Information Retrieval

An essential and key advantage of NLP is 'information retrieval' from the vast amount of information available (nowadays), google search is the perfect example of this. Based on tf-idf matrix (vectors), searching for the closest documents within the list of documents, leveraging 'cosine similarity' (illustrated here image).

How it works:

Based on the concept of matrix multiplication which is really fast, I thought of representing every document/text/queries in the numeric form (else you can't perform matrix multiplication -> core idea of deep learning as well). Tf-idf is one of the 'text vectorization' technique which is being used in order to transform corpus into an equivalent matrix (tf-dif matrix : a numeric representation) form . Once you have the tf-idf vector (metrix) representation, you can leverage this metrix to compute cosine similarity with any other similar vector which can be used to find most closest doc/text/query present in the corpus. This enables us to find those closest/similar patterns in a very short duration. This method does not consider semantic/syntatic similarities between documents/queries/sentences hence it won't work where the objective is to retrieve look-alike documents/queries/sentences, based on the context.

Getting started:

create virtual environment
install requirements

For building tf-idf matrix and getting similar documents:

python find_similar_text.py

(Above script supports user interactions, do check by typing text/pattern and the output would be the closest articles present based on the input)

Where to use:

1. Browsing through huge number of articles, to find the most closest one w.r.t. your task, you can use tf-idf based matrix multiplication to achieve it very quickly
2. For queries where your model fail to predict the right class (sentiments/intents etc.), this repo can help you identify queries which is of similar pattern or use same type of vocab
3. To merge similar type of patterns/documents since it's always better to have more discernible group

Further improvements:

1. By leveraging text pre-processing options like stemming, lemmitization. The advantage of using them upfront is to not only reduce the matrix sparsity but also help with more informed similarity matching using lemma/root of the word (which might get missed out due to inflectional endings)
2. Using different techniques of text vectorization

• Binary Term Frequency
• Bag of Words(BoW) Term Frequency
• L1 Normalized Term Frequency
• L2 Normalized tf-idf (currently in use)
• Word2Vec (kindly have a look at genesim for pre-trained embeddings)
• flairNLP library (entials most of the available pre-learned embeddings)

3. Using semantic similarity from pre-trained language models (bert, albert , transformer based model etc). Find implemention with BERT👉 https://github.com/Rishabbh-Sahu/semantic_lookalike_transformers

References:

Data set is all about "Medium Articles", A collection of articles on ML, AI and data science. Download link 👉 https://www.kaggle.com/hsankesara/medium-articles?select=articles.csv