From Word Representations to Sentiment Analysis – Talk by Dr. Feng Ao

On the afternoon of August 6 at 17:00, a CDRD‑TALK titled “From Word Representations to Sentiment Analysis” was held in the 6th‑floor training room, presented by Dr. Feng Ao.

Dr. Feng, a Ph.D. in Science and associate professor, focuses on artificial intelligence, data mining, and information retrieval. He currently serves as deputy dean of the School of Computer Science at Chengdu University of Information Technology and has previously worked as a software development engineer at Amazon and a senior researcher at Lenovo Chengdu Research Institute.

His academic background includes a bachelor's degree in Automation from Tsinghua University (1999), a master’s in Pattern Recognition and Intelligent Systems (2001), and a Ph.D. in Computer Science from the University of Massachusetts Amherst (2008), where he researched information‑retrieval models and topic detection and tracking.

The talk covered the development history of natural‑language understanding, using sentiment classification examples to illustrate the progression from tokenization, syntactic‑tree construction, and classic classifiers to end‑to‑end neural models that directly ingest raw text. Dr. Feng described the evolution of pre‑training from NNLM to XLNet, highlighting each model’s characteristics.

He also shared his own experiments with convolutional approaches, first applying convolutions over word dimensions and later over embedding rows, presenting experimental results that offer valuable insights for industrial classification tasks.

Finally, Dr. Feng discussed the impact of large‑scale models such as BERT and XLNet, noting that while they have become powerful tools, their resource‑intensive nature makes them less suitable for all industrial scenarios. He emphasized that model choice should balance data size, hardware constraints, performance needs, and application requirements, whether using lightweight methods like SVM or Naïve Bayes or heavyweight pretrained deep networks.