Computational Modeling Of Language Research Articles

Navigating the semantic space: Unraveling the structure of meaning in psychosis using different computational language models

Speech in psychosis has long been ascribed as involving ‘loosening of associations’. We pursued the aim to elucidate its underlying cognitive mechanisms by analysing picture descriptions from 94 subjects (29 healthy controls, 18 participants at clinical high risk, 29 with first-episode psychosis, and 18 with chronic schizophrenia), using five language models with different computational architectures: FastText, which represents meaning non-contextually/statically; BERT, which represents contextual meaning sensitive to grammar and context; Infersent and SBERT, which provide sentential representations; and CLIP, which evaluates speech relative to a visual stimulus. These models were used to quantify semantic distances crossed between successive tokens/sentences, and semantic perplexity indicating unexpectedness in continuations. Results showed that, among patients, semantic similarity increased when measured with FastText, Infersent, and SBERT, while it decreased with CLIP and BERT. Higher perplexity was observed in first-episode psychosis. Static semantic measures were associated with clinically measured impoverishment of thought and referential semantic measures with disorganization. These patterns indicate a shrinking conceptual semantic space as represented by static language models, which co-occurs with a widening in the referential semantic space as represented by contextual models. This duality underlines the need to separate these two forms of meaning for understanding mechanisms involved in semantic change in psychosis.

Towards adaptive support for self‐regulated learning of causal relations: Evaluating four Dutch word vector models

Advances in computational language models increasingly enable adaptive support for self‐regulated learning (SRL) in digital learning environments (DLEs; eg, via automated feedback). However, the accuracy of those models is a common concern for educational stakeholders (eg, policymakers, researchers, teachers and learners themselves). We compared the accuracy of four Dutch language models (ie, spaCy medium, spaCy large, FastText and ConceptNet NumberBatch) in the context of secondary school students' learning of causal relations from expository texts, scaffolded by causal diagram completion. Since machine learning relies on human‐labelled data for the best results, we used a dataset with 10,193 students' causal diagram answers, compiled over a decade of research using a diagram completion intervention to enhance students' monitoring of their text comprehension. The language models were used in combination with four popular machine learning classifiers (ie, logistic regression, random forests, support vector machine and neural networks) to evaluate their performance on automatically scoring students' causal diagrams in terms of the correctness of events and their sequence (ie, the causal structure). Five performance metrics were studied, namely accuracy, precision, recall, F1 and the area under the curve of the receiver operating characteristic (ROC‐AUC). The spaCy medium model combined with the neural network classifier achieved the best performance for the correctness of causal events in four of the five metrics, while the ConceptNet NumberBatch model worked best for the correctness of the causal sequence. These evaluation results provide a criterion for model adoption to adaptively support SRL of causal relations in DLEs. Practitioner notesWhat is already known about this topic Accurate monitoring is a prerequisite for effective self‐regulation. Students struggle to accurately monitor their comprehension of causal relations in texts. Completing causal diagrams improves students' monitoring accuracy, but there is room for further improvement. Automatic scoring could be used to provide adaptive support during diagramming. What this paper adds Comparison of four Dutch word vector models combined with four machine learning classifiers for the automatic scoring of students' causal diagrams. Five performance metrics to evaluate the above solutions. Evaluation of the word vector models for estimating the semantic similarity between student and model answers. Implications for practice and/or policy High‐quality word vector models could (em)power adaptive support during causal diagramming via automatic scoring. The evaluated solutions can be embedded in digital learning environments (DLEs). Criteria for model adoption to adaptively support SRL of causal relations in DLEs. The increased saliency of (in)correct answers via automatic scoring might help to improve students' monitoring accuracy.

Unraveling lexical semantics in the brain: Comparing internal, external, and hybrid language models.

To explain how the human brain represents and organizes meaning, many theoretical and computational language models have been proposed over the years, varying in their underlying computational principles and in the language samples based on which they are built. However, how well they capture the neural encoding of lexical semantics remains elusive. We used representational similarity analysis (RSA) to evaluate to what extent three models of different types explained neural responses elicited by word stimuli: an External corpus-based word2vec model, an Internal free word association model, and a Hybrid ConceptNet model. Semantic networks were constructed using word relations computed in the three models and experimental stimuli were selected through a community detection procedure. The similarity patterns between language models and neural responses were compared at the community, exemplar, and word node levels to probe the potential hierarchical semantic structure. We found that semantic relations computed with the Internal model provided the closest approximation to the patterns of neural activation, whereas the External model did not capture neural responses as well. Compared with the exemplar and the node levels, community-level RSA demonstrated the broadest involvement of brain regions, engaging areas critical for semantic processing, including the angular gyrus, superior frontal gyrus and a large portion of the anterior temporal lobe. The findings highlight the multidimensional semantic organization in the brain which is better captured by Internal models sensitive to multiple modalities such as word association compared with External models trained on text corpora.

Languages with more speakers tend to be harder to (machine-)learn

Computational language models (LMs), most notably exemplified by the widespread success of OpenAI's ChatGPT chatbot, show impressive performance on a wide range of linguistic tasks, thus providing cognitive science and linguistics with a computational working model to empirically study different aspects of human language. Here, we use LMs to test the hypothesis that languages with more speakers tend to be easier to learn. In two experiments, we train several LMs—ranging from very simple n-gram models to state-of-the-art deep neural networks—on written cross-linguistic corpus data covering 1293 different languages and statistically estimate learning difficulty. Using a variety of quantitative methods and machine learning techniques to account for phylogenetic relatedness and geographical proximity of languages, we show that there is robust evidence for a relationship between learning difficulty and speaker population size. However, contrary to expectations derived from previous research, our results suggest that languages with more speakers tend to be harder to learn.

Computational Analysis of Superfood Representations in News Media

ABSTRACT What do berries, avocado, quinoa, and ginger have in common? These food items are often regarded as superfoods, a marketing term that overstates the importance of single food items for one’s health and wellbeing. In the present paper, we set out to investigate how purported superfoods are represented in the discourse of online news. We use computational language models to extract the unique topics and terms used to discuss superfoods. Our results show that news coverage is dominated by many specific claims about the healing properties of superfoods. The structural topic model further demonstrates that articles mentioning superfoods are more likely to include topics about a) nutrients, physical appearance, and health in the same context, b) retail strategies, and c) scientific research about the health benefits of superfoods. These results illustrate complex representations of superfoods in news media.

So Cloze Yet So Far: N400 Amplitude Is Better Predicted by Distributional Information Than Human Predictability Judgements

More predictable words are easier to process—they are read faster and elicit smaller neural signals associated with processing difficulty, most notably, the N400 component of the event-related brain potential. Thus, it has been argued that prediction of upcoming words is a key component of language comprehension, and that studying the amplitude of the N400 is a valuable way to investigate the predictions we make. In this study, we investigate whether the linguistic predictions of computational language models or humans better reflect the way in which natural language stimuli modulate the amplitude of the N400. One important difference in the linguistic predictions of humans versus computational language models is that while language models base their predictions exclusively on the preceding linguistic context, humans may rely on other factors. We find that the predictions of three top-of-the-line contemporary language models—GPT-3, RoBERTa, and ALBERT—match the N400 more closely than human predictions. This suggests that the predictive processes underlying the N400 may be more sensitive to the statistics of language than previously thought.

An eye-tracking-with-EEG coregistration corpus of narrative sentences

We present the Radboud Coregistration Corpus of Narrative Sentences (RaCCooNS), the first freely available corpus of eye-tracking-with-EEG data collected while participants read narrative sentences in Dutch. The corpus is intended for studying human sentence comprehension and for evaluating the cognitive validity of computational language models. RaCCooNS contains data from 37 participants (3 of which eye tracking only) reading 200 Dutch sentences each. Less predictable words resulted in significantly longer reading times and larger N400 sizes, replicating well-known surprisal effects in eye tracking and EEG simultaneously. We release the raw eye-tracking data, the preprocessed eye-tracking data at the fixation, word, and trial levels, the raw EEG after merger with eye-tracking data, and the preprocessed EEG data both before and after ICA-based ocular artifact correction.

Removing Biases in Communication of Severity Assessments of Intimate Partner Violence: Model Development and Evaluation

To support a victim of violence and establish the correct penalty for the perpetrator, it is crucial to correctly evaluate and communicate the severity of the violence. Recent data have shown these communications to be biased. However, computational language models provide opportunities for automated evaluation of the severity to mitigate the biases. We investigated whether these biases can be removed with computational algorithms trained to measure the severity of violence described. In phase 1 (P1), participants (N=71) were instructed to write some text and type 5 keywords describing an event where they experienced physical violence and 1 keyword describing an event where they experienced psychological violence in an intimate partner relationship. They were also asked to rate the severity. In phase 2 (P2), another set of participants (N=40) read the texts and rated them for severity of violence on the same scale as in P1. We also quantified the text data to word embeddings. Machine learning was used to train a model to predict the severity ratings. For physical violence, there was a greater accuracy bias for humans (r2=0.22) compared to the computational model (r2=0.31; t38=-2.37, P=.023). For psychological violence, the accuracy bias was greater for humans (r2=0.058) than for the computational model (r2=0.35; t38=-14.58, P<.001). Participants in P1 experienced psychological violence as more severe (mean 6.46, SD 1.69) than participants rating the same events in P2 (mean 5.84, SD 2.80; t86=-2.22, P=.029<.05), whereas no calibration bias was found for the computational model (t134=1.30, P=.195). However, no calibration bias was found for physical violence for humans between P1 (mean 6.59, SD 1.81) and P2 (mean 7.54, SD 2.62; t86=1.32, P=.19) or for the computational model (t134=0.62, P=.534). There was no difference in the severity ratings between psychological and physical violence in P1. However, the bias (ie, the ratings in P2 minus the ratings in P1) was highly negatively correlated with the severity ratings in P1 (r2=0.29) and in P2 (r2=0.37), whereas the ratings in P1 and P2 were somewhat less correlated (r2=0.11) using the psychological and physical data combined. The results show that the computational model mitigates accuracy bias and removes calibration biases. These results suggest that computational models can be used for debiasing the severity evaluations of violence. These findings may have application in a legal context, prioritizing resources in society and how violent events are presented in the media.

Informational content vs. discourse orientation: experimental and computational perspectives

The aim of this study is to investigate how human speakers and computational language models process (i) the informational content and (ii) the discourse orientation of natural language sentences. These two dimensions of meaning have received little attention outside theoretical literature, especially in the computational linguistics domain. To help fill this void, we present the results of four experiments that exploit the specific semantics of two French adverbs, namely presque (≃ ’almost’) and à peine (≃ ’barely’), which put these two dimensions of meaning at odds. Each experiment focuses on one kind of population (humans or language models), and one kind of meaning (informational content or discourse orientation). Our results show that humans are indeed sensitive to informational content and discourse direction, as assumed in the theoretical literature. Language models exhibit a less transparent behavior. Their performances in dealing with the semantics of presque appear in line with predictions based on the way these models are trained, but this does not extend to à peine.

A synchronized multimodal neuroimaging dataset for studying brain language processing

We present a synchronized multimodal neuroimaging dataset for studying brain language processing (SMN4Lang) that contains functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG) data on the same 12 healthy volunteers while the volunteers listened to 6 hours of naturalistic stories, as well as high-resolution structural (T1, T2), diffusion MRI and resting-state fMRI data for each participant. We also provide rich linguistic annotations for the stimuli, including word frequencies, syntactic tree structures, time-aligned characters and words, and various types of word and character embeddings. Quality assessment indicators verify that this is a high-quality neuroimaging dataset. Such synchronized data is separately collected by the same group of participants first listening to story materials in fMRI and then in MEG which are well suited to studying the dynamic processing of language comprehension, such as the time and location of different linguistic features encoded in the brain. In addition, this dataset, comprising a large vocabulary from stories with various topics, can serve as a brain benchmark to evaluate and improve computational language models.

Computational Topic Models for Theological Investigations

ABSTRACT Sallie McFague’s theological models construct a tensive relationship between conceptual structures and symbolic, metaphorical language to interpret the defining and elusive aspects of theological phenomena and loci. Computational models of language can extend and formalize the conceptual structures of theological models to develop computer-augmented interpretations of theological texts. Previously unclear is whether computational models can retain the tensive symbolism essential for theological investigation. I demonstrate affirmatively by constructing a computational topic model of the moral theology of Thomas Aquinas from Summa Theologica (Second Part, in English translation) useful for interpreting not only the Thomistic text but also recent papal encyclicals.

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).

Hierarchy in language interpretation: evidence from behavioural experiments and computational modelling

ABSTRACT It has long been recognised that phrases and sentences are organised hierarchically, but many computational models of language treat them as sequences of words without computing constituent structure. Against this background, we conducted two experiments which showed that participants interpret ambiguous noun phrases, such as second blue ball, in terms of their abstract hierarchical structure rather than their linear surface order. When a neural network model was tested on this task, it could simulate such “hierarchical” behaviour. However, when we changed the training data such that they were not entirely unambiguous anymore, the model stopped generalising in a human-like way. It did not systematically generalise to novel items, and when it was trained on ambiguous trials, it strongly favoured the linear interpretation. We argue that these models should be endowed with a bias to make generalisations over hierarchical structure in order to be cognitively adequate models of human language.

Linguistic Variation and Change in 250 Years of English Scientific Writing: A Data-Driven Approach.

We trace the evolution of Scientific English through the Late Modern period to modern time on the basis of a comprehensive corpus composed of the Transactions and Proceedings of the Royal Society of London, the first and longest-running English scientific journal established in 1665. Specifically, we explore the linguistic imprints of specialization and diversification in the science domain which accumulate in the formation of “scientific language” and field-specific sublanguages/registers (chemistry, biology etc.). We pursue an exploratory, data-driven approach using state-of-the-art computational language models and combine them with selected information-theoretic measures (entropy, relative entropy) for comparing models along relevant dimensions of variation (time, register). Focusing on selected linguistic variables (lexis, grammar), we show how we deploy computational language models for capturing linguistic variation and change and discuss benefits and limitations.

Evaluating Computational Language Models with Scaling Properties of Natural Language

In this article, we evaluate computational models of natural language with respect to the universal statistical behaviors of natural language. Statistical mechanical analyses have revealed that natural language text is characterized by scaling properties, which quantify the global structure in the vocabulary population and the long memory of a text. We study whether five scaling properties (given by Zipf’s law, Heaps’ law, Ebeling’s method, Taylor’s law, and long-range correlation analysis) can serve for evaluation of computational models. Specifically, we test n-gram language models, a probabilistic context-free grammar, language models based on Simon/Pitman-Yor processes, neural language models, and generative adversarial networks for text generation. Our analysis reveals that language models based on recurrent neural networks with a gating mechanism (i.e., long short-term memory; a gated recurrent unit; and quasi-recurrent neural networks) are the only computational models that can reproduce the long memory behavior of natural language. Furthermore, through comparison with recently proposed model-based evaluation methods, we find that the exponent of Taylor’s law is a good indicator of model quality.

The influence of place and time on lexical behavior: A distributional analysis.

We measured and documented the influence of corpus effects on lexical behavior. Specifically, we used a corpus of over 26,000 fiction books to show that computational models of language trained on samples of language (i.e., subcorpora) representative of the language located in a particular place and time can track differences in people's experimental language behavior. This conclusion was true across multiple tasks (lexical decision, category production, and word familiarity) and provided insight into the influence that language experience imposes on language processing and organization. We used the assembled corpus and methods to validate a new machine-learning approach for optimizing language models, entitled experiential optimization (Johns, Jones, & Mewhort in Psychonomic Bulletin & Review, 26, 103-126, 2019).

Lexical Predictability During Natural Reading: Effects of Surprisal and Entropy Reduction.

What are the effects of word-by-word predictability on sentence processing times during the natural reading of a text? Although information complexity metrics such as surprisal and entropy reduction have been useful in addressing this question, these metrics tend to be estimated using computational language models, which require some degree of commitment to a particular theory of language processing. Taking a different approach, this study implemented a large-scale cumulative cloze task to collect word-by-word predictability data for 40 passages and compute surprisal and entropy reduction values in a theory-neutral manner. A separate group of participants read the same texts while their eye movements were recorded. Results showed that increases in surprisal and entropy reduction were both associated with increases in reading times. Furthermore, these effects did not depend on the global difficulty of the text. The findings suggest that surprisal and entropy reduction independently contribute to variation in reading times, as these metrics seem to capture different aspects of lexical predictability.

Using stochastic language models (SLM) to map lexical, syntactic, and phonological information processing in the brain.

Language comprehension involves the simultaneous processing of information at the phonological, syntactic, and lexical level. We track these three distinct streams of information in the brain by using stochastic measures derived from computational language models to detect neural correlates of phoneme, part-of-speech, and word processing in an fMRI experiment. Probabilistic language models have proven to be useful tools for studying how language is processed as a sequence of symbols unfolding in time. Conditional probabilities between sequences of words are at the basis of probabilistic measures such as surprisal and perplexity which have been successfully used as predictors of several behavioural and neural correlates of sentence processing. Here we computed perplexity from sequences of words and their parts of speech, and their phonemic transcriptions. Brain activity time-locked to each word is regressed on the three model-derived measures. We observe that the brain keeps track of the statistical structure of lexical, syntactic and phonological information in distinct areas.

Language and other complex behaviors: Unifying characteristics, computational models, neural mechanisms

Similar to other complex behaviors, language is dynamic, social, multimodal, patterned, and purposive, its purpose being to promote desirable actions or thoughts in others and self (Edelman, 2017b). An analysis of the functional characteristics shared by complex sequential behaviors suggests that they all present a common overarching computational problem: dynamically controlled constrained navigation in concrete or abstract situation spaces. With this conceptual framework in mind, I compare and contrast computational models of language and evaluate their potential for explaining linguistic behavior and for elucidating the brain mechanisms that support it.

Reference Production as Search: The Impact of Domain Size on the Production of Distinguishing Descriptions.

When producing a description of a target referent in a visual context, speakers need to choose a set of properties that distinguish it from its distractors. Computational models of language production/generation usually model this as a search process and predict that the time taken will increase both with the number of distractors in a scene and with the number of properties required to distinguish the target. These predictions are reminiscent of classic findings in visual search; however, unlike models of reference production, visual search models also predict that search can become very efficient under certain conditions, something that reference production models do not consider. This paper investigates the predictions of these models empirically. In two experiments, we show that the time taken to plan a referring expression-as reflected by speech onset latencies-is influenced by distractor set size and by the number of properties required, but this crucially depends on the discriminability of the properties under consideration. We discuss the implications for current models of reference production and recent work on the role of salience in visual search.