0
$\begingroup$

I have a big data-set of sentences (tens of thousands) which some of these sentences are big but some are short. The main problem is that you should classify some sentences according to previous or sometimes next sentences!!! for example, a sentence individually maybe belongs to class 1 but if you look at previous sentence(s) it goes for other class like class 2. Simple text combination makes the results worst

I have implemented lots of traditional models and approaches like TFIDF, Ngrams and ... with all common classifiers and all types of ensemble classifiers, And even lots of deep learning approaches (even state-of-the-art) like CNN,Fast-Text,all types of LSTM (like BiLSTM,CLSTM,...), TextRNN Bert, XLnet and so on...

But no one gaves me a good result(the best one was 75%). I did lots of text mining projects but this one really drives me crazy :(((

Any suggestion would be appreciated. Thank you...

$\endgroup$
0
$\begingroup$

Build document/sentence vectors as features. In my experience the model is less important (although a simple Naive Bayes typically performs quite well in document classification. SVMs are also well suited.) than providing it the right embeddings/vector representations.

There are pre-compiled "dictionaries" for generating document vectors, such as Doc2Vec, but if you want to be sure to capture the "context" of the sentence as part of the embedding, you can generate your own document embeddings with an rnn or lstm.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.