I am developing a generative AI in Python using an LSTM model to handle queries related to legal texts in French. I’m facing outputs issues as they are either not correctly processed in French or inaccurate.

I’ve tried adding more LSTM layers, keeping only minimisation as a preprocessing but my generated outputs lacked quality and accuracy as you can notice in the below screenshot.

Do you have any specific advices you recommend for improving my text generation?

Thanks.

https://imgur.com/a/UMwmWLm

I’ve trained the below models that generated the outputs present in the imgur link:

• model0_2l with 2 LSTM layers and with preprocessing (I removed the punctuation, the stop words, and I minimized the training texts)

• model1_3l with 3 LSTM layers and with the same preprocessing.

• model2_min_3l with 3 LSTM layers and with text minimization as preprocessing.

• model3_min_4l with 4 LSTM layers and with text minimization as preprocessing.