Integration of the new modules of data augmentation with the pipeline

[farinamhz]

In this step, We will add the modules of Back-translation, Word-alignment, and Similarity-comparison to the pipeline to use them for augmenting the data, determine which language is better, which reviews should be added to the original English dataset, and also find the related aspect labels of each augmented or new review.

Edit: In this step, we will add the back-translated reviews to the original dataset and retrain the model.

[farinamhz]

The modules are integrated with the pipeline, and the back-translated dataset is joined with the original one. However, we have a memory problem in the code that leads to finishing the code in the first fold before reaching all the folds. I am searching for the reason. @hosseinfani

[farinamhz]

Hi @hosseinfani, These are the results after back-translation. There is just a little bit of improvement (+0.029. at Success@1)

[farinamhz]

Also, we had a conversation about changing the evaluation to hit the predictions, which are a word with similar meanings to the aspect, so in case of similar words in the dictionary, we consider that the prediction was correct. However, I am not sure how to do it as we have a golden truth and a dictionary of words with the highest probability to the lowest to find the latent aspect, and I am confused about how to tell the pytrec library that even if the word is not exactly the one in the golden truth but has a similar meaning say that this one was a true prediction. @hosseinfani

[farinamhz]

Also, dictionary tokens of LDA improved from 203 to 422 words, which seems to be a good result. So probably we need to change sth, or maybe there is a mistake that results seem not improved so much. Or even if I need a proper interpretation of the results. @hosseinfani

[hosseinfani]

@farinamhz let's see the result for the btm.

[farinamhz]

@hosseinfani Also, these are the results for different languages and without back-translation for LDA:

[farinamhz]

@hosseinfani This are the results for BTM in German back-translation

[hosseinfani]

@farinamhz Good results. Can you put the lda and btm results for german in one figure?

[farinamhz]

@hosseinfani Here are the results:

[hosseinfani]

@farinamhz Good! BTM is better than LDA?? You reported otherwise before, right? If this is the correct trend, it's very nice. Please fix the line coloring, pick solid lines for before and dashed for after. Also, pick red for lda and blue for btm.

[farinamhz]

Hi @hosseinfani, Please take a look at these results. I am not sure which language is working better as it seems different in different topic models. However, based on the assumption, French and German should work better.

[farinamhz]

@hosseinfani These results seem good 😁 (Specifically in the right one for the first 5)

[hosseinfani]

@farinamhz thanks. remove success_from x-ticks value -> success remove metric

Pls draw for ndcg and recal too.

[farinamhz]

Hi @hosseinfani, These are all the results for back-translation using the French language for all the metrics and after updating your comments.

[hosseinfani]

@farinamhz Good. I have an idea to connect the right figures to left ones. We'll discuss it today at lab.

[hosseinfani]

@farinamhz I'm closing this issue as it's done already. Let me know otherwise.