BERT Model Architectures For Semantic Similarity

Рет қаралды 23,645

Күн бұрын

In this video, I discuss different kinds of model architectures that you can use for #SemanticSimilarity using #BERT or any other #Transformers based model.
Please subscribe and like the video to help me keep motivated to make awesome videos like this one. :)
To buy my book, Approaching (Almost) Any Machine Learning problem, please visit: bit.ly/buyaaml
Follow me on:
Twitter: / abhi1thakur
LinkedIn: / abhi1thakur
Kaggle: kaggle.com/abhishek

Пікірлер: 31

@username-notfound9841 3 жыл бұрын

I came here after reading the SBERT paper, and now it starts making sense. Good one.

@rbhambriiit 2 жыл бұрын

Happy to see how you have evolved since the 2017 Berlin talk on Quora duplicate question talk.

@shreyasdantkale 3 жыл бұрын

I had directly used SBERT in my project(assumed it will be very complex to fine tune 2 BERT models). Here this all made sense with your such easy understandable explanation.

@sushasuresh1687 3 жыл бұрын

How are the results ? I mean is avert better than the Bert ?

@SurjeetSingh-mz4ne 3 жыл бұрын

thank you, Abhishek for creating this video... I am working on that model and will let you know the results when it is done...

@maunishdave2573 3 жыл бұрын

I was trying bert on kaggle's new contradictory watson competition but was having little trouble with how to format the data and you uploaded this video. Thank you.

@AlexeyMatushevsky 3 жыл бұрын

You improved your results?

@maunishdave2573 3 жыл бұрын

@@AlexeyMatushevsky Actually I am trying various thing for many days. I tried both the format shown in this video. with every possible combination of learning_rate and epochs but my training_loss is always around 1.05 (crossentropyloss) it's because it is giving same output every time. I tired same architecture in TF and it worked but I don't know what is wrong with my pytorch model still trying to find out.

@shaurabhsaha2082 3 жыл бұрын

Hi @abhishek Just a small request from your side can you please make a playlist for all your videos uploaded till now it will be really helpful and informative if you align them it will be really easy to follow the right sequence to watch. Appreciate your content really grateful to see your videos ❤️

@ayoubbariki7951 3 жыл бұрын

Very informative ! Thanks u

@adityakumarmishra6931 3 жыл бұрын

Thanks for the video sir. very informative. 😁

@shishirkumar1450 3 жыл бұрын

Hey @abhishek, Do you have a tutorial on the NLP Question Answering system for any custom domain?

@raghavavinay8660 3 жыл бұрын

Thanks

@luis96xd Жыл бұрын

Excellent video, great explanation, you taught me a lot of things, thanks 😁💯

@bryancc2012 3 жыл бұрын

great video. thanks. is there any way to train a document similarity without using those dataset ? like an autoencoder? any way to create an bert based autoencoder ? thanks

@depenz 3 жыл бұрын

When adding two bert heads as bert_1 and bert_2, is that the same thing as a siamese network? :)

@oltionpreka3542 3 жыл бұрын

What if one would like to create a similarity model that compares the order of similarity among, say, 1000 questions? Is there any alternative than training the model on 1 by 1 similarity as it would grow more than exponentially with the number of questions of the dataset.

@depenz 3 жыл бұрын

Is the pooled output ”o2” the same as the mean of each word token. Or if not, how is it computed and what information gain does it give thats better than the mean of tokens / or simply using the CLS token. Thanks :)

@boscojay1381 3 жыл бұрын

I trained SBert from transformers library on allNLI data and used to build sentence embeddings for a custom dataset. I cluster those embeddings and find that the representations are of poor quality. InferSent model from face book gave better results. What could be the problem?