Project #4 of the NLP: What do you like in boardgames?

For the project I chose to elaborate the proposal number 4 as I have a personal interest in the topic of the domain, being boardgames.

All proposals are present in the repository under resources/

How to run

To run the solution dependencies shall be installed. They are listed in the requirements.
Also install:

python -m spacy download encorewebmd
python -m spacy download encorewebsm

No script has been written as I believed notebooks to be better at guiding the thought process.

First run the main/dataset/bgg_corpus_service.ipynb or download the dataset directly from: https://www.kaggle.com/datasets/jacopofichera/bgg-scrapped-reviews

To run preprocessing go to main/dataset/pre_processing.ipynb. This will generate various pre-processed datasets based on the starting one.

For LDA simply refer to main/lda/final_model.ipynb to launch training on the best found configuration of hyperparameters. The model is then created under \output in the same directory being an LdaMulticore instance of Gensim that can be reloaded.

For ABAE it is the same but in the abae folder. It creates more files being one for the word embeddings model, one for the initialization of aspect weight matrix before training and the keras instance model. To load and manipulate the model please refer to the ABAEManager class that holds methods based on what output is needed (if classify or loss evaluation).

Inference is left to be done by hand but using class #todo you can save it as part of the model output definition to be reloaded and used with #todo class to infer correct labels

References

My reference paper I think:

Paper: https://aclanthology.org/P17-1036.pdf
Repo :https://github.com/ruidan/Unsupervised-Aspect-Extraction/blob/master/code/train.py

Another interesting useful reference for an indepth application:

https://www.kaggle.com/code/nkitgupta/aspect-based-sentiment-analysis
Explains well how to do all. Nice insight on Emojis and Unicode normalization

Approach?

In an unsupervised paradigm for aspect extraction, you don't rely on labeled data. Instead, you can use clustering and topic modeling techniques to identify and extract aspects. Heres how you can approach it:

Data Collection and Preprocessing:
    Collect Data: Gather a large corpus of text related to your domain.
    Preprocess Text: Tokenize the text, remove stop words, and perform other cleaning steps.

Text Representation:
    Word Embeddings: Use pre-trained embeddings like Word2Vec, GloVe, or contextual embeddings like BERT embeddings.
    Document Embeddings: Represent each document as a vector, for instance by averaging word embeddings or using sentence embeddings from models like Sentence-BERT.

Aspect Extraction Techniques: ABAE, LDA

Project Setup and Instllation

python -m spacy download encoreweb_trf