Publication
Chronological Inference with Latent Gaussian Distributions
Thomas Achim Schmeyer; Remi Oguni; Julia Bayer; Boris Brandherm; Ichiro Kobayashi
In: 2026 Joint 14th International Conference on Soft Computing and Intelligent Systems and 26th International Symposium on Advanced Intelligent Systems (SCIS&ISIS). International Symposium on Soft Computing and Intelligent Systems (SCIS) (SCIS&ISIS-2026), November 2-5, Tokyo, Japan, IEEE, 2027.
Abstract
Compared with languages in which tense information is extensively encoded through verbal morphology, Japanese often relies more heavily on contextual cues for temporal interpretation, posing challenges for automatic temporal relation extraction. Previous work on Japanese text corpora modeled the chronological inference between two events using Gaussian distributions of Allen’s interval relations. These approaches rely on probability-inspired heuristic scores. The proposed approach generalizes these scores by introducing a separable fuzzy continuous distribution to incorporate global temporal knowledge and explicitly addresses the discrete classification problem. This work presents a stable probabilistic approach, describing the chronological event distribution via fuzzy-temporal relations. Since continuous modeling of event distributions does not natively allow for discrete classification, an additional neural decision tree classifier is introduced. It performs a hierarchical defuzzification process to generate a discrete probability distribution while strictly conserving total probability mass. The classification architecture leverages pretrained BERT embeddings combined with fine-tuned MLP heads. The framework is evaluated on a generative Japanese text corpus across five temporal relation classes. Beyond its mathematical foundation, the results demonstrate that a robust 2-parameter formulation outperforms 4-parameter architectures and provides a physically interpretable alternative to standard softmax classifiers. Its explainability opens a wide field for temporal reasoning in natural language processing.