Your videos are of high quality and cover quite a range of topics. But i wonder why the subscribers are so few relatively. My personal take is that you lay a very good foundation- easy to understand, then dive right into coding which is very practical. i feel there's something missing in-between.
hi thanks for such an informative video , what about the scenerio if we extract numeric features from our datasets like sentiments etc then how can we input them for transformer specially T5, Albert without doing masking
Thanks for making this video. Learnt a lot. Follow up question: Can the question and answering more of a chat format where you can build questions and follow ups? Let’s say I am embedding the text, create vector of the text. When a question is asked, it’s converted to vector and then using cosine similarity, fetch the response. Can it be done with any of the models this way? Could you please make a video or share feedback if possible? Thanks.
What do you mean build questions? Maybe like form a digital identity based on the questions you ask? You can definitely do that. In fact, the new buzz: Chat GPT 3 can do just that.
@@SpencerPaoHere Apologies for lack of clarity from my side. Yeah, when I meant follow up questions: I meant very similar to how a chatbot work taking all the context and answer like a conversation rather than asking a question , it returns a response, starting over again. Trying to understand how to make it conversational - open source way (cause openAI one’s are costs :(
@@sriramkrishna6853 I see! I'd recommend this site for more info: www.geeksforgeeks.org/chatbots-using-python-and-rasa/ I think what you are asking is Conversational AI. (Natural Language Understanding) This is an entire sub-industry. But there are many resources out there! I definitley recommend diving deeper in chat gpt because that definitely answers your question.
So we give question as input from prompt then our model picks up a random context from our dataset and gives random answer...(if we didn't fine tune the model)
What is the for a custom dataset, the question for a context has answers coming from multiple section of the paragraph? I believe for the dataset here you only have one answer per question from a context but how to handle multiple start index for a question?
I unfortunately do not; however, you can think of MCQ as just multiple entries - dictionary where key: (list record of strings)- then you provide the answers in the QA model tuning. I think that’s what you mean ? Otherwise there are a few articles that are present - happy to share
HI ! Thanks about the data format I read the link and it mainly explain that the data have to be in form of json or list or dictionnaries does it mean that if I have a pandas dataframe with column question, answer, answer_start and answer_end it won't work ?
If you are using the pandas library, there should be a read_json(). So, you should be fine! And from what you just described, if your features are structured then you should be okay.
It's on my github: github.com/SpencerPao/Natural-Language-Processing/blob/main/Question%20Answering%20Modeling/Question_Answering_Modeling_colab.ipynb Try clicking on the "Open in Colab" button.
What's the use of the model in question answering system, if the dataset contains answer column already? Simple search will also work for SQUAD then there's no need to finetune a model for that. Correct me if I'm wrong about squad dataset
At a high level, the QA model is basically a search function, attempting to find the relationships between the given question and a given answer. Now, in practice, you are going to have many "questions". And, a QA model uses the weightage from its training sets to see which is a good answer for your given question. The beauty is that you do not need to already have a predefined answer. The QA model learns from previous Question-answer pairs and you can ask new questions (not previously defined) and perhaps get a good answer.
@@loading757 have you consiedered upgrading to get more storage? An alternative would also be to access an external storage (another cloud or personal computer) and do the computations via chunking.