Jul 2023 – Apr 2024 · 9 mos · Remote

May 2022 – Apr 2023 · 11 mos · Mountain View, California, United States

• Led a team of 3 AI data scientists in developing a multilingual intents and slots classifier using PyTorch and XLM-Roberta, achieving a 15% and 25% improvement in accuracy.
• Presented a zero-shot learning method that incorporated generated translated data and delexicalization for intent and slots classifiers.
• Achieved 3rd place in the MMNLU Competition and published the work in EMNLP 2022.
• Established an experimental environment on AWS m5.xlarge machine for model evaluation, exploring architectures such as XLM-Roberta and mT5 for the intents and slots classifier project.
• Developed a prototype product for a multilingual intents and slots classifier on Samsung Bixby, expanding its language capabilities from processing and responding in a single language to 51 languages.
• Optimized Large Language Models (LLMs) such as Dolly and GPT-J for improved intent and slot classification, implementing the LoRA algorithm to reduce trainable parameters by 10,000 times and GPU memory requirements by 3 times while maintaining accuracy.
• Executed prompt engineering in inference with ChatGPT and Dolly for automated data annotation, achieving an 80% reduction in time and cost for data annotation.

Jun 2020 – Dec 2020 · 6 mos · Taipei City, Taiwan

• Built a framework to train Empathetic Chatbots (CheerBots), based on BERT and OpenAI-GPT models, to make interlocutors feel more sympathy while chatting using Pytorch.
• Utilized BERT-based emotion predictor to examine empathy valence between interlocutors’ emotional state as rewards and optimized rewards to arouse sympathy using Reinforcement Learning.
• Developed a framework whereby several CheerBots are based on understanding users' implied feelings and replying empathetically for multiple dialogues turns.
• Published "CheerBots: Chatbots toward Empathy and Emotion using Reinforcement Learning" in NIPS 2020 workshop.
• Received the Best Paper Award and achieved SOTA performance on the EmpatheticDialogues dataset.

Jun 2020 – Dec 2020 · 6 mos · Taipei City, Taiwan

• Concentrated on a controllable chatbot to guide human beings combining expertise in Natural Language Processing and Meta-reinforcement Learning.
• Proposed a framework to train Guiding Chatbots based on DialoGPT using Reinforcement Learning on Pytorch to induce the interlocutors to reply with intentions, such as long responses, joyful responses, and responses with specific words.
• Established three different chatbots, GoogleBot, fixed-weight DialoGPT chatbot, and BERT-based retrieval chatbot, to simulate human beings’ behaviors.
• Validated the effectiveness of chatbots to improve rewards by an average of 20 times over 3 intentions.
• Published the work “Put Chatbot into its Interlocutor's Shoes: New Framework to learn Chatbot responding with Intention” on NAACL 2021.

Jul 2018 – Sep 2018 · 2 mos · Hsinchu City, Taiwan

• Led 7 software engineers in a project, called Central Control Unit (CCU), to automate software verification pipelines on modem testing by using Android Debug Bridge and Python.
• Developed an internal web tool to visualize modem testing results, resulting in a 50% time savings for testers and analysts during result checks and analysis.
• Saved 60% of human resources (16 people to 5 people) and 20% of the testing time (18hrs to 14hrs).

Jul 2017 – Sep 2017 · 2 mos · Taipei City, Taiwan

• Researched DNA sequence assembly in a group that combines bioinformatics expertise
• Saved 250% of the time to perform the DNA sequences assembly using Canu against using PBcR Algorithm.
• Accelerated the execution time of the read classification in the Sam file by 100% in a larger file (70G) and 20~50% in a smaller file (20G)