Reducing outliers and updating topics when using OpenAI model as representation model

[serenalotreck]

I've trained a model that generates the topic representations with gpt-3.5-turbo, using the following code:

sentence_model = SentenceTransformer('allenai/scibert_scivocab_cased')
umap_model = UMAP(n_neighbors=15, n_components=5, min_dist=0.0, metric='cosine', random_state=42)
vectorizer_model = CountVectorizer(stop_words="english", ngram_range=(1, 3), min_df=10)
client = openai.OpenAI(api_key=API_KEY)
representation_model = OpenAI(client, "gpt-3.5-turbo", chat=True)

topic_model = BERTopic(embedding_model=sentence_model, umap_model=umap_model, representation_model=representation_model, vectorizer_model=vectorizer_model)
topics, probs = topic_model.fit_transform(docs)

The output of topic_model.get_topic_info()[:10]:

However, when I try and run reduce_outliers and then update_topics, it reverts the topic representations to the c-tf-idf representations. I've tried these three things thus far:

# Passing the vectorizer with update topics
new_topics = topic_model.reduce_outliers(docs, topics)
topic_model.update_topics(docs, topics=new_topics, vectorizer_model=vectorizer_model)

# Not passing the vectorizer
new_topics = topic_model.reduce_outliers(docs, topics)
topic_model.update_topics(docs, topics=new_topics)

# Using embedding strategy
new_topics = topic_model.reduce_outliers(docs, topics, strategy='embeddings')
topic_model.update_topics(docs, topics=new_topics)

For all three options, the output looks something like this:

I also get the following warning upon running update_topics:

2024-01-31 18:21:22,790 - BERTopic - WARNING: Using a custom list of topic assignments may lead to errors if topic reduction techniques are used afterwards. Make sure that manually assigning topics is the last step in the pipeline. Note that topic embeddings will also be created through weighted c-TF-IDF embeddings instead of centroid embeddings.

I assume that the issue is that the ChatGPT topics are assigned as custom topics when I first do fit_transform.

How can I reduce outliers while maintaining the OpenAI representations?

[MaartenGr]

Check the docstrings of .update_topics, you will notice that you are missing some vital parameters in your code, like representation_model and vectorizer_model.

[serenalotreck]

Passing representation_model fixed it, thank you!