ABOUT THE TALK

We propose to investigate the feasibility of interpreting different neural models for representing text in a probabilistic manner. A key hurdle for interpretability studies is that the observation and its analysis may be subjective to the investigator. We initiate our investigation by examining Neural Topic Models (NTM), proposing an alternate angle of interpreting its word-topic distribution, producing better topic representations for interpretation. Our method maps the problem of finding these better interpretations to classical NP-hard graph problems, enabling examination of topic distributions in a composite manner. Noting that how we make are observations is critical to interpretability evaluations, our next target of investigation is on how are text represented in the human mental modal. We propose and formulate a large-scale correlation analysis and accompanying user studies to examine automated coherence metrics and human evaluations. Our results show that automated coherence metrics are correlated to human evaluations on Wikipedia English text. We further examine the human responses, anchoring its analysis on word statistics, obtaining some insights applicable to future user studies. Finally, with the rising popularity in Large Language Models (LLM) and in interpreting its mechanisms, we apply our previous findings to mine interpretations from the weights of Transformer-based LLM. It is widely theorized that the superposition of concepts in the neurons of LLMs is a barrier to interpretability. We propose to disentangle these superposed concepts, mapping this problem to a classical NP-hard graph problem, optimizing on automated coherence metric scores.

ABOUT THE SPEAKER

Jia Peng is entering his fourth year of studies under the supervision of Associate Professor Hady Lauw, working on problems in the Natural Language Processing (NLP) domain with an interest in interpretability. His research work has been published in ACL, CL, and EMNLP.