Polysemous words pose a challenge for computational approaches to language change. We extend a recent hypothesis-driven, prototype-based framework to estimate word sense prevalence in diachronic text corpora and apply it to 109,940 usages of schizophrenia drawn from U.S. news media (1985–2025). Our extensions include a contextual dispersion measure (Breadth), robust prototype construction, and human-calibrated prototype-similarity thresholds for conservative sense assignment at scale. Across four decades, distributional semantic change indices commonly used in lexical semantic change detection (LSCD) show significant increases in Breadth and baseline-relative semantic drift (APD), while changes in the central usage prototype (PRT) are influenced by term frequency. In contrast, threshold-calibrated sense assignments reveal stable sense proportions: the psychiatric sense remains dominant, with split-personality and metaphorical senses consistently marginal. Together, these results demonstrate that dispersion- and drift-based LSCD metrics can increase even under stable sense prevalence, indicating that such increases can occur without sense redistribution and primarily reflect broad shifts in usage distributions rather than evidence of polysemization or sense loss. We introduce a threshold-calibrated, prototype-based sense-tracking pipeline that enables conservative sense prevalence estimation at scale and clarifies whether rising distributional LSCD metrics reflect sense redistribution or increasing contextual diversity when historical sense annotation is limited.

Polysemy enables a single word to convey multiple related meanings, reflecting the conceptual and emotional aspects of the evolution of the senses. We introduce the first sense-level benchmark, SenseRel, for modeling semantic relations between word senses, uniting denotational and connotational aspects of meaning. SenseRel distinguishes denotational relations, such as generalization or metaphor, as well as two connotational dimensions: valence and arousal. We evaluate large language models (LLMs), GPT-4o, Llama 3.1, and DeepSeek, in zero-shot and fine-tuned settings. Results show that GPT-4o best aligns with human affective judgments, while a fine-tuned RoBERTa model excels at classifying denotational relations.

Mental health stigma has profoundly harmful impacts but its complexity makes it difficult to evaluate. Stigma may involve explicit derogation, but also subtler forms of blame, fear, paternalistic pity, social distancing, structural exclusion, and discrimination. We introduce a theory-grounded benchmark for automatic evaluation of mental health stigma in online communication, consisting of naturally occurring online news and social media text annotated with a fine-grained taxonomy of stigma across multiple mental health conditions. Our annotation framework comprises a binary stigma-detection task and a multi-level taxonomy covering (i) stigma mode, (ii) domain, and (iii) specific components of certain forms of stigma. We apply this framework to texts mentioning six mental health conditions and evaluate large language models alongside stigma-related classifiers for detecting sentiment, toxicity, and hate speech. Results show that mental health stigma is not well captured by models trained to detect these neighboring constructs, and that LLMs often overpredict stigma unless given explicit operational rules - mirroring the importance of decision rules in human annotation

Lexical Semantic Change (LSC) provides insight into cultural and social dynamics. Yet, the validity of methods for measuring different kinds of LSC remains unestablished due to the absence of historical benchmark datasets. To address this gap, we propose LSC-Eval, a novel three-stage general-purpose evaluation framework to: (1) develop a scalable methodology for generating synthetic datasets that simulate theory-driven LSC using In-Context Learning and a lexical database; (2) use these datasets to evaluate the sensitivity of computational methods to synthetic change; and (3) assess their suitability for detecting change in specific dimensions and domains. We apply LSC-Eval to simulate changes along the Sentiment, Intensity, and Breadth (SIB) dimensions, as defined in the SIBling framework, using examples from psychology. We then evaluate the ability of selected methods to detect these controlled interventions. Our findings validate the use of synthetic benchmarks, demonstrate that tailored methods effectively detect changes along SIB dimensions, and reveal that a state-of-the-art LSC model faces challenges in detecting affective dimensions of LSC. LSC-Eval offers a valuable tool for dimension- and domain-specific benchmarking of LSC methods, with particular relevance to the social sciences.

Historical linguists have identified multiple forms of lexical semantic change. We present a three-dimensional framework for integrating these forms and a unified computational methodology for evaluating them concurrently. The dimensions represent increases or decreases in semantic 1) sentiment (valence of a target word’s collocates), 2) intensity (emotional arousal of collocates or the frequency of intensifiers), and 3) breadth (diversity of contexts in which the target word appears). These dimensions can be complemented by evaluation of shifts in the frequency of the target words and the thematic content of its collocates. This framework enables lexical semantic change to be mapped economically and systematically and has applications in computational social science. We present an illustrative analysis of semantic shifts in mental health and mental illness in two corpora, demonstrating patterns of semantic change that illuminate contemporary concerns about pathologization, stigma, and concept creep.

The present study evaluates semantic shifts in mental health-related concepts in two diachronic corpora spanning 1970-2016, one academic and one general. It evaluates whether their meanings have broadened to encompass less severe phenomena and whether they have become more pathology related. It applies a recently proposed methodology (Baes et al., 2023) to examine whether words collocating with a sample of mental health concepts have become less emotionally intense and develops a new way to examine whether the concepts increasingly co-occur with pathology-related terms. In support of the first hypothesis, mental health-related concepts became associated with less emotionally intense language in the psychology corpus (addiction, anger, stress, worry) and in the general corpus (addiction, grief, stress, worry). In support of the second hypothesis, mental health-related concepts came to be more associated with pathology-related language in psychology (addiction, grief, stress, worry) and in the general corpus (grief, stress). Findings demonstrate that some mental health concepts have become normalized and/or pathologized, a conclusion with important social and cultural implications.