Yamini Prakash

Semantic Textual Similarity

Goal of the Project was to determine how similar two sentences are and assigning a similarity score. We advanced our machine learning architecture by using various word embedding models to capture some sort of relationship in constructing word embedding space, be it meaning, morphology, context, or some other kind of relationship. Five techniques were used in achieving this task - Word2Vec, GloVe, FastText, ELMo and BERT. BERT Model gave us the best results with .86 Pearson Correlation (State Of The Art Performance is .92 on Leader board records) We ideated and designed “Similitude” a new app using the BERT Sentence Similarity model to enhance the online education evaluation system challenges and ease automated evaluation of “Descriptive Answers”. This app measures similarity between the student’s answers to the “expected answers” and provides a similarity score which can be used to automatically grade the student. It provided an excellent hands-on experience in coding an end-to-end real world solution, starting from pre-processing the data, to constructing word embeddings and testing the model and finally deploying the model in Cloud. The final user interface accurately identified the similarity between 2 given sentences and confirmed the similarity score.