MaartenGr
the 6 steps of BERTopic
Have you searched existing issues? 🔎
- [X] I have searched and found no existing issues
Desribe the bug
Hello,
I have a question. According to the document, I understand that BERT-Topic consists of six steps, with the representation tuning step being optional. I have read many articles in a specific field that used BERTopic, but my question is why they don’t all include these five steps. For example, some articles only include embedding, dimension reduction, clustering, and weighting scheme (c-TF-IDEF). I’d like to know if each step can be omitted and whether using all five remaining steps is necessarily required?,
Reproduction
from bertopic import BERTopic
BERTopic Version
0.16.3
For example, some articles only include embedding, dimension reduction, clustering, and weighting scheme (c-TF-IDEF).
These are actually five steps:
- Embedding
- Dim reduction
- Clustering
- Tokenization
- c-TF-IDF
Although tokenization isn't mentioned, it is definitely used.
Typically, you would see those five steps with the optional representation step. If you would want to remove a step, the only that you could potentially remove is the dimensionality reduction step. All other are needed.
Many papers just implement the basic BERTopic functionality and compare it with that, which is a shame considering the representation models often improve the output significantly. I can't say their reasoning, but I wished the representing step would be included more often.
Hello MaartenGr,
In my case, I used all six steps with three different respresentation models (gpt 3.5, MMR and KeyBert) with the same other five steps. I evaluated the result using OCTIS npmi and topic diversity., but the result was somewhat different from what I expected. Do you think it makes sense?,
with gpt 3.5 : (npmi_score: 0.1267 , diversity : 0.9851) with MMR : (npmi_score: 0.2625 , diversity : 0.7263) with KeyBERT: (npmi_score: 0.3027 , diversity : 0.6421)
I had intended for the NPMI value for gpt to be higher.
It may be worthwhile to do a deep-dive into how topic coherence (and diversity) metrics work. They assume we have a list of keywords as the main representation for topics. This is true for MMR and KeyBERT but not for GPT-3.5 since that only generates a single label and not a mixture of words.
Hello MaartenGr, I have a question. I understand that in LDA, the input data is typically based on the Bag of Words (BoW) representation. My question is: if we change the vector representation from BoW to TF-IDF or SentenceTransformers and then feed it to the LDA model, does this approach make sense? I am looking for ways to improve the results of LDA.
@TalaN1993 LDA is quite a different method compared to BERTopic and I don't think it would work that easily with embeddings without any significant changes. I believe there is something called LDA2Vec or something similar that you could research.