license: apache-2.0
library_name: onnx
tags:
- punctuation
- sentence boundary detection
- truecasing
language:
- af
- am
- ar
- bg
- bn
- de
- el
- en
- es
- et
- fa
- fi
- fr
- gu
- hi
- hr
- hu
- id
- is
- it
- ja
- kk
- kn
- ko
- ky
- lt
- lv
- mk
- ml
- mr
- nl
- or
- pa
- pl
- ps
- pt
- ro
- ru
- rw
- so
- sr
- sw
- ta
- te
- tr
- uk
- zh
Model Overview
This model accepts as input lower-cased, unpunctuated, unsegmented text in 47 languages and performs punctuation restoration, true-casing (capitalization), and sentence boundary detection (segmentation).
All languages are processed with the same algorithm with no need for language tags or language-specific branches in the graph. This includes continuous-script and non-continuous script languages, predicting language-specific punctuation, etc.
Model Details
This model generally follows the graph shown below, with brief descriptions for each step following.
Encoding: The model begins by tokenizing the text with a subword tokenizer. The tokenizer used here is a
SentencePiecemodel with a vocabulary size of 64k. Next, the input sequence is encoded with a base-sized Transformer, consisting of 6 layers with a model dimension of 512.Post-punctuation: The encoded sequence is then fed into a classification network to predict "post" punctuation tokens. Post punctuation are punctuation tokens that may appear after a word, basically most normal punctuation. Post punctation is predicted once per subword - further discussion is below.
Re-encoding All subsequent tasks (true-casing, sentence boundary detection, and "pre" punctuation) are dependent on "post" punctuation. Therefore, we must conditional all further predictions on the post punctuation tokens. For this task, predicted punctation tokens are fed into an embedding layer, where embeddings represent each possible punctuation token. Each time step is mapped to a 4-dimensional embeddings, which is concatenated to the 512-dimensional encoding. The concatenated joint representation is re-encoded to confer global context to each time step to incorporate puncuation predictions into subsequent tasks.
Pre-punctuation After the re-encoding, another classification network predicts "pre" punctuation, or punctation tokens that may appear before a word. In practice, this means the inverted question mark for Spanish and Asturian,
¿. Note that a¿can only appear if a?is predicted, hence the conditioning.Sentence boundary detection Parallel to the "pre" punctuation, another classification network predicts sentence boundaries from the re-encoded text. In all languages, sentence boundaries can occur only if a potential full stop is predicted, hence the conditioning.
Shift and concat sentence boundaries In many languages, the first character of each sentence should be upper-cased. Thus, we should feed the sentence boundary information to the true-case classification network. Since the true-case classification network is feed-forward and has no context, each time step must embed whether it is the first word of a sentence. Therefore, we shift the binary sentence boundary decisions to the right by one: if token
N-1is a sentence boundary, tokenNis the first word of a sentence. Concatenating this with the re-encoded text, each time step contains whether it is the first word of a sentence as predicted by the SBD head.True-case prediction Armed with the knowledge of punctation and sentence boundaries, a classification network predicts true-casing. Since true-casing should be done on a per-character basis, the classification network makes
Npredictions per token, whereNis the length of the subtoken. (In practice,Nis the longest possible subword, and the extra predictions are ignored). This scheme captures acronyms, e.g., "NATO", as well as bi-capitalized words, e.g., "MacDonald".
Post-Punctuation Tokens
This model predicts the following set of "post" punctuation tokens:
| Token | Description | Relavant Languages |
|---|---|---|
| . | Latin full stop | Many |
| , | Latin comma | Many |
| ? | Latin question mark | Many |
| ？ | Full-width question mark | Chinese, Japanese |
| ， | Full-width comma | Chinese, Japanese |
| 。 | Full-width full stop | Chinese, Japanese |
| 、 | Ideographic comma | Chinese, Japanese |
| ・ | Middle dot | Japanese |
| । | Danda | Hindi, Bengali, Oriya |
| ؟ | Arabic question mark | Arabic |
| ; | Greek question mark | Greek |
| ። | Ethiopic full stop | Amharic |
| ፣ | Ethiopic comma | Amharic |
| ፧ | Ethiopic question mark | Amharic |
Pre-Punctuation Tokens
This model predicts the following set of "post" punctuation tokens:
| Token | Description | Relavant Languages |
|---|---|---|
| ¿ | Inverted question mark | Spanish |
Usage
This model is released in two parts:
- The ONNX graph
- The SentencePiece tokenizer
Training Details
This model was trained in the NeMo framework.
Training Data
This model was trained with News Crawl data from WMT.
1M lines of text for each language was used, except for a few low-resource languages which may have used less.
Languages were chosen based on whether the News Crawl corpus contained enough reliable-quality data as judged by the author.
Limitations
This model was trained on news data, and may not perform well on conversational or informal data.
This model predicts punctuation only once per subword. This implies that some acronyms, e.g., 'U.S.', cannot properly be punctuation. This concession was accepted on two grounds:
- Such acronyms are rare, especially in the context of multi-lingual models
- Punctuated acronyms are typically pronounced as individual characters, e.g., 'U.S.' vs. 'NATO'. Since the expected use-case of this model is the output of an ASR system, it is presumed that such pronunciations would be transcribed as separate tokens, e.g, 'u s' vs. 'us' (though this depends on the model's pre-processing).
Further, this model is unlikely to be of production quality. Though trained to convergence, it was trained with "only" 1M lines per language, and the dev sets may have been noisy due to the nature of web-scraped news data. This is also a base-sized model with many languages and many tasks, so capacity may be limited.
Evaluation
In these metrics, keep in mind that
- That data is noisy
- Sentence boundaries and true-casing is conditioned on predicted punctuation
English
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 98.71 98.69 98.70 107750
. (label_id: 1) 87.82 88.89 88.36 6005
, (label_id: 2) 67.90 67.24 67.57 3571
? (label_id: 3) 80.51 78.19 79.33 486
？ (label_id: 4) 0.00 0.00 0.00 0
， (label_id: 5) 0.00 0.00 0.00 0
。 (label_id: 6) 0.00 0.00 0.00 0
、 (label_id: 7) 0.00 0.00 0.00 0
・ (label_id: 8) 0.00 0.00 0.00 0
। (label_id: 9) 0.00 0.00 0.00 0
؟ (label_id: 10) 0.00 0.00 0.00 0
، (label_id: 11) 0.00 0.00 0.00 0
; (label_id: 12) 0.00 0.00 0.00 0
። (label_id: 13) 0.00 0.00 0.00 0
፣ (label_id: 14) 0.00 0.00 0.00 0
፧ (label_id: 15) 0.00 0.00 0.00 0
-------------------
micro avg 97.15 97.15 97.15 117812
macro avg 83.74 83.25 83.49 117812
weighted avg 97.15 97.15 97.15 117812
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 99.62 99.49 99.56 362399
UPPER (label_id: 1) 89.11 91.75 90.41 16506
-------------------
micro avg 99.15 99.15 99.15 378905
macro avg 94.37 95.62 94.98 378905
weighted avg 99.17 99.15 99.16 378905
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.29 99.43 99.36 111466
FULLSTOP (label_id: 1) 89.69 87.49 88.58 6346
-------------------
micro avg 98.78 98.78 98.78 117812
macro avg 94.49 93.46 93.97 117812
weighted avg 98.77 98.78 98.78 117812