Siddhant Arora

I am currently a Ph.D. student at Carnegie Mellon University's Language Technology Institute. My research interests are in the field of Natural Language (NLP) and Speech Processing, particularly in Spoken Language Understanding and Spoken Dialog Systems. My vision is to create robust spoken language understanding frameworks for improving performance in real-world scenarios. I have been lucky to find the support of strong advisor in Prof. Shinji Watanabe, with whom I have made some initial investigations in these directions.

I am working with Prof Shinji Watanabe on Spoken Language Understanding (SLU) with an emphasis on robust testing of SOTA SLU models and creating an open source benchmark for various datasets. We are further interested in enhancing semantic modeling in SLU systems and performing intent classification for spoken conversations by capturing the context in which utterance is spoken.

During my Master's, I worked with Prof. Graham Neubig, Prof. Zachary Lipton and William Cohen at Google AI on research projects involving rethinking user study design for evaluating explanations. This research taught me the importance of generating better interpretability frameworks and more actionable explanations. I also worked with Prof. Norman Sadeh at the Usable Privacy Policy Project to empower internet users to take back control of their privacy.

Previously, I was an undergrad at IIT Delhi, where I worked with Prof. Srikanta Bedathur on applying machine learning to the structured text, i.e., knowledge bases and relational databases. During my undergrad, I got an opportunity to work as a research scholar at Max Planck Institute of Informatics in the field of Neural Information Retrieval with Prof. Andrew Yates. I have also interned at Adobe Research Labs, Bangalore on toxic speech detection and Australia National University with Prof. Tom Gedeon on Emotion Recognition.

As a part of my research, I have published papers at leading conferences like ICASSP, InterSpeech, AAAI, AKBC, WSDM, and workshops at leading conferences like NeurIPS and ECIR. I am a recipient of the CMU Graduate Research Fellowship. I have also been awarded an Institute silver medal for securing Department Rank 1 in my undergrad at IIT Delhi.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  Twitter  /  Github

Devised the toolkit ESPnet-SLU, which is designed for quick development of spoken language understanding models inside ESPnet. Released an open-source toolkit and the trained models that match or even significantly outperform the SOTA by upto 8% on SLU benchmarks.

Developed a 2-pass SLU system that makes low latency prediction using acoustic information from the few seconds of the audio in the first pass and makes higher quality prediction in the second pass by combining semantic and acoustic representations.

Proposed a design where users have query access to the model and can observe how model predictions and explanations change in real time. Investigated extending the simulation task by prompting users to edit examples to reduce the model confidence towards the predicted class.

Proposed a framework to construct robust test sets over sub-task specific utility functions and observed more actionable comparisons between different architectures.

Introduce the first bilingual corpus of mobile app privacy policies consisting of 64 privacy policies in English (292K words) and 91 privacy policies in German (478K words), respectively with manual annotations for 8K and 19K fine-grained data practices.

Utilized probabilistic soft logic for removing noise from Knowledge graph based on ontology and Knowledge Graph embedding for inferring new facts together in a feedback manner.

Designed toolkit with implementations of prominent neural ranking integrated with the Anserini toolkit.

Used embedding to enable semantic queries on relational database.

Combined IR models based on instance level query features that are inspired by IR axiom.

Understand the various strengths and weaknesses of the proposed methodology based on GNN for Knowledge Graph Completion and try to find new exciting research problems in this area that require further investigation.

Performed hate speech detection & analysed how events impact hate content with synthetic time series. Predicted potential hate instigators for events to prevent toxic content from spreading online.

Last Updated: Nov. 16, 2021 Site Template