offer a substantial advantage in this specific task.
  • Orthography's Etymological Advantage: For languages like English, orthography might even have an advantage due to its reflection of etymological information, which can be helpful for predicting inflectional behavior, especially concerning vowels.

Key Results

  • Overall Accuracy: In the "High" data setting, the orthographic, IPA, and feature-based models performed comparably to each other and to the best shared task systems. The average accuracy for orthographic, IPA, and feature-based models was 86.85%, 76.21%, and 72.73% respectively in the "Medium" setting, and 95.0%, 92.66%, and 89.36% in the "High" setting.
  • IPA/Feature vs. Orthography: The IPA and feature-based models showed a slight decrease in accuracy compared to orthographic models, particularly for languages like English and German. This difference was attributed to the complexities of English orthography and pronunciation (e.g., vowel variations) and potentially etymological cues present in spelling.
  • IPA and Feature Similarity: A paired permutation test showed no significant difference in accuracy between IPA segment and distinctive feature representations (p-value of 0.43).

Limitations and Future Work

  • Orthographic Advantage: The authors acknowledge that orthographic inputs sometimes perform better, especially for languages with complex spelling-pronunciation relationships like English. They suggest that etymological information embedded in orthography might be beneficial.
  • Data Sparsity: The performance gap between phonetic and orthographic inputs was more pronounced in the "Medium" data setting, highlighting the need for more data for phonetic representations to shine.
  • Future Directions: The paper suggests exploring scenarios where distinctive features might offer an advantage, such as in transfer learning for unseen languages or for segments that rarely appear in training data. Combining IPA and feature representations (e.g., through addition or concatenation) did not yield significant improvements, but further exploration of hybrid representations could be fruitful.

Summary updated

BibTeX
@inproceedings{wiemerslage2018phonological,
  author = {Adam Wiemerslage and Miikka Silfverberg and Mans Hulden},
  title = {{Phonological Features for Morphological Inflection}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure