PVG at WASSA 2021: A Multi-Input, Multi-Task, Transformer-Based Architecture for Empathy and Distress Prediction
[abstract] [paper] [bibtex] [code]
- Developed a multi-input, multi-task learning framework for predicting Batson’s empathic concern and personal distress scores.
- The framework aggregates information from textual (using RoBERTa), categorical (using entity embeddings), and numeric data to generate robust representations, and simultaneously predicts the text’s primitive emotion (multi-class classification), presence of empathy (binary classification), and corresponding empathy score (regression).
- The multi-input, multi-task paradigm is further bolstered with the addition of NRC lexicons and Empath features for distress score prediction.
- Our system ranked 1st for the WASSA 2021 Shared Task. [scoreboard link]
Cluster Analysis of Online Mental Health Discourse using Topic-Infused Deep Contextualized Representations
[paper] [abstract] [bibtex] [code]
- Conducted research on discourse themes mining from online mental health communities.
- Proposed topic-infused deep-contextualized representations, a novel data representation technique that uses concatenated denoise autoencoder to combine deep contextual embeddings (RoBERTa) with topical information (LDA) for generating robust representations for text clustering.
- Employed UMAP for dimensionality reduction and HDBSCAN to draw out prominent clusters.
- Collaborated with clinical psychologists to conduct qualitative and quantitative characterization of each cluster, unraveling the thematic overlap, similarities, and differences amongst them.
- Creating a novel and comprehensive dataset of PTSD related posts on Reddit from 2015-2020.
An Attention Ensemble Approach for Efficient Text Classification of Indian Languages
[abstract] [paper] [bibtex] [code]
- Developed a Hybrid CNN-BiLSTM Attention Ensemble model for the task of coarse-grained automatic technical domain identification of short texts in the Marathi Language.
- The model generates latent representation that has useful temporal features from the sequences generated by the BiLSTM according to the context generated by the CNN.
- Our system ranked 1st for the TechDoFication Shared Task organized at ICON 2020. [scoreboard link]
Smart Cap: A Deep Learning and IoT Based Assistant for the Visually Impaired
[abstract] [paper] [bibtex]
- Developed ”Smart Cap” for the visually impaired providing features like scene description, text recognition, object detection and face recognition. [video demo]
- Implemented image captioning using attention-based encoder-decoder model, face recognition based on dlib’s face recognition, and OCR using Google Vision.
Other Selected Projects
- Implementation of Meta-Word-Embeddings, a combination of word2vec, GloVe, and fassttext word embeddings using different types of autoencoders namely: Decopuled Autoencoder, Concatenated Autoencoder, Average Autoencoder.
- The meta-embeddings contain common as well as complementary infomration from the input embedding set and outperfrom them in many tasks.
- Implementation based on the paper "Learning Word Meta-Embeddings by Autoencoding". [paper link]
Dynamic Sea Route Optimization
- Developed a graph-based strategy to connect all the lat-long coordinates in a shipping lane.
- Designed Algorithms for finding the distance-based optimal sea route using Depth First Search (DFS), Dynamic Programming (DP), and Beam Search.