zhiyin121@gmail.com
- Extracting structured knowledge from scientific articles (topic modelling, ontology supplementation, etc.) under the Hybrint project.
- Processing unstructured data (intents & slots annotation, syntax & discourse correction and diversity, ASR annotations) used for dialogue system training and testing, and analysing errors to improve intent and slot recognition accuracy by 10%.
- Significantly increase the speed of annotating new data by developing an annotation tool using Windows Presentation Foundation (WPF). It assembles multiple reference files on a single page with click-to-link and obtains the required tags with clickable use cases replacing manually typing the tags.
- Create the potential for marketing departments to move from purely manual collection of competitors' product and sales information to real-time automated information scraping by leaving a set of data scraping Python code and workflows.
- Prepare teaching materials, organize classroom group discussions, provide feedback.
- Assist in organizing several large-scale academic confluence for greater than 100 participants, and fully responsible for publicity planning (The 10th Cross-Strait Symposium on Modern Chinese Language, ULS15&SDP2, etc.).
- Analyzed social media text data to gauge public opinion, emphasizing the enhancement of granularity in sentiment analysis model classifications.
- Write opinion analysis reports for delivery to clients, participation in competitions, and write brochures "Data Mining".
- Mandarin (Native), Cantonese (Native)
- English (C1), German (A2)
- Master thesis
- Utilized SPARQL to extract artificial product entity data from DBpedia, and associated types and properties from Wikidata.
- Evaluated and enhanced performance using models including CNN, LSTM, BERT, ALBERT, and RoBERTa through Contrastive Learning.
- Take the chu-Liu-Edmonds algorithm as the baseline, Bi-LSTM model and Deep Biaffine Attention as the boosting model. The accuracy was obtained on English and German datasets with UAS 87.1%, LAS 85% and UAS 85.9%, LAS 81.9% respectively.
- Crawl user comments and user self-tagged personality MBTI tags from personality type forums, collecting a total of 22,422 texts.
- Four binary Bi-LSTM models are used for replacing one sixteen classification model to achieve an average accuracy of 20%.
- Using BERT as a pre-trained model, obtained f1 31% for the 16 classification models, f1 70% for the four binary classification models.
- Optimizing RL policies by capture emotion signals from the logs file of rule-based simulator-system conversations in the dialogue system.
- Build eight classification models using Naïve Bayes and Ordered Neuron LSTM with Stance Sentiment Emotion Corpus (SSEC) as training and testing data respectively, obtain an average accuracy of 60%.
- Using Django (HTML, CSS, JS, Python) implement a webpage application allows users to learn the basic vocabulary in Arabic, Chinese and German.
- Collaborate with AI company SONEAN on a real-time platform for using AI technology to improve the efficiency of machine supply chain operations.
- Propose a global intelligent business system solution for building partnership networks and ESG indices for each entity:
- Establish real-time crawling of business news and news for real-time tagging of environmental, social, and government (ESG) signal indices and sentiment indices for each company entity to facilitate the selection of suppliers by buyers.
- Use docker and Neo4j to display company network graphs on the web to show competitors' partnership networks and suppliers' various indices in real time.
- Personal responsibility:
- Write Python code for crawling news and business reports and building a sentiment analysis system.
- Build graphs and write Cypher code for Neo4j queries.
- Create an animated Pitching video.
- A nine-month training program for twelve selected female students in computer science-related disciplines on topics including AI, data science and technology consulting.
- Graphic Design, Painting (>20 years), Presentation (Demo video).
- News Interview (5 years).
