LAMBO segmenter

LAMBO (Layered Approach to Multi-level BOundary identification) is a segmentation tool that is able to divide text on several levels:

1. Dividing the original text into turns according to the provided list of separators. Turns can correspond to seperate utterences in a dialogue, paragraphs in a continuous text, etc.
2. Splitting each turn into sentences.
3. Finding tokens in sentences. Most tokens correspond to words, but multi-word tokens are also detected. LAMBO also supports special tokens that should be kept separate regardless of context, such as emojis and pause markers.

LAMBO is a machine learning model, which means it was trained to recognise boundaries of tokens and sentences from real-world text. It is implemented as a PyTorch deep neural network, including embeddings and recurrent layers operating at the character level. At the same time, LAMBO contains rule-based elements to allow a user to easily adjust it to one's needs, e.g. by adding custom special tokens or turn division markers.

LAMBO was developed in context of dependency parsing. Thus, it includes models trained on Universal Dependencies treebanks, uses .conllu as the training data format and supports integration with COMBO, a state-of-the-art system for dependency parsing and more. However, you can use LAMBO as the first stage of any NLP process.

LAMBO currently includes models trained on 98 corpora in 53 languages. The full list is available in languages.txt. For each of these, two model variants are available:

• simple LAMBO, trained on the UD corpus
• pretrained LAMBO, same as above, but starting from weights pre-trained on unsupervised masked character prediction using multilingual corpora from OSCAR.

Installation

Installation of LAMBO is easy.

1. First, you need to prepare an environment with Python, at least 3.6.9,

2. Then, download LAMBO from this repository:

git clone https://gitlab.clarin-pl.eu/syntactic-tools/lambo.git

3. Install LAMBO:

pip install ./lambo

You now have LAMBO installed in your environment.

Using LAMBO

To use LAMBO, you first need to import it:

from lambo.segmenter.lambo import Lambo

Now you need to create a segmenter by providing the language your text is in, e.g. English:

lambo = Lambo.get('English')

This will (if necessary) download the appropriate model from the online repository and load it. Note that you can use any language name (e.g. Ancient_Greek) or ISO 639-1 code (e.g. fi) from languages.txt.

Alternatively, you can select a specific model by defining LAMBO variant (LAMBO or LAMBO_no_pretraining) and training dataset from languages.txt:

lambo = Lambo.get('LAMBO-UD_Polish-PDB')

Once the model is ready, you can perform segmentation of a given text:

text = "A simple sentence might be enough... But some of us just ❤️ emojis. They should be tokens even when (yy) containing many characters, such as 👍🏿."
document = lambo.segment(text)

The document object contains a list of turns, each composed of sentences, which in turn include tokens. The structured could be explored in the following way:

for turn in document.turns:
    print('======= TURN =======')
    print('TEXT: ' + turn.text[:100] + '...')
    for sentence in turn.sentences:
        print('======= SENTENCE =======')
        print('TEXT: "' + sentence.text + '"')
        formatted = ''
        for token in sentence.tokens:
            if token.is_multi_word:
                formatted += '((' + token.text + '))'
            else:
                formatted += '(' + token.text + ')'
        print('TOKENS: ' + formatted)

Chech out if the special tokens, i.e. emojis and pause ((yy)) were properly recognised.

Using LAMBO with COMBO

You can use LAMBO to segment text that is going to be processed by COMBO. To do that, first you will need to install both COMBO and LAMBO in the same environment. Make sure you have the COMBO version supporting LAMBO.

Once both tools are ready, you need to import COMBO with LAMBO tokeniser:

from combo.predict import COMBO
from combo.utils import lambo

You can now create COMBO instance using LAMBO for segmentation. Make sure both use models appropriate for the language of text:

nlp_new = COMBO.from_pretrained("polish-herbert-base-ud29",tokenizer=lambo.LamboTokenizer("pl"))

Now, there are two ways for interacting with COMBO-LAMBO duet. You can provide a list of sentences as input and LAMBO will tokenise each of them and pass to COMBO for further processing:

text = ["To zdanie jest OK 👍🏿.", "To jest drugie zdanie."]
sentences = nlp_new(text)

Alternatively, you can provide a single string, using the fact that LAMBO is more than a tokeniser and can split sentences on its own:

text="To zdanie jest OK 👍🏿. To jest drugie zdanie."
sentences = nlp_new(text)

In any case, you should get a full dependency parsing output, which you can print via:

print("{:5} {:15} {:15} {:10} {:10} {:10}".format('ID', 'TOKEN', 'LEMMA', 'UPOS', 'HEAD', 'DEPREL'))
	for sentence in sentences:
		for token in sentence.tokens:
			print("{:5} {:15} {:15} {:10} {:10} {:10}".format(str(token.id), token.token, token.lemma, token.upostag, str(token.head), token.deprel))
		print("\n")

Extending LAMBO

You don't have to rely on the models trained so far in COMBO. You can use the included code to train on new corpora and languages, tune to specific usecases or simply retrain larger models with more resources. The scripts in examples include examples on how to do that:

• run_training.py -- train simple LAMBO models. This script was used with UD treebanks to generate LAMBO_no_pretraining models.
• run_pretraining.py -- pretrain unsupervised LAMBO models. This script was used with OSCAR.
• run_training_pretrained.py -- train LAMBO models on UD training data, starting from pretrained models. This script was used to generate LAMBO models.
• run_tuning.py -- tune existing LAMBO model to fit new data.
• run_evaluation.py -- evaluate existing models using UD gold standard.

Credits

If you use LAMBO in your research, please cite it as software:

@software{LAMBO,
  author = {{Przyby{\l}a, Piotr}},
  title = {LAMBO: Layered Approach to Multi-level BOundary identification},
  url = {https://gitlab.clarin-pl.eu/syntactic-tools/lambo},
  version = {1.0.0},
  year = {2022},
}

License

This project is licensed under the GNU General Public License v3.0.