Researchers have developed an algorithm called MIL-BERT that can classify large text documents by selecting relevant excerpts. The approach is inspired by multiple instance learning and has been shown to be scalable, even with texts containing nearly a million tokens. The team evaluated their method on seven datasets, achieving state-of-the-art results in three areas: identifying political bias in news outlets, trigger warnings in long stories, and demographic characteristics of authors in tweet collections. MIL-BERT can also generalize from weakly-labeled text collections to accurately classify smaller instances. --- Why it matters: This matters because it provides a new approach for classifying large texts, which is essential for applications such as information retrieval, sentiment analysis, and topic modeling. The scalability and performance guarantees of MIL-BERT make it a valuable tool for researchers and practitioners working with large text datasets. Source: https://arxiv.org/abs/2608.20636