metadata

license: apache-2.0
language:
  - en
tags:
  - punctuation
  - true casing
  - sentence boundary detection
  - token classification
  - nlp

Model Overview

This model accepts as input lower-cased, unpunctuated English text and performs in one pass punctuation restoration, true-casing (capitalization), and sentence boundary detection (segmentation).

In contast to many similar models, this model can predict punctuated acronyms (e.g., "U.S.") via a special "acronym" class, as well as arbitarily-capitalized words (NATO, McDonald's, etc.) via multi-label true-casing predictions.

Usage

The easy way to use this model is to install punctuators:

pip install punctuators

Running the following script should load this model and run some texts:

Example Usage

from punctuators.models import PunctCapSegModelONNX

# Instantiate this model
# This will download the ONNX and SPE models. To clean up, delete this model from your HF cache directory.
m = PunctCapSegModelONNX.from_pretrained("pcs_en")

# Define some input texts to punctuate
input_texts: List[str] = [
    "hello friend how's it going it's snowing outside right now in connecticut a large storm is moving in",
    "i live in the us where george hw bush was once president"
]
results: List[List[str]] = m.infer(input_texts)
for input_text, output_texts in zip(input_texts, results):
    print(f"Input: {input_text}")
    print(f"Outputs:")
    for text in output_texts:
        print(f"\t{text}")
    print()

Expected Output

Note that "Friend" in this context is a proper noun, which is why the model consistently upper-cases tokens in similar contexts.

Model Details

This model implements 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 SentencePiece model with a vocabulary size of 32k. Next, the input sequence is encoded with a base-sized Transformer, consisting of 6 layers with a model dimension of 512.
Punctuation: The encoded sequence is then fed into a feed-forward classification network to predict punctuation tokens. Punctation is predicted once per subword, to allow acronyms to be properly punctuated. An indiret benefit of per-subword prediction is to allow the model to run in a graph generalized for continuous-script languages, e.g., Chinese.
Sentence boundary detection For sentence boundary detection, we condition the model on punctuation via embeddings. Each punctuation prediction is used to select an embedding for that token, which is concatenated to the encoded representation. The SBD head analyzes both the encoding of the un-punctuated sequence and the puncutation predictions, and predicts which tokens are sentence boundaries.
Shift and concat sentence boundaries In English, 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-1 is a sentence boundary, token N is the first word of a sentence. Concatenating this with the 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 N predictions per token, where N is the length of the subtoken. (In practice, N is 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".

The model's maximum length is 256 subtokens, due to the limit of the trained embeddings. However, the punctuators package as described above will transparently predict on overlapping subgsegments of long inputs and fuse the results before returning output, allowing inputs to be arbitrarily long.

Punctuation Tokens

This model predicts the following set of punctuation tokens:

Token	Description
NULL	Predict no punctuation
ACRONYM	Every character in this subword ends with a period
.	Latin full stop
,	Latin comma
?	Latin question mark

Training Details

Training Framework

This model was trained on a forked branch of the NeMo framework.

Training Data

This model was trained with News Crawl data from WMT.

Approximately 10M lines were used from the years 2021 and 2012. The latter was used to attempt to reduce bias: annual news is typically dominated by a few topics, and 2021 is dominated by COVID discussions.

Limitations

Domain

This model was trained on news data, and may not perform well on conversational or informal data.

Noisy Training Data

The training data was noisy, and no manual cleaning was utilized.

Acronyms and abbreviations are especially noisy; the table below shows how many variations of each token appear in the training data.

Token	Count
Mr	115232
Mr.	108212

Token	Count
U.S.	85324
US	37332
U.S	354
U.s	108
u.S.	65

Thus, the model's acronym and abbreviation predictions may be a bit unpredictable.

Further, an assumption for sentence boundary detection targets is that each line of the input data is exactly one sentence. However, a non-negligible portion of the training data contains multiple sentences in one line. Thus, the SBD head may miss an obvious sentence boundary if it's similar to an error seen in the training data.

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 metrics are much higher w.r.t. the reference targets.
Punctuation can be subjective. E.g.,

Hello Frank, how's it going?

or

Hello Frank. How's it going?

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 10 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 was used, generating 2,000 unique examples of 10 sentences each.

Examples longer than the model's maximum length (256) were truncated. The number of affected sentences can be estimated from the "full stop" support: with 2,000 sentences and 10 sentences per example, we expect 20,000 full stop targets total.

Results

Fun Facts

Some fun facts are examined in this section.

Embeddings

Let's examine the embeddings (see graph above) to see if the model meaningfully employed them.

We show here the cosine similarity between the embeddings of each token:

	NULL	ACRONYM	.	,	?
NULL	1.00
ACRONYM	-0.93	1.00
.	-1.00	0.94	1.00
,	1.00	-0.94	-1.00	1.00
?	-1.00	0.93	1.00	-1.00	1.00

Recall that these embeddings are used to predict sentence boundaries... thus we should expect full stops to cluster.

Indeed, we see that NULL and "," are exactly the same, because neither have an implication on sentence boundaries.

Next, we see that "." and "?" are exactly the same, because w.r.t. SBD these are exactly the same: strong full stop implications. (Though, we may expect some difference between these tokens, given that "." is predicted after abbreviations, e.g., 'Mr.', that are not full stops.)

Further, we see that "." and "?" are exactly the opposite of NULL. This is expected since these tokens typically imply sentence boundaries, whereas NULL and "," do not.

Lastly, we see that ACRONYM is very, but not totally, similar to the full stops "." and "?", and almost, but not totally, the opposite of NULL and ",". Intuition suggests this is because acronyms can be full stops ("I live in the northern U.S. It's cold here.") or not ("It's 5 a.m. and I'm tired.").