This repository contains a Python implementation of the many-to-many aligner proposed by [1]. The aligner was proposed to align graphemes to phonemes as a preprocessing step in order to learn grapheme-to-phoneme conversion models.
The aligner learns to align graphemes to phonemes from a provided set of (grapheme, phoneme) pairs. Alignment probabilities are learned by Expectation Maximisation (EM).
The main innovation of many-to-many aligners is that they allow alignment of grapheme n-grams to phoneme n-grams. Previous work mostly used 1-to-1 alignments, i.e. only single graphemes were aligned to single phonemes. However, this an unrealistic assumption for most languages. For example, consider the word "phoenix" with IPA /'fiːnɪks'. Here, the alignments should be
| ph | oe | n | i | x |
|---|---|---|---|---|
| 'f | iː | n | ɪ | ks |
while any 1-to-1 alignment would be implausible.
The main class EMAligner is in em_aligner.py.
It takes the following parameters:
max_source_ngram: The maximum n-gram length of grapheme n-grams that may be alignedmax_target_ngram: The maximum n-gram length of phoneme n-grams that may be alignednormalisation_mode: Choices are:"conditional": Learns a distribution over phoneme n-grams for each grapheme n-gram individually"joint": Learns a joint distribution of grapheme n-grams and phoneme n-grams
many2many: Whether to allow grapheme n-grams of length >= 2 to be aligned to phoneme n-grams of length >= 2. IfFalse, only 1-to-many and many-to-1 alignments are allowedallow_delete_graphemes: IfTrue, allows grapheme n-grams to not be aligned to any phoneme n-gram. In edit operation terminology, this corresponds to Deletionallow_delete_phonemes: IfTrue, allows phoneme n-grams to not be aligned to any grapheme n-gram.epochs: Number of iterations of the EM-Algorithm.
Please note that memory complexity is number of source n-grams times number of target
n-grams. Therefore, setting max_source_ngram or max_target_ngram may result in
performance issues.
Once the aligner is instantiated, it can be trained by calling
aligner.fit(source=source, target=target)
where source is a list of strings representing the graphemes and target is a nested list
where each phoneme sequence is represented by a list containing phonemes as strings.
The method align(source, target) can be used to align graphemes and phonemes.
Arguments are required to have identical format like arguments to the fit method.
Return format is a list of (aligned graphemes, aligned phonemes) tuples, where both
aligned graphemes and aligned phonemes are nested lists of equal length.
Each sublist contains the aligned n-gram at the corresponding position.
Models can be saved by calling aligner.save(path).
path is a string specifying where to save the model to.
The model is saved in two files called, one called parameters.pickle containing the
aligner's arguments, and one called parameters.npy containing the aligner's learned
parameters.
Models can be loaded by calling EMAligner.load(path) where path is a string
specifying a location containing a file called parameters.pickle and a file
called parameters.npy created by saving a trained aligner model.
[1] Sittichai Jiampojamarn, Grzegorz Kondrak, and Tarek Sherif. 2007. Applying Many-to-Many Alignments and Hidden Markov Models to Letter-to-Phoneme Conversion. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 372–379, Rochester, New York. Association for Computational Linguistics.