Research

As hate speech on online platforms increasingly harms individuals and communities, counterspeech that can reduce social conflict becomes critically important. To better understand how users on Chinese social media platforms naturally use counterspeech, we manually annotated the CRetor-Human corpus, which consists of 824 gold-standard post–response pairs containing counter-speech. To enable large-scale analysis, we developed a human-in-the-loop annotation pipeline to create the CRetor-Machine corpus, which contains 114,479 machine-annotated items. We then propose a framework to explain the discontinuity between our results and Western studies. We posit that, unlike the American public sphere, the Chinese public sphere evaluates speech through a lens of social belonging, specifically group-innocence versus otherness. The overwhelming cross-platform preference in Chinese data for the “General Denouncement” rhetorical device can be interpreted as a boundary-marking cue that flags a stance as belonging to the out-group and therefore illegitimate. We further show how platform affordances constrain rhetorical choices. We argue that the construction of the public sphere, together with platform-specific values, jointly determines which strategies emerge and which are likely to succeed.

Conversational AI tutors produce rich records of how learners demonstrate their knowledge and skills, but assessing learning attainment from these conversations requires substantial human effort. We evaluate automated annotation of learner attainment across Bloom’s taxonomy using 199 human-AI conversations annotated by five trained humans. We assess prompted LLMs and a fine-tuned encoder on three objectives: attainment ratings, induced conversation rankings, and supporting-evidence selection. Using the Alternative Annotator Test (Alt-Test) and complementary metrics, we find that automated annotators achieve human-comparable agreement on ratings and rankings. Notably, zero-shot Claude Sonnet 5.5 exceeds matched human references in rating agreement (QWK: 0.377 vs. 0.341) and ranking correlation (ρ: 0.572 vs. 0.501), without task-specific training. Evidence selection remains more challenging, highlighting the importance of evaluating annotator substitutability across complementary dimensions. Our findings demonstrate the potential of automated annotators as scalable alternatives for cognitive assessment objectives in human-AI learning.

Understanding a narrative requires readers to connect evidence, infer relationships, and evaluate competing interpretations. Collaborative discussion can support these processes, but agreement alone does not establish understanding. We introduce Timeline Detective, an AI-supported game for narrative comprehension and collaborative reasoning. Teachers prepare and review AI-generated illustrations, which learners sequence and connect to supporting passages in a story. Learners first record independent judgments and explanations, then discuss their interpretations with peers and revise their responses. Structured discussion and group feedback are designed to make disagreements visible while preserving opportunities for independent reflection. Two exploratory playtests with university participants examined how learners used the platform and informed successive design revisions. Participants articulated interpretations, reconsidered textual and visual evidence, and incorporated peers’ observations. However, greater agreement did not consistently coincide with greater accuracy, and interface difficulties sometimes obscured participants’ understanding. These observations identify design challenges in making reasoning visible, communicating feedback, and distinguishing productive deliberation from convergence alone. The playtests establish the feasibility of studying these processes within the platform rather than demonstrating improvements in reading ability. Timeline Detective provides a setting for investigating how visual representations, explanation, and peer interaction shape narrative inference and collective learning.

This paper examines the relationship between executive compensation and firm size, with a particular focus on the mediating role of research and development (R&D) investment. Using panel data from Chinese A-share listed firms on the Shanghai and Shenzhen Stock Exchanges that received government innovation subsidies between 2008 and 2015, we find that higher executive compensation contributes to firm expansion, with approximately 15% of the total effect operating indirectly through increased R&D investment. These findings suggest that, in the context of China’s innovation-driven development policies, executive compensation may provide incentives for managers to increase R&D investment, thereby facilitating firm expansion. The results also highlight the potential role of government innovation subsidies in supporting the development and viability of domestic innovative firms.

Many academic studies on stock markets provide evidence of style rotation and methods of style investing. However, most of the researches are based on U.S. Domestic equity mutual funds. In this paper, we present that the existence of this phenomenon in Chinese stock markets and propose a new index, named stock market characteristics index (SMCI) to investigate style rotation, which is also the fundamental of style investing strategies. For a large sample of data, our backtesting results suggest that, compared to the benchmark, these strategies can achieve superior performance, which is still robust when investing in different stock indexes between the two equity segments. Hence, it is important for delegated portfolio management to consider style rotation when investing in Chinese stock market.