Text Classification | Sentiment Analysis with BERT using huggingface, PyTorch and Python Tutorial

hey guys I hope that you're doing great and in this video we are going to do some sentiment analysis using birth finally so we have the Google place reviews we have the text be processed and now we are going to use the hugging face library and the birth model from there to do the sentiment analysis itself but first we are going to have a look at what birth actually is so let's get started so birth was introduced in 2018 in this paper which is available on archive and the paper is called birth returning of deep bi-directional transformers for language understanding and this is a mouthful and if you don't know what any of this means that is quite all right but if we open the paper itself you see that there is a comparison between the open iya AI GPT which is the first version of the very popular model GPT 2 and we have Elmo here so compared to the elmo you can see that the bird is using this thing called transformer which is an architecture which is introduced in another paper called attention is all you need also by the Google team of course and compared to the ELMO we are not using our streams and Elmo has introduced contextual embeddings so what this means is that the word in a sentence might have different meaning depending on the words that surrounded so for example the word nails might have two different meanings for example finger nails or metal nails so the next thing that is such a big idea in the NLP community is the transformer act architecture itself and this one is used in the original version of the DBT model and the transformer is very different compared to the HTM the main thing about it is that it feeds the input in at once compared to the sequential reading that the htm' models required and it turns out that in the real world at least the transformer is much better at handling long term that dependencies between words so you might have the task of question answering or reading comprehension in general and when you do that you have to basically remember some of the context of the stuff that you've read before or even after sometimes depending on the text that you are reading and in those cases the transformer models are actually doing better job compared to the other streams when looking at information vastly in the previous words or the next words so the next thing that actually bert introduced was the by directionality so if we have a look over here as well the GPT model is doing left to right understanding of the text and sometimes when you're answering questions it might be the case that the answer is actually after the question was posed into the text so the bird model is actually bi-directional actually it's non directional but that's a bit well it's a bit more advanced topic and we will not go into that so the big thing about the bird it's bi-directional or non directional model it uses transformers and it does have contextual word embeddings so different so the same word might have a different meaning depending on the words surrounding it and the most most important thing about the bird model is that it can be actually very easily fine-tuned on the tasks that you might be interested in so you might do tech specification with it you might do question answering you might do sentiment analysis as we are doing in our case entity recognition etc etc but how was the bird model actually trained well the thing about the board model is that it had to in task one was the masked language modeling and the other one was the next sentence predictions but what this means if we go into the roughly the beginning of the paper you can see that we have let me just scroll up we have the masked LM task in the next sentence prediction so roughly speaking about 15 percent of the text that we had in the corpus and the corpus was comprised of the Toronto Book corpus and the Wikipedia English corpus so had a large large data set and in that data set some of the words at least fifteen percent of the words was were masked and the task of the bird model was to uncover or guess the words that were masked the next task was called next sentence prediction and given a pair of sentences the task was to say whether the second sentence should or follows the first one so I know that this might be a bit confusing so let's have a look at some examples and we are going to start with the mask language modeling let me open up a text editor for you and in here we might have the input which is masked that's mask she masks so given this input the task of the bird model was to uncover or guess the words that were behind the mask so in this case this should be that's what she said alright so this is the masked language modeling task alright for the next task we have the next sentence prediction so this is basically yeah the next sentence prediction we might have the following input this should include the classification special token that we found that we talked about in the previous video so we are asking birth to do some classification and we have basically the same thing next we have a separator which is again a special token that we've talked into the previous video and then we might have something like ahaha nice and we have another separator here so the label for this one should be yes this is next sentence okay in the other case we might have something like the same input but this time we might have something like Dwight you ignorant and we might have a mask right here of course the label for this should be not next and basically those are the two tasks that were used for training the bird model of course we are not going to use the bird model in the original paper for which the source code was provided by the Google team they also provided some pre-trained models which is just amazing but we are going to use of course the holding face library and the face library has the sorry the hugging face library code to the Transformers and this library has a whole section on both models of course you can do that with pay torch or 10:02 and we are going to use the Python version in this one so there is a very good summary in here you can go ahead and read when you start working with birth using the huggy face library you might find out that there are a couple of different different flavors of birth of course you might be confused at first but let's get right into understanding what those types might be so the first thing about birth was that I am showing you the industry birth by Jay our mark here which is a great article and there will be a reference into the source code in the tutorial of this video to this article if you have if you wanna have a look at it and here we have the bird watch and the bird base another thing that you might notice in here is that the bird model is basically a stack of transformer encoders and the number of those transformer encoders is labeled as bird base for which we have twelve encoders and for the bird watch we have 24 encoders of course the watch model is more accurate it has more parameters but it can be a bit slow to Train even on P 100 password or whatever you might want to train with the or the word one you might want to fine tune the word one when you have multiple GPUs at least that's my experience of it of course we have a bit more attention layers in the attention thing refers to the idea that when for example you are saying something let me open up the text editor right here of course so you might have something like jim is pretty cool his ideas are amazing or his let's say pranks are amazing okay so in this case the attention mechanism of course the attention mechanism and the sail rotation mechanism is very let's say complex think into the NLP community and you might wanna give a deeper dive into the transformer architecture in the attention mechanisms from the papers or on other tutorials but in our case I'm going to give you a simple intuition right here so let's say that we have this example text or these two sentences and you might have a look at this thing this word his so the attention mechanism is basically going to show you that there is a high relationship between his and Jim so his is referring to Jim at least in this case so this is what the attention mechanism does for you okay so we have the attention mechanism we have the bird base and the bird watch models and we have the hugging face library so the next thing that I'm going to do is to what the bird model itself from the hugging face library I'm going to open up the notebook from the previous video and I'm going to what the bird model from the human face library and this is the notebook with a bit of changes actually so in our case I'm using the version of bird which is the base model so with the 12 sorry with the 12 transformer in colors here and I'm using the cased version so in the last video I've shown you how to work the encased version but in our case at least what I found by experimenting or empirically is that the casing of the words matter when we are doing some sentiment analysis for reviews of course this might be the case because intuitively when you write something like that in all uppercase might convey more sentiment or that something is really bad compared to when you use the bat into work is letters so I'm basically pre-loading the pre-training model with the bird base cased version so if we go ahead and look at the tokens of our tokenizer you can see that the casing was preserved so this is a bit different from what I've shown you into the previous video another thing that I've changed in here is the fact that now our GP reviewed data set is also returning the review text and this will be helpful when we are trying to evaluate the model itself the last thing that I've changed was to actually increase the seek the maximum sequence length of our reviews and I'm using 160 tokens right now and I'm going to train with a part size of 16 and I'm going to Train 450 box okay so this is pretty much the new picture of course in the worst video I've also had extra dimension right here which was one so I've removed that as well so now we have the batch size times the sequence length and all this is now into the correct shape so given that I'm going to use the hugging face library and initialize a bird model from that alright so I'm going to create the model and I'm going to call bird model from pre-trained and I'm going to pass in the pre trained model name and this will go ahead and download the model weights from the report in this may take a bit so during this time I'm going to basically call the resulting model and I'm going to call it using the input IDs from the example that we had of the encoding and I'm going to pass in the attention mask as well all right so this is great and this will return two things the last hidden state from the topmost in color which is this one in our case and it's going to return a put output so we are going to have a look at both of these and note that we are still running on the CPU for this one and now that we have the outputs of the Bert model we can have a look at those and the last hidden state is the sequence of hidden states of the worst layer of the model so it's basically a sequence from the worst encoder right here from the break base model and we can have a look at the shape of it and we have one example seven six eight hidden units we are going to have a look at this so this is okay but the put output is the one that we are going to use and it's basically a pulling procedure over the 32 elements into this dimension and this is coming from the birth pore which is defined into the github repo of the whole unified library so this is taking the hidden size of the seven six eight right here and it's basically applying a tonnage function over it so this is the pooling that the hanging face library is doing for you so you're basically getting a summary of this hidden state and you might go ahead and try something better but in our case did this will suffice all right so where you can see this information about the hidden size we have the birth model it has a config and you can have a look at the hidden size right here hidden size alright so yeah this is the number of hidden units right here okay so with this in mind we are going to build a sentiment classifier building sentiment classifier using birth so we are going to create a new class and this class is going to extend from the torch and in module and we are going to accept the number of classes that we have we are going to feed the data to our birth model we are going to apply some dropout for regularization and then we are going to append an output layer which will be of course fully connected layer and call a soft max function of the output of that one so this will basically give us the classification that we are going to need in order to classify a review as neutral positive or negative one so let's start with the cross for the sentiment classifier and we are going to have a constructor which is going to accept the number of classes that we have and I'm going to call the super constructor and I'm going to pass in the self right here next I'm going to initialize the birth model the same way that we did into the demo above next I'm going to create a drop out layer and I'm going to specify the probability of drop out as 0.3 and I'm going to specify the out put layer this will be the number of hidden size units that we have into our birth model and I'm going to use the config that I've shown you and this way you can basically do some pretty generic stuff and implement the birth model using the watch or the base model and I'm going to output the number of classes right here finally I'm going to apply a soft max function so I'm going to create instance of that and we are going to do I want the first dimension of course because we want to classify each example next I'm going to define a forward function which will implement the base forward function and here we have the self as the first parameter we have the input IDs of our tokens in the attention mask so right here I'm going to skip the worst hidden States and get just the boot output from our birth model I'm going to pass in the input IDs in the attention mask next I'm going to get the output from the dropout and I'm going to apply the output layer crossfire and finally I'm going to return the result of the softmax function so this should pretty much give us a classifier based on the birth model right here and I'm going to create an instance of that so I'm going to take the length of the class names that we have and I'm going to move the model to the GPU device which in at least in this notebook I'm going to use let me see test what t4 so this is the device that I'm using to this quad notebook at least and next I'm going to have a look at the data or a sample data from our data Waller and this data is going to contain the input IDs I'm going to take that from the data or sample data I'm going to put that onto the GPU and I'm going to basically do the same thing for the attention mask let's have a look at the shapes of these and this should correspond to the bud size which is 16 and the number of tokens that we have in each example which is 160 all right so this is work this works at least as I expected and let's basically run our sentiment classifier model through this of course our model is still untrained so we are not going to have a deeper look at the accuracy of this model and I'm going to pass in the attention mask so this should be the output and we have 16 examples right here and for each of the three classes that we have we have the probabilities of how much our model things about each class and you can see that this tensor is basically on the GPU device and we have we can apply backward propagation on the result of this model next we are going to train our model and hopefully it's going to be good at predicting sentiment so let's start with the training process and this one is going to be a bit different compared to the standard stuff that you might already know how to do using pi torch so I'm going to go ahead and create a training section right here and the first interesting bit about the training at least using the holding face library and the birth model is that we are going to use a bit of a different optimizer and this one is called Adam W so this one is basically implementation of the Adam algorithm with weight K fix and this one is just the same thing that the original birth paper did and I'm going to pass him the model parameters I'm going to pass in a warning grade which is going to be this number right here because basically the birth paper is using or recommending some warning rates which are doing good on fine-tuning the birth model and I'm going to pass him that I don't want to correct the bias of the algorithm because this is what the original paper also did next we are going to use a linear schedule which is going to decay the warning rate and for that I'm going to need to calculate the steps that we have and this will be the number of batches that we have in to the training data water times the number of epochs which is 50 in our case so using that I'm going to create schedule get linear schedule with warm-up which is provided by the hanging face library I'm going to pass in an optimizer and I don't want to have any warm-up steps just as the bird paper recommends and I'm going to pass in the total steps that we have finally I'm going to define hours function and this is going to be cross entropy was because we are doing classification cross-entropy was I guess the out complete is not very good today and I'm going to put that onto the GPU okay so now that we have all of this already we are going to define a helper function which is going to be called train epoch and as the name of the function suggests this will go ahead and go over all the training data at least for one ebook and use that to apply backpropagation to our sentiment classifier so for this one I'm going to let me just move this away from our screen I'm going to go ahead and define the function and this one will take a lot of parameters I know that's a bit of a bad design but yeah let's go with it we'll accept the model the data water the worst function the optimizer the device on which we are going to do the training the schedule and the number of examples because we will calculate the was the training goals that is and the training accuracy okay so the first thing that I'm going to do here is to put the model into training mode so that for example our dropout is actually applied during this training process I'm going to store the losses and the correct predictions okay so now for the meat of this function and this is going to be a bit of a length one I'm going to iterate over each example into our data water and I'm going to take the input IDs the attention masks in the targets or the labels water which we are predicting the sentiment I'm going to take that from the input IDs and I'm going to move that over to the GPU I'm going to do the same thing for the attention mask and for the targets of the sentiment I'm going to take the outputs of the model and attention masks alright now that we have this I'm going to take the predictions which are going to be basically the classes that have the highest probability and I'm going to do that using torch max I want the first dimension and I'm going to apply a loss function over the model outputs and I'm going to use the targets to actually calculate the cross entropy rows and the number of correct predictions right here will be this one using torch dot some I'm going to get all the predictions that are basically the same as the targets so this will tell us the number of correct predictions and I'm going to append the current Wars next I'm going to do the back propagation steps but we have something a bit different right here I'm going to apply the backward propagation then I'm going to basically do some gradient clipping on our model and we are going to use the normalization of 1.0 so this is basically duplicating the steps that original birth paper has described during the final recommendations for the training core fine-tuning the birth model and to do that I'm going to use a name dot u2's doc clip dropped norm and I'm going to do this in place so I'm going to use the underscore version of it I'm going to pass in the model parameters and maximum equal to 1.0 so if you're not familiar with the gradient clipping Huck if you will this is basically a way to combat the exploding gradients so if the gradients are to watch for example your model will or your training procedure will go unstable and you will basically have no not a good model so the gradient clipping is basically doing a hack to not allow the gradients to become so much the next thing I'm going to call here is to do a step of the optimization I'm going to do the same for the schedule not that this schedule is called every time we are going through an example but in our data water and finally I'm going to zero the gradients right here so the last thing in this function that we are going to do is to return the number of correct predictions and the was for this training key book and I'm going to take the correct predictions the double tensor from that and I'm going to divide that by the number of examples and I'm going to take the losses and calculate the mean was from that one so this is pretty much the training training epoch helper method or function okay so we have that and I'm going to write another one for evaluating the current model this one will take a model data water was function device and the number of examples as well so during a vow mode to basically evaluate our model we are going to put it into a vowel mode so now that the stuff like button normal is and drop out are not enabled I'm going to also store the losses and the correct predictions right here and the first thing that I'm going to do before looping or going over the examples from the data folder are going to be to disable the gradient function so torches a bit faster and next I'm going to iterate over our examples into the data water and I'm going to basically do the same thing that we did over here so I'm going to go ahead and copy and paste that at me alright so we have pretty much the input IDs the attention must targets we have the outputs of the model we have the predictions the Wars and we are storing the correct predictions and the losses so we're not doing any training right here so nothing else pretty much we are done and I'm going to copy and paste the return statement which is going to be the same thing right here okay so the vowel model function is ready and mixed I'm going to execute this cell and I'm going to basically create a training group with the help of both of those functions of course I want something to store the training and validation losses and accuracies so we define our default dictionary which contains lists and we'll call that history and this is very similar to the way that basically the Charis library is storing the history and I want to store the best accuracy so I'll be able to basically discard or save the models which are the best at doing the valley basically has the best accuracy on the validation set and I'm going to iterate over the number of epochs I'm going to start with printing the current epoch out of all the a box and I'm going to print some characters so this is a bit nicer to look at and now we are going to basically call ball to call the train a balk helper function in this will determine the training accuracy and the training course for this epoch and I'm going to pass in the parameters which are model the train data water the worst function the optimizer the device schedule and length of the number of training examples which is stored into the training data frame and now that we have this I'm going to print out train was which will be the train was in the accuracy it's going to be the Train accuracy all right I'm going to do pretty much the same thing for the evaluation modes sorry on evaluating the model model validation data water the worst function the device and the number of examples into the validation dataset so I'm going to paste to get this paste it down here and I'm going to form up this so we have pretty much the same visual representation of this and here I'm going to put in the validation what's in the validation occurs and finally I'm going to print a new wine right here okay so now that we have the training can validation accuracies and what's this we are going to basically store these into the history that we've created I'm going to do that for the worse and next I'm going to do the same thing for the validation accuracy in the relational alright so the final thing is that we are going to save the model with the best accuracy so I'm going to check the validation accuracy and if this is better than the best accuracy for which were starting with zero so we are storing the first-ever model that we have I'm going to save the model to model dot but and I'm going to save the best accuracy into the validation occurs I'm going to set the best accuracy to the validation occurs so the worst thing that I'm going to do is to basically time this using the magic command percent percent time so I'm going to run this and Brett's is not defined so I have an error okay so I have this arrow into the this name so this is pets and this is bred for predictions so if I execute this again and wait for the training to start okay I guess that this should work out and we will basically go ahead and wait for the model to try so we train the model for a little bit over three and a half hours and I'm going to show you the walk of the training process and it's a different notebook that I'm preparing for the tutorial to this video and here is the result you can see I've trained for 50 a box and at the final epoch we have a validation accuracy of 84% and training accuracy of roughly 92% and I've actually did better in the previous epochs you can see that we have validation accuracy of 84.6% but I'm currently just saving the worst model because I've been saving the model using the model itself and I found that you have to watch this thing with the state dict so I'll save it with the static and I'm going to what the static otherwise the ringing face Larry is not loading the model properly I guess so a little note on how I found the parameters that are work for our case I did lots of experiments with the base in the watch model I've read through the recommendations of the bird paper model and basically I had to run roughly let's say 15 experiments and I did that for a different amount of epochs after started with five epochs and after that after I've filtered most of the parameters and just let the model train for 50 box I guess that if you continue training this model you might get even better results but I'm going to basically give you the Google Drive check point to the model that I found that works best so let's continue with this and this is our original notebook and now that I've changed that I am saving the state instant of the model right here so I'm going to go ahead and download the pre trained model and I'm going to take the ID to my Google Drive so this should be available for you as well and this should go ahead and save this model based case state and I'm saying that the validation accuracy for this one is eighty four point one percent and I'm going to create a new instance of our model I'm going to specify the length of the class names and I'm going to what the state dict using torch dot Watt and I'm going to pass in the model base case state dict 88 for one and I'm going to move the model to the GPO and while this is running I'm going to open up the file browser and you can see that this model is 413 megabytes which is quite a large model but keep in mind that this is just the base birth model included right here so if you are using the word run it might be a lot bigger okay so we'll have to do some evaluation and I'm going to write the function called get predictions so this function will basically written the text reviews the predictions of our models for model the prediction probabilities for each prediction and the real values or the real sentiments for each review and this is going to accept a model in a date water right here I'm going to start with going to evaluation mode and I'm going to create four lists which will store the data that we need so you have a list for the review texts the predictions the prediction probabilities and the real values or the sentiments that are assigned to each review so I'm going to go ahead let me just save us some time and go ahead and copy most of the above model stuff that we have right here I'm going to fix the indentation and we are not interested into the walls right here at least and I'm going to take the review text from the water and I'm going to viewing the lists that we've created I'm going to extend that those lists because the texts and input ideas etc all of it are just lists I'm going to do the same thing for the predictions and the prediction probabilities and I'm going to just take the route outputs because those are the outputs of the softmax function from our model and I'm going to save the targets okay so now that we have all the data in our lists I'm going to convert those lists into tensors and I'm going to return their values so the predictions will be torch stuck so this will convert the list of pencils to tensors to a single tensor and we'll basically add a new dimension to our tensor I'm going to move that to the CPU I'm going to do the same thing for the prediction props or probabilities and I'm going to do exactly the same thing for the targets or the sentiments finally I'm going to return the review texts and we didn't do any pre-processing on those because they are just a list of texts or strings and now to have this rather watch helper function I'm going to call it and I'm going to pass in the model in the test data water so this should take a while and while this is running we might be let me just show you how you can do evaluation on the test data water actually so let's get the test accuracy and the test was I believe and I'm going to call a vowel model and I'm going to pass in the model the test data water the reverse function the device and the number of examples is going to be e F dot test I believe and let me just go ahead and check the eval model so it's the accuracy and the Wars okay so if I run this I should get the test accuracy for a model and you should be this number right here okay so we have 82 percent 82.2% accuracy on the test set which is quite good because the reviews themself are quite hard to sometimes classify even as a human I mean like neutral a reviews can be really challenging and we will see that ever are pretty much state of the art sentiment crossfire model is having hard times with some of the neutral reviews at least and let me start our evaluation or continue our evaluation using the classification report from the psychic warned library I'm going to pass him the test values the predicted values and the target names which will be the class names that we have and for each class we have a precision recall f1 score so right here you can see that the positive and the negative reviews have relatively high precision and recall and the neutral reviews at least in our case the ball is having a hard time with those which is expected because yeah those even as a human are hard I can tell you that I've read through many many many reviews on Google Play or the App Store on Apple and guessing reviews with three stars it's really hot stuff okay so let's see let's have a look at the confusion matrix for our model and I'm going to go ahead and copy and paste a helper function that I've shown multiple times right now so this show confusion matrix function accepts a confusion matrix in the format of data frame so I'm going to prepare the confusion matrix by calling that confusion matrix from socket worm then I'm going to convert that into updated frame and the index is going to contain the class names and the columns are going to be again the class names and finally I'm going to call show confusion matrix so you can see that once again our model is having relatively hard time with the classifying the neutral reviews otherwise it's going to drink pretty okay job on the positive and negative reviews which is once again expected so if you continue training this you might get even better results but I wouldn't expect that anytime soon a model is going to be perfect at this task you can have a look at the training accuracy walk and you can see that even on the training data our model is not able to achieve close to 100% accuracy next I'm going to basically show you what unmod our model thinks about a single review and I'm going to choose an index for that review and we'll start with index of two I'm going to take the review text for that so this is where the review tags are really useful I'm going to take the truth sentiment and I'm going to create a prediction data frame which is going to tell us the probabilities for each class and the values for those probabilities at least the predicted ones alright so next I'm going to print out the text review and I'm going to use the app function provided by the text wrap module in Python 3 and if I bring this you are going to see this is the review text I'm going to print another wine and I'm going to print out the true sentiment for this review and you can spend some time and read through this and guess the sentiment of that one I guess you might say that it's pretty hard thing to do so if I run this again you can see that true sentiment for this is neutral so I used to use habitable and I must say this is a great step up I'd like to see more social features such as sharing tasks only one person has to perform setbacks for it to be checked off otherwise etc etc I could easily justify roughly $1 per month or eternal subscription if that place could be met as well as fine tuning this would be easily worth 5 stars so this is pretty much you might say that this user is kind of a bully because he wants some features and after some price reductions or something like that then he or she might eventually give 5 stars to this up but if your developer don't succumb to this kind of social bullying or user bullying so next I'm going to show you what our model things about this review and does it actually predict correctly that this is this is really a neutral sentiment so I'm going to put the predictions or the probabilities for each class name I'm going to pass in the prediction data frame and I'm going to orient the chart horizontally so it's a bit easier to understand what's happening I'm going to label the y-axis a sentiment and the x-axis as probability and I'm going to do a limit or normalization on the x-axis so it's always 0 / 1 in in our case our model thinks that this indeed is a neutral sentiment let's go ahead and try another example so let's see example 42 why not and this is once again neutral sentiment let's try something else 33 we have some negative sentiment what the heck I used to wonder this happily for seven years complete BS they make you sign in on account to use this if no deleting this up immediately so this is pretty much pretty negative stuff right here and if we rerun the prediction of our model it is highly certain that this is really a negative sentiment so our model is quite good actually at least from what I've expected so next I'm going to predict sentiment on route texts so let's give it some review texts that we are going to think of I love completing my to dos best app ever let's do uppercase here so we have this review and obviously this is like positive review or five-star review and if we go ahead and encode dust this I'm going to use the tokenizer and cold plus so this is pretty much the same thing that we did into the data frame in the previous video for how you can encode your data and I'm going to pass in the max length of the sequence which is 160 I'm going to speche up the special tokens a return token type IDs I don't want those and I'm going to pass the sequence I want the attention mask and I'm going to convert those to dancers or fighters dancers so after the encoding is complete I have the input IDs in the attention mask so I'm going to move those to the GPU the input IDs to the device in the same thing for the attention mask okay so finally I'm going to run the input to our model and get output or the prediction and I'm going to use torch max to get the most certain prediction and next I'm going to print the review text and I'm going to print out the sentiment that our model has predicted at least for this one and I'm going to print the quest name of the sentiment and if we run this you can see that our model is predicting that this is a positive sentiment which is great let's try another example horrible no let's see very bad so expensive developers are trash let's try this one of course those are some easy examples and you can see that the sentiment is negative right here so this is pretty much how you can train your model I will create or finish up the tutorial and this one will be linked into the description of this video I'm going to continue with this series and in the next video I'm going to show you how you can deploy this model behind REST API and we are going to build a simple web app in which you're inputting the text that you're interested in instead of using this text right here into the group eternal book so it will be a more user friendly stuff and the web app that we are going to build probably in JavaScript is going to make a post or it's going to make a rest request or HTTP request to the server and we are going to predict the sentiment using the model that we've already pre trained we are going to return the result and the web app is going to present some sort of well I hope beautiful design or UX so the user can predict the sentiment for whatever text he is interested in thanks for watching guys I hope you enjoyed this series and if you liked this please like subscribe to this channel turn the notification bell on so you know when you videos are landing thank you bye stay safe

Channel: Venelin Valkov

Views: 45,831

Rating: 4.9476333 out of 5

Keywords: Machine Learning, Artificial Intelligence, Data Science, Python, BERT, Sentiment Analysis, huggingface, PyTorch, NLP, Dataset, Google Play, Apps, User Reviews, Pandas, Tokenizer, Preprocessing

Id: 8N-nM3QW7O0

Length: 62min 23sec (3743 seconds)

Published: Sun Apr 19 2020