Similar Embedding Research Articles

GraphRec-based Korean expert recommendation using author contribution index and the paper abstracts in marine

Expert recommendation systems recommend specialized experts in a particular field to users based on the knowledge of those experts. However, these systems are limited by the number of experts available and the potential for subjective evaluation, which may result in inappropriate recommendations. Furthermore, we explore the evolution from traditional to deep learning-based recommendation systems, emphasizing graph-based recommendation systems. Nonetheless, deep learning-based systems require large amounts of data, and marine expert recommendation training data are scarce. To address these issues, we constructed and utilized marine expert data in this study. The dataset contains abstracts of marine-related papers and information on their authors. Graphs were generated by assessing the similarity among the abstracts, representing them in a graph format indicative of this similarity, and using the author contribution index to depict the relationship between the abstracts and their respective authors. Various similarity methods and abstract embedding techniques were experimentally explored to realize performance optimization. In the experiments, the optimized model achieved a mean absolute error of 0.7556 and a root-mean-squared error of 1.0421. Notably, this study highlights the limitations of traditional evaluation metrics and proposes the averaged mean reciprocal rank as a suitable alternative. This metric facilitates the quantitative evaluation of model performance on newly created data, obviating a comparison model. Finally, applying the newly constructed data to the GraphRec model by using their graphical representation significantly improves the system performance.

Continuous Training of Recommendation System for Airbnb Listings Using Graph Learning

Recommender systems are getting increasingly important nowadays as they can boost user engagement and benefit businesses. However, there remain some unsolved problems. This paper will address two key performance issues. First, the limited ability to identify and leverage intrinsic relationships between data points. Second, the inability to adapt to new data. The first issue is proposed to be addressed through a Graph Neural Network (GNN) to curate better recommendations. GNN will be trained with Airbnb’s review data to utilize its outstanding expressive power to represent complex user-listing interactions at scale, followed by generating embeddings to compute the relevant recommendations to the users. With the generated embeddings, the recommender system will compute a recommendation list to every user based on the embedding similarity between the user and listings or the user’s first-ever reviewed listing and listings. The second issue is proposed to be resolved by incorporating Continuous Training. The proposed recommender system employs GraphSAGE with a customized Rating-Weighted Triplet Ranking Loss function, which outperformed unsupervised GraphSAGE. Offline simulation validated the recommender system's ability to learn from the latest data and improve over time. Overall, the proposed user-to-item (U2I) recommendation rating-weighted GraphSAGE substantially increased by 99.88% in hit-rate@5 and 98.15% in coverage. This offers an effective solution for enhancing the recommender system for Airbnb listings. This research validates the efficacy of GNN-based recommendations in capturing user-item relationships to aid in predicting relevant recommendations, thus significantly driving up the adoption of GNN-based recommender systems.

Bridging the Semantic Latent Space between Brain and Machine: Similarity Is All You Need

How our brain encodes complex concepts has been a longstanding mystery in neuroscience. The answer to this problem can lead to new understandings about how the brain retrieves information in large-scale data with high efficiency and robustness. Neuroscience studies suggest the brain represents concepts in a locality-sensitive hashing (LSH) strategy, i.e., similar concepts will be represented by similar responses. This finding has inspired the design of similarity-based algorithms, especially in contrastive learning. Here, we hypothesize that the brain and large neural network models, both using similarity-based learning rules, could contain a similar semantic embedding space. To verify that, this paper proposes a functional Magnetic Resonance Imaging (fMRI) semantic learning network named BrainSem, aimed at seeking a joint semantic latent space that bridges the brain and a Contrastive Language-Image Pre-training (CLIP) model. Given that our perception is inherently cross-modal, we introduce a fuzzy (one-to-many) matching loss function to encourage the models to extract high-level semantic components from neural signals. Our results claimed that using only a small set of fMRI recordings for semantic space alignment, we could obtain shared embedding valid for unseen categories out of the training set, which provided potential evidence for the semantic representation similarity between the brain and large neural networks. In a zero-shot classification task, our BrainSem achieves an 11.6% improvement over the state-of-the-art.

A Novel Entity and Relation Joint Interaction Learning Approach for Entity Alignment

Entity alignment (EA) aims to find equivalent entities in knowledge graphs (KGs) from multiple data sources and is a crucial step in integrating KGs. Recent studies learn the similarity of entity embeddings by aggregating neighboring entities. However, these methods solely compare neighboring entities and do not incorporate the connected relation between an entity and its neighbors. In this paper, we propose a novel Entity and Relation joint Interaction Learning (ERIL) approach, which effectively captures the interaction between entities and relations, enhancing the precision of alignment across different KGs. Specifically, the ERIL model jointly learns the neighborhood features of entities and the spatial structure of relations to train a shared permutation matrix, capturing comprehensive associative relations within KGs. Moreover, a semi-supervised iterative framework is designed to leverage the positive interactions between entities and relations to identify more aligned entities. Extensive experiments are conducted on five benchmark datasets to demonstrate the effectiveness of ERIL compared with existing state-of-the-art EA methods. On DBP15K, our model ERIL outperforms currently available EA methods by 1.9% on Hits@10.

Meta-path-based key node identification in heterogeneous networks

Identifying key nodes in complex networks remains challenging. Whereas previous studies focused on homogeneous networks, real-world systems comprise multiple node and edge types. We propose a meta-path-based key node identification (MKNI) method in heterogeneous networks to better capture complex interconnectivity. Considering that existing studies ignore the differences in propagation probabilities between nodes, MKNI leverages meta-paths to extract semantics and perform node embeddings. Trust probabilities reflecting propagation likelihoods are derived by calculating embedding similarities. Node importance is calculated by using metrics incorporating direct and indirect influence based on trust. The experimental results on three real-world network datasets, DBLP, ACM and Yelp, show that the key nodes identified by MKNI exhibit better information propagation in the Susceptible Infected (SI) and susceptibility-influence model (SIR) model compared to other methods. The proposed method provides a reliable tool for revealing the topological structure and functional mechanisms of the network, which can guide more effective regulation and utilization of the network.

Topic modeling for conversations for mental health helplines with utterance embedding

Conversations with topics that are locally contextual often produces incoherent topic modeling results using standard methods. Splitting a conversation into its individual utterances makes it possible to avoid this problem. However, with increased data sparsity, different methods need to be considered. Baseline bag-of-word topic modeling methods for regular and short-text, as well as topic modeling methods using transformer-based sentence embeddings were implemented. These models were evaluated on topic coherence and word embedding similarity. Each method was trained using single utterances, segments of the conversation, and on the full conversation. The results showed that utterance-level and segment-level data combined with sentence embedding methods performs better compared to other non-sentence embedding methods or conversation-level data. Among the sentence embedding methods, clustering using HDBScan showed the best performance. We suspect that ignoring noisy utterances is the reason for better topic coherence and a relatively large improvement in topic word similarity.

Unsupervised graph anomaly detection with discriminative embedding similarity for viscoelastic sandwich cylindrical structures

In the detection of slipping anomalies in viscoelastic sandwich cylindrical structures (VSCS), conventional methods may encounter challenges due to the extremely rare and weak nature of slipping signals. This study focuses on normal signals and introduces unsupervised graph representation learning (UGRL) with discriminative embedding similarity for VSCS's detection. UGRL involves data preprocessing, model embedding, and matrix reconstructing. Association graphs are constructed based on sample similarities for yielding adjacency and attribute matrices. Subsequently, the matrices undergo embedding and reconstruction via various network modules to enhance graph data characterization. Detection indicators are derived by calculating embedding similarities and reconstruction errors, and thresholds are constructed using these indicators to enable efficient anomaly detection. The experiments in VSCS slipping dataset effectively indicate the superiority of the proposed method.

Deep Correspondence Matching-Based Robust Point Cloud Registration of Profiled Parts

Due to ability to estimate the spatial transformation of coordinate frames, point cloud registration is a fundamental technique in manufacturing. Previous methods prone to converge to wrong local minima, in the cases of large initialization, noise, outliers, and partiality. This study presents a new learning-based robust point cloud registration approach to predict a rigid transformation in a one-shot way. Our network aims to determine a matchability matrix to yield an accurate registration result. Each element of the matchability matrix refers to similarity of learned per-point embeddings and represents the probability of a potential correspondence. Two major blocks are developed to guide the matchability matrix to represent correct correspondences: 1) an attention block is introduced to enhance the discriminativeness of learned per-point embeddings; 2) a zero-mean Gaussian based annealing layer and a differentiable Sinkhorn normalization layer are designed to enforce a permutation matchability matrix. With the matchability matrix, an intuitive solution is integrated to obtain the relative transformation of the source and target point clouds. Different from existing work, our network can handle partially overlapped point-cloud pairs effectively. Experimental results demonstrate the superiority of the proposed approach over the state-of-the-art registration approaches in terms of accuracy and robustness.

Supervised Hierarchical Online Hashing for Cross-modal Retrieval

Online cross-modal hashing has gained attention for its adaptability in processing streaming data. However, existing methods only define the hard similarity between data using labels. This results in poor retrieval performance, as they fail to exploit the semantic structure information of labels and miss the high-quality hash codes guided by the hierarchical relevance between labels. In addition, they ignore the bit-flipping problem, which leads to sub-optimal cross-modal retrieval performance. To address these issues, we propose Supervised Hierarchical Online Hashing (SHOH) for cross-modal retrieval. Our approach acquires hierarchical similarity via cross-layer affiliation of labels and explores its application to online hashing. We design a hierarchical similarity learning method in the online learning framework, which includes virtual center learning and hierarchical similarity embedding. Labels with soft similarity bridge the label hierarchy and cross-modal hash embedding. Furthermore, we propose a Weighted Retrieval Strategy (WRS) to mitigate the impact caused by bit-flipping errors. Extensive experiments and verification on hierarchical and non-hierarchical datasets demonstrate that SHOH preserves accurate inter-class distances and achieves performance improvements compared to state-of-the-art methods. The source code is available at https://github.com/HUST-IDSM-AI/SHOH .

Adaptive instance similarity embedding for online continual learning

We study the online continual learning (CL) paradigm, where the learner must continually learn a sequence of tasks. In this setting, improving the learning ability of the model and mitigating catastrophic forgetting are two pivotal factors. Note that most existing approaches for online continual learning are based on the experience replay strategy. In this type of method, a memory buffer is applied to store a subset of previous tasks to prevent catastrophic forgetting. The samples from between the current task and the memory buffer are jointly trained to update the network parameters. Consider that most methods only generate the feature embeddings via a shared feature extractor and then train the network via cross-entropy loss. We argue that such methods fail to explore the feature embedding in its entirety and neglect the similar relations between samples, thus leading to lower discriminant performance, especially in an online learning setting. To this end, we propose the Adaptive Instance Similarity Embedding for Online Continual Learning (AISEOCL) framework, which further takes all the sample relations in a given batch into account. In detail, firstly, the experience replay strategy is used to avoid catastrophic forgetting. Then, during training, we apply the adaptive similar embedding to obtain additional valuable similar information from the current training samples composed of the current and previous tasks. Since not all samples are equally important to make a prediction, we further weigh the importance of each instance accordingly resorting to the attention mechanism. Importantly, we further impose a similarity distillation loss on the distributions of the similarity relationship between current and previous models. Such operation can transfer the similarity relationship between different samples from the old model to the current model to alleviate catastrophic forgetting. With this strategy, AISEOCL can further improve the learning ability of the model while enhancing the discriminant power, which is also beneficial to stably resist forgetting. The experiments on several existing benchmarks validate the effectiveness of our proposed approach.

MOASL: Predicting drug mechanism of actions through similarity learning with transcriptomic signature

Understanding the mechanisms of actions (MOAs) of compounds is crucial in drug discovery. A common step in drug MOAs annotation is to query the dysregulated gene signatures induced by drugs in a reference library of pre-defined signatures. However, traditional similarity-based computational strategies face challenges when dealing with high-dimensional and noisy transcriptional signature data. To address this issue, we introduce MOASL (MOAs prediction via Similarity Learning), a novel approach that contrastive to learn similarity embeddings among signatures with shared MOAs automatically. We evaluated the accuracy of signature matching on various transcriptional activity score (TAS) datasets and individual cell lines by using MOASL. The results show MOASL achieved higher performance over several statistical and machine learning methods. Furthermore, we provided the rationale of our model by visualizing the signature annotation procedure. Using MOASL, the MOAs label of query signature could be conveniently defined by calculating the similarity between the query embedding and the reference embeddings. Finally, we applied MOASL to repurpose thousands of compounds as glucocorticoid receptor (GR) agonists, accurately identifying 8 out of the top 10 compounds. MOASL is conveniently accessible on GitHub at https://github.com/jianglikun/MOASL, empowering researchers and practitioners in the field of drug discovery to predict the MOAs of drug.

Obtaining clinical term embeddings from SNOMED CT ontology

Clinical term embeddings are traditionally obtained using corpus-based methods, however, these methods cannot incorporate knowledge about clinical terms which is already present in medical ontologies. On the other hand, graph-based methods can obtain embeddings of clinical concepts from ontologies, but they cannot obtain embeddings for clinical terms and words. In this paper, a novel method is presented to obtain embeddings for clinical terms and words from the SNOMED CT ontology. The method first obtains embeddings of clinical concepts from SNOMED CT using a graph-based method. Next, these concept embeddings are used as targets to train a deep learning model to map clinical terms to concepts embeddings. The learned model then provides embeddings for clinical terms and words as well as maps novel clinical terms to their embeddings. The embeddings obtained using the method out-performed corpus-based embeddings on the task of predicting clinical term similarity on five benchmark datasets. On the clinical term normalization task, using these embeddings simply as a means of computing similarity between clinical terms obtained accuracy which was competitive to methods trained specifically for this task. Both corpus-based and ontology-based embeddings have a limitation that they tend to learn similar embeddings for opposite or analogous terms. To counter this, we also introduce a method to automatically learn patterns that indicate when two clinical terms represent the same concept and when they represent different concepts. Supplementing the normalization process with these patterns showed improvement. Although clinical term embeddings obtained from SNOMED CT incorporate ontological knowledge which is missed by corpus-based embeddings, they do not incorporate linguistic knowledge which is needed for sentence-based tasks. Hence combining ontology-based embeddings with corpus-based embeddings is an avenue for future work.

Optimising speech‐testing to predict prodromal Alzheimer’s disease: head‐to‐head comparison study of tasks and analysis methods

AbstractBackgroundChanges in speech occur in early‐stage Alzheimer’s disease. A range of approaches have been used for eliciting and automatically analysing speech, but limited research has directly compared these methods.MethodParticipants from the AMYPRED‐UK (NCT04828122) and AMYPRED‐US (NCT04928976) studies completed four speech‐elicitation tasks: the Automatic Story Recall Task (ASRT), the Logical Memory Test (LMT), Semantic (Animals, Vegetables, Fruit) and Phonemic Verbal Fluency (F, A, S) Tasks. Responses were recorded and automatically transcribed. Analyses were completed with Novoic’s speech analysis software, using four key approaches: (1) feature extraction, (2) representations from large language models (LLM), (3) text‐similarity evaluation, and (4) autoscoring. Together, these evaluated audio, linguistic, and temporal speech domains. Outputs were entered into logistic regression models predicting Mild Cognitive Impairment (MCI)/mild AD and Amyloid positivity in MCI/mild AD. Area under the ROC curves (AUC) were evaluated via 5‐fold cross‐validation.Result165 older adults (including N = 74 MCI and 9 mild AD; of which N = 38 MCI/mild AD and amyloid beta positive) provided data for all tasks. Strength of model predictions varied by task, analytic approach and domain (Fig 1). AUCs predicting MCI/mild AD were generally higher for ASRTs compared to other tasks (Fig 1A). Text‐similarity approaches (G‐match: similarity of word embeddings between the source text and retelling, and V‐match: string‐based matching), produced the highest AUC for ASRTs (G‐match AUC = 0.87‐0.88), and the LMT (V‐match AUC = 0.83‐85). Autoscoring tasks produced the strongest predictor for the semantic fluency (AUC = 0.85), and also one of the strongest for phonemic fluency (AUC = 0.72). Changes in linguistic and temporal characteristics were more sensitive to MCI/mild AD, compared to audio‐only metrics. Amyloid in MCI/mild AD predictions were overall less strong (Fig 1B), with the most consistent pattern being better performance of text‐similarity approaches on story recall tasks (up to AUC = 0.77 for ASRTs, 0.73 for LMT).ConclusionSelection of task and analytic approach is important for developing more sensitive speech‐based testing. The best performing tasks (ASRT, Semantic Fluency) and metrics (text‐similarity metrics and autoscoring) have been incorporated in Novoic’s Storyteller automated remote speech‐based test.

Optimising speech‐testing to predict prodromal Alzheimer’s disease: head‐to‐head comparison study of tasks and analysis methods

AbstractBackgroundChanges in speech occur in early‐stage Alzheimer’s disease. A range of approaches have been used for eliciting and automatically analysing speech, but limited research has directly compared these methods.MethodParticipants from the AMYPRED‐UK (NCT04828122) and AMYPRED‐US (NCT04928976) studies completed four speech‐elicitation tasks: the Automatic Story Recall Task (ASRT), the Logical Memory Test (LMT), Semantic (Animals, Vegetables, Fruit) and Phonemic Verbal Fluency (F, A, S) Tasks. Responses were recorded and automatically transcribed. Analyses were completed with Novoic’s speech analysis software, using four key approaches: (1) feature extraction, (2) representations from large language models (LLM), (3) text‐similarity evaluation, and (4) autoscoring. Together, these evaluated audio, linguistic, and temporal speech domains. Outputs were entered into logistic regression models predicting Mild Cognitive Impairment (MCI)/mild AD and Amyloid positivity in MCI/mild AD. Area under the ROC curves (AUC) were evaluated via 5‐fold cross‐validation.Result165 older adults (including N = 74 MCI and 9 mild AD; of which N = 38 MCI/mild AD and amyloid beta positive) provided data for all tasks. Strength of model predictions varied by task, analytic approach and domain (Fig 1). AUCs predicting MCI/mild AD were generally higher for ASRTs compared to other tasks (Fig 1A). Text‐similarity approaches (G‐match: similarity of word embeddings between the source text and retelling, and V‐match: string‐based matching), produced the highest AUC for ASRTs (G‐match AUC = 0.87‐0.88), and the LMT (V‐match AUC = 0.83‐85). Autoscoring tasks produced the strongest predictor for the semantic fluency (AUC = 0.85), and also one of the strongest for phonemic fluency (AUC = 0.72). Changes in linguistic and temporal characteristics were more sensitive to MCI/mild AD, compared to audio‐only metrics. Amyloid in MCI/mild AD predictions were overall less strong (Fig 1B), with the most consistent pattern being better performance of text‐similarity approaches on story recall tasks (up to AUC = 0.77 for ASRTs, 0.73 for LMT).ConclusionSelection of task and analytic approach is important for developing more sensitive speech‐based testing. The best performing tasks (ASRT, Semantic Fluency) and metrics (text‐similarity metrics and autoscoring) have been incorporated in Novoic’s Storyteller automated remote speech‐based test.

On the Power of Gradual Network Alignment Using Dual-Perception Similarities.

Network alignment (NA) is the task of finding the correspondence of nodes between two networks based on the network structure and node attributes. Our study is motivated by the fact that, since most of existing NA methods have attempted to discover all node pairs at once, they do not harness information enriched through interim discovery of node correspondences to more accurately find the next correspondences during the node matching. To tackle this challenge, we propose Grad-Align, a new NA method that gradually discovers node pairs by making full use of node pairs exhibiting strong consistency, which are easy to be discovered in the early stage of gradual matching. Specifically, Grad-Align first generates node embeddings of the two networks based on graph neural networks along with our layer-wise reconstruction loss, a loss built upon capturing the first-order and higher-order neighborhood structures. Then, nodes are gradually aligned by computing dual-perception similarity measures including the multi-layer embedding similarity as well as the Tversky similarity, an asymmetric set similarity using the Tversky index applicable to networks with different scales. Additionally, we incorporate an edge augmentation module into Grad-Align to reinforce the structural consistency. Through comprehensive experiments using real-world and synthetic datasets, we empirically demonstrate that Grad-Align consistently outperforms state-of-the-art NA methods.

Improving Deforestation Detection Accuracy in Noisy Satellite Images with Contrastive Learning-based Approach

Deforestation, the large-scale destruction of trees, has far-reaching biological and environmental consequences that pose a significant threat to the environment. Accurate deforestation detection is crucial for successful conservation initiatives and effective land management. Over the last decade, numerous deforestation detection methods utilizing spaceborne photography have been proposed. However, these methods tend to be sensitive to unique image noise in the satellite domain by virtue of the diverse aerial characteristics and air qualities in different regions. To solve this problem, we propose a novel noise-robust deforestation detection framework with a contrastive-learning based approach. The proposed framework consists of two phases: contrastive learning, which aims to extract similar feature embeddings for the same category, proceeded with transfer learning in order to develop the deforestation classifier. Remarkably, the proposed contrastive learning approach successfully handles noisy input satellite images during the feature extraction process. Upon conducting validation, we have found that the proposed method outperforms existing deforestation detection methods by a significant performance gap, highlighting the effectiveness of the proposed contrastive learning approach.

Standard Model in conformal geometry: Local vs gauged scale invariance

We discuss comparatively local versus gauged Weyl symmetry beyond Standard Model (SM) and Einstein gravity and their geometric interpretation. The SM and Einstein gravity admit a natural embedding in Weyl integrable geometry which is a special limit of Weyl conformal (non-metric) geometry. The theory has a local Weyl scale symmetry but no associated gauge boson. Unlike previous models with such symmetry, this embedding is truly minimal i.e. with no additional fields beyond SM and underlying geometry. This theory is compared to a similar minimal embedding of SM and Einstein gravity in Weyl conformal geometry (SMW) which has a full gauged scale invariance, with an associated Weyl gauge boson. At large field values, both theories give realistic, Starobinsky–Higgs like inflation. The broken phase of the current model is the decoupling limit of the massive Weyl gauge boson of the broken phase of SMW, while the local scale symmetry of the current model is part of the larger gauged scale symmetry of SMW. Hence, the current theory has a gauge embedding in SMW. Unlike in the SMW, we note that in models with local scale symmetry the associated current is trivial, which is a concern for the physical meaning of this symmetry. Therefore, the SMW is a more fundamental UV completion of SM in a full gauge theory of scale invariance that generates Einstein gravity in the (spontaneously) broken phase, as an effective theory.

CL3: Generalization of Contrastive Loss for Lifelong Learning.

Lifelong learning portrays learning gradually in nonstationary environments and emulates the process of human learning, which is efficient, robust, and able to learn new concepts incrementally from sequential experience. To equip neural networks with such a capability, one needs to overcome the problem of catastrophic forgetting, the phenomenon of forgetting past knowledge while learning new concepts. In this work, we propose a novel knowledge distillation algorithm that makes use of contrastive learning to help a neural network to preserve its past knowledge while learning from a series of tasks. Our proposed generalized form of contrastive distillation strategy tackles catastrophic forgetting of old knowledge, and minimizes semantic drift by maintaining a similar embedding space, as well as ensures compactness in feature distribution to accommodate novel tasks in a current model. Our comprehensive study shows that our method achieves improved performances in the challenging class-incremental, task-incremental, and domain-incremental learning for supervised scenarios.

Predicting the transmission trend of respiratory viruses in new regions via geospatial similarity learning

The outbreak and spread of COVID-19 remind us again of the devastating attack that human-to-human transmitted respiratory infectious diseases (H-HRIDs) bring to global economics and public health. Predicting the trend of H-HRIDs yields important suggestions for making response strategies. Existing methods predict the future trend of H-HRIDs based on the prophase epidemiological data. However, it is also crucial to predict the potential outbreak trend of H-HRIDs in regions without prophase epidemiological data based on those with epidemiological data available. This prediction can provide constructive pharmaceutical strategies and non-pharmaceutical interventions such as pre-allocating the limited preventions to regions with different potential trends, especially at the early stage of an H-HRID. In this paper, transform the problem of predicting the early-stage trend of an H-HRID in new regions without prophase epidemiological data into the problem of learning regions’ geospatial features via a Contrastive learning-based Hierarchical Graph Convolutional Neural Network (CHGCN). Specifically, CHGCN first generates training data from regions with epidemiological data available via a Relaxed Graph Edit Distance (RGED) algorithm. CHGCN then uses a constructive learning schema to learn region embeddings, where regions with similar geospatial features and H-HRID trends exhibit similar embeddings. Once trained, CHGCN searches for the region with epidemiological data available that has the highest embedding similarity to the given query region without prophase epidemiological data, so as to achieve early-stage H-HRID trend prediction for the query region. Experimental results on COVID-19 demonstrate that the proposed CHGCN achieves state-of-the-art performance in predicting the early-stage trend of H-HRIDs, compared with the baselines.

Identifying and predicting stereotype change in large language corpora: 72 groups, 115 years (1900-2015), and four text sources.

The social world is carved into a complex variety of groups each associated with unique stereotypes that persist and shift over time. Innovations in natural language processing (word embeddings) enabled this comprehensive study on variability and correlates of change/stability in both manifest and latent stereotypes for 72 diverse groups tracked across 115 years of four English-language text corpora. Results showed, first, that group stereotypes changed by a moderate-to-large degree in manifest content (i.e., top traits associated with groups) but remained relatively more stable in latent structure (i.e., average cosine similarity of top traits' embeddings and vectors of valence, warmth, or competence). This dissociation suggests new insights into how stereotypes and their consequences may endure despite documented changes in other aspects of group representations. Second, results showed substantial variability of change/stability across the 72 groups, with some groups revealing large shifts in manifest and latent content, but others showing near-stability. Third, groups also varied in how consistently they were stereotyped across texts, with some groups showing divergent content, but others showing near-identical representations. Fourth, this variability in change/stability across groups was predicted from a combination of linguistic (e.g., frequency of mentioning the group; consistency of group stereotypes across texts) and social (e.g., the type of group) correlates. Groups that were more frequently mentioned in text changed more than those rarely mentioned; sociodemographic groups changed more than other group types (e.g., body-related stigmas, mental illnesses, occupations), providing the first quantitative evidence of specific group features that may support historical stereotype change. (PsycInfo Database Record (c) 2023 APA, all rights reserved).