Review rule, for dataset:
- [overall > 3.0] - positive
- [overall <= 3.0] - negative
- Generate term statistics: -- Vocabulary size with word frequencies -- N-grams -- POS collections
- Verify Zipf’s law – what is the best fit for your corpus?
- Which set of terms best describe your corpus?
Sentiment Analysis using statistical NLP
- Use the following vector space models -- CountVectorizer. -- TF-IDF. -- Any external vectorizer (cite the original paper).
- Do sentiment analysis using all (a,b,c) using classical ML techniques -- Naive Bayes Model. -- Decision Tree. -- Logistic Regression.
- Report metrics [accuracy, f1 score, confusion matrix] for all the combinations in (1 and 2)
- Analyse the results. [Report clearly which vector space model is giving better results on each model used]
Topic analysis and topic (attribute) wise sentiment analysis
- Extract the topics from the reviews using any topic extraction technique of your choice.
- Report sentences under each topic.
- Analyse whether the topics extracted make sense. Justify your claim with some examples.
- Report topic wise sentiment distribution for the whole repository. Explain the method that you used. Give complete reference of any paper that you use for the purpose.
- Run Stanford NER and Spacy NER and extract Named Entities from the data.
- Find the top 100 LOC and PERSON entities from the data set. What is the degree of correlation between the two systems? Consider partial and full matches.
- Generate Word2vec, Glove, Fasttext and BERT word vectors for the above corpus.
- Use the NERs detected in previous file to create annotated documents for NER detection. Divide the document collection into training, validation and test data sets. Implement a custom NER system (for all 4 vector embedding techniques mentioned in (c)) using LSTM. Compare the results of all models obtained (namely, (i) LSTM with word2vec, (ii) LSTM with glove, (iii) LSTM with fast text,
Develop an
- Bi-LSTM based sentiment analysis model using (a) word2vec embeddings (b) glove embeddings.
- BERT based sentiment analysis model.
- Compare the performance on the test set among these models. (Bi-LSTM (with word2vec, glove), BERT)
- Also, compare with Traditional ML Models developed in Task_1.