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
The data is noisy
Sentence boundaries and true-casing are conditioned on predicted punctuation, which is the most difficult task and sometimes incorrect. When conditioning on reference punctuation, true-casing and SBD is practically 100% for most languages.
Punctuation can be subjective. E.g.,
Hola mundo, ¿cómo estás?or
Hola mundo. ¿Cómo estás?When the sentences are longer and more practical, these ambiguities abound and affect all 3 analytics.
Selected Language Evaluation Reports
Each test example was generated using the following procedure:
- Concatenate 5 random sentences
- Lower-case the concatenated sentence
- Remove all punctuation
The data is a held-out portion of News Crawl, which has been deduplicated. 2,000 lines of data per language was used, generating 2,000 unique examples of 5 sentences each. The last 4 sentences of each example were randomly sampled from the 2,000 and may be duplicated.
English
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 98.71 98.66 98.68 156605
. (label_id: 1) 87.72 88.85 88.28 8752
, (label_id: 2) 68.06 67.81 67.93 5216
? (label_id: 3) 79.38 77.20 78.27 693
？ (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 97.13 97.13 97.13 171266
macro avg 83.46 83.13 83.29 171266
weighted avg 97.13 97.13 97.13 171266
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 99.63 99.49 99.56 526612
UPPER (label_id: 1) 89.19 91.84 90.50 24161
-------------------
micro avg 99.15 99.15 99.15 550773
macro avg 94.41 95.66 95.03 550773
weighted avg 99.17 99.15 99.16 550773
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.37 99.42 99.39 162044
FULLSTOP (label_id: 1) 89.75 88.84 89.29 9222
-------------------
micro avg 98.85 98.85 98.85 171266
macro avg 94.56 94.13 94.34 171266
weighted avg 98.85 98.85 98.85 171266
Spanish
punct_pre test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.94 99.92 99.93 185535
¿ (label_id: 1) 55.01 64.86 59.53 296
-------------------
micro avg 99.86 99.86 99.86 185831
macro avg 77.48 82.39 79.73 185831
weighted avg 99.87 99.86 99.87 185831
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 98.74 98.86 98.80 170282
. (label_id: 1) 90.07 89.58 89.82 9959
, (label_id: 2) 68.33 67.00 67.66 5300
? (label_id: 3) 70.25 58.62 63.91 290
？ (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 97.39 97.39 97.39 185831
macro avg 81.84 78.51 80.05 185831
weighted avg 97.36 97.39 97.37 185831
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 99.62 99.60 99.61 555041
UPPER (label_id: 1) 90.60 91.06 90.83 23538
-------------------
micro avg 99.25 99.25 99.25 578579
macro avg 95.11 95.33 95.22 578579
weighted avg 99.25 99.25 99.25 578579
[NeMo I 2023-02-22 17:24:04 punct_cap_seg_model:427] seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.44 99.54 99.49 175908
FULLSTOP (label_id: 1) 91.68 89.98 90.82 9923
-------------------
micro avg 99.03 99.03 99.03 185831
macro avg 95.56 94.76 95.16 185831
weighted avg 99.02 99.03 99.02 185831
Chinese
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 98.82 97.34 98.07 147920
. (label_id: 1) 0.00 0.00 0.00 0
, (label_id: 2) 0.00 0.00 0.00 0
? (label_id: 3) 0.00 0.00 0.00 0
？ (label_id: 4) 85.77 80.71 83.16 560
， (label_id: 5) 59.88 78.02 67.75 6901
。 (label_id: 6) 92.50 93.92 93.20 10988
、 (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 96.25 96.25 96.25 166369
macro avg 84.24 87.50 85.55 166369
weighted avg 96.75 96.25 96.45 166369
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 97.07 92.39 94.67 394
UPPER (label_id: 1) 70.59 86.75 77.84 83
-------------------
micro avg 91.40 91.40 91.40 477
macro avg 83.83 89.57 86.25 477
weighted avg 92.46 91.40 91.74 477
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.58 99.53 99.56 156369
FULLSTOP (label_id: 1) 92.77 93.50 93.13 10000
-------------------
micro avg 99.17 99.17 99.17 166369
macro avg 96.18 96.52 96.35 166369
weighted avg 99.17 99.17 99.17 166369
Hindi
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.58 99.59 99.59 176743
. (label_id: 1) 0.00 0.00 0.00 0
, (label_id: 2) 68.32 65.23 66.74 1815
? (label_id: 3) 60.27 44.90 51.46 98
？ (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) 96.45 97.43 96.94 10136
؟ (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 99.11 99.11 99.11 188792
macro avg 81.16 76.79 78.68 188792
weighted avg 99.10 99.11 99.10 188792
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 98.25 95.06 96.63 708
UPPER (label_id: 1) 89.46 96.12 92.67 309
-------------------
micro avg 95.38 95.38 95.38 1017
macro avg 93.85 95.59 94.65 1017
weighted avg 95.58 95.38 95.42 1017
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.87 99.85 99.86 178892
FULLSTOP (label_id: 1) 97.38 97.58 97.48 9900
-------------------
micro avg 99.74 99.74 99.74 188792
macro avg 98.62 98.72 98.67 188792
weighted avg 99.74 99.74 99.74 188792
Amharic
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.58 99.42 99.50 236298
. (label_id: 1) 0.00 0.00 0.00 0
, (label_id: 2) 0.00 0.00 0.00 0
? (label_id: 3) 0.00 0.00 0.00 0
？ (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) 89.79 95.24 92.44 9169
፣ (label_id: 14) 66.85 56.58 61.29 1504
፧ (label_id: 15) 67.67 83.72 74.84 215
-------------------
micro avg 98.99 98.99 98.99 247186
macro avg 80.97 83.74 82.02 247186
weighted avg 98.99 98.99 98.98 247186
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 96.65 99.78 98.19 1360
UPPER (label_id: 1) 98.90 85.13 91.50 316
-------------------
micro avg 97.02 97.02 97.02 1676
macro avg 97.77 92.45 94.84 1676
weighted avg 97.08 97.02 96.93 1676
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.85 99.74 99.80 239845
FULLSTOP (label_id: 1) 91.72 95.25 93.45 7341
-------------------
micro avg 99.60 99.60 99.60 247186
macro avg 95.79 97.49 96.62 247186
weighted avg 99.61 99.60 99.61 247186