2
$\begingroup$

I am considering two tasks:

  • Dialog Act Classification from Text (e.g. classify to: question; opinion; ...)
  • Emotion Recognition from Speech (e.g. happy; calm; sad; ...)

Which DL model should perform better for such tasks? I am planning to use CNN which should work for both of them, however not sure how well. Can I apply LSTM or some other methods? I used Keras before.

Is it good to apply attention mechanism or some other approaches for these 2 tasks?

$\endgroup$
2
$\begingroup$

Welcome to the site. I'm a little disturbed by the other two answers you received here. It sounds like you are skipping a whole lot of steps and wanting to jump right into modeling - that's a massive mistake!

You are a scientist! Your role is to create the most fair, unbiased environment possible to let the data speak to you (not the other way around!). What worked before (LSTM) may or may not be the best approach to this completely new data set. Therefore, after doing your EDA phase, you should keep an "open field" view to the multiple models that you will examine and test prior to making any decisions about which model to proceed with. The answer may not even be a neural network, it may be a whole different approach.

Please, be responsible your data science practice. You cannot jump into modeling right away. Let the data speak to you.

$\endgroup$
0
$\begingroup$

You can use both 1D Convolutions and RNNs like LSTM for text classification task. It is hard to say which one is better because it depends on your neural network and dataset structure.

Take a smaller sample from your dataset, then train and evaluate both networks. Pick the best model. Train with bigger data on this model. I think the most convenient method is this.

I suggest you to read this and this articles to understand how LSTM works and what you can do with it. There are some examples and use cases. Decide is it appropriate for your data or not.

$\endgroup$
  • $\begingroup$ I'm disappointed by this answer. You make no inquiries as to the nature of their data but are willing to jump straight into modeling? That's not a sound methodology in data science. $\endgroup$ – I_Play_With_Data Feb 20 '19 at 22:30
  • $\begingroup$ @I_Play_With_Data He did not ask the methodology, he asked for model alternatives. You can not judge anybody like this. Please read the question firstly. $\endgroup$ – Abdüssamet ASLAN Feb 21 '19 at 23:18
  • $\begingroup$ Are you really trying to defend your use of bad methodology? I’m disappointed in you. $\endgroup$ – I_Play_With_Data Feb 21 '19 at 23:31
  • $\begingroup$ @I_Play_With_Data My methodology is not best practice also please remember you are not a judge. You can feel disappointed about whatever you want. I only answered his question. He did not asked first steps, he asked for models. Please consider this website is not wikipedia, it is a q&a platform. $\endgroup$ – Abdüssamet ASLAN Feb 21 '19 at 23:39
  • $\begingroup$ Very disappointed in you $\endgroup$ – I_Play_With_Data Feb 21 '19 at 23:45
0
$\begingroup$

A typical model you could use is shown below-

Input Text -> Word Embedding -> Bidirectional LSTM -> Dense output layer

Word embedding layer - maps the words from the vocabulary into vectors of real numbers.

Bidirectional LSTM - since they can preserve information from both the past and the future they can understand context better as compared to unidirectional LSTM.

Checkout the following links for more details-

https://machinelearningmastery.com/what-are-word-embeddings/ https://machinelearningmastery.com/develop-bidirectional-lstm-sequence-classification-python-keras/

$\endgroup$
  • $\begingroup$ I'm disappointed by this answer. You make no inquiries as to the nature of their data but are willing to jump straight into modeling? That's not a sound methodology in data science. $\endgroup$ – I_Play_With_Data Feb 20 '19 at 22:30

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.