license: apache-2.0
library_name: generic
tags:
- text2text-generation
- punctuation
- sentence-boundary-detection
- truecasing
- true-casing
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 is an xlm-roberta fine-tuned to restore punctuation, true-case (capitalize),
and detect sentence boundaries (full stops) in 47 languages.
Model Architecture
This model implements the following graph, which allows punctuation, true-casing, and fullstop prediction in every language without language-specific behavior:
We start by tokenizing the text and encoding it with XLM-Roberta, which is the pre-trained portion of this graph.
Then we predict punctuation before and after every subtoken. Predicting before each token allows for Spanish inverted question marks. Predicting after every token allows for all other punctuation, including punctuation within continuous-script languages and acronyms.
We use embeddings to represent the predicted punctuation tokens to inform the sentence boundary head of the punctuation that'll be inserted into the text. This allows proper full stop prediction, since certain punctuation tokens (periods, questions marks, etc.) are strongly correlated with sentence boundaries.
We then shift full stop predictions to the right by one, to inform the true-casing head of where the beginning of each new sentence is. This is important since true-casing is strongly correlated to sentence boundaries.
For true-casing, we predict N predictions per subtoken, where N is the number of characters in the subtoken.
In practice, N is the maximum subtoken length and extra predictions are ignored. Essentially, true-casing is
modeled as a multi-label problem. This allows for upper-casing arbitrary characters, e.g., "NATO", "MacDonald", "mRNA", etc.
Applying all these predictions to the input text, we can punctuate, true-case, and split sentences in any language.
Tokenizer
Instead of the hacky wrapper used by FairSeq and strangely ported (not fixed) by HuggingFace, the xlm-roberta SentencePiece model was adjusted to correctly encode
the text. Per HF's comments,
# Original fairseq vocab and spm vocab must be "aligned":
# Vocab | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9
# -------- | ------- | ------- | ------ | ------- | --- | --- | --- | ----- | ----- | ----
# fairseq | '<s>' | '<pad>' | '</s>' | '<unk>' | ',' | '.' | '▁' | 's' | '▁de' | '-'
# spm | '<unk>' | '<s>' | '</s>' | ',' | '.' | '▁' | 's' | '▁de' | '-' | '▁a'
The SP model was un-hacked with the following snippet (SentencePiece experts, let me know if there is a problem here):
from sentencepiece import SentencePieceProcessor
from sentencepiece.sentencepiece_model_pb2 import ModelProto
m = ModelProto()
m.ParseFromString(open("/path/to/xlmroberta/sentencepiece.bpe.model", "rb").read())
pieces = list(m.pieces)
pieces = (
[
ModelProto.SentencePiece(piece="<s>", type=ModelProto.SentencePiece.Type.CONTROL),
ModelProto.SentencePiece(piece="<pad>", type=ModelProto.SentencePiece.Type.CONTROL),
ModelProto.SentencePiece(piece="</s>", type=ModelProto.SentencePiece.Type.CONTROL),
ModelProto.SentencePiece(piece="<unk>", type=ModelProto.SentencePiece.Type.UNKNOWN),
]
+ pieces[3:]
+ [ModelProto.SentencePiece(piece="<mask>", type=ModelProto.SentencePiece.Type.USER_DEFINED)]
)
del m.pieces[:]
m.pieces.extend(pieces)
with open("/path/to/new/sp.model", "wb") as f:
f.write(m.SerializeToString())
Now we can use just the SP model without a wrapper.
Post-Punctuation Tokens
This model predicts the following set of punctuation tokens after each subtoken:
| Token | Description | Relevant Languages |
|---|---|---|
| <NULL> | No punctuation | All |
| <ACRONYM> | Every character in this subword is followed by a period | Primarily English, some European |
| . | 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 punctuation tokens before each subword:
| Token | Description | Relevant Languages |
|---|---|---|
| <NULL> | No punctuation | All |
| ¿ | Inverted question mark | Spanish |
Training Details
This model was trained in the NeMo framework.
This model was trained on an A100 for approximately 9 hours.
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. Notably, when presented with many short sentences, the model misses obvious sentence boundaries since the model was trained on relatively-long sentences.
Further, this model is unlikely to be of production quality. 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 model over-predicts Spanish question marks, especially the inverted question mark ¿ (see metrics below).
Since ¿ is a rare token, especially in the context of a 47-language model, Spanish questions were over-sampled
by selecting more of these sentences from additional training data that was not used. However, this seems to have
"over-corrected" the problem and a lot of Spanish question marks are predicted.
The model may also over-predict commas.
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.
Test Data and Example Generation
Each test example was generated using the following procedure:
- Concatenate 11 random sentences (1 + 10 for each sentence in the test set)
- Lower-case the concatenated sentence
- Remove all punctuation
Targets are generated as we lower-case letters and remove punctuation.
The data is a held-out portion of News Crawl, which has been deduplicated.
3,000 lines of data per language was used, generating 3,000 unique examples of 11 sentences each.
We generate 3,000 examples, where example i begins with sentence i and is followed by 10 random
sentences selected from the 3,000 sentence test set.
For measuring true-casing and sentence boundary detection, reference punctuation tokens were used for conditioning (see graph above). If we use predicted punctuation instead, then incorrect punctuation will result in true-casing and SBD targets not aligning correctly and these metrics will be artificially low.
Selected Language Evaluation Reports
For now, metrics for a few selected languages are shown below. Given the amount of work required to collect and pretty-print metrics in 47 languages, I'll add more eventually.
Expand any of the following tabs to see metrics for that language.
English
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.25 98.43 98.84 564908
<ACRONYM> (label_id: 1) 63.14 84.67 72.33 613
. (label_id: 2) 90.97 93.91 92.42 32040
, (label_id: 3) 73.95 84.32 78.79 24271
? (label_id: 4) 79.05 81.94 80.47 1041
？ (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
፧ (label_id: 16) 0.00 0.00 0.00 0
-------------------
micro avg 97.60 97.60 97.60 622873
macro avg 81.27 88.65 84.57 622873
weighted avg 97.77 97.60 97.67 622873
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 99.72 99.85 99.78 2134956
UPPER (label_id: 1) 96.33 93.52 94.91 91996
-------------------
micro avg 99.59 99.59 99.59 2226952
macro avg 98.03 96.68 97.34 2226952
weighted avg 99.58 99.59 99.58 2226952
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.99 99.98 99.99 591540
FULLSTOP (label_id: 1) 99.61 99.89 99.75 34333
-------------------
micro avg 99.97 99.97 99.97 625873
macro avg 99.80 99.93 99.87 625873
weighted avg 99.97 99.97 99.97 625873
Spanish
punct_pre test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.94 99.89 99.92 636941
¿ (label_id: 1) 56.73 71.35 63.20 1288
-------------------
micro avg 99.83 99.83 99.83 638229
macro avg 78.34 85.62 81.56 638229
weighted avg 99.85 99.83 99.84 638229
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.19 98.41 98.80 578271
<ACRONYM> (label_id: 1) 30.10 56.36 39.24 55
. (label_id: 2) 91.92 93.12 92.52 30856
, (label_id: 3) 72.98 82.44 77.42 27761
? (label_id: 4) 52.77 71.85 60.85 1286
？ (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
፧ (label_id: 16) 0.00 0.00 0.00 0
-------------------
micro avg 97.40 97.40 97.40 638229
macro avg 69.39 80.44 73.77 638229
weighted avg 97.60 97.40 97.48 638229
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 99.82 99.86 99.84 2324724
UPPER (label_id: 1) 95.92 94.70 95.30 79266
-------------------
micro avg 99.69 99.69 99.69 2403990
macro avg 97.87 97.28 97.57 2403990
weighted avg 99.69 99.69 99.69 2403990
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.99 99.96 99.98 607057
FULLSTOP (label_id: 1) 99.31 99.88 99.60 34172
-------------------
micro avg 99.96 99.96 99.96 641229
macro avg 99.65 99.92 99.79 641229
weighted avg 99.96 99.96 99.96 641229
Amharic
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.83 99.28 99.56 729664
<ACRONYM> (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) 0.00 0.00 0.00 0
። (label_id: 14) 91.27 97.90 94.47 25341
፣ (label_id: 15) 61.93 82.11 70.60 5818
፧ (label_id: 16) 67.41 81.73 73.89 1177
-------------------
micro avg 99.08 99.08 99.08 762000
macro avg 80.11 90.26 84.63 762000
weighted avg 99.21 99.08 99.13 762000
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 98.40 98.03 98.21 1064
UPPER (label_id: 1) 71.23 75.36 73.24 69
-------------------
micro avg 96.65 96.65 96.65 1133
macro avg 84.81 86.69 85.73 1133
weighted avg 96.74 96.65 96.69 1133
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.99 99.85 99.92 743158
FULLSTOP (label_id: 1) 95.20 99.62 97.36 21842
-------------------
micro avg 99.85 99.85 99.85 765000
macro avg 97.59 99.74 98.64 765000
weighted avg 99.85 99.85 99.85 765000
Chinese
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.53 97.31 98.41 435611
<ACRONYM> (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) 81.85 87.31 84.49 1513
， (label_id: 6) 74.08 93.67 82.73 35921
。 (label_id: 7) 96.51 96.93 96.72 32097
、 (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
፧ (label_id: 16) 0.00 0.00 0.00 0
-------------------
micro avg 97.00 97.00 97.00 505142
macro avg 87.99 93.81 90.59 505142
weighted avg 97.48 97.00 97.15 505142
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 94.89 94.98 94.94 2951
UPPER (label_id: 1) 81.34 81.03 81.18 796
-------------------
micro avg 92.02 92.02 92.02 3747
macro avg 88.11 88.01 88.06 3747
weighted avg 92.01 92.02 92.01 3747
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.99 99.97 99.98 473642
FULLSTOP (label_id: 1) 99.55 99.90 99.72 34500
-------------------
micro avg 99.96 99.96 99.96 508142
macro avg 99.77 99.93 99.85 508142
weighted avg 99.96 99.96 99.96 508142
Japanese
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.34 95.90 97.59 406341
<ACRONYM> (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) 70.55 73.56 72.02 1456
， (label_id: 6) 0.00 0.00 0.00 0
。 (label_id: 7) 94.38 96.95 95.65 32537
、 (label_id: 8) 54.28 87.62 67.03 18610
・ (label_id: 9) 28.18 71.64 40.45 1100
। (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
፧ (label_id: 16) 0.00 0.00 0.00 0
-------------------
micro avg 95.51 95.51 95.51 460044
macro avg 69.35 85.13 74.55 460044
weighted avg 96.91 95.51 96.00 460044
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 92.33 94.03 93.18 4174
UPPER (label_id: 1) 83.51 79.46 81.43 1587
-------------------
micro avg 90.02 90.02 90.02 5761
macro avg 87.92 86.75 87.30 5761
weighted avg 89.90 90.02 89.94 5761
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.99 99.92 99.96 428544
FULLSTOP (label_id: 1) 99.07 99.87 99.47 34500
-------------------
micro avg 99.92 99.92 99.92 463044
macro avg 99.53 99.90 99.71 463044
weighted avg 99.92 99.92 99.92 463044
Hindi
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.75 99.44 99.59 560358
<ACRONYM> (label_id: 1) 0.00 0.00 0.00 0
. (label_id: 2) 0.00 0.00 0.00 0
, (label_id: 3) 69.55 78.48 73.75 8084
? (label_id: 4) 63.30 87.07 73.31 317
？ (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) 96.92 98.66 97.78 32118
؟ (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
፧ (label_id: 16) 0.00 0.00 0.00 0
-------------------
micro avg 99.11 99.11 99.11 600877
macro avg 82.38 90.91 86.11 600877
weighted avg 99.17 99.11 99.13 600877
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 97.19 96.72 96.95 2466
UPPER (label_id: 1) 89.14 90.60 89.86 734
-------------------
micro avg 95.31 95.31 95.31 3200
macro avg 93.17 93.66 93.41 3200
weighted avg 95.34 95.31 95.33 3200
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 100.00 99.99 99.99 569472
FULLSTOP (label_id: 1) 99.82 99.99 99.91 34405
-------------------
micro avg 99.99 99.99 99.99 603877
macro avg 99.91 99.99 99.95 603877
weighted avg 99.99 99.99 99.99 603877
Arabic
punct_post test report:
label precision recall f1 support
<NULL> (label_id: 0) 99.30 96.94 98.10 688043
<ACRONYM> (label_id: 1) 93.33 77.78 84.85 18
. (label_id: 2) 93.31 93.78 93.54 28175
, (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) 65.93 82.79 73.40 860
، (label_id: 12) 44.89 79.20 57.30 20941
; (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
፧ (label_id: 16) 0.00 0.00 0.00 0
-------------------
micro avg 96.29 96.29 96.29 738037
macro avg 79.35 86.10 81.44 738037
weighted avg 97.49 96.29 96.74 738037
cap test report:
label precision recall f1 support
LOWER (label_id: 0) 97.10 99.49 98.28 4137
UPPER (label_id: 1) 98.71 92.89 95.71 1729
-------------------
micro avg 97.55 97.55 97.55 5866
macro avg 97.90 96.19 96.99 5866
weighted avg 97.57 97.55 97.52 5866
seg test report:
label precision recall f1 support
NOSTOP (label_id: 0) 99.99 99.97 99.98 710456
FULLSTOP (label_id: 1) 99.39 99.85 99.62 30581
-------------------
micro avg 99.97 99.97 99.97 741037
macro avg 99.69 99.91 99.80 741037
weighted avg 99.97 99.97 99.97 741037