Google Summer of Code 2022 Final Report.md

Google Summer of Code 2022 Final Report

Name: Hatma Suryotrisongko • Project: OWASP Maryam

Proposal Topic

Deep Learning for OWASP Maryam's NLP operations

Work was Done

Designed and implemented a topic modeling module (PR #269) (Personal repo link).

Milestones Achieved

• Learning and implementing all existing Maryam's modules. (Jupyter Notebooks).

• The first prototype.

  • Designed and implemented the first prototype for the core/util/iris/topicmodeling.py: K-Means algorithm and Wordcloud. (notebook),

    output = tm.perform_kmeans()

    

• The second prototype.

  • Designed and implemented topic modeling using Mr. Kaushik's dataset, with NMF and LDA algorithm (notebook),

    tm = topicmodeling(jsonfile)

    tm.perform_nmf1()

    

    tm.perform_nmf2()

    )

    tm.perform_lda()

    

• The third prototype.

  • Implemented BERT algorithm using BERTopic and SentenceTransformer. (notebook), users will be able to choose their preferred word embedding model, from repositories, as for a particular dataset user may need to select the best word embedding model which has the most superior performance.

    

    tm.run_topic_modeling("paraphrase-distilroberta-base-v1")

    

• The fourth prototype.

  • Adding CVF file support and stopwords removal. (notebook),

    tm = topicmodeling(inputfile="testdataset.csv", filetype="csv", verbose=True)

    

• The initial creation of the PR for topicmodeling module (PR).

  • Maryam/maryam/core/util/iris/topic.py

  • Maryam/maryam/modules/iris/topicmodeling.py

• Fix bug and add new feature (commit).

  • Renaming the plotly.py to solve the dependencies problem. In addition, I also added the --output and --api functionality. Tested by using =

    topicmodeling -i mixed.json -t json -m all-distilroberta-v1 --output

    report json testreport iris/topicmodeling

    topicmodeling -i mixed.json -t json -m all-distilroberta-v1 --api

• Added a new feature:

  • Search Topics Using a Keyword (Top2Vec). (notebook),

    topic_words, word_scores, topic_scores, topic_nums = model.search_topics(keywords=["music"], num_topics=2) for topic in topic_nums: model.generate_topic_wordcloud(topic)

    

    

To Continue My Work

• Top2Vec has a limitation: that it might not work well for a small dataset. Therefore, when using a small dataset (such as Mr. Kaushik's mixed.json) sometimes it will not work. See the demonstration on jupyter notebook.

  • The same notebook as the previous notebook, but when we re-run the notebook sometimes it will fail.

  • Top2Vec algorithm has a good performance, but still has this problem when using a small dataset. To continue my work, we need to find an alternative algorithm or improving the current Top2Vec implementation.