Zijian Jin (Alex)

The first executor Supervised by Prof. Mrinmaya Sachan at ETH.

• An information-theoretic probe is designed to understand how language models encode commonsense knowledge by estimating the mutual information between commonsense knowledge graph and word representation from the language models.

• Using different GNN and Random Walk methods to get graph embeddings and process the knowledge graph in order to get pre-trained representations

• A paper is in preparation for submission to TACL.

The second executor Supervised by Prof. Kewei Tu at ShanghaiTech University.

• Combine symbolic rules (regular expressions, REs) and neural networks by converting REs into bidirectional Recurrent Neural Networks via finite-state transducers, enabling the neural network to generalize in zero-shot or cold-start scenarios.

• On the slot filling task, our model achieves a superior zero-shot and few-shot performance while remains competitive in rich-resource settings.

• A paper was submitted to EMNLP 2021.

The first executor Supervised by Prof. Mrinmaya Sachan at ETH.

• Generating pseudo labels using neural networks transformed from Regular Expressions(REs).

• A self-training framework is developed leveraging the generated pseudo labels for fine-tuning language models with weak supervision.

• Still working on it.

Implemented a demo that aims for monitoring the abnormal oil tube conditions (i.e., fire detection) based on Microsoft Azure using Single Shot Multi-Box Detector (SSD) enabling it to be able to process video at real time.