Future prediction with automatically extracted morphosemantic patterns. (January 2020)

• Record Type: Journal Article
• Authors: Nakajima, Yoko; Ptaszynski, Michal; Masui, Fumito; Honma, Hirotoshi
• Is Part Of: Cognitive systems research. Volume 59(2020)
• Page Start: 37
• Page End: 62
• Publication Date: 2020-01
• Subjects: Future prediction -- Future reference sentences -- Morphosemantic patterns -- Pattern extraction -- Text mining; Cognition -- Periodicals; Cognitive engineering (System design) -- Periodicals; Artificial intelligence -- Periodicals; 153.05
• DOI: 10.1016/j.cogsys.2019.09.004
• Languages: English
• ISSNs: 1389-0417

Abstract

Abstract: In the following paper, we investigated the usefulness of future reference sentence patterns in the prediction of the unfolding of future events. To obtain such patterns we first collected sentences that have any reference to the future from newspapers and Web news. Based on this collection, we developed a novel method for automatic extraction of frequent patterns from such sentences. The extracted patterns, consisting of multilayer semantic information and morphological information, were implemented in the formation of a general model of linguistically expressed future. To fully assess the performance of the proposed method we performed a number of evaluation experiments. In the first experiment, we evaluated the automatic extraction of future reference sentence patterns with the proposed extraction algorithm. In the second set of experiments, we estimated the effectiveness of those patterns and applied them to automatically classify sentences into future referring and other. The final model was then tested for performance in retrieving a new set of future reference sentences from a large news corpus. The obtained results confirmed that the proposed method outperformed state-of-the-art method in fully automatic retrieval of future reference sentences. Lastly, we applied the method in practice to confirm its usefulness in two tasks. The first is to support human readers in the everyday prediction of unfolding future events. In the second task, we developed a fully automatic prototype method for future prediction and tested its performance using the tasks included in the official Future Prediction Competence Test. The results indicate that the prototype system outperforms natural human foreseeing capability.