Yuming Liu Email & Phone Number

Kirkland, Wa, Us

Greater Seattle Area

Redmond, Washington, United States

Distilled NLG model for Copiloto End-to-end domain specific NLG model development.o Prompt engineering for data argumentation and automatic metrics. Build prompt engineering based data science solution, utilize GPT-4 to generate and re-sampling the training data; Compose automatic evaluation pipeline, design comprehensive metrics for data quality improvement and model evaluation in large-scale.o NLG model development. Introduce multi-stage supervised fine-tuning and reinforcement learning.o Cross team collaboration to enable task specific models in several products, including Edge, enterprise search and Windows. Graph Enhanced Text Understandingo Graph + text representation comprises of the intent from user behavior and search result, which is critical for the search engine. With the help of the graph, natural language understanding models can get more useful information and improve accuracy.o Graph empowered nature language understanding, proposed TextGNN and AdsGNN to incorporate text w/ user behaviors.o Cooperation with engineering team to implement offline GNN training and online serving platform.

Text understanding for Ads relevanceo The Ads Relevance Model is for Ad quality control. It measures how much an advertisement would meet a user’s search query in several aspects. Ads quality improvement via semantic similarity score and embeddings.o Ads specific model development on SOTA transformer-based model, which contains multi-task fine-tune, single task fine-tune (adaption), feature-enrichment and knowledge distillation.o Oversaw the data refresh pipeline, which publishes the latest data to online environments to make ads relevance models work as expected and monitor data correctness.Embedding based retrieval modelo Twin-tower transformer based retrieval model. Comprehensive optimization for model structure, including vocab, transformer layer and pooling layer. o Two-stage optimization. Adopted contrastive learning and mixture-of-teacher for pre-training stage, and mining fine-grained positive and hard negative samples for second-stage.

Keyword Opportunitieso One of the key features for Microsoft advertiser services. Advertisers get high quality bid-keyword, which helps to reach more potential users and expand their business, thereby boosting Microsoft revenue advertising with the increase of bid-keyword.o In charge of the end-to-end backend data pipelines for this product. First, I mined candidate keywords via log analysis, semantic models, and competitors before leveraging models and concluded rules for advertiser-aware quality control to filter out irrelevant candidates. Lastly, I designed the KPI estimation formula to predict impressions and clicks for each of the keyword suggestions.

Master'S Degree, Ml - Nature Language Processing

Bachelor'S Degree, Information Engineering