Contextualized Topic Models (CTM) are a family of topic models that use pre-trained representations of language (e.g., BERT) to support topic modeling. See the papers for details:
- Bianchi, F., Terragni, S., Hovy, D., Nozza, D., & Fersini, E. (2021). Cross-lingual Contextualized Topic Models with Zero-shot Learning. EACL. https://arxiv.org/pdf/2004.07737v2.pdf
- Bianchi, F., Terragni, S., & Hovy, D. (2020). Pre-training is a Hot Topic: Contextualized Document Embeddings Improve Topic Coherence https://arxiv.org/pdf/2004.03974.pdf
Our new topic modeling family supports many different languages (i.e., the one supported by HuggingFace models) and comes in two versions: CombinedTM combines contextual embeddings with the good old bag of words to make more coherent topics; ZeroShotTM is the perfect topic model for task in which you might have missing words in the test data and also, if trained with muliglingual embeddings, inherits the property of being a multilingual topic model!
ZeroShotTM is going to appear at EACL2021! If you want to replicate our results, you can use our code. You will find the W1 dataset in the colab and here: https://github.com/vinid/data, if you need the W2 dataset, send us an email (it is a bit bigger than W1 and we could not upload it on github).
You can look at our medium blog post or start from one of our Colab Tutorials:
- In CTMs we have two models. CombinedTM and ZeroShotTM, which have different use cases.
- CTMs work better when the size of the bag of words has been restricted to a number of terms that does not go over 2000 elements. This is because we have a neural model that reconstructs the input bag of word, Moreover, in CombinedTM we project the contextualized embedding to the vocab space, the bigger the vocab the more parameters you get, with the training being more difficult and prone to bad fitting. This is NOT a strict limit, however, consider preprocessing your dataset. We have a preprocessing pipeline that can help you in dealing with this.
- Check the contextual model you are using, the multilingual model one used on English data might not give results that are as good as the pure English trained one.
- Preprocessing is key. If you give a contextual model like BERT preprocessed text, it might be difficult to get out a good representation. What we usually do is use the preprocessed text for the bag of word creating and use the NOT preprocessed text for BERT embeddings. Our preprocessing class can take care of this for you.
- Free software: MIT license
- Documentation: https://contextualized-topic-models.readthedocs.io.
- Super big shout-out to Stephen Carrow for creating the awesome https://github.com/estebandito22/PyTorchAVITM package from which we constructed the foundations of this package. We are happy to redistribute this software again under the MIT License.
If you find this useful you can cite the following papers :)
ZeroShotTM
@inproceedings{bianchi2021crosslingual, title={Cross-lingual Contextualized Topic Models with Zero-shot Learning}, author={Federico Bianchi and Silvia Terragni and Dirk Hovy and Debora Nozza and Elisabetta Fersini}, booktitle={EACL}, year={2021} }
CombinedTM
@article{bianchi2020pretraining, title={Pre-training is a Hot Topic: Contextualized Document Embeddings Improve Topic Coherence}, author={Federico Bianchi and Silvia Terragni and Dirk Hovy}, year={2020}, journal={arXiv preprint arXiv:2004.03974}, }
Important: If you want to use CUDA you need to install the correct version of the CUDA systems that matches your distribution, see pytorch.
Install the package using pip
pip install -U contextualized_topic_models
An important aspect to take into account is which network you want to use: the one that combines BERT and the BoW or the one that just uses BERT. It's easy to swap from one to the other:
ZeroShotTM:
ZeroShotTM(input_size=len(qt.vocab), bert_input_size=embedding_dimension, n_components=number_of_topics)
CombinedTM:
CombinedTM(input_size=len(qt.vocab), bert_input_size=embedding_dimension, n_components=number_of_topics)
But remember that you can do zero-shot cross-lingual topic modeling only with the ZeroShotTM
model. See cross-lingual-topic-modeling
The examples below use a multilingual embedding model distiluse-base-multilingual-cased
. This means that the representations you are going to use are mutlilinguals (16 languages). However you might need a broader coverage of languages. In that case, you can check SBERT to find a model you can use.
If you are doing topic modeling in English, you SHOULD use an English sentence-bert model, for example paraphrase-distilroberta-base-v1. In that case, it's really easy to update the code to support monolingual English topic modeling. If you need other models you can check SBERT for other models.
qt = TopicModelDataPreparation("bert-base-nli-mean-tokens")
In general, our package should be able to support all the models described in the sentence transformer package and in HuggingFace. You need to take a look at HuggingFace models and find which is the one for your language. For example, for Italian, you can use UmBERTo. How to use this in the model, you ask? well, just use the name of the model you want instead of the english/multilingual one:
qt = TopicModelDataPreparation("Musixmatch/umberto-commoncrawl-cased-v1")
Our ZeroShotTM can be used for zero-shot topic modeling. It can handle words that are not used during the training phase. More interestingly, this model can be used for cross-lingual topic modeling! See the paper (https://arxiv.org/pdf/2004.07737v1.pdf)
from contextualized_topic_models.models.ctm import ZeroShotTM
from contextualized_topic_models.utils.data_preparation import TopicModelDataPreparation
from contextualized_topic_models.utils.data_preparation import bert_embeddings_from_file
from contextualized_topic_models.datasets.dataset import CTMDataset
text_for_contextual = [
"hello, this is unpreprocessed text you can give to the model",
"have fun with our topic model",
]
text_for_bow = [
"hello unpreprocessed give model",
"fun topic model",
]
qt = TopicModelDataPreparation("distiluse-base-multilingual-cased")
training_dataset = qt.create_training_set(text_for_contextual, text_for_bow)
ctm = ZeroShotTM(input_size=len(qt.vocab), bert_input_size=512, n_components=50)
ctm.fit(training_dataset) # run the model
ctm.get_topics()
As you can see, the high-level API to handle the text is pretty easy to use; text_for_bert should be used to pass to the model a list of documents that are not preprocessed. Instead, to text_for_bow you should pass the preprocessed text used to build the BoW.
Advanced Notes: in this way, SBERT can use all the information in the text to generate the representations.
Once you have trained the cross-lingual topic model, you can use this simple pipeline to predict the topics for documents in a different language (as long as this language is covered by distiluse-base-multilingual-cased).
# here we have a Spanish document
testing_text_for_contextual = [
"hola, bienvenido",
]
testing_dataset = qt.create_test_set(testing_text_for_contextual)
# n_sample how many times to sample the distribution (see the doc)
ctm.get_doc_topic_distribution(testing_dataset, n_samples=20) # returns a (n_documents, n_topics) matrix with the topic distribution of each document
Advanced Notes: We do not need to pass the Spanish bag of word: the bag of words of the two languages will not be comparable! We are passing it to the model for compatibility reasons, but you cannot get the output of the model (i.e., the predicted BoW of the trained language) and compare it with the testing language one.
You can also create a word cloud of the topic!
ctm.get_wordcloud(topic_id=47, n_words=15)
Here is how you can use the CombinedTM. This is a standard topic model that also uses contextualized embeddings. The good thing about CombinedTM is that it makes your topic much more coherent (see the paper https://arxiv.org/abs/2004.03974).
from contextualized_topic_models.models.ctm import CombinedTM
from contextualized_topic_models.utils.data_preparation import TopicModelDataPreparation
from contextualized_topic_models.utils.data_preparation import bert_embeddings_from_file
from contextualized_topic_models.datasets.dataset import CTMDataset
qt = TopicModelDataPreparation("bert-base-nli-mean-tokens")
training_dataset = qt.create_training_set(list_of_unpreprocessed_documents, list_of_preprocessed_documents)
ctm = CombinedTM(input_size=len(qt.vocab), bert_input_size=768, n_components=50)
ctm.fit(training_dataset) # run the model
ctm.get_topics()
Advanced Notes: Combined TM combines the BoW with SBERT, a process that seems to increase the coherence of the predicted topics (https://arxiv.org/pdf/2004.03974.pdf).
training_dataset = qt.create_test_set(testing_text_for_contextual, testing_text_for_bow)
# n_sample how many times to sample the distribution (see the doc)
ctm.get_doc_topic_distribution(testing_dataset, n_samples=20)
Sure, here is a snippet that can help you. You need to create the embeddings (for bow and contextualized) and you also need to have the vocab and an id2token dictionary (maps integers ids to words).
qt = TopicModelDataPreparation()
training_dataset = qt.load(contextualized_embeddings, bow_embeddings, id2token)
ctm = CombinedTM(input_size=len(vocab), bert_input_size=768, n_components=50)
ctm.fit(training_dataset) # run the model
ctm.get_topics()
You can give a look at the code we use in the TopicModelDataPreparation object to get an idea on how to create everything from scratch. For example:
vectorizer = CountVectorizer() #from sklearn
train_bow_embeddings = vectorizer.fit_transform(text_for_bow)
train_contextualized_embeddings = bert_embeddings_from_list(text_for_contextual, "chosen_contextualized_model")
vocab = vectorizer.get_feature_names()
id2token = {k: v for k, v in zip(range(0, len(vocab)), vocab)}
We have also included some of the metrics normally used in the evaluation of topic models, for example you can compute the coherence of your topics using NPMI using our simple and high-level API.
from contextualized_topic_models.evaluation.measures import CoherenceNPMI
with open('preprocessed_documents.txt', "r") as fr:
texts = [doc.split() for doc in fr.read().splitlines()] # load text for NPMI
npmi = CoherenceNPMI(texts=texts, topics=ctm.get_topic_lists(10))
npmi.score()
Do you need a quick script to run the preprocessing pipeline? We got you covered! Load your documents and then use our SimplePreprocessing class. It will automatically filter infrequent words and remove documents that are empty after training. The preprocess method will return the preprocessed and the unpreprocessed documents. We generally use the unpreprocessed for BERT and the preprocessed for the Bag Of Word.
from contextualized_topic_models.utils.preprocessing import WhiteSpacePreprocessing
documents = [line.strip() for line in open("unpreprocessed_documents.txt").readlines()]
sp = WhiteSpacePreprocessing(documents)
preprocessed_documents, unpreprocessed_documents, vocab = sp.preprocess()
- Federico Bianchi <f.bianchi@unibocconi.it> Bocconi University
- Silvia Terragni <s.terragni4@campus.unimib.it> University of Milan-Bicocca
- Dirk Hovy <dirk.hovy@unibocconi.it> Bocconi University
This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template. To ease the use of the library we have also included the rbo package, all the rights reserved to the author of that package.
Remember that this is a research tool :)