Not Created Well Topic From BERTopic for Nepali Dataset

[hari-chalise]

This is the plot of derived topic and these are the code

from bertopic import BERTopic
from sentence_transformers import SentenceTransformer
embeddinge_model = SentenceTransformer('sentence-transformers/paraphrase-MiniLM-L3-v2')
model = BERTopic(verbose=True, embedding_model= embeddinge_model, min_topic_size= 7)
headline_topics, _ = model.fit_transform(data.story)

and Topic identified are
![image](https://github.com/user-attachments/assets/973c9dc2-692a-4f35-a2d6-a62579860ef0)

 

Hello Everyone and  @MaartenGr , there required Preprocessing required before embedding or not i am using Nepali Text Dataset. While i tried with English dataset its work well but not in Nepali Dataset.

[MaartenGr]

Generally, there is no need to preprocess the data. Whenever you use a different language than English, make sure you:

• Use an embedding model that is suited for that language (you are using an English one, so that's not suited)
• [OPTIONAL] - Use a different tokenizer in the CountVectorizer (e.g., for mandarin)

Also, make sure to read through the official docs. There is a lot of examples there that might help you out.

[hari-chalise]

Oky, Thank you @MaartenGr for you Response. I have working on Nepali Dataset and Using sentence-transformers/paraphrase-MiniLM-L3-v2 as embedding and try to use Tokenizer that support Nepali. But i am struggling to discover relevant topic for my findings.

[hari-chalise]

This is code

[MaartenGr]

As mentioned, you will also need to have an embedding model that supports Nepali. I believe the model you are using is an English-only model. Other than that, I would highly advise going through the best practices.

[hari-chalise]

Thank you @MaartenGr For your quick response. I think paraphrase-multilingual-MiniLM-L12-v2 is Embedding Model for Multilingual,According to Official Document. I have used the same model for Nepali Text.

[MaartenGr]

Also note that it is typically not advised to use nr_topics as that will create very abstract topics if set too low. Make sure you first use hdbscan_model with min_cluster_size to control the number of topics.

Moreover, min_topic_size will not do anything since you already set hdbscan_model, so you can remove that.

[hari-chalise]

Thank you, @MaartenGr . Your advice has been really helpful for continuing my research. I will do my best to follow your suggestions.

Additionally, I have another question. I have proposed research focusing solely on BERTopic, titled "Topic Modeling for Nepali Text Using BERTopic". My plan is to evaluate the model using metrics such as Coherence Score, Topic Diversity, and Silhouette Score.

Is this a correct approach for evaluating the model, or are there other evaluation methods I should consider? I would greatly appreciate your guidance.

[MaartenGr]

I would advise reading through this thread. There, you will find my thoughts on evaluation. In general, I suggest using evaluation metrics to suit a specific use case rather than following the traditional coherence scores just because others are using those.

[zorasdangol]

@hari-chalise @MaartenGr I am able to figure out this issue and generate Nepali topics.
I am working on this as my Graduation Thesis. We can communicate with me through email.

[hari-chalise]

Oh Great, in which model did you work? From Which university?

…

On Sun, Jun 1, 2025 at 10:18 PM Saroj Dangol ***@***.***> wrote: @hari-chalise <https://github.com/hari-chalise> @MaartenGr <https://github.com/MaartenGr> I am able to figure out this issue and generate Nepali topics. I am working on this as my Graduation Thesis. image.png (view on web) <https://github.com/user-attachments/assets/830a0c10-01fb-44f5-ae1f-180e68eb5a4a> — Reply to this email directly, view it on GitHub <#2265 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AM7ILK7FOW4ZZT76AXRX24T3BMTL7AVCNFSM6AAAAABVBHMT5SVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGMZTGU2TMNA> . You are receiving this because you were mentioned.Message ID: ***@***.***>