Speaker:Jie Yuan

Tensor-Product-Representation-Based Topic Modeling

Most statistical and neural topic models seek to discover one level/grain of latent topic concepts from textual data; however, they ignore to leverage distinct levels/grains of latent topic concepts to augment the semantic information of word or document representations. To bridge the gap, we propose a novel topic model based on Tensor Product Representations, called TPR-TM, which employs TPR ‘binding’ of high- and low-level latent topic concepts to encode words in vector space, and the decoder reconstructs the input document using the averaged TPRs of all words in the document. Comprehensive quantitative and qualitative evaluations illustrate the superior performance of TPR-TM compared to state-of-the-art models. Through a unique case study on forecasting market volatility, we further demonstrate the power of the TPR-TM in downstream tasks.

Committee: Zhu Zhang (major professor), Forrest Sheng Bao, Qi Li, Wallapak Tavanapong, and Jin Tian

Join on WebEx: https://iastate.webex.com/meet/jieyuan