0
$\begingroup$

I'm beginner in deep learning and I'd like to cluster text-based software requirements by themes (words similarities/frequency of words) using neural networks. Is there any example/tutorial/github code of unsupervised neural network that groups texts based on themes and words similarities?

Thank you very much for your answers!

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

I recommend using word2vec as feature vector of words and LSTM autoencoder to encode a sentence (or text) . After you get a vector for each sentence (or text), you can cluster your sentences (or texts) using a variety of clustering techniques like k-means or dbscan and represent them using t-sne or u-map. Start from here: https://blog.myyellowroad.com/unsupervised-sentence-representation-with-deep-learning-104b90079a93

Improve this answer
$\endgroup$
3
  • $\begingroup$ thank you for your answer ! regarding the sentence encoding, there is an existing implementation of the TF-IDF algorithm in sklearn, here is the tutorial (pythonprogramminglanguage.com/kmeans-text-clustering). Thus, i am wondering why it's recommended to use encoding techniques such as word2vec and LSTM. Can you please explain the advantages of using such techniques compared to the one implemented in sklearn for instance? $\endgroup$
    – Takwa
    Apr 19, 2019 at 14:12
  • $\begingroup$ You’re welcome. Actually, the first advantage of using word2vec over tf-idf is that, word2vec contains contextual information but tf-idf does not. The second advantage is that, word2vec uses information from a large dataset (pre-training), so it better models the language than tf-idf. And for the third advantage you should consider that as the vocabulary size increases, the tf-idf size increases, too. However, pre-trained word2vec vectors have fixed size, regardless of vocabulary size. $\endgroup$
    – pythinker
    Apr 19, 2019 at 18:45
  • $\begingroup$ @Takwa You’re welcome $\endgroup$
    – pythinker
    Apr 25, 2019 at 9:19

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.