Timezone: »

exBERT: A Visual Analysis Tool to Explain BERT's Learned Representations
Benjamin Hoover · Hendrik Strobelt · Sebastian Gehrmann

Tue Dec 10 05:30 PM -- 07:30 PM (PST) @ East Exhibition Hall B + C #801

Large language models can produce powerful contextual representations that lead to improvements across many NLP tasks. Though these models can comprise undesired inductive biases, it is challenging to identify what information they encode in their learned representations. Since the model-internal reasoning process is often guided by a sequence of learned self-attention mechanisms, it is paramount to be able to explore what the attention has learned. While static analyses for this can lead to targeted insights, interactive tools can be more dynamic and help humans gain an intuition for the model-internal reasoning process. We present exBERT, a tool that helps gather insights into the meaning of contextual representations. exBERT matches a human-specified input to similar contexts in a large annotated dataset. By aggregating these annotations across all similar contexts, exBERT can help to explain what each attention-head has learned.

Author Information

Benjamin Hoover (IBM)
Hendrik Strobelt (IBM Research)
Hendrik Strobelt

Hendrik Strobelt is Senior Research Scientist at IBM Research (Cambridge, MA) and the Explainability Lead at the MIT-IBM Watson AI Lab. His recent research is on visualization for and human collaboration with AI models to foster explainability and intuition. His work involves NLP models and generative models while he is advocating to utilize a mix of data modalities to solve real-world problems. His research is applied to tasks in machine learning, in NLP, in the biomedical domain, and in chemistry. Hendrik joined IBM in 2017 after postdoctoral positions at Harvard SEAS and NYU Tandon. He received a Ph.D. (Dr. rer. nat.) from the University of Konstanz in computer science (Visualization) and holds an MSc (Diplom) in computer science from TU Dresden. His work has been published at venues like IEEE VIS, ICLR, ACM Siggraph, ACL, NeurIPS, ICCV, PNAS, Nature BME, or Science Advances. He received multiple best paper/honorable mention awards at EuroVis, BioVis, VAST, CHI, ACL Demo, or NeurIPS demo. He received the Lohrmann medal from TU Dresden as the highest student honor. Hendrik has served in program committees and organization committees for IEEE VIS, BioVis, EuroVis. He served on organization committees for IEEE VIS, VISxAI, ICLR, ICML, NeurIPS. Hendrik is visiting researcher at MIT CSAIL.

Sebastian Gehrmann (Harvard)

More from the Same Authors