About Me

| | | | |

I am a Senior AI/ML Manager at Apple Knowledge Platform where I work on Natural Language Processing and Machine Learning. Prior to this I was a Senior Staff Applied Scientist and Engineering Manager at IBM Watson where I designed and developed algorithms for IBM's conversational AI product - Watson Assistant. I got my Masters degree at the Language Technologies Institute at Carnegie Mellon University in 2014.

What I’ve been working on

I work on the Apple Knowledge Platform which powers interactions across Siri and Spotlight. I currently work on Knowledge Graphs, Entity Linking, Question Answering and Synthetic Query Generation.

Prior to this I worked on the Natural Language Understanding components of Waston Assistant which includes intent classification, entity recognition, spellcheck and irrelevant detection across multiple languages. The algorithms are designed to be custom-trained for customers globally, deployed at scale with hundreds of thousands of models in production and serves more than 1.9% of the world’s population every month.

Awards & Recognition

MIT 35 Innovators under 35 - Semi Finalist
IBM Master Inventor - Title recognizing sustained and outstanding contributions to IP
Corporate Award 2021 - Watson Assistant
Outstanding Technical Achievement Award 2021 - State-of-the-art algorithms for intent classification and entity recognition in Watson Assistant
Research AI Accomplishment (A-level) 2020 - Meta-Learning for Low-Resource NLP
Corporate Award 2020 - Language enablement for Watson Services
Women in Leadership Program 2019 - I was one of the 25 women leaders across IBM selected for the fully funded eCornell certificate program
Corporate Award 2018 - Watson Conversation Service
Best of IBM 2018 - Awarded to less than 1000 employees globally contributions to IBM’s business
Eminence and Excellence Award 2017
Outstanding Technical Achievement Award 2016 - Watson Conversation Service
Invention Achievement Awards

Selected Patents

Complete list on Google Patents

US Patent 10937416 Cross-domain multi-task learning for text classification
US Patent 10977445 Weighting features for an intent classification system
US Patent 11023683 Out-of-domain sentence detection
US Patent 11093707 Adversarial training data augmentation data for text classifiers
US Patent 11120225 Updating an online multi-domain sentence representation generation module of a text classification system
US Patent 11068656 Displaying text classification anomalies predicted by a text classification model

Selected Publications

Complete list on Google Scholar

[pdf]

Distinguish Sense from Nonsense: Out-of-Scope Detection for Virtual Assistants

EMNLP 2022

Cheng Qian, Haode Qi, Gengyu Wang, Ladislav Kunc, Saloni Potdar

An OOS detection method that outperforms standard OOS detection methods in a real-world deployment of virtual assistants.

[pdf]

Benchmarking Language-agnostic Intent Classification for Virtual Assistant Platforms

Workshop on Multilingual Information Access (MIA) - NAACL 2022

Gengyu Wang, Cheng Qian, Lin Pan, Haode Qi, Ladislav Kunc, Saloni Potdar

A benchmark for evaluating language-agnostic intent classification across 29 public and internal datasets across 10 low-resource languages.

[pdf]

Improved text classification via contrastive adversarial training

AAAI 2022

Lin Pan, Chung-Wei Hang, Avi Sil, Saloni Potdar

We propose a simple and general method to regularize the fine-tuning of Transformer-based encoders for text classification tasks.

[pdf] [blog]

Benchmarking Commercial Intent Detection Services with Practice-Driven Evaluations

NAACL 2021 Industry Track

Haode Qi, Lin Pan, Atin Sood, Abhishek Shah, Ladislav Kunc, Mo Yu, Saloni Potdar

We propose a methodology to assess the performance of intent detection "in the wild" in task-oriented dialog systems.

[pdf]

Multilingual BERT Post-Pretraining Alignment

NAACL 2021

Lin Pan, Chung-Wei Hang, Haode Qi, Abhishek Shah, Saloni Potdar, Mo Yu

We propose a simple method to align multilingual contextual embeddings as a postpretraining step for improved cross-lingual transferability of the pretrained language models.