Can we use precomputed embeddings when using ZeroShot modeling

[joshedOpenAi]

Is there a way we can use precomputed embeddings when using ZeroShot modeling.Can I get an example ?

[MaartenGr]

You can only use precomputed embeddings for the documents, not the labels themselves. To use the embeddings, simply pass them to .fit_transform.

[joshedOpenAi]

This is the error I get when trying to use precomputed embeddings stored in csv file. Any help will be greatly appreciated

---

2024-01-08 21:24:17,954 - BERTopic - Dimensionality - Fitting the dimensionality reduction algorithm
2024-01-08 21:25:18,437 - BERTopic - Dimensionality - Completed ✓
2024-01-08 21:25:18,441 - BERTopic - Cluster - Start clustering the reduced embeddings
2024-01-08 21:25:28,804 - BERTopic - Cluster - Completed ✓
2024-01-08 21:25:28,828 - BERTopic - Representation - Extracting topics from clusters using representation models.

---

AttributeError Traceback (most recent call last)
d:\NPSv2\zeroshottopics.ipynb Cell 17 line 3
24 topic_model = BERTopic(
25 min_topic_size=5,
26 zeroshot_topic_list=zeroshot_topics,
27 zeroshot_min_similarity=.80, verbose=True,n_gram_range=(1,2),
28 representation_model=KeyBERTInspired()
29 )
30 #topics, _ = topic_model.fit_transform(ZeroReviews)
---> 31 topics, probs = topic_model.fit_transform(ZeroReviews,ZeroMatrix)

File d:\npsv3\npsenv\lib\site-packages\bertopic_bertopic.py:433, in BERTopic.fit_transform(self, documents, embeddings, images, y)
430 self._save_representative_docs(custom_documents)
431 else:
432 # Extract topics by calculating c-TF-IDF
--> 433 self._extract_topics(documents, embeddings=embeddings, verbose=self.verbose)
435 # Reduce topics
436 if self.nr_topics:

File d:\npsv3\npsenv\lib\site-packages\bertopic_bertopic.py:3637, in BERTopic.extract_topics(self, documents, embeddings, mappings, verbose)
3635 documents_per_topic = documents.groupby(['Topic'], as_index=False).agg({'Document': ' '.join})
3636 self.c_tf_idf, words = self.c_tf_idf(documents_per_topic)
-> 3637 self.topic_representations = self._extract_words_per_topic(words, documents)
3638 self._create_topic_vectors(documents=documents, embeddings=embeddings, mappings=mappings)
...
3294 elif documents[0] is None and images is None:
3295 raise ValueError("Make sure to use an embedding model that can either embed documents"
3296 "or images depending on which you want to embed.")

AttributeError: 'NoneType' object has no attribute 'embed_documents'

[joshedOpenAi]

from bertopic.representation import KeyBERTInspired

for i, sentence in enumerate(ZeroReviews):
if not isinstance(sentence, str):
ZeroReviews[i] = str(sentence)

zeroshot_topic_list = ["Clustering", "Topic Modeling", "Large Language Models"]

umap_model = UMAP(n_neighbors=15, n_components=5, min_dist=0.0, metric='cosine', random_state=42)

aspect_model1 = [MaximalMarginalRelevance(diversity=.3)]
main_representation = KeyBERTInspired
representation_model = {
"Aspect1":main_representation,
"Aspect2": aspect_model1
}

topic_model = BERTopic(
min_topic_size=5,
zeroshot_topic_list=zeroshot_topics,
zeroshot_min_similarity=.80, verbose=True,n_gram_range=(1,2),
representation_model=KeyBERTInspired()
)
topics, probs = topic_model.fit_transform(ZeroReviews,ZeroMatrix)

[MaartenGr]

You should also pass the embedding model to BERTopic. Without it, it cannot create the word/phrase embeddings needed for MMR and KeyBERTInspired.