README.md

---
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 a fine-tuned `xlm-roberta` model that restores punctuation, true-cases (capitalizes), 
and detects 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: 

![graph.png](https://s3.amazonaws.com/moonup/production/uploads/62d34c813eebd640a4f97587/jpr-pMdv6iHxnjbN4QNt0.png)

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,

```python
# 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):

```python
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())
```


## 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.

## 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.

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.



# Evaluation
In these metrics, keep in mind that
1. The data is noisy
2. 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.
4. 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:

1. Concatenate 10 random sentences
2. Lower-case the concatenated sentence
3. Remove all 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 10 sentences each.
The last 4 sentences of each example were randomly sampled from the 3,000 and may be duplicated.

Examples longer than the model's maximum length were truncated. 
The number of affected sentences can be estimated from the "full stop" support: with 3,000 
sentences and 10 sentences per example, we expect 30,000 full stop targets total.

## Selected Language Evaluation Reports