Phonetic Vector Representations for Sound Sequence Alignment

Pavel Sofroniev, Çağrı ÇöltekinEMNLP 2018

TL;DR

This paper explores data-driven methods for creating vector representations of phonetic segments to improve sound sequence alignment, a crucial task in computational historical linguistics. They demonstrate that these learned representations, particularly those capturing sequential information and multilingual context, outperform traditional symbolic methods and linguistically-motivated features, achieving results comparable to state-of-the-art approaches.

What Problem Does This Paper Address?

Aligning sound sequences is fundamental to many fields, including computational linguistics and bioinformatics. In computational historical linguistics, accurate alignment is essential for tasks like inferring word cognacy and identifying sound correspondences across languages. Traditional alignment algorithms often treat phonetic segments as discrete, symbolic units, ignoring their inherent similarities. This simplification can lead to suboptimal alignment accuracy. While some prior work has used linguistically-motivated features or reduced alphabets, this paper investigates whether data-driven vector representations, learned from large datasets, can capture phonetic similarities more effectively and thus improve alignment performance.

Methods

The authors propose and evaluate three data-driven methods for learning phonetic vector representations:

  • phon2vec: An adaptation of the word2vec model, treating phonetic segments within a word as a "bag of words" to learn dense vector representations. This method ignores the sequential order of segments.
  • Position Sensitive Neural Network Embeddings (NN Embeddings): A feed-forward neural network that learns representations by predicting context from a given segment, crucially considering the positional information of segments within a context.
  • RNN Embeddings: A sequence-to-sequence recurrent neural network (RNN) trained on multilingual word pairs (cognates). This method leverages the fact that words representing the same concept in different languages can provide valuable contextual information, and it inherently captures sequential dependencies.

These learned vector representations are then used with the Needleman-Wunsch algorithm, employing cosine similarity between vectors as the scoring function for segment alignment. They also compare their methods against baselines: a one-hot encoding (representing purely symbolic units) and PHOIBLE Online feature vectors (linguistically-motivated). Additionally, they compare against two state-of-the-art scoring functions: the ASJP-based PMI method and the sound-class-based SCA.

Datasets

  • Training Data: NorthEuraLex, a comprehensive lexicostatistical database containing IPA-encoded lexical data for over 100 languages, resulting in over 121,000 IPA transcriptions. For RNN embeddings, cognate pairs were identified based on normalized Levenshtein distance.
  • Evaluation Data: The Benchmark Database for Phonetic Alignments (BDPA), containing 7198 aligned IPA sequence pairs from 12 source datasets across 6 language families. They also used a small set of 82 pairs from Covington (1996).

Key Insights and Ideas

  • Data-driven beats symbolic: Learned vector representations consistently outperform the one-hot encoding baseline, demonstrating that capturing phonetic similarity is crucial for alignment.
  • Sequential order matters: NN embeddings, which consider positional information, generally outperform phon2vec, which treats context as a bag of words, suggesting that the order of phonetic segments is important.
  • Multilingual context is powerful: RNN embeddings, trained on multilingual cognates, achieve strong performance, indicating that leveraging cross-linguistic information can significantly improve phonetic representations.
  • Task adaptation is beneficial: While models are optimized for their respective prediction tasks (e.g., context prediction), adapting them for alignment yields competitive results.
  • PHOIBLE vs. Data-driven: PHOIBLE's linguistically-informed features perform comparably to phon2vec but are generally outperformed by the NN and RNN embeddings, suggesting data-driven approaches can learn useful linguistic generalizations.

Key Results

  • Overall Performance: The data-driven methods (phon2vec, NN, RNN) generally achieve higher alignment accuracy than the one-hot encoding and PHOIBLE features across most datasets.
  • RNN Embeddings: These consistently perform well, often on par with or exceeding the PMI method and yielding better results than phon2vec and NN embeddings in many cases.
  • SCA: The sound-class-based alignment (SCA) method often achieves the highest scores, particularly on datasets with segments not well-represented in the training data. However, the variance in SCA's performance across datasets is notable.
  • PMI: The PMI method also performs strongly, comparable to RNN embeddings.
  • Specific Dataset Performance: For example, on the Sinitic dataset, the performance difference between SCA and other methods can be as high as 26 percent points.

Limitations and Future Work

  • Data Coverage: The performance of data-driven methods can be limited by the coverage of IPA segments in the training data. For instance, tones present in some datasets but not in the training data (e.g., NorthEuraLex) can hinder model performance.
  • Training Data Size and Diversity: The authors suggest that larger and more diverse training datasets would likely improve the performance of data-driven methods.
  • Hyperparameter Tuning: While they performed a random search, more systematic hyperparameter optimization could yield better results.
  • Future Directions:
    • Investigating the use of larger and more diverse training data.
  • Exploring embeddings specific to language families or macro-areas.
  • Further comparing and evaluating these methods in larger applications like cognacy identification or phylogenetic inference.

Summary updated

BibTeX
@inproceedings{sofroniev2018phonetic,
  author = {Pavel Sofroniev and Çağrı Çöltekin},
  title = {{Phonetic Vector Representations for Sound Sequence Alignment}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure