Sentiment analysis on emoji data.

[PrashanthAadepu]

I am using BERT to do sentiment classification. I am currently classifying into positive, negative and neutral.

I have some data with emojis and it is always classifying them as neutral. I think I am missing something here.

Could someone explain to me how to deal with emoji data to classify them correctly.

Thanks in advance.

[aditya-malte]

Hello, Could you elaborate on what you mean by "data with emojis". Do you mean the emoji alone or with some surrounding text? Because as far as I remember the author of this repo has added emoji support. Thanks

[PrashanthAadepu]

Hi,

I have data like the below sentences.

😍 Love your service period! 😂😂😂😉🤗💕

When I classify the sentences with only emojis its always predicting them as neutral.

Thanks.

[aditya-malte]

Interesting. Does your training data consists of a mixture of emoji and emoji less text? Or do all of them have emojis?

[PrashanthAadepu]

Data has below variants 1, Sentence with no emoji. Ex: Very useful for customers 2, Sentence with text and emoji. Ex: 😍 Love your service period! 3, Sentence with the only emoji. Ex: 😂😂😂😉🤗💕

Thanks.

[aditya-malte]

That's very surprising. Could you share the hyperparameters that you have used so that I can see if something is wrong.

[PrashanthAadepu]

Hello

I am using the below notebook. Tweaked it to classify neutral sentiments also.

https://github.com/google-research/bert/blob/master/predicting_movie_reviews_with_bert_on_tf_hub.ipynb

I am using below tokenizer to properly tokenize the emojis. https://github.com/google-research/bert/blob/master/tokenization.py

And I also added some emojis in vocab.txt and passing it to model training.

Thanks.

[dataislife]

Hey! Any improvements on that aspect? Seems surprising since emojis should be taken into account now by Bert tokenizer. Older version was considering an emoji as UNK token.

[bobbyinfj]

Hello

I am using the below notebook. Tweaked it to classify neutral sentiments also.

https://github.com/google-research/bert/blob/master/predicting_movie_reviews_with_bert_on_tf_hub.ipynb

I am using below tokenizer to properly tokenize the emojis. https://github.com/google-research/bert/blob/master/tokenization.py

And I also added some emojis in vocab.txt and passing it to model training.

Thanks.

Hi Prashanth,

Could you share your code tweaks to classify neutral sentiments? I am starting with the same notebook and am actively struggling with making the same tweaks.

Thank you.

[freeIsa]

Hi, I am also trying to use BERT on data containing emojis but they are always encoded as <UNK> from the Tokenizer. Has there been any progress in making emojis correctly processed?

[Vithurshana]

Hello.. could you please share the dataset with emojis.. that would be more helpful..