How to Build Q&A Models in Python (Transformers)

[@rashmisingh3914]

Video is helpful. Can you please tell me something about the "score", a confidence score, how they calculate it? Any kind of formula or based on some metrics?

[@prashun6957]

I want to get answers from a table in the image, how to do that?

[@flaskapp9885]

subscribed :)

[@khaoyaa]

Is it possible to implement this as a mobile app?

[@alfafa6657]

How do I enter (input) a question directly? I want to know how to connect this model to social media like telegram or whatsapp, do you know it?

[@niloozh6098]

thank you for your helpful video😍. how can we recognize that a question is not answerable?

[@xuejingfu3861]

I want to ask what will happen if the question we ask has nothing to do with the context? Then if we ask the same question multiple times how to make the program recognize it

[@jamesbriggs]

the answer score will (hopefully) be very low, and it will likely return nonsense - recognizing the same question multiple times I suppose you could keep track of past questions and perform a check for repetition

[@dato007]

is there way to pre-load a context? i.e. i don't want the model to read in the entire context (e.g. a book) everytime i wan to ask a question

[@imdadood5705]

I am enrolled to your NLP in Python. I haven’t been actively participating yet. It is really good!

[@jamesbriggs]

hey that's awesome to hear, glad you're enjoying it!

[@shanefeng3215]

Can you talk a bit about how to further train (fine-tune) the model? Do I just put the start index and end index as the labels to BertForQuestionAnswering? I followed the hugging face page's example but it doesn't work like the loss is decreasing but the accuracy is very low. I calculated the accuracy as if only the predicted start index match with the label start index AND the predicted end index match with the label end index.

[@charmz973]

M just enjoying the channel

[@jamesbriggs]

Haha that's awesome, reading all your comments now!

[@charmz973]

@@jamesbriggs we can't thank you enough for this gold mine

[@mariiii97]

Great video! Do you know how to increase the number of words in the answer? Thanks :)

[@jamesbriggs]

It depends on how the model has been pretrained and fine-tuned, you would need to fine-tune the model on your Q&A dataset (with longer answers) for it to start outputting longer answers - which you'll need the transformers train_qa.py script for - and this guide will help too huggingface.co/transformers/custom_datasets.html#qa-squad Good luck!

[@abhishekprakash9803]

@@jamesbriggs i always stuck to generate long answers.....i saw ur webpages..pinecone. and read abstractive model...but i wana ask u he only generate the long answers only for sqad datasets, how i will perform to generate the long answers for my documents .......long answers generation is so challengeing even i tired to serach ..in internet, i hope u will help me

[@meylyssa3666]

great introduction to Q&A!

[@jamesbriggs]

Happy you liked it :)

[@mamunurrahman5341]

This video is so simple to understand and so helpful to reach my idea project.

[@user-wr4yl7tx3w]

video idea, what are the challenges for doing transformer on scale.

[@siamakshams1923]

Excellent work. Even with some minor changes to HF code recommendation I found the tutorial very helpful. Thank you.

[@vitocorleone1991]

Tanq

[@shakshuka94]

What if we want long answers? For example if we need to extract all preferred qualifications from a job description context then what to do? I tried this but it gives a very short answer and doesn’t pull all the details or requirements. This is just one example. Any suggestions here?

[@jamesbriggs]

You can try either pulling just the context, eg a doc_store -> retriever pipeline or you can look into long-form question-answering (LFQA) which using a doc_store -> retriever -> generator, where the generator is trained to produce multi-sentence answers. You should be able to find info on how to do both of these online, I'll also be publishing a video on LFQA around tuesday next week

[@张尼玛-n6l]

Good job, Bro!