Algorithm
This is heavily inspired from the English g2p.
- Spells out arabic numbers and some currency symbols, e.g.
Rp 200,000 -> dua ratus ribu rupiah. This is borrowed from Cahya's code. - Attempts to retrieve the correct pronunciation for homographs based on their POS (part-of-speech) tags.
- Looks up a lexicon (pronunciation dictionary) for non-homographs. This list is originally from ipa-dict, and we later made a modified version.
- For OOVs, we predict their pronunciations using either a BERT model or an LSTM model.
Phoneme and Grapheme Sets
graphemes = ['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z']
phonemes = ['a', 'b', 'd', 'e', 'f', 'ɡ', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'r', 's', 't', 'u', 'v', 'w', 'z', 'ŋ', 'ə', 'ɲ', 'tʃ', 'ʃ', 'dʒ', 'x', 'ʔ']
Homographs
Indonesian words (as far as we know) only have one case of homograph, that is, differing ways to pronounce the letter e. For instance, in the word apel (meaning: apple), the letter e is a mid central vowel ə. On the other hand, the letter e in the word apel (meaning: going to a significant other's house; courting), is a closed-mid front unrounded vowel e. Sometimes, a word might have >1 es pronounced in both ways, for instance, mereka (meaning: they) is pronounced as məreka. Because of this, there needs a way to disambiguate homographs, and in our case, we used their POS (part-of-speech) tags. However, this is not a foolproof method since homographs may even have the same POS tag. We are considering a contextual model to handle this better.
OOV Prediction
Initially, we relied on a sequence2sequence LSTM model for OOV (out-of-vocabulary) prediction. This was a natural choice given that it can "automatically" learn the rules of grapheme-to-phoneme conversion without having to determine the rules by hand. However, we soon noticed that despite its validation results, the model performed poorly on unseen words, especially on longer ones. We needed a more controllable model that makes predictions on necessary characters only. We ended up with a customized BERT that predicts the correct pronunciation of the letter e while keeping the rest of the string unchanged. We then apply a hand-written g2p conversion algorithm that handles the other characters.
You can find more detail in this blog post.
POS Tagging
We trained an NLTK PerceptronTagger on the POSP dataset, which achieved 0.956 and 0.945 F1-score on the valid and test sets, respectively. Given its performance and speed, we decided to adopt this model as the POS tagger for the purpose of disambiguating homographs, which is just like the English g2p library.
| tag | precision | recall | f1-score |
|---|---|---|---|
| B-$$$ | 1.000000 | 1.000000 | 1.000000 |
| B-ADJ | 0.904132 | 0.864139 | 0.883683 |
| B-ADK | 1.000000 | 0.986667 | 0.993289 |
| B-ADV | 0.966874 | 0.976987 | 0.971904 |
| B-ART | 0.988920 | 0.978082 | 0.983471 |
| B-CCN | 0.997934 | 0.997934 | 0.997934 |
| B-CSN | 0.986395 | 0.963455 | 0.974790 |
| B-INT | 1.000000 | 1.000000 | 1.000000 |
| B-KUA | 0.976744 | 0.976744 | 0.976744 |
| B-NEG | 0.992857 | 0.972028 | 0.982332 |
| B-NNO | 0.919917 | 0.941288 | 0.930480 |
| B-NNP | 0.917685 | 0.914703 | 0.916192 |
| B-NUM | 0.997358 | 0.954488 | 0.975452 |
| B-PAR | 1.000000 | 0.851064 | 0.919540 |
| B-PPO | 0.991206 | 0.991829 | 0.991517 |
| B-PRI | 1.000000 | 0.928571 | 0.962963 |
| B-PRK | 0.793103 | 0.851852 | 0.821429 |
| B-PRN | 0.988327 | 0.988327 | 0.988327 |
| B-PRR | 0.995465 | 1.000000 | 0.997727 |
| B-SYM | 0.999662 | 0.999323 | 0.999492 |
| B-UNS | 0.916667 | 0.733333 | 0.814815 |
| B-VBE | 1.000000 | 0.985714 | 0.992806 |
| B-VBI | 0.929119 | 0.877034 | 0.902326 |
| B-VBL | 1.000000 | 1.000000 | 1.000000 |
| B-VBP | 0.926606 | 0.933457 | 0.930018 |
| B-VBT | 0.939759 | 0.953333 | 0.946498 |
| --------- | --------- | -------- | -------- |
| macro avg | 0.966490 | 0.946937 | 0.955913 |
Attempts that Failed
- Parsed online PDF KBBI, but it turns out that it has very little phoneme descriptions.
- Scraped online Web KBBI, but it had a daily bandwidth which was too low to be used at this level.
Potential Improvements
There is a ton of room for improvements, both from the technical and the linguistic side of the approaches. Consider that a failure of one component may cascade to an incorrect conclusion. For instance, an incorrect POS tag can lead to the wrong phoneme, ditto for incorrect OOV prediction. We propose the following future improvements.
- Use a larger pronunciation lexicon instead of having to guess.
- Find a larger homograph list.
- Use contextual model instead of character-level RNNs.
- Consider hand-written rules for g2p conversion.
- Add to PyPI.