--- 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: ![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' | '▁de' | '-' # spm | '' | '' | '' | ',' | '.' | '▁' | '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="", type=ModelProto.SentencePiece.Type.CONTROL), ModelProto.SentencePiece(piece="", type=ModelProto.SentencePiece.Type.CONTROL), ModelProto.SentencePiece(piece="", type=ModelProto.SentencePiece.Type.CONTROL), ModelProto.SentencePiece(piece="", type=ModelProto.SentencePiece.Type.UNKNOWN), ] + pieces[3:] + [ModelProto.SentencePiece(piece="", 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 | | ---: | :---------- | :----------- | | \ | No punctuation | All | | \ | 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 | | ---: | :---------- | :----------- | | \ | 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 slightly longer than 7 hours. For validation and train metrics, see the [Tensorboard Logs](https://tensorboard.dev/experiment/xxnULI1aTeK37vUDL4ejiw/). ## 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. 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. This can be fixed by exposing prior probabilities, but I'll fine-tune it later to fix this the right way. # 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 11 random sentences (1 + 10 for each sentence in the test set) 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 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. ## 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 ```text punct_post test report: label precision recall f1 support (label_id: 0) 99.18 98.47 98.82 538769 (label_id: 1) 66.03 78.63 71.78 571 . (label_id: 2) 90.66 93.68 92.14 30581 , (label_id: 3) 74.18 82.93 78.31 23230 ? (label_id: 4) 78.10 80.08 79.07 1024 ？ (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 97.56 97.56 97.56 594175 macro avg 81.63 86.76 84.03 594175 weighted avg 97.70 97.56 97.62 594175 ``` ```text cap test report: label precision recall f1 support LOWER (label_id: 0) 99.71 99.85 99.78 2036824 UPPER (label_id: 1) 96.40 93.27 94.81 87747 ------------------- micro avg 99.58 99.58 99.58 2124571 macro avg 98.06 96.56 97.30 2124571 weighted avg 99.57 99.58 99.58 2124571 ``` ```text seg test report: label precision recall f1 support NOSTOP (label_id: 0) 99.97 99.98 99.98 564228 FULLSTOP (label_id: 1) 99.73 99.54 99.64 32947 ------------------- micro avg 99.96 99.96 99.96 597175 macro avg 99.85 99.76 99.81 597175 weighted avg 99.96 99.96 99.96 597175 ```
Spanish ```text punct_pre test report: label precision recall f1 support (label_id: 0) 99.96 99.76 99.86 609200 ¿ (label_id: 1) 39.66 77.89 52.56 1221 ------------------- micro avg 99.72 99.72 99.72 610421 macro avg 69.81 88.82 76.21 610421 weighted avg 99.83 99.72 99.76 610421 ``` ```text punct_post test report: label precision recall f1 support (label_id: 0) 99.17 98.44 98.80 553100 (label_id: 1) 23.33 43.75 30.43 48 . (label_id: 2) 91.92 92.58 92.25 29623 , (label_id: 3) 73.07 82.04 77.30 26432 ? (label_id: 4) 49.44 71.84 58.57 1218 ？ (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 97.39 97.39 97.39 610421 macro avg 67.39 77.73 71.47 610421 weighted avg 97.58 97.39 97.47 610421 ``` ```text cap test report: label precision recall f1 support LOWER (label_id: 0) 99.82 99.86 99.84 2222062 UPPER (label_id: 1) 95.96 94.64 95.29 75940 ------------------- micro avg 99.69 99.69 99.69 2298002 macro avg 97.89 97.25 97.57 2298002 weighted avg 99.69 99.69 99.69 2298002 ``` ```text seg test report: label precision recall f1 support NOSTOP (label_id: 0) 99.99 99.97 99.98 580519 FULLSTOP (label_id: 1) 99.52 99.81 99.66 32902 ------------------- micro avg 99.96 99.96 99.96 613421 macro avg 99.75 99.89 99.82 613421 weighted avg 99.96 99.96 99.96 613421 ```
Amharic ```text punct_post test report: label precision recall f1 support (label_id: 0) 99.81 99.40 99.60 729695 (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) 91.44 97.78 94.50 25288 ፣ (label_id: 15) 66.93 80.45 73.07 5774 ፧ (label_id: 16) 72.14 77.01 74.49 1170 ------------------- micro avg 99.17 99.17 99.17 761927 macro avg 82.58 88.66 85.42 761927 weighted avg 99.24 99.17 99.19 761927 ``` ```text cap test report: label precision recall f1 support LOWER (label_id: 0) 98.50 97.22 97.86 1150 UPPER (label_id: 1) 56.16 70.69 62.60 58 ------------------- micro avg 95.94 95.94 95.94 1208 macro avg 77.33 83.95 80.23 1208 weighted avg 96.47 95.94 96.16 1208 ``` ```text seg test report: label precision recall f1 support NOSTOP (label_id: 0) 99.97 99.91 99.94 743103 FULLSTOP (label_id: 1) 97.16 99.04 98.09 21824 ------------------- micro avg 99.89 99.89 99.89 764927 macro avg 98.57 99.48 99.02 764927 weighted avg 99.89 99.89 99.89 764927 ```
Chinese ```text punct_post test report: label precision recall f1 support (label_id: 0) 99.47 97.46 98.45 414383 (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) 81.41 85.80 83.55 1444 ， (label_id: 6) 74.93 92.79 82.91 34094 。 (label_id: 7) 96.35 96.86 96.60 30668 、 (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 97.05 97.05 97.05 480589 macro avg 88.04 93.23 90.38 480589 weighted avg 97.47 97.05 97.19 480589 ``` ```text cap test report: label precision recall f1 support LOWER (label_id: 0) 94.82 93.97 94.39 2786 UPPER (label_id: 1) 79.23 81.76 80.48 784 ------------------- micro avg 91.29 91.29 91.29 3570 macro avg 87.03 87.87 87.44 3570 weighted avg 91.40 91.29 91.34 3570 ``` ```text seg test report: label precision recall f1 support NOSTOP (label_id: 0) 99.99 99.98 99.98 450589 FULLSTOP (label_id: 1) 99.75 99.81 99.78 33000 ------------------- micro avg 99.97 99.97 99.97 483589 macro avg 99.87 99.89 99.88 483589 weighted avg 99.97 99.97 99.97 483589 ```
Japanese ```text punct_post test report: label precision recall f1 support (label_id: 0) 99.32 95.84 97.55 387103 (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) 75.12 68.14 71.46 1378 ， (label_id: 6) 0.00 0.00 0.00 0 。 (label_id: 7) 93.30 97.44 95.33 31110 、 (label_id: 8) 53.91 87.52 66.72 17710 ・ (label_id: 9) 29.33 64.60 40.35 1048 । (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 95.46 95.46 95.46 438349 macro avg 70.20 82.71 74.28 438349 weighted avg 96.81 95.46 95.93 438349 ``` ```text cap test report: label precision recall f1 support LOWER (label_id: 0) 92.64 92.67 92.65 4036 UPPER (label_id: 1) 80.75 80.70 80.73 1539 ------------------- micro avg 89.36 89.36 89.36 5575 macro avg 86.70 86.68 86.69 5575 weighted avg 89.36 89.36 89.36 5575 ``` ```text seg test report: label precision recall f1 support NOSTOP (label_id: 0) 99.98 99.95 99.97 408349 FULLSTOP (label_id: 1) 99.36 99.78 99.57 33000 ------------------- micro avg 99.94 99.94 99.94 441349 macro avg 99.67 99.86 99.77 441349 weighted avg 99.94 99.94 99.94 441349 ```
Hindi ```text punct_post test report: label precision recall f1 support (label_id: 0) 99.73 99.47 99.60 533761 (label_id: 1) 0.00 0.00 0.00 0 . (label_id: 2) 0.00 0.00 0.00 0 , (label_id: 3) 70.69 76.48 73.47 7713 ? (label_id: 4) 65.41 74.75 69.77 301 ？ (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) 96.46 98.81 97.62 30641 ؟ (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 572416 macro avg 83.07 87.38 85.11 572416 weighted avg 99.15 99.11 99.13 572416 ``` ```text cap test report: label precision recall f1 support LOWER (label_id: 0) 97.46 96.50 96.98 2346 UPPER (label_id: 1) 89.01 91.84 90.40 723 ------------------- micro avg 95.41 95.41 95.41 3069 macro avg 93.23 94.17 93.69 3069 weighted avg 95.47 95.41 95.43 3069 ``` ```text seg test report: label precision recall f1 support NOSTOP (label_id: 0) 100.00 100.00 100.00 542437 FULLSTOP (label_id: 1) 99.92 99.97 99.95 32979 ------------------- micro avg 99.99 99.99 99.99 575416 macro avg 99.96 99.98 99.97 575416 weighted avg 99.99 99.99 99.99 575416 ```