Chinnadhurai Sankar

Email  /  Google Scholar  /  LinkedIn  /  Twitter

Research Lead at SliceX AI.

- Prev: Senior Research Scientist at Meta AI for End-to-End Dialog Research Efforts - Project CAIRaoke[CNET]
- Research Focus: NLP, Conversational AI, On-device AI, DeepLearning, Deep RL.
- Best Paper award, SIGDIAL 2019 and Best Paper Honorable Mention, EACL2021 & ACL 2019.
- Area Chair for ACL ARR 2021, Senior PC - CoLLAs 2022
- Workshop co-organizer - NILLI @ EMNLP 2022, NeuCAIR @ ICLR 2021.

[Education] I completed my Ph.D. at the University of Montreal, MILA lab advised by Prof. Yoshua Bengio. I earned my bachelor’s degree at IIT Madras, where I majored in Electrical engineering and minored in Physics and my master's degree in ECE at Purdue University.

[Jul 2024] 📣 Excited to Launch Llama3.1-ELM-Turbo models - 3B, 4B, 6B model sizes with MMLU - 56 @ 16000 tokens/sec.[Blog post]

[Apr 2024] Launched ELM-V1 model collection - LLMs with a custom efficient transformer architecure that once pre-trained, yielded multiple smaller models for inference. [Blog post]

[Jan 2024] Serving as a Senior PC for CoLLAs 2024.

[Dec 2023] Co-organizing NILLI 2023 at EMNLP2023! Checkout our 2022 edition - here.

[Dec 2023] Our Continual Dialogue State Tracking paper, a novel way of combining continual learning with in-context learning to solve dialog state tracking is accepted to EMNLP-23

[Jul 2023] Serving as a Senior PC for CoLLAs 2023.

[Dec 2022] Started a new journey at SliceX AI!.

[Oct 2022] Our CheckDST paper on comprehensive Dialogue State Tracking metrics is accepted to EMNLP-22, findings.

[Oct 2022] Our paper on data efficiency of instruction tuning vs prompting accepted to ENLSP @ NeurIPS2022.

[Aug 2022] Our work - Data Augmented Invariant Regularization (DAIR) accepted to TMLR2022.

[Apr 2022] Two of our papers have been accepted NAACL 2022 - DSR-ambiguity to model ambiguities in task oriented dialogs and KETOD to combine Task-Oriented and Chit-Chat dialogs grounded in external knowledge.

[Mar 2022] Serving as a Senior PC for CoLLAs 2022.

[Feb 2022] Glad to have incepted and worked on foundational conversational AI tech - Project CAIRaoke presented during the Meta AI day 2022 [ Video][CNET]

[Dec 2021] Two preprints out - DSR-ambiguity to model ambiguities in task oriented dialogs and CheckDST, a checklist to measure real-world generalization of dialog models

[Nov 2021] Co-organizing Learning to learn through interaction (NILLI @ EMNLP-22).

[Nov 2021] Serving as an area chair for ACL ARR 2021.

[Oct 2021] Checkout our work - a novel regularizing loss function to improve model robustness beyong Data Augmentation

[Jul 2021] Two of our recent works on dialog - DialogStitch: a framework to synthetically create longer dialogs by stitching existing ones and an improved version of the MultiWOZ2.2 dataset by fixing inconsistent annotations are accepted to SIGDIAL2021.

[May 2021] Our work on multi-step reasoning in video grounded dialogue has been accepted to ACL2021.

[Apr 2021] Our paper on on-device LSH based Transformers won Best Paper Award, Honorable Mention at EACL 2021 (oral presentation).

[Jan 2021] Two of our recent works on the robustness of LSH based text representations and on-device LSH based Transformers are accepted to EACL2021.

[Jan 2021] I am co-organizing ICLR 2021 Workshop on Neural Conversational AI (CAIR @ ICLR-21).

[Oct 2020] Our paper which proposes a new RL based data augmentation for modelling open domain dialogs has been accepted to JAIR.

[Mar 2020] Joined as a Research Scientist at Facebook AI to work on conversational AI.

[Jan 2020] Two of our research efforts cited as important conversational AI papers from 2019. Check out this Forbes article

[Sep 2019] Our paper on modelling chit-chat dialog with discrete attributes using Deep RL won Best Paper Award at SIGDIAL 2019 (oral presentation).

[Aug 2019] Our paper, TaskMaster Dialog Corpus: Toward a Realistic and Diverse Dataset, accepted to EMNLP 2019 (oral presentation). Check out Google AI blog post, Venturebeat

[Jul 2019] Our paper on analyzing context representation in neural dialog systems has been accepted to ACL 2019 (oral presentation, Best Paper Award nomination).

[Mar 2019] Our paper on training embedding-less word representations has been accepted to NAACL 2019.

[Jan 2019] Started internship with Google Brain to work on Task oriented dialog systems. Papers in review at EMNLP 2019 and NeurIPS 2019.

[Jan 2019] Gave a contributed talk at Deep-Dial, AAAI 2019 about our work on modelling chit-chat dialog. Also, accepted to Conv-AI, NeurIPS 2018.

[Oct 2018] Our paper which proposes a NEW simple recurrent architecure for modelling long term dependencies has been accepted to AAAI 2019 (spotlight presentation).

[Dec 2017] Our dialog bot, MILABOT won 2nd prize in NeurIPS 2017 demonstration track.

Know Thy Strengths: Comprehensive Dialogue State Tracking Diagnostics
[EMNLP 2022, findings][paper]

DAIR: Data Augmented Invariant Regularization
[TMLR 2022][paper]

Database Search Results Disambiguation for Task-Oriented Dialog Systems
[NAACL 2022][paper]

KETOD: Knowledge-Enriched Task-Oriented Dialogue
[NAACL 2022, findings][paper]

Annotation Inconsistency and Entity Bias in MultiWOZ
[SIGDIAL 2021][paper]

DialogStitch: Synthetic Deeper and Multi-Context Task-Oriented Dialogs
[SIGDIAL 2021][paper]

DVD: A Diagnostic Dataset for Multi-step Reasoning in Video Grounded Dialogue.
[ACL 2021][paper]

ProFormer: Towards On-Device LSH Projection Based Transformers.
[EACL 2021, oral, Best Paper Award, Honorable Mention][paper]

On-Device Text Representations Robust To Misspellings via Projections.
[EACL 2021][paper]

Deep Reinforcement Learning For Modeling Chit-Chat Dialog With Discrete Attributes.
[SIGDIAL 2019, oral, Best Paper Award ][arxiv]

TaskMaster-1 Dialog Corpus: Toward a Realistic and Diverse Dataset.
[EMNLP 2019, oral][Google AI blog post][arxiv][data][Media Coverage]

Do Neural Dialog Systems Use the Conversation History Effectively? An Empirical Study.
[ACL 2019, oral, Best Paper Award nomination][arxiv][code]

Transferable Neural Projection Representations.
[NAACL 2019][arxiv]

Towards Non-saturating Recurrent Units for Modelling Long-term Dependencies.
[AAAI 2019, spotlight][arxiv]

Please refer to my Google Scholar page for a detailed list of my publications.