Ying-Jhe, Tang

演算法工程師

Asus 華碩電腦股份有限公司

四月 2023 - Present
Taipei, Taiwan

• Design Error Correction algorithm for speech recognition 

  • Correct the error word of speech recognition model output

    • Arpa model based Error Correction

    • Traverse the Arpa word graph which is constructed by the customer sentences in beam search manner.

  • Correct the error word in ChatBot conversation

    • Context Entity based Error Correction

    • Extract the entity by the NER model, then correct error word base on the extracted entities

• Text-to-Speech technology application

• Create the API for the Error Correction algorithm as service

• Create the offline version for Speech Recognition System

• Large Language Model application

• Build web crawler and data cleaning to obtain the model training data

研究助理

中央研究院 資訊科學研究所

十一月 2021 - Present
Taipei, Taiwan

• Multi-news abstractive summarization

  • Leverage discourse relationship tree to help the model learn the news discourse and structure in order to generate an accurate and coherent summary 

• News knowledge update task

  • Given an existing article about a related news event, then generate an updated article according to the information from the news event

  • Leverage web crawler to collect original articles, news and updated articles triples from the website and generate the news event triggered knowledge update dataset 

  • Published in CIKM 2022 [1]:

    • Paper: https://dl.acm.org/doi/abs/10.1145/3511808.3557537

    • Github: https://github.com/hhhuang/NetKu

學習型研究助理

KKBOX香港商科科科技股份有限公司台灣分公司

八月 2020 - 六月 2021
Taipei, Taiwan

• Predict the sentiment of Celebrity reviews on websites

  • Solve the problem of low resource labeled data

  • Reduce the noises from the out-of-domain data in transfer learning 

  • Construct the teacher-student model to increase model generalization ability 

  • Published in WI-IAT 2021 [2]:

    • Paper: https://dl.acm.org/doi/abs/10.1145/3486622.3494009

    • Github: https://github.com/isaacaaa/Cross-domain-sentiment-analysis