Version incompatibility issues are rampant when reusing or reproducing deep learning models and applications. Existing techniques are limited to library dependency specifications declared in PyPI or open-source projects on GitHub. Therefore, they cannot account for latent errors due to undocumented version constraints or dependency to non-Python libraries such as CUDA and cuDNN. Meanwhile, Stack Overflow (SO) offers abundant and up-to-date discussions of version issues spanning across the deep learning stack, e.g., libraries, runtime, OS, and GPU.

We propose to extract the rich knowledge from SO and build a knowledge graph to support the detection of version compatibility issues. Specifically, we develop a novel indirect supervision method that uses a pre-trained Question-Answering (QA) model to extract logic facts from online discussions. Extracted logic facts are further consolidated into a probabilistic knowledge graph to resolve duplicates and conflicts in online discussions. Our evaluation results show that (1) our system can accurately extract version-related facts with 81% precision and 88% recall, and (2) our system can accurately identify 65% while two state-of-the-art approaches can only detect 18% and 6% on a benchmark of 10 popular DL projects.