A simple implementation method for classifying multiple Japanese sentences.

The benefits of using TF-IDF are

• You can use a library (such as scikit-learn) to calculate it, so you can easily design features.
• If the model trained with these feature vectors retains the importance of coefficients and feature quantities, you can see things like, "Which word was important to prediction?"
• Simple machine learning algorithms such as logistic regression, naive Bayes, and decision trees can be applied.

There is a point that(I think there are other things, but I mentioned it as something that comes to mind.)

Since you use naive Bayes, the importance of feature quantity can be obtained by using the link below.
https://stackoverflow.com/questions/50526898/how-to-get-feature-importance-in-naive-bayes

If you want to use rails, calling python from rails is relatively easy.
https://github.com/mrkn/pycall.rb

The disadvantage is

• More data and more words increase the dimension of the feature vector, making memory inefficient.
• I don't have any potential information directly (word A and word B are similar).
• Adding learning to accommodate new words in the new data will incur a feature vector recalculation cost.

There are points such as .(I think there are other things, but I mentioned it as something that comes to mind.)

Other options include:

• Use words2vec and tensorflow-hub's nnlm, universal sentence encoder, laser to represent sentences in vectors lower than tf-idf, and then enter them into models created in keras to learn.
• Sentencepiece is used to divide each sentence into subwords (converting those subwords into IDs) with a BPE trained from the sentence set and input it into a typical sentence classification model created by keras (a model consisting of layers such as Embedding, LSTM, etc.).
• Fine-tuning pre-learned language models such as BERT