Automated Code Comment Generation
Analysis of automated code comment generation with zero-shot learning, in-context learning and prompt tuning on LLMs
This project aims to investigate and enhance the automated code comment generation process using Large Language Models (LLMs). Leveraging techniques such as zero-shot learning, in-context learning, and prompt tuning, the project seeks to improve the quality and relevance of generated code comments. Through comprehensive experimentation and evaluation, the research aims to provide valuable insights into the capabilities of LLMs in the context of software development, ultimately contributing to more efficient and effective code documentation practices.
github.com
December 13, 2023
Generating Morals for Fables
Unlocking Wisdom's Codes: Decoding Morals in Fables with NLP
Fables are short narratives that teach us valuable lessons in the form of morals; understanding these morals is important because they are the central messages of their fables. This thesis explores the challenging task of understanding the morals of fables using natural language processing. We propose features, focused on the characters and events of a fable, that include both semi-structured subject-predicate-object triplets, as well as with the Story Intention Graph; we develop the first fully automated pipeline for building SIGs from narratives. Using the triplet features, we train a neural classification model to label a fable with one of twelve abstract categories of morals, demonstrating that neural networks are capable of extracting moral information. Using SIGs, we tackle the more challenging, open-ended task of generating a fable’s moral from scratch. We develop graph-to-sequence models for this task, proposing a novel global meta-node for initializing a sequence decoder from a graph encoder. Our best-performing model achieves state-of-the-art results for the task of moral generation.
November 19, 2023
Prompting political
Classifying English and Arabic news articles using Prompt-Based Techniques
The objective of this study was to enhance the performance of language models in identifying political violence by fine-tuning BERT, RoBERTa using prompting techniques. Two types of prompting were employed: tuning-free prompting and fixed prompt language model tuning. The results showed that fixed prompt language model tuning was the most effective approach, achieving an average F1-score of 0.784 across both models on the BBC violence classification dataset. These findings highlight the potential of prompting techniques in enhancing the ability of language models to analyze and identify political violence in news articles. The study also applied the same fine-tuning techniques on a Arabic language dataset for fake news classification and demonstrated that fixed prompt language model tuning was more effective, resulting in an average F1-score of 0.62.
github.com
April 25, 2023
DDS: Document Decoding Service
An structured information extraction system.
An end-to-end structured information extraction system from receipt and invoice images. As part of AI research and development team, I was responsible for continuous developement of the solution and timely integration with the production. I worked on GraphNN with textual, image, geometrical features of the words in the document to determine the tabular structure in the document. Trained and refined the model to achieve production level results. Improved the workflow by optimizing the pre/post-processing components.
July 20, 2021
Match
A system to match professionals to correct job positions.
An AI driven system to match the skilled employees to correct job openings. As a part of AI incubation team, I was resposible for introducing ML solutions that would improve the system. Developed a healthcare recruitment specific lexicon model using TransitionBasedParser to retrieve important information, like skills, experience, certifications, from free text.
June 10, 2021
ReLIE: Representation Learning for Information Extraction
Unofficial implementation of google's representation learning for information extraction.
A transformer based entity recogtnition algorithm from form-like documents. Trained a base transformer model to recognize predefined entities in document images and then extract the correponding values.
research.google
github.com
May 20, 2021
Pufferfish
Open-source fully featured mechanical ventilator.
Open source fully featured mechanical ICU ventillator designed with the idea of low cost and rapid production. Worked on building asynchronous python server for highly efficient and isolated communication between the UI and micro-controller.
pez-globo.org
github.com
November 15, 2020
HRMS: Human Resource Management System
A realtime video analytics patform.
A system to monitor COVID violations and report it to the admin. Worked on NVIDIA Deepstream with TensorRT models to implement 40 simultaneous highly efficient stream processing pipelines. Also worked on extracting person’s temperature information using thermal camera data. These solutions have been filed for patent and are waiting for approval.
blackstraw.ai
September 10, 2020
Speech Translation
End-to-end active speech translation system.
As a proof of concept, developed and deployed a bi-directional Mandarin-to-English active translation pipeline. The pipeline consisted of three components speech-to-text, neural machine translation and text-to-speech. Worked with Mozilla’s DeepSpeech 2 for speech recognition, output of which was processed by Fairseq’s convolutional encoder + BiLSTM decoder neural machine translation model and then converted back to audio using Tacotron 2 for speech synthesis.
March 25, 2020
Abstractive summarization
As a proof of concept, developed a abstractive summarization system to capture data from various sources and generate a brief and comprehensive summary. Worked on PreSumm, which uses a modified BERT encoder and a regular transformer decoder with different optimizers for efficient fine-tuning. While the encoder was able to capture the representations of salient contexts, the decoder lacked in generating cohesive text. Improved it by using GPT-2 as the decoder. The GPT-2 based decoder was fine-tuned to use BERT encoder representations to generate new text.
November 12, 2019
Sentence semantic similarity using dependency parsing.
A novel approach to measure semantic closeness between sentences.
Developed an novel metric to calculate the semantic closeness between given sentences. Compared the dependency parse trees of the sentences along with the WordNet semantic closness of the corresponding words to generate the closeness rating.
ieeexplore.ieee.org
github.com
August 10, 2019
Rehabo: Physical Rehabilitation with Computer Vision
An AI aided tool to help patients with their physical rehabilitation.
Developed ML powered game for physical rehabilitation of patients. A plug play device whic guides the user to through simple exercises, hence working on their recovery and tracking their progress. Worked on VNECT 3d pose estimation model on NVIDIA Jetson Nano and custom algorithm for the interactive game. This project was launched as a startup, which was supported by SSIP.
linkedin.com
April 14, 2019
MedSearch: A medical image search platform.
Content based image retrieval as a platform for medical personnel.
Developed an end-to-end image search platform specifically for people in medical domain to share information on pecular cases so that the infomartion could be used in future incidents. Used SIFT image feature extraction and KNN matching to find similar features while comparing the images. Implemented this as a service using Apache Spark cluster, MongoDB and an Android application.
github.com
June 04, 2018