projects | Kaili Huang

School

Engineering Effective In-Context Inputs for GPT-3 in OpenQA

Designed and evaluated novel in-context learning strategies to improve GPT-3’s performance on OpenQA without access to gold passages. Explored lexical, syntactic, and semantic similarity-based example selection methods, and introduced reverse ordering to enhance contextual relevance. The semantic similarity + reverse order strategy achieved the best performance (F1=0.57), yielding a 5% improvement over the random baseline. Findings highlight the impact of example amount, quality, similarity, and ordering on large language model effectiveness.

Optimizing Dialogue History Encoding for Multilingual Task-Oriented Systems

Proposed an efficient training strategy for multilingual virtual assistants by replacing natural language dialogue history with structured dialogue states. Built on the BiToD architecture to reduce reliance on weak natural language encoders and improve slot value extraction. The work investigated the impact of history length (number of previous turns) on model performance, identified diminishing returns, and explored few-shot learning for hard examples. It is designed to improve scalability and robustness in low-resource languages.

Which Trick Works Best in GNN?

In this blog, we reproduced the work of "Residual Network and Embedding Usage - New Tricks of Node Classification with Graph Convolutional Networks." and deliver a thorough introduction and analysis. In detail, we implemented the Graph Attention Network and several commonly used tricks to challenge the node classification task on the ogbn-arxiv dataset. Our experiments show satisfactory results on this dataset. At the same time, we explored the effectiveness of different tricks proposed or introduced in the paper. From our experiments, we found that Node2Vec embeddings, label usage, and C&S achieved noticeable improvements on the task while changing the network structure of GAT and leveraging self-KD barely increases the accuracy.