Word Archipelagoes for Explaining Contextual Shifts in Sequential Actions and Events

Text, as data which can represent actors’ behaviors with the reflection of emotional dimension, is expected to be used combined with physical mobility data for explaining contextual continuity and transitions. Even in the era of LLM, we still desire methods for the preparation of essential parts of texts for finely obtaining the contextual shift which may go from and to interleaving meaningful key-terms rather than sheer change of topics corresponding to word clusters. In this study, the explanatory relay of words in text is represented in the form of archipelagoes. Here, each archipelago is a sequence of islands composed of the occurrences of a certain word. An island here is interpreted as the local sequence where the word is emphasized, and an archipelago of a length comparable to the target text is extracted by using the variation of entropy type-A, the window-based entropy based on the distribution of the word’s occurrences with the width of each time window. The results show that the parts of the target text including the words forming archipelagoes thus extracted, without pre-learned knowledge, form an explanatory part of the text that is of smaller entropy B than the parts extracted by the baseline methods. Here, entropy type-B is the graph-based entropy representing the contextual dispersion of the text. This result means archipelagoes work, for extracting contextual flow with discarding noisy words in the text without a prepared knowledge about stop words in text or noises in sequential data.