Embedding Models¶

BERTopic starts with transforming our input documents into numerical representations. Although there are many ways this can be achieved, we typically use sentence-transformers ("all-MiniLM-L6-v2") as it is quite capable of capturing the semantic similarity between documents.

However, there is not one perfect embedding model and you might want to be using something entirely different for your use case. Since BERTopic assumes some independence among steps, we can allow for this modularity:

This modularity allows us not only to choose any embedding model to convert our documents into numerical representations, we can use essentially any data to perform our clustering. When new state-of-the-art pre-trained embedding models are released, BERTopic will be able to use them. As a result, BERTopic grows with any new models being released. Out of the box, BERTopic supports several embedding techniques. In this section, we will go through several of them and how they can be implemented.

Sentence Transformers¶

You can select any model from sentence-transformers here and pass it through BERTopic with embedding_model:

from bertopic import BERTopic
topic_model = BERTopic(embedding_model="all-MiniLM-L6-v2")

Or select a SentenceTransformer model with your parameters:

from sentence_transformers import SentenceTransformer

sentence_model = SentenceTransformer("all-MiniLM-L6-v2")
topic_model = BERTopic(embedding_model=sentence_model)

Tip 1!

This embedding back-end was put here first for a reason, sentence-transformers works amazing out of the box! Playing around with different models can give you great results. Also, make sure to frequently visit this page as new models are often released.

Tip 2!

New embedding models are released frequently and their performance keeps getting better. To keep track of the best embedding models out there, you can visit the MTEB leaderboard. It is an excellent place for selecting the embedding that works best for you. For example, if you want the best of the best, then the top 5 models might the place to look.

Many of these models can be used with SentenceTransformers in BERTopic, like so:

from sentence_transformers import SentenceTransformer

embedding_model = SentenceTransformer("BAAI/bge-base-en-v1.5")
topic_model = BERTopic(embedding_model=embedding_model)

🤗 Hugging Face Transformers¶

To use a Hugging Face transformers model, load in a pipeline and point to any model found on their model hub (https://huggingface.co/models):

from transformers.pipelines import pipeline

embedding_model = pipeline("feature-extraction", model="distilbert-base-cased")
topic_model = BERTopic(embedding_model=embedding_model)

Tip!

These transformers also work quite well using sentence-transformers which has great optimizations tricks that make using it a bit faster.

Flair¶

Flair allows you to choose almost any embedding model that is publicly available. Flair can be used as follows:

from flair.embeddings import TransformerDocumentEmbeddings

roberta = TransformerDocumentEmbeddings('roberta-base')
topic_model = BERTopic(embedding_model=roberta)

You can select any 🤗 transformers model here.

Moreover, you can also use Flair to use word embeddings and pool them to create document embeddings. Under the hood, Flair simply averages all word embeddings in a document. Then, we can easily pass it to BERTopic to use those word embeddings as document embeddings:

from flair.embeddings import WordEmbeddings, DocumentPoolEmbeddings

glove_embedding = WordEmbeddings('crawl')
document_glove_embeddings = DocumentPoolEmbeddings([glove_embedding])

topic_model = BERTopic(embedding_model=document_glove_embeddings)

Spacy¶

Spacy is an amazing framework for processing text. There are many models available across many languages for modeling text.

To use Spacy's non-transformer models in BERTopic:

import spacy

nlp = spacy.load("en_core_web_md", exclude=['tagger', 'parser', 'ner', 
                                            'attribute_ruler', 'lemmatizer'])

topic_model = BERTopic(embedding_model=nlp)

Using spacy-transformer models:

import spacy

spacy.prefer_gpu()
nlp = spacy.load("en_core_web_trf", exclude=['tagger', 'parser', 'ner', 
                                             'attribute_ruler', 'lemmatizer'])

topic_model = BERTopic(embedding_model=nlp)

If you run into memory issues with spacy-transformer models, try:

import spacy
from thinc.api import set_gpu_allocator, require_gpu

nlp = spacy.load("en_core_web_trf", exclude=['tagger', 'parser', 'ner', 
                                             'attribute_ruler', 'lemmatizer'])
set_gpu_allocator("pytorch")
require_gpu(0)

topic_model = BERTopic(embedding_model=nlp)

Universal Sentence Encoder (USE)¶

The Universal Sentence Encoder encodes text into high-dimensional vectors that are used here for embedding the documents. The model is trained and optimized for greater-than-word length text, such as sentences, phrases, or short paragraphs.

Using USE in BERTopic is rather straightforward:

import tensorflow_hub
embedding_model = tensorflow_hub.load("https://tfhub.dev/google/universal-sentence-encoder/4")
topic_model = BERTopic(embedding_model=embedding_model)

Gensim¶

BERTopic supports the gensim.downloader module, which allows it to download any word embedding model supported by Gensim. Typically, these are Glove, Word2Vec, or FastText embeddings:

import gensim.downloader as api
ft = api.load('fasttext-wiki-news-subwords-300')
topic_model = BERTopic(embedding_model=ft)

Tip!

Gensim is primarily used for Word Embedding models. This works typically best for short documents since the word embeddings are pooled.

Scikit-Learn Embeddings¶

Scikit-Learn is a framework for more than just machine learning. It offers many preprocessing tools, some of which can be used to create representations for text. Many of these tools are relatively lightweight and do not require a GPU. While the representations may be less expressive than many BERT models, the fact that it runs much faster can make it a relevant candidate to consider.

If you have a scikit-learn compatible pipeline that you'd like to use to embed text then you can also pass this to BERTopic.

from sklearn.pipeline import make_pipeline
from sklearn.decomposition import TruncatedSVD
from sklearn.feature_extraction.text import TfidfVectorizer

pipe = make_pipeline(
    TfidfVectorizer(),
    TruncatedSVD(100)
)

topic_model = BERTopic(embedding_model=pipe)

Warning

One caveat to be aware of is that scikit-learns base Pipeline class does not support the .partial_fit()-API. If you have a pipeline that theoretically should be able to support online learning then you might want to explore the scikit-partial project. Moreover, since this backend does not generate representations on a word level, it does not support the bertopic.representation models.

OpenAI¶

To use OpenAI's external API, we need to define our key and explicitly call bertopic.backend.OpenAIBackend to be used in our topic model:

import openai
from bertopic.backend import OpenAIBackend

client = openai.OpenAI(api_key="sk-...")
embedding_model = OpenAIBackend(client, "text-embedding-ada-002")

topic_model = BERTopic(embedding_model=embedding_model)

Cohere¶

To use Cohere's external API, we need to define our key and explicitly call bertopic.backend.CohereBackend to be used in our topic model:

import cohere
from bertopic.backend import CohereBackend

client = cohere.Client("MY_API_KEY")
embedding_model = CohereBackend(client)

topic_model = BERTopic(embedding_model=embedding_model)

Multimodal¶

To create embeddings for both text and images in the same vector space, we can use the MultiModalBackend. This model uses a clip-vit based model that is capable of embedding text, images, or both:

from bertopic.backend import MultiModalBackend
model = MultiModalBackend('clip-ViT-B-32', batch_size=32)

# Embed documents only
doc_embeddings = model.embed_documents(docs)

# Embedding images only
image_embeddings = model.embed_images(images)

# Embed both images and documents, then average them
doc_image_embeddings = model.embed(docs, images)

Custom Backend¶

If your backend or model cannot be found in the ones currently available, you can use the bertopic.backend.BaseEmbedder class to create your backend. Below, you will find an example of creating a SentenceTransformer backend for BERTopic:

from bertopic.backend import BaseEmbedder
from sentence_transformers import SentenceTransformer

class CustomEmbedder(BaseEmbedder):
    def __init__(self, embedding_model):
        super().__init__()
        self.embedding_model = embedding_model

    def embed(self, documents, verbose=False):
        embeddings = self.embedding_model.encode(documents, show_progress_bar=verbose)
        return embeddings 

# Create custom backend
embedding_model = SentenceTransformer("all-MiniLM-L6-v2")
custom_embedder = CustomEmbedder(embedding_model=embedding_model)

# Pass custom backend to bertopic
topic_model = BERTopic(embedding_model=custom_embedder)

Custom Embeddings¶

The base models in BERTopic are BERT-based models that work well with document similarity tasks. Your documents, however, might be too specific for a general pre-trained model to be used. Fortunately, you can use the embedding model in BERTopic to create document features.

You only need to prepare the document embeddings yourself and pass them through fit_transform of BERTopic:

from sklearn.datasets import fetch_20newsgroups
from sentence_transformers import SentenceTransformer

# Prepare embeddings
docs = fetch_20newsgroups(subset='all',  remove=('headers', 'footers', 'quotes'))['data']
sentence_model = SentenceTransformer("all-MiniLM-L6-v2")
embeddings = sentence_model.encode(docs, show_progress_bar=False)

# Train our topic model using our pre-trained sentence-transformers embeddings
topic_model = BERTopic()
topics, probs = topic_model.fit_transform(docs, embeddings)

As you can see above, we used a SentenceTransformer model to create the embedding. You could also have used 🤗 transformers, Doc2Vec, or any other embedding method.

TF-IDF¶

As mentioned above, any embedding technique can be used. However, when running UMAP, the typical distance metric is cosine which does not work quite well for a TF-IDF matrix. Instead, BERTopic will recognize that a sparse matrix is passed and use hellinger instead which works quite well for the similarity between probability distributions.

We simply create a TF-IDF matrix and use them as embeddings in our fit_transform method:

from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import TfidfVectorizer

# Create TF-IDF sparse matrix
docs = fetch_20newsgroups(subset='all',  remove=('headers', 'footers', 'quotes'))['data']
vectorizer = TfidfVectorizer(min_df=5)
embeddings = vectorizer.fit_transform(docs)

# Train our topic model using TF-IDF vectors
topic_model = BERTopic(stop_words="english")
topics, probs = topic_model.fit_transform(docs, embeddings)

Here, you will probably notice that creating the embeddings is quite fast whereas fit_transform is quite slow. This is to be expected as reducing the dimensionality of a large sparse matrix takes some time. The inverse of using transformer embeddings is true: creating the embeddings is slow whereas fit_transform is quite fast.