dimits-ts / text_analytics

Language Modelling (text generation, spell correction) and Sentiment Analysis / POS Tagging with MLP, RNN, CNN and BERT models and LLM prompting

Text Analytics

Language Moddeling

We create bigram, trigram and linear interpolation language models which are used for language generation and spell correction.

Source code Report

Sentiment Classification and POS Tagging tasks

We create deep learning models using the Transformers\Datasets, Pytorch and Tensorflow libraries. We also use the keras_tuner / transformers_trainer frameworks to optimize hyperparameters and model architecture.

We briefly mention additional tasks carried out:

• Sentiment Analysis: Dataset selection, exploratory analysis, custom stopwords, data augmentation.
• POS Taggging: Dataset selection, exploratory analysis, custom parsing, custom baseline ("smart dummy") model, local caching of heavy computations, automated results generation (python -> LaTeX).

Each task features two IPython notebooks containing the executed code, python source files for repeated custom tasks and a unified report.

The reports discuss in detail the design decisions for each classifier and include graphs and aggregated results comparing the current model to the previous models.

Simple MLP model

Sentiment classification POS Tagging Report

RNN Model

Sentiment classification POS Tagging Report

CNN Model

Sentiment classification POS Tagging Report

BERT Model

Sentiment classification POS Tagging Report