Projects

Causality in NLP Modeling [Slides]

• We present papers about how causal learning can be applied in different machine learning areas (E.g., meta-learning and representation learning) while I focus on modeling various NLP tasks from a causality perspective.

Comparison of Transformer Models from Topological Perspective [Paper]

• We extract mention representations from the whole sentence encoding outputted by each transformer model (E.g. BERT, GPT2) and build mapper graphs as the model’s topological summary; analyzed the mapper graph and determine the consistency of graph similarity between two models, corresponding to their architectural similarity.

Diff-Transferer: Any-Speaker Adaptive Text-to-Speech with Diffusion [Paper]

• We proposed Diff-Transferer, an any-speaker text-to-speech model with a shallow diffusion mechanism; stabilized the traning process of the diffusion model and reduced the number of training steps needed to reach convergence.

NLP Research Hotspots Analysis with Structural and Unstructured Data [Slides]

• We crawled information of NLP papers in recent three years from semantic scholar API; extracted keywords from the abstract of each paper and conducted unsupervised topic clustering; analyzed the authors’ collaboration relationship with Neo4j; analyzed the combination of different topics with co-occurrence heatmap

Mental Support Conversational Recommendation Chatbot

• We sampled the triplets in C3KG to simulate users’ emotional cause chains; use sampled data as the training data of the conversational recommendation system and built a progressive mental support conversational recommendation chatbot to reason user’s emotional causes and provide suggestions. Then we reproduced a production version code of the conversational recommendation model with TensorFlow; optimized and deployed the model in the cloud server; created an API to call the multi-turn model. We also built a demo to interact with the backend model; built a dialog component with Vue.js, including message boxes, input boxes, and interactive buttons; created a visualization box based on D3.js to interact with the model dynamically.

Automated Essay Scoring (AES) system [Github]

AES is a composition automatic scoring system to get feedback to users with wording recommendation and data visualization. We built it Flask and Vue.