Research
I am interested in Creative NLP and Language Generation. Language Model's ability to generate narratives with controllable parameters reflects their generation capability. I have also researched building NLP models efficient in parameters and robust to variations in data. I consider interpreting models to be a crucial step in achieving this. Only when we find out what ML models learn; can we guide them to learn better. I also contribute to improving resource-scarce languages' representation by curating resources in manners that require minimal human intervention.
I have applied my research interests across NLP applications such as Text Generation, Sentiment Analysis, Reading Comprehension, Question Generation, and WER Estimation. Representative papers are highlighted.
|
|
Narrative Language Understanding and Generation
Story Generation has applications in education and interactive learning. I proposed ReLiST [1] for story generation with control over relationships. ReLiST addresses the most important challenge of this task - writing a story coherently while reflecting the given relationships.
Relevant Publications:
[1] PARROT, a zero-shot approach for enhancing narrative reading comprehension via parallel reading
Chao Zhao,
Anvesh Rao Vijjini,
Snigdha Chaturvedi
EMNLP, 2023 (Findings) [link]
[2] Towards Inter-character Relationship-driven Story Generation
Anvesh Rao Vijjini,
Faeze Brahman,
Snigdha Chaturvedi
EMNLP, 2022 [arXiv] [anthology] [ppt] [poster]
|
|
Model Interpretability
Interpreting Machine Learning models is the key to identifying potential challenges and improvements while also moving towards Responsible AI.
I have researched and identified the behavior of CNNs, especially what convolutional layers are good at learning and what they are not. To address this, I propose SCARN - a model effectively utilizing recurrent and convolutional structures for text classification [1]. I have also researched and identified when Curriculum Learning (CL) works, especially in sentiment analysis. Using BERT’s attention visualizations, we give qualitative explanations of how CL works; by breaking down a more challenging problem into multiple easier subproblems [2].
Relevant Publications:
[1] Sequential Learning of Convolutional Features for Effective Text Classification
Avinash Madasu, Anvesh Rao Vijjini
EMNLP-IJCNLP, 2019 [arXiv] [anthology] [ppt]
[2] Analyzing Curriculum Learning for Sentiment Analysis along Task Difficulty, Pacing and Visualization Axes
Anvesh Rao Vijjini*,
Kaveri Anuranjana*,
Radhika Mamidi
WASSA at EACL, 2021 [arXiv] [anthology] [ppt]
|
|
Efficient Natural Language Processing
Lately, performance improvements in NLP tasks are being made based on
increasing parameters rather than efficiently training a single model. I am interested in achieving efficiency in NLP models where performance improvements can be made while maintaining model sizes hence moving towards Green NLP.
Efficiency can be achieved in multiple ways; my work explores some of these methods. Curriculum Learning introduces an ordering to the training while using the same model [2]. We have introduced self-normalizing layers within CNN for text classification to achieve better generalization while reducing parameters [4]. I have proposed parameter less decay based weighting layers to weight words closer to aspect terms more than farther ones for Aspect-Based Sentiment Analysis [3]. More recently, I proposed WER-BERT for Automatic WER Estimation, which uses a custom loss function for exploiting the ordinal nature of the WER classification task [1]. I am also interested in distilling models and semi-supervised learning.
Relevant Publications:
[1] Curricular Next Conversation Prediction Pretraining for Transcript Segmentation
Anvesh Rao Vijjini, Hanieh Deilamsalehy, Franck Dernoncourt, Snigdha Chaturvedi
EACL, 2023 (Findings) [anthology]
[2] WER-BERT: Automatic WER Estimation with BERT in a Balanced Ordinal Classification Paradigm
Anvesh Rao Vijjini*, Akshay Krishna Sheshadri*, Sukhdeep Kharbanda
EACL, 2021 [arXiv] [anthology] [ppt] [poster]
[3] A SentiWordNet Strategy for Curriculum Learning in Sentiment Analysis
Anvesh Rao Vijjini*,
Kaveri Anuranjana*,
Radhika Mamidi
NLDB, 2020 [arXiv] [springer] [ppt] [code]
[4] A Position Aware Decay Weighted Network For Aspect Based Sentiment Analysis
Avinash Madasu, Anvesh Rao Vijjini
NLDB, 2020 [arXiv] [springer] [ppt]
[5] Effectiveness of Self Normalizing Neural Networks for Text Classification
Avinash Madasu, Anvesh Rao Vijjini
CICLing, 2019 [arXiv] [ppt] [poster]
|
|
Robust Natural Language Processing
My research at Samsung focused on Domain Adaptation, which helps build models that scale across domains. I have proposed gated convolutional architectures for this problem. The gated mechanism filters out domain-specific information [2]. I further refined this idea to present a Sequential Domain Adaptation framework [1]. I am also interested in exploring Robust NLP for building models that are impervious to noise and are unbiased.
Relevant Publications:
[1] Sequential Domain Adaptation through Elastic Weight Consolidation for Sentiment Analysis
Avinash Madasu, Anvesh Rao Vijjini
ICPR, 2020 [arXiv] [ppt] [poster]
[2] Gated Convolutional Neural Networks for Domain Adaptation
Avinash Madasu, Anvesh Rao Vijjini
NLDB, 2019 [arXiv] [springer]
|
|
Resources for Under-Represented Languages
Modern machine learning algorithms rely heavily on data, and this reliance has adverse effects in Resource-Scarce Settings. I improved two pre-existing lexicon-based resources in Telugu [3,4,6]. We also created Hindi resources for Question Answering. To aid Educational Applications, we divided the resource into academic grades based on difficulty [5].
While resource creation for scarce languages is essential, hand-curating datasets can be time and labor-intensive. Moving towards Automated Resource Creation, we developed a solution to exploit the massive emoji-rich Social Media presence of Indian languages to automate Multilingual Sentiment Analysis datasets [1]. We also made a linguistically informed rule-based system that relies on semantic roles identified via dependency parsing to convert Hindi sentences into questions to facilitate HindiQA [2].
Relevant Publications:
[1] Twitter corpus of Resource-Scarce Languages for Sentiment Analysis and Multilingual Emoji Prediction
Nurendra Choudary,
Rajat Singh,
Anvesh Rao Vijjini,
Manish Shrivastava
COLING, 2018 [anthology] [data]
[2] Hindi Question Generation Using Dependency Structures
Anvesh Rao Vijjini*,
Kaveri Anuranjana*,
Radhika Mamidi
Humanizing AI (HAI) at IJCAI, 2019 [arXiv] [poster]
[3] Towards Automation of Sense-type Identification of Verbs in OntoSenseNet
Sreekavitha Parupalli,
Anvesh Rao Vijjini,
Radhika Mamidi
SocialNLP at ACL, 2018 [arXiv] [anthology]
[4] Towards Enhancing Lexical Resource and Using Sense-annotations of OntoSenseNet for Sentiment Analysis
Sreekavitha Parupalli,
Anvesh Rao Vijjini,
Radhika Mamidi
SemDeep-3 at COLING, 2018 [arXiv] [anthology] [code]
[5] HindiRC: A Dataset for Reading Comprehension in Hindi
Anvesh Rao Vijjini*,
Kaveri Anuranjana*,
Radhika Mamidi
CICLing, 2019 [researchgate] [poster] [data]
[6] BCSAT: A Benchmark Corpus for Sentiment Analysis in Telugu UsingWord-level Annotations
Sreekavitha Parupalli,
Anvesh Rao Vijjini,
Radhika Mamidi
Student Research Workshop (SRW) at ACL, 2018 [arXiv] [anthology]
|
|
Reviewer: ACL 2020, EMNLP 2020, EACL 2021, ACL 2021, EMNLP 2021
Sessions Chair: NLDB 2019
|
|
Teaching Assistant, Natural Language Processing, Monsoon 2016, Dr. Manish Shrivastava
Teaching Assistant, Natural Language Processing, Monsoon 2017, Dr. Manish Shrivastava
Teaching Assistant, Natural Language Applications, Spring 2017, Dr. Manish Shrivastava
|
|