Limitations to optimal search in naturalistic active learning.

Optimality in active learning is under intense debate in numerous disciplines. We introduce a new empirical paradigm for studying naturalistic active learning, as well as new computational tools for jointly modeling algorithmic and rational theories of information search. Participants in our task can ask questions and learn about hundreds of everyday items but must retrieve queried items from memory. To maximize information gain, participants need to retrieve sequences of dissimilar items. In eight experiments (N = 795), we find that participants are unable to do this. Instead, associative memory mechanisms lead to the successive retrieval of similar items, an established memory effect known as semantic congruence. The extent of semantic congruence (and thus suboptimality in question asking) is unaffected by task instructions and incentives, though participants can identify efficient query sequences when given a choice between query sequences. Overall, our results indicate that participants can distinguish between optimal and suboptimal search if explicitly asked to do so, but have difficulty implementing optimal search from memory. We conclude that associative memory processes may place critical restrictions on people's ability to ask good questions in naturalistic active learning tasks. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Extracting social determinants of health events with transformer-based multitask, multilabel named entity recognition.

OBJECTIVE: Social determinants of health (SDOH) are nonclinical, socioeconomic conditions that influence patient health and quality of life. Identifying SDOH may help clinicians target interventions. However, SDOH are more frequently available in narrative notes compared to structured electronic health records. The 2022 n2c2 Track 2 competition released clinical notes annotated for SDOH to promote development of NLP systems for extracting SDOH. We developed a system addressing 3 limitations in state-of-the-art SDOH extraction: the inability to identify multiple SDOH events of the same type per sentence, overlapping SDOH attributes within text spans, and SDOH spanning multiple sentences. MATERIALS AND METHODS: We developed and evaluated a 2-stage architecture. In stage 1, we trained a BioClinical-BERT-based named entity recognition system to extract SDOH event triggers, that is, text spans indicating substance use, employment, or living status. In stage 2, we trained a multitask, multilabel NER to extract arguments (eg, alcohol "type") for events extracted in stage 1. Evaluation was performed across 3 subtasks differing by provenance of training and validation data using precision, recall, and F1 scores. RESULTS: When trained and validated on data from the same site, we achieved 0.87 precision, 0.89 recall, and 0.88 F1. Across all subtasks, we ranked between second and fourth place in the competition and always within 0.02 F1 from first. CONCLUSIONS: Our 2-stage, deep-learning-based NLP system effectively extracted SDOH events from clinical notes. This was achieved with a novel classification framework that leveraged simpler architectures compared to state-of-the-art systems. Improved SDOH extraction may help clinicians improve health outcomes.

Free association in a neural network.

Free association among words is a fundamental and ubiquitous memory task. Although distributed semantics (DS) models can predict the association between pairs of words, and semantic network (SN) models can describe transition probabilities in free association data, there have been few attempts to apply established cognitive process models of memory search to free association data. Thus, researchers are currently unable to explain the dynamics of free association using memory mechanisms known to be at play in other retrieval tasks, such as free recall from lists. We address this issue using a popular neural network model of free recall, the context maintenance and retrieval (CMR) model, which we fit using stochastic gradient descent on a large data set of free association norms. Special cases of CMR mimic existing DS and SN models of free association, and we find that CMR outperforms these models on out-of-sample free association data. We also show that training CMR on free association data generates improved predictions for free recall from lists, demonstrating the value of free association for the study of many different types of memory phenomena. Overall, our analysis provides a new account of the dynamics of free association, predicts free association with increased accuracy, integrates theories of free association with established models of memory, and shows how large data sets and neural network training methods can be used to model complex cognitive processes that operate over thousands of representations. (PsycInfo Database Record (c) 2023 APA, all rights reserved).

Transformer networks of human conceptual knowledge.

We present a computational model capable of simulating aspects of human knowledge for thousands of real-world concepts. Our approach involves a pretrained transformer network that is further fine-tuned on large data sets of participant-generated feature norms. We show that such a model can successfully extrapolate from its training data, and predict human knowledge for new concepts and features. We apply our model to stimuli from 25 previous experiments in semantic cognition research and show that it reproduces many findings on semantic verification, concept typicality, feature distribution, and semantic similarity. We also compare our model against several variants, and by doing so, establish the model properties that are necessary for good prediction. The success of our approach shows how a combination of language data and (laboratory-based) psychological data can be used to build models with rich world knowledge. Such models can be used in the service of new psychological applications, such as the modeling of naturalistic semantic verification and knowledge retrieval, as well as the modeling of real-world categorization, decision-making, and reasoning. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Classifying social determinants of health from unstructured electronic health records using deep learning-based natural language processing.

OBJECTIVE: Social determinants of health (SDOH) are non-medical factors that can profoundly impact patient health outcomes. However, SDOH are rarely available in structured electronic health record (EHR) data such as diagnosis codes, and more commonly found in unstructured narrative clinical notes. Hence, identifying social context from unstructured EHR data has become increasingly important. Yet, previous work on using natural language processing to automate extraction of SDOH from text (a) usually focuses on an ad hoc selection of SDOH, and (b) does not use the latest advances in deep learning. Our objective was to advance automatic extraction of SDOH from clinical text by (a) systematically creating a set of SDOH based on standard biomedical and psychiatric ontologies, and (b) training state-of-the-art deep neural networks to extract mentions of these SDOH from clinical notes. DESIGN: A retrospective cohort study. SETTING AND PARTICIPANTS: Data were extracted from the Medical Information Mart for Intensive Care (MIMIC-III) database. The corpus comprised 3,504 social related sentences from 2,670 clinical notes. METHODS: We developed a framework for automated classification of multiple SDOH categories. Our dataset comprised narrative clinical notes under the "Social Work" category in the MIMIC-III Clinical Database. Using standard terminologies, SNOMED-CT and DSM-IV, we systematically curated a set of 13 SDOH categories and created annotation guidelines for these. After manually annotating the 3,504 sentences, we developed and tested three deep neural network (DNN) architectures - convolutional neural network (CNN), long short-term memory (LSTM) network, and the Bidirectional Encoder Representations from Transformers (BERT) - for automated detection of eight SDOH categories. We also compared these DNNs to three baselines models: (1) cTAKES, as well as (2) L2-regularized logistic regression and (3) random forests on bags-of-words. Model evaluation metrics included micro- and macro- F1, and area under the receiver operating characteristic curve (AUC). RESULTS: All three DNN models accurately classified all SDOH categories (minimum micro-F1 = 0.632, minimum macro-AUC = 0.854). Compared to the CNN and LSTM, BERT performed best in most key metrics (micro-F1 = 0.690, macro-AUC = 0.907). The BERT model most effectively identified the "occupational" category (F1 = 0.774, AUC = 0.965) and least effectively identified the "non-SDOH" category (F = 0.491, AUC = 0.788). BERT outperformed cTAKES in distinguishing social vs non-social sentences (BERT F1 = 0.87 vs. cTAKES F1 = 0.06), and outperformed logistic regression (micro-F1 = 0.649, macro-AUC = 0.696) and random forest (micro-F1 = 0.502, macro-AUC = 0.523) trained on bag-of-words. CONCLUSIONS: Our study framework with DNN models demonstrated improved performance for efficiently identifying a systematic range of SDOH categories from clinical notes in the EHR. Improved identification of patient SDOH may further improve healthcare outcomes.

Process and content in decisions from memory.

Information stored in memory influences the formation of preferences and beliefs in most everyday decision tasks. The richness of this information, and the complexity inherent in interacting memory and decision processes, makes the quantitative model-driven analysis of such decisions very difficult. In this article we present a general framework that can address the theoretical and methodological barriers to building formal models of naturalistic memory-based decision making. Our framework implements established theories of memory search and decision making within a single integrated cognitive system, and uses computational language models to quantify the thoughts over which memory and decision processes operate. It can thus describe both the content of the information that is sampled from memory, as well as the processes involved in retrieving and evaluating this information in order to make a decision. Furthermore, our framework is tractable, and the parameters that characterize memory-based decisions can be recovered using thought listing and choice data from existing experimental tasks, and in turn be used to make quantitative predictions regarding choice probability, length of deliberation, retrieved thoughts, and the effects of decision context. We showcase the power and generality of our framework by applying it to naturalistic binary choices from domains such as risk perception, consumer behavior, financial decision making, ethical decision making, legal decision making, food choice, and social judgment. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Similarity Judgment Within and Across Categories: A Comprehensive Model Comparison.

Similarity is one of the most important relations humans perceive, arguably subserving category learning and categorization, generalization and discrimination, judgment and decision making, and other cognitive functions. Researchers have proposed a wide range of representations and metrics that could be at play in similarity judgment, yet have not comprehensively compared the power of these representations and metrics for predicting similarity within and across different semantic categories. We performed such a comparison by pairing nine prominent vector semantic representations with seven established similarity metrics that could operate on these representations, as well as supervised methods for dimensional weighting in the similarity function. This approach yields a factorial model structure with 126 distinct representation-metric pairs, which we tested on a novel dataset of similarity judgments between pairs of cohyponymic words in eight categories. We found that cosine similarity and Pearson correlation were the overall best performing unweighted similarity functions, and that word vectors derived from free association norms often outperformed word vectors derived from text (including those specialized for similarity). Importantly, models that used human similarity judgments to learn category-specific weights on dimensions yielded substantially better predictions than all unweighted approaches across all types of similarity functions and representations, although dimension weights did not generalize well across semantic categories, suggesting strong category context effects in similarity judgment. We discuss implications of these results for cognitive modeling and natural language processing, as well as for theories of the representations and metrics involved in similarity.

The spatial arrangement method of measuring similarity can capture high-dimensional semantic structures.

Psychologists collect similarity data to study a variety of phenomena including categorization, generalization and discrimination, and representation itself. However, collecting similarity judgments between all pairs of items in a set is expensive, spurring development of techniques like the Spatial Arrangement Method (SpAM; Goldstone, Behavior Research Methods, Instruments, & Computers, 26, 381-386, 1994), wherein participants place items on a two-dimensional plane such that proximity reflects perceived similarity. While SpAM greatly hastens similarity measurement, and has been successfully used for lower-dimensional, perceptual stimuli, its suitability for higher-dimensional, conceptual stimuli is less understood. In study 1, we evaluated the ability of SpAM to capture the semantic structure of eight different categories composed of 20-30 words each. First, SpAM distances correlated strongly (r = .71) with pairwise similarity judgments, although below SpAM and pairwise judgment split-half reliabilities (r's > .9). Second, a cross-validation exercise with multidimensional scaling fits at increasing latent dimensionalities suggested that aggregated SpAM data favored higher (> 2) dimensional solutions for seven of the eight categories explored here. Third, split-half reliability of SpAM dissimilarities was high (Pearson r = .90), while the average correlation between pairs of participants was low (r = .15), suggesting that when different participants focus on different pairs of stimulus dimensions, reliable high-dimensional aggregate similarity data is recoverable. In study 2, we show that SpAM can recover the Big Five factor space of personality trait adjectives, and that cross-validation favors a four- or five-dimension solution on this dataset. We conclude that SpAM is an accurate and reliable method of measuring similarity for high-dimensional items like words. We publicly release our data for researchers.

Modeling the emergence of lexicons in homesign systems.

It is largely 4ledged that natural languages emerge not just from human brains but also from rich communities of interacting human brains (Senghas, ). Yet the precise role of such communities and such interaction in the emergence of core properties of language has largely gone uninvestigated in naturally emerging systems, leaving the few existing computational investigations of this issue at an artificial setting. Here, we take a step toward investigating the precise role of community structure in the emergence of linguistic conventions with both naturalistic empirical data and computational modeling. We first show conventionalization of lexicons in two different classes of naturally emerging signed systems: (a) protolinguistic "homesigns" invented by linguistically isolated Deaf individuals, and (b) a natural sign language emerging in a recently formed rich Deaf community. We find that the latter conventionalized faster than the former. Second, we model conventionalization as a population of interacting individuals who adjust their probability of sign use in response to other individuals' actual sign use, following an independently motivated model of language learning (Yang, , ). Simulations suggest that a richer social network, like that of natural (signed) languages, conventionalizes faster than a sparser social network, like that of homesign systems. We discuss our behavioral and computational results in light of other work on language emergence, and other work of behavior on complex networks.