Skip to main content
NCBI home page
Scientific Data logoLink to Scientific Data
. 2024 Oct 9;11:1104. doi: 10.1038/s41597-024-03915-8

DERCo: A Dataset for Human Behaviour in Reading Comprehension Using EEG

Boi Mai Quach 1,2,✉, Cathal Gurrin 1,3, Graham Healy 1,3
PMCID: PMC11464549  PMID: 39384587

Abstract

This paper introduces the DERCo (Dublin EEG-based Reading Experiment Corpus), a language resource combining electroencephalography (EEG) and next-word prediction data obtained from participants reading narrative texts. The dataset comprises behavioral data collected from 500 participants recruited through the Amazon Mechanical Turk online crowd-sourcing platform, along with EEG recordings from 22 healthy adult native English speakers. The online experiment was designed to examine the context-based word prediction by a large sample of participants, while the EEG-based experiment was developed to extend the validation of behavioral next-word predictability. Online participants were instructed to predict upcoming words and complete entire stories. Cloze probabilities were then calculated for each word so that this predictability measure could be used to support various analyses pertaining to semantic context effects in the EEG recordings. EEG-based analyses revealed significant differences between high and low predictable words, demonstrating one important type of potential analysis that necessitates close integration of these two datasets. This material is a valuable resource for researchers in neurolinguistics due to the word-level EEG recordings in context.

Subject terms: Computational neuroscience, Scientific data

Background & Summary

The availability of publicly accessible EEG datasets for semantic-level information in reading is limited1. Researchers often design and collect data with limited consideration of reuse of the data for other researchers to conduct investigations or test hypotheses, making these datasets difficult to reuse in other studies. For example, Dufau et al.2 released an EEG dataset with a thousand words to examine the time course of orthographic, lexical, and semantic influences on word-level information. Thus, it could not be reused for studies related to decision-making or reading behaviors. In contrast, Davis et al.3 conducted a reading-based experiment with complex decision-making (CDM) tasks to indicate how brain activity in CDM was distributed across frequencies and brain regions. Therefore, this dataset cannot be potentially used to explore semantic-level hypotheses.

Moreover, designing and conducting reading experiments wherein neural responses are captured is painstaking due to the complexity and time-consuming nature of successfully deploying EEG experiments. These reasons drive us to seize an opportunity to build an EEG-based reading dataset, paving the way for cutting-edge insights into cognitive processes.

The study of brain activity during reading is a complex challenge for neuroscientists4–8, as it involves disentangling complex cognitive processes from continuously measured brain signals that are often contaminated with various sources of non-neural activity. Three widely used measurement techniques to study brain function include functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG), and electroencephalography (EEG).

Although fMRI can effectively map brain responses in different areas to linguistic stimuli at a high spatial resolution9, it is limited in terms of precisely measuring the timing of cognitive processes (one scan for every 1 to 3 seconds10). Particularly in reading, neural events often occur within milliseconds. On the other hand, MEG and EEG provide dense snapshots of brain processing at a higher temporal resolution, in the order of milliseconds range but with poorer spatial resolution than that of fMRI. Compared to EEG, MEG provides a much better balance of temporal and spatial resolution, a higher signal-to-noise ratio (SNR), and shorter procedural time than EEG11. However, EEG is often preferred as it is portable and less expensive, hence it has been used in a large number of neurolinguistic studies to date.

Event-related potentials (ERPs) in a continuous EEG recording are voltage fluctuations that are time-locked to particular types of events, such as the presentation time of stimuli, where averaging of the time-locked EEG signals is typically employed to improve SNR. ERPs have been extensively utilised in many language studies, especially the N400 ERP12–18, which typically occurs 300-500 ms after the word onset. Studies investigating the N400 have established an association with semantic complexity17, as an index of semantic processing. Several studies have shown that the N400 is sensitive to both linguistic and non-linguistic characteristics, including expectancy effects12,18, frequency13, orthographic neighbourhood effects15,16, and lexical association14.

Over time, EEG has become an important measurement technique not only for neuroscientific research but also for those in the field of Artificial Intelligence (AI), in particular for natural language processing (NLP). Working with EEG and AI in combination has enabled researchers to explore the similarities and differences between brain activity and AI models in language processing19–21, mapping neural signals to natural language8,22, decoding linguistic information from the brain5,23, or encoding language into neural activity representations24.

Two notable EEG-based reading datasets made available for other researchers are the Kilo-word ERP dataset and ZuCo dataset. Dufau et al2 published the Kilo-word ERP dataset in 2015, consisting of EEG data from 75 subjects performing a lexical decision task on approximately 1,000 English words. This dataset has been reused in neurolinguistic studies25 and for teaching ERPs26. However, the paradigm used to capture this data presents word stimuli independent of each other without context, i.e. word stimuli presentations are not part of a sentence. The ZuCo 1.0 dataset27 was first released in 2018, updated to ZuCo 2.028 in 2019, and has been applied in numerous studies8,23,29,30. This dataset comprises EEG measures co-registered to eye-tracking during natural reading. However, the dataset’s focus is on NLP tasks such as entity recognition, relation extraction, and sentiment analysis using machine learning algorithms. Hence, the ZuCo dataset does not completely focus on single-trial responses at the word level.

To balance the disadvantages and advantages of these two datasets, our dataset focuses on single-trial neural responses collected in a narrative context. Single-trial responses in our experiment are synonymous with word-level EEG responses. This approach is particularly important as it allows us to perform analyses such as comparing high versus low predictable words. While responses may be averaged within a category (e.g. high probability), word-level (single-trial) responses are still ultimately required to perform analysis on language-related ERP components31.

Moreover, our dataset includes data from two types of experiments. The first one is a typical word-by-word presentation, also known as Rapid Serial Visual Presentation (RSVP), which requires participants to maintain their gaze on a location on screen where words are displayed one by one in the fovea at a fixed pace32. This classic RSVP ignores some properties of natural reading, namely parafoveal perception, which occurs in natural reading where when a person is fixating on a word they still perceive properties (e.g. word length) of the surrounding words in a sentence. Horizontal EOG (Electrooculargraphy) was encluded as a part of dataset, enabling researchers to examine horizontal eye movements should they wish to further investigate this. From our observation of the data, there was minimal eye movements, but users of the data may decide to impose stricter thresholds regarding minor eye movements (e.g. microsaccades). Also, this paradigm uses a fixed presentation rate for each word, effectively driving the speed at which participants read33–35.

A major benefit of using the RSVP word stimulus presentation approach is that it mitigates against cornea-retinal dipole activity related to eye movements that would occur during natural reading from contaminating the EEG signal recordings, which are particularly difficult to effectively remove after the fact36. Nevertheless, an important drawback of the RSVP approach is the reduction in visual acuity from the foveal region to the parafovea and then the periphery37. To overcome this, we designed the second experiment called RSVP-with-flanker paradigm. In this type of experiment, the word at fixation always reveals while the words on neighbouring sides are flanked by letter strings. Many of the studies using RSVP-with-flankers also incorporate ERPs analysis related to semantic integration38–41.

An important feature of this dataset is human annotation data via a behavioural word-prediction experiment conducted using Amazon Mechanical Turk42 for 500 participants. The word “behaviour” was intended to indicate that participants were required to predict upcoming words43. Behavioural experiments are designed to test a belief or prediction and to discover new insights44. They are frequently conducted on crowdsourcing platforms for tasks related to psychology, neuroscience, and cognition45. In our research, this experiment captures the predictions for next words from 500 participants, allowing us to measure the predictability of words in the context. Combining this dataset with the EEG-based reading data for 22 participants, we demonstrate ERP differences using high-and-low cloze probabilities for next-word predictions, demonstrating the complementary integration of both datasets.

We anticipate that this dataset will facilitate AI-related research, including the development of brain encoding and decoding models, research in cognitive science and linguistics, and advancing the capabilities of natural language processing systems. For instance, it will enable researchers to study the nature of brain signals underlying semantic-level information in reading contexts. Additionally, by standardising EEG data formats and providing thorough metadata annotation, this dataset will support data reuse in the co-registration of EEG with different modalities such as fMRI, MEG, eye tracking data, etc. The technical validation of this dataset, described further below, serves a proof of the quality of the datasets.

Methods

In our experiments, we used the same materials for both the EEG-based reading and the behavioural word-prediction experiment. The script of the five articles are reported in Supplementary File 1. The study was approved by the Dublin City University Research Ethics Committee (Reference: DCUREC/2023/085) for both experiments. Participants’ responses were de-identified i.e. each individual data was represented by a unique random ID with no recording of name. They also provided permission for their data to be shared as a part of informed consent process.

Materials

The primary purpose of this dataset was to enable researchers to investigate reading behaviours and comprehension, necessitating the selection of materials that avoided using advanced/unfamiliar words and did not require specific domain knowledge, such as that needed for understanding a scientific article. Many previous studies of brain activity during reading experiments have utilised naturalistic stories for the collection of neural data, including the Alice datasets46, Le Petit Prince (The Little Prince) corpus47, the Harry Potter datasets48,49, and the Adventures of Sherlock Holmes datasets19,50. Therefore, we found that the use of narrative stories, particularly fairy tales, fully satisfied the mentioned requirements.

Based on these considerations, five short stories from Grimms’ Fairy Tales were selected, namely: “The Mouse, the Bird, and the Sausage” (Article 1), “Straw, Coal, and Bean” (Article 2), “Poverty and Humility Lead to Heaven” (Article 3), “The Death of the Little Hen” (Article 4), and “The Wolf and the Fox” (Article 5). These fairy tales include a large, rich set of verbal descriptions without the use of advanced vocabulary, making them comprehensible to individuals with varying degrees of knowledge.

Additionally, these stories are not the commonly known, thereby mitigating the impact of familiarity. When creating the descriptive table, we removed currency symbols, numerals, and punctuation. Table 1 presents the overall descriptive statistics of the materials split by story. It shows the variability in article length, sentence structure, and word usage, with distinct word counts ranging from 224 to 301, and sentence counts from 23 to 36. The table also includes mean and standard deviation values for words per sentence and word length, indicating the average and variability within these metrics across the articles.

Table 1.

Descriptive statistics of reading materials (M = mean, SD = standard deviation).

Article 1 Article 2 Article 3 Article 4 Article 5
Number of words 564 466 493 660 776
Number of distinct words 243 235 226 224 301
Number of sentences 33 32 23 36 34
Word per sentences M 17.09 14.56 21.43 18.33 22.82
SD 8.43 7.16 7.89 13.61 11.41
Word length M 4.17 4.09 3.89 3.83 3.97
SD 1.94 2.01 1.92 1.65 1.91
Open in a new tab

Behavioural Word-Prediction Experiment

Participants

A total of 500 participants completed an upcoming word prediction task (behavioural word-prediction experiment) on Mechanical Turk42, a crowdsourcing marketplace on Amazon. Each HIT (Human Intelligence Task) corresponded to one of the five stories, meaning that we had 100 subjects per story. Demographic information was not collected, and participants remained completely anonymous. The HITs were only available for participants in the USA, Canada, Australia, England, and Ireland, aged 18 years or older. Potential participants saw the HIT on the Amazon Mechanical Turk platform and then decided whether to complete the tasks for compensation. Before beginning the HIT, participants were required to sign a digital consent form with three conditions: 1. You voluntarily agree to participate, 2. You are 18 years of age or older, 3. You are a native English speaker (as shown in Supplementary Figure S1). Volunteers participating in our experiment were compensated with a fee of $1.2 to $1.5 per story, depending on the story’s length. Each HIT (per story) lasted 25 minutes on average.

Stimuli & Experimental Design

The experiment was presented to participants as an HTML webpage with the task flow controlled by jsPsych51, which is a JavaScript framework for creating behavioral experiments running in a web browser. The words were displayed in 21-px Arial font with a black color at the centre of the webpage, while the page had a white background color.

This experiment is also known as cloze procedure52, a traditional approach which involves asking participants to predict and complete unfinished sentences or passages based on the accumulated preceding context. For each story, the title was initially displayed, and then the first two words in the content were revealed, where participants were then asked to predict (by typing) the next word. After providing their prediction, the actual upcoming word was disclosed, where participants were then asked to predict the next word in the sequence. Once ten words were displayed on the screen, the left-most word was then removed, and the upcoming word was presented. This procedure was repeated until participants provided their predictions for all upcoming missing words. The ten-word sliding window is illustrated in Fig. 1, and Fig. 2 shows the display used on the website.

Fig. 1.

Fig. 1

Open in a new tab

The stimulus was transcribed in article 1 for the behavioral experiment. (a) Transcript (b) A ten-word sliding window example.

Fig. 2.

Fig. 2

Open in a new tab

A ten-word sliding window was presented on the website for the article 1. (a) The next word is “the” (b) The next word is “next”.

Since our main focus is on next-word predictability, we derived the experimental design from papers using the cloze procedure and its application in brain activity analyses for reading and audio experiments. Goldstein et al.21 conducted a cloze experiment using a sliding-window behavioural paradigm similar to ours. In their experiment, participants initially saw two words from a story, predicted the next word, and then continued to predict subsequent words as they were revealed one by one. After displaying ten words, the earliest word was removed to make space for the next word in the sequence. Varda et al.53 implemented a similar experimental design. Additionally, they conducted a cloze experiment based on the original EEG-based reading dataset54,55. The materials used included a total of 205 English sentences, ranging from 5 to 15 words, with an average of 10 words. This average of ten words was also the number we chose for our experiment. Although participants had access to more context in the cloze experiment compared to the EEG-based experiment, both studies achieved significant results in analysing predictability in human brain activity based on cloze probability.

Procedure

Participants on Mechanical Turk42 found the experiment by browsing for HITs. After accepting the HIT, the participants were provided with task instructions and a web link to navigate to the experiment’s website. An informed consent form was displayed after the welcome screen. Participants then proceeded with a test trial to familiarise themselves with the experiment. Before moving on to the next step, we emphasised three important points to ensure data quality and prevent technical interruptions during the experiment: 1. Please do not skip words, 2. Please do not use online resources and, 3. Please do not reload the page. After completing the entire story (i.e. providing a complete set of upcoming word predictions), they were then asked to answer some short questions about the text they had just read to test their comprehension (included in Supplementary File 2). Finally, a survey code was provided at the end of the experiment. The participants could then use this code to complete the HIT task on the Mechanical Turk platform. A task was considered successful if the participant adhered to the rules and used the exact survey code provided by the experimenters. After experiment completion, participants were not allowed to take part in further experiments in this study.

EEG-based Reading Experiment

Participants

EEG data was recorded from 22 participants who were native English speakers, originating from the UK, Ireland, and the USA. All subjects had normal or corrected-to-normal vision and no neurological disorders, by self-reporting. The participants were employees or students at Dublin City University (18-51 years of age, mean age 23.86; 7 women, 15 men). Of these, two participants (males) were excluded from data analysis due to contaminated EEG from excessive eye movements. Every participant received a voucher gratuity of €25 upon completion of the experiment, which lasted approximately two hours. All participants provided informed consent for their participation, the re-use of the data prior to the start of the data collection, and were aware that they could withdraw participation at any point during the experiment.

Stimuli & Experimental Design

Two types of RSVP paradigm were used for collecting the DERCo dataset for each participant: classic RSVP and RSVP-with-flanker, with a stimulus presentation rate of 200 ms. This rate is due to the fact that when a person reads a text, a word is fixated on for around 200 to 250 ms before a saccade occurs and the next word is fixated on56,57. Both experimental conditions were similar, except that the RSVP-with-flanker used triple-word stimuli, whereas the classic RSVP presented only one word at a time.

While the cloze procedure was suitable for the online experiment, it was not appropriate for use in the EEG experiment. For accurate time-locking in EEG analyses, it was crucial to present stimuli at a steady rate, minimising the impact of eye movements on the EEG signals i.e. using RSVP36. More importantly, participants in the EEG-based reading experiment were not explicitly made aware that their word predictability was being investigated. We allocated 25 minutes per story in the cloze experiment because participants needed time to type their predictions. There were five short articles, each corresponding to one session. Each session lasted approximately 4 minutes (the shortest story) to 7 minutes (the longest story).

In the experiment, the title was displayed at the beginning of each session. The session then began with a blank screen and a fixation cross in the centre of the screen for 1000 ms. After that, words were presented as white letters in lowercase 60-px Arial font on a gray background on a 1920 x 1080 Full HD monitor. Each stimulus word was presented for 200 ms, followed by a 300 ms blank screen. In the RSVP-with-flanker, the preceding word and the target word were revealed, while the succeeding word was hidden by a string of Xs, corresponding in length to that of the masked word. PsychoPy58, an open-source package for creating experiments in behavioral science, was used for stimulus presentation. The RSVP experiment implementation is illustrated in Fig. 3.

Fig. 3.

Fig. 3

Open in a new tab

The procedure of word presentation for the first three words “Once upon a…” in (a) RSVP experiment (b) RSVP-with-flanker experiment.

Procedure

The EEG data collection took place at the School of Computing in Dublin City University. After participants had been given an overview of the experiment procedure, they were provided with a Plain Language Statement outlining the study’s objectives, the attention-related metrics to be recorded, the usage of the collected experimental data, the extent of their involvement in the research study, and details about the associated risks, benefits, and inclusion criteria. Subsequently, they were required to sign a written informed consent to participate in the study.

Participants were seated in a comfortable chair, approximately 100 cm in front of a 1920 x 1080 Full HD monitor in a semi-darkened sound-attenuated room. They were instructed to read (via RSVP) the five short stories in silence, where the first three stories were presented using RSVP (i.e. word-by-word display), and the last two stories were presented using RSVP-with-flankers (i.e. three consecutive words displayed at a time with the upcoming word represented with X’s). Participants completed a test trial before conducting the real experiment. Participants were asked to remain relaxed, and to minimise their movements during RSVP presentation, a common instruction to minimize movement artefacts in EEG recordings.

The stimuli were displayed following the description in the Stimuli & Experimental Design. After the final word disappeared, the participants answered comprehension questionnaires related to what they had just read. The primary purpose of these basic questions on the story content was to ensure the subject maintained their attention during the task i.e. the participants needed to pay attention in order to be able to answer these questions afterwards.

Sessions were self-paced i.e. participants could take a 5-10 minute rest between sessions until they were ready to proceed. After completing five sessions, the participants were required to answer certain questions about their experience during the experiment. The questionnaire for the five articles are included in Supplementary File 2.

Data acquisition

Behavioural Word-Prediction Dataset

We used the AWS service to save collected data from the HITs. Amazon S359 was used for storage to save the raw data after participants completed the task. Each survey code corresponded to a folder for each participant, and the raw data was saved in a CSV file.

EEG Dataset

The EEG data was recorded using an ActiCHamp EEG system60 with a 32-channel active electrode cap, with electrode positions following the international 10-20 system61. The signals were collected at a sampling rate of 1,000 Hz with a 500 Hz IIR (Infinite Impulse Response) low-pass filter, referenced to an internal reference (virtual ground) of the amplifier. The impedance of each electrode was carefully checked before recording to ensure good conductivity, and were checked again before each session to maintain all electrode impedances below 5kΩ. EEG signals were continuously recorded during each session. A photodiode connected to an external circuit was affixed to the top right hand corner of the screen and was used to generate markers (corresponding with word presentations) on the EEG recording62. This allowed for correctly aligned epochs corresponding to word presentations to be extracted after the experiment. Participants were not able to see the visual marker used to trigger the photodiode i.e. it was physically occluded.

Data preprocessing

Behavioural Word-Prediction Dataset

For each response in the prediction data, we remove punctuation and convert them to lowercase. Additionally, we also conducted checks to ensure the data quality. Namely, any spelling errors were corrected; for instance, ‘mornin’ is a typo for ‘morning’, ‘brige’ is a typo for ‘bridge’. Another scenario involves participants typing too quickly, pressing adjacent keys on their keyboards, resulting in grammatically incorrect words such as ‘sround’ instead of ‘around’ because the character ‘s’ is next to ‘a’. Importantly, these errors were not due to participants predicting incorrectly but would ultimately impact subsequent analyses. Generally, if a (misspelled) word only differed by one letter to its correct version (i.e. including a letter, removing a letter, or changing a single letter), the misspelling was rectified in the preprocessing stage.

Each response corresponded to a word at a unique location in the article. Keeping the words as indices ostensibly results in duplicates for analysis. Importantly, however, even though the words are the same, people may have different predictions or thoughts based on the context of the sentence, the topic, or their knowledge63,64. Thus, we combined the position and story index for each word to create new indices.

In order to organise the necessary information for use, we designed three schemas corresponding to questions and predictions. All 100 results for each story were also saved in CSV files. We have published the dataset (both raw and preprocessed data) on the Open Science Framework (OSF). Please refer to the OSF project65 for more information about data organisation, an introduction, size, authors, and schema of the DERCo dataset.

EEG Dataset

The raw EEG data was filtered with bandpass cutoffs at 0.1 Hz to 45 Hz using an FIR (Finite Impulse Response) filter with a Hamming window. Subsequently, using the triggers at stimulus onsets, we epoched the raw EEG data into several segments. These raw epochs were then input into an EEG epochs data preprocessing pipeline (as described in Fig. 4) to create preprocessed epochs. The description of the preprocessing is outlined below.

Fig. 4.

Fig. 4

Open in a new tab

The EEG epochs data preprocessing pipeline.

First of all, the raw EEG epochs were imported into the pipeline using the MNE package version 1.466. Common average referencing (CAR) was applied to all data. The next stage of the pipeline aimed to mitigate artifacts in the EEG recordings. The initial step involves outlier removal using Fully Automated Statistical Thresholding for EEG artifact Rejection (FASTER)67. Epochs were rejected based on a global threshold on the z-score, exceeding 3 standard deviations for epoch amplitude, deviation, and variance range. The outputs were then preprocessed by applying independent component analysis (ICA)68 to further improve the signal-to-noise ratio (SNR). The number of principal components generated from the the pre-whitening principal component analysis (PCA) step was selected based on the percentage of explainable variance, 99% was used in this study, which was accounted for by the principal components retained after dimensionality reduction.

ICA allows the decomposition of EEG signals into independent components (ICs), and then the corrected EEG signals can be obtained by discarding ICs containing artifacts69. The Picard algorithm70 was used for maximum likelihood independent component analysis. To automatically determine which ICs contain artifacts, we applied the ICLabel classifier71. The remaining ICs were then reconstructed to create clean EEG epochs.

Following the ICA process, Autoreject72 was applied for automatically identifying and removing artifacts. Autoreject uses unsupervised learning to estimate the rejection threshold for the epochs. In order to reduce computation time, which increased with the number of epochs and channels, we fitted Autoreject on a representative subset of epochs randomly picked from 25% of the total epochs. The number of cleaned epochs ultimately was less than or equaled the number of raw epochs. Importantly, this means for some word indices in stories there are no corresponding EEG epochs after trial rejection for some participants.

Data Records

The dataset is available on the OSF website65. There are two main folders corresponding to two experiments: the Behavioural Word-Prediction Experiment and the EEG-based Reading Experiment. The behavioural word-prediction experiment folder consists of three subfolders: “experienced_question”,"prediction”, and “question”. Each subfolder contains 5 CSV files corresponding to the participants’ responses across 5 articles. Supplementary File 2 contains all the questions asked in the experiment.

In the EEG-based reading experiment folder, there is an “answers” folder containing the answers of all participants collected after they completed reading each article. The main folder “EEG_data” has two subfolders: “raw” and “preprocessed”, which have the same structure. They contain 22 subfolders, each corresponding to a participant. Each participant’s folder has five subfolders for five articles, named “article_0”, “article_1”, etc. In each article folder, there is an epoch file in fif format. The directory tree for the repository of the DERCo project is shown in Fig. 8.

Fig. 8.

Fig. 8

Open in a new tab

The structure of the dataset.

Technical Validation

In this section, we examine next-word predictability - a key aspect in reading behaviour73 - to validate that the word-level predictability measures calculated in the behavioural word-prediction experiment dataset can be meaningfully integrated with the EEG-based reading experiment dataset. To achieve this, we created the DERCo dataset using data collected from two experiments: the “Behavioral Word-Prediction Experiment” and the “EEG-based Reading Experiment”. After applying the techniques described in the Data Preprocessing section, we obtained neural responses based on word-level and semantic-level information. However, we could not collect the exact predictions of the next words for each participant using the EEG-based reading experiment. Therefore, the behavioral word-prediction experiment was carried out with a large sample size to measure the predictability of each word (at a group level) in the texts. By combining the data collected from both experiments, we have EEG responses and predictability values for each word in the articles. The data in the behavioral word-prediction dataset was validated to demonstrate participants’ proficiency in predicting subsequent words and to assess participants’ reading comprehension of the articles. Given the significant results at the semantic level, we then analyse next-word predictability based on EEG signals at the word level using the predictability measures calculated from the behavioural word-prediction experiment dataset. This validation exercise is crucial to demonstrate that these two datasets are meaningfully aligned, ensuring they can be used together for effective analysis and investigation.

Behavioural Word-Prediction Dataset

Using the general questions in each story (Supplementary File 2), we estimated a participant’s prior knowledge of the recently read narratives and identified whether the articles used excessively advanced or unusual vocabulary. The result indicated that a significant majority of subjects, ranging from 98% to 100% of the total, had not read the stories before. Despite the unfamiliar stories, most participants agreed that no advanced or uncommon words were used. Only the final and longest story had agreement from 89% of participants, whereas the other articles achieved percentages between 95% and 98%.

As an estimate of participants’ overall reading comprehension (as a proxy measure to ensure participants correctly read the texts), a proportion was calculated based on the number of correct answers to multiple-choice questions for each subject. As a result, a substantial number of participants, between 98 and 100 participants, showed a performance greater than or equal to 60%, indicating their ability to comprehend and recall the stories’ content to provide accurate answers after reading.

The ability to anticipate upcoming words is commonly linked with the “cloze probability” or “cloze value”12,74,75. It is calculated as the proportion of participants that picked the same target word for the next word prediction. Cloze values range from 0 to 1, where when 0 indicates that no participant predicted the same next word, and 1 indicating all participants predicted the same next word. Notably, high predictability was observed for all words from all parts of speech in a sentence, and not just the last words in a sentence. More than 1000 words out of about 3,000 words in all articles had a cloze probability higher than 0.5. Of these, 62.74% of words were functions words, 20.32% were nouns, 6.24% were verbs, 5.35% were adjectives and verbs, and 5.35% were classified as other. The parts of speech were identified using the Python library IPA76. For further details, Table 2 provides further information about top-1 accuracy for each story and word categories, including content words (nouns, verbs, adjectives, and adverbs), function words, and others. The accuracy refers to the percentage of times participants correctly predicted the upcoming words in the story. We found that the number of sentences decreases in the order of articles 4, 5, 1, 2, and 3, corresponding to a reduction in the percentages of content words in the same sequence (38.77%, 35.48%, 31.67%, 24.46%, and 19.59%). As a result, articles with more sentences show higher accuracy in predicting content words (nouns, verbs, adjectives, and adverbs). For all words, we included the descriptive statistics in Supplementary Table S1. These findings suggest that humans are proficient in anticipating next words in narrative texts when asked to do so.

Table 2.

Human next-word prediction performance.

Story Accuracy No. words (p ≥ 0.5) % nouns % verbs % adjs & advs % function words % others
1 45.02% 180 23.89% 5.00% 2.78% 60.56% 7.78%
2 40.95% 139 15.83% 2.16% 6.47% 72.66% 2.88%
3 41.34% 148 14.19% 2.70% 2.70% 75.00% 5.41%
4 55.32% 294 21.77% 7.14% 9.86% 56.80% 4.42%
5 41.60% 248 22.18% 10.48% 2.82% 58.47% 6.05%
Total 45.17% 1009 20.32% 6.24% 5.35% 62.74% 5.35%
Open in a new tab

% shows the percentage of correctly predicted words for each story and part of speech. “No. words (p ≥ 0.5)” shows the number of words with a cloze probability above.5.

Neural Responses in Reading

The experimental design differed between the behavioural word-prediction experiment and the EEG-based reading experiment due to their distinct requirements, as mentioned in the procedure sections for both experiments. For the online experiment, the cloze procedure was appropriate, allowing participants to type their predictions for upcoming words. However, this method was unsuitable for the EEG-based reading experiment, which required precise time-locking of stimulus presentation to enable ERP analysis77. Despite the differences in the stimulus presentation paradigms between both experiments, we show significant results, demonstrating that word-level predictability measures calculated in the behavioural word-prediction experiment dataset can be meaningfully integrated with the EEG-based reading experiment dataset, as outlined below.

EEG data was segmented into epochs, starting 50 ms before and lasting 550 ms after the word-onset. Each grand-averaged word ERP was calculated by averaging the neural responses across participants for each channel generated by a given word. Figure 5 shows the EEG averages across all words for 32 electrodes in the first article between all subjects and a selected individual (user-id = JPY86). Consistent with the findings of the previous study by Dimigen et al.78, we can see a positive peak at around 100 ms post-word onset, also known as P100. A second positive peak (P200) can be identified at around 200 ms after word-onset.

Fig. 5.

Fig. 5

Open in a new tab

Butterfly plots for EEG averages across all words in article 1 (a) across subjects (b) single subject (JPY86). Each time-series line graph corresponds to an electrode.

For a technical validation, that incorporated both the behavioural word-prediction dataset (Mechanical Turk) with the EEG dataset, we investigated N400-like components within the 300 - 500 ms time window. The ERP was examined across midline electrodes Fz (Midline Frontal), Cz (Midline Central), Pz (Midline Parietal), and Oz (Midline Occipital). Using midline electrodes in EEG studies provides balanced recordings of brain activity. They capture significant neural activity from central brain regions involved in cognitive processes, reduce hemispheric lateralisation effects, and ensure consistency and comparability due to their standardized use across studies79.

We used two non-overlapping groups of low predictable words (cloze p ≤ 0.06, n = 989), and high predictable words (cloze p > 0.52, n = 962), calculated from the Mechanical Turk dataset. Figure 6 provides a more detailed view of the topographies of response differences between conditions across time, computed averages of 300 - 500 ms range in RSVP and RSVP with flankers, respectively. The color intensity indicates the mean ERP amplitude across all participants for the low and high predictability groups. As shown in Figure 6, the effect of word predictability actually modulated the ERP evoked after the word onset.

Fig. 6.

Fig. 6

Open in a new tab

Scalp distributions of the response for low predictable words, high predictable words, and the difference between low and high predictable words (low minus high) in (a) RSVP experiment and (b) RSVP-with-flanker experiment. Amplitudes (μV) were averaged in 300-500 time window (for all electrodes).

In order to capture variability in cloze values, we applied an independent samples t-test on the high and low predictability, and plotted the ERP waveforms between these conditions to examine ERP activity in the time window of 300 - 500 ms. Results from the t-test on midline electrodes and the corresponding visualisations are presented in Fig. 7. Post-hoc t-tests revealed significant differences (p < 0.05) between high and low predictability at all mid-line electrodes in between 300 and 500 ms post stimulus in both RSVP and RSVP-with-flankers paradigms. If conducting t-tests for all channels, approximately 70% electrodes showed significant differences (p < 0.05) in both experiments. In the RSVP paradigm, nine electrodes (FC1, C3, TP9, P7, TP10, T8, FT10, FC6, and F4) had no predictability effect in the 300 - 500 ms range between two conditions. The predictability effect was not found for ten channels (Fp1, FC1, C3, P7, TP10, T8, FT10, FC6, F4, and Fp2) in the RSVP-with-flanker experiment at the same time window.

Fig. 7.

Fig. 7

Open in a new tab

ERP activity for midline electrodes in RSVP experiment (left) and RSVP with flankers experiment (right). The time range of significant differences (p < 0.05) between 300ms and 500ms post-stimulus are indicated.

These validation results show one potential analysis trajectory that necessitates the use of both the behavioural word-prediction dataset and the EEG data. Moreover, it validates that the word-level predictability measures calculated in the behavioural word-prediction experiment dataset can be meaningfully integrated with the EEG-based reading experiment dataset. Importantly, such analyses could not be performed with one dataset alone.

Limitations

Although the EEG and behavioral word-prediction datasets provide a novel perspective on reading behaviors, some limitations still exist. Firstly, while the experimental design for EEG collection simulates a closer approximation of the natural reading scenario, it is still not completely naturalistic. However, because we focused on word-level EEG responses, a free-viewing paradigm was deemed unsuitable, and moreover, due to how eye movements generate artefacts in EEG. This is because people tend to skip easily predictable80 or high-frequency words81. Additionally, temporal overlap occurs between the ERPs during longer saccades or regressions in a sentence when using free-viewing paradigms82. Thus, researchers need to consider this when proposing hypotheses using this dataset.

Secondly, another limitation of the experimental setup is that instead of using two different groups of participants for the EEG experiment, one for the classic RSVP experiment and another for the RSVP-with-flankers experiment, we used the same participants for both. We used RSVP for the first three stories and RSVP-with-flankers for the remaining two, which reduced the total number of words available for analysis in each condition. This decision allowed for the EEG dataset to support analysis of the two RSVP conditions, whilst maintaining a sufficient number of participants for each.

Lastly, our EEG experiment participant pool was limited due to an unbalanced division of gender (approximately 68% male and 32% female). This gender imbalance should be remedied in future to achieve equal representation of each gender, because this gender distribution could have skewed results or prevented an effect from becoming significant due to lack of power83. However, our research primarily aimed to analyse next-word prediction behavior in adults aged 18 years and older, rather than focusing on gender differences. It should be acknowledged that several similar brain activity studies in reading experiments have also had unbalanced sample sizes for gender2,27,46–48.

Supplementary information

Questionnaire for the five articles (52.5KB, docx)
Transcripts of five articles (20.1KB, docx)
41597_2024_3915_MOESM3_ESM.jpg (536.1KB, jpg)

Online Consent Form for Behavioural Word-Prediction Experiment

Human next-word prediction performance (9.1KB, xlsx)

Acknowledgements

We thank all the participants who generously participated in this research. We thank all the research staff involved in participant recruitment, data collection. This publication has emanated from research conducted with the financial support of Science Foundation Ireland under Grant number 18/CRT/6183 and 13/RC/2106_P2. For the purpose of Open Access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission.

Code availability

All the scripts for preprocessing can be found at https://github.com/Tayerquach/DERCo.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

The online version contains supplementary material available at 10.1038/s41597-024-03915-8.

References

  • 1.Deniz, F., Nunez-Elizalde, A. O., Huth, A. G. & Gallant, J. L. The representation of semantic information across human cerebral cortex during listening versus reading is invariant to stimulus modality. Journal of Neuroscience39, 7722–7736 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Dufau, S., Grainger, J., Midgley, K. J. & Holcomb, P. J. A thousand words are worth a picture: Snapshots of printed-word processing in an event-related potential megastudy. Psychological science26, 1887–1897 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Davis, C. E., Hauf, J. D., Wu, D. Q. & Everhart, D. E. Brain function with complex decision making using electroencephalography. International journal of psychophysiology79, 175–183 (2011). [DOI] [PubMed] [Google Scholar]
  • 4.Pfeiffer, C., Hollenstein, N., Zhang, C. & Langer, N. Neural dynamics of sentiment processing during naturalistic sentence reading. NeuroImage218, 116934 (2020). [DOI] [PubMed] [Google Scholar]
  • 5.Murphy, A., Bohnet, B., McDonald, R. & Noppeney, U. Decoding Part-of-Speech from Human EEG Signals. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2201–2210 (2022).
  • 6.Antúnez, M., Milligan, S., Hernández-Cabrera, J. A., Barber, H. A. & Schotter, E. R. Semantic parafoveal processing in natural reading: Insight from fixation-related potentials & eye movements. Psychophysiology59, e13986 (2022). [DOI] [PubMed] [Google Scholar]
  • 7.Troyer, M., Kutas, M., Batterink, L. & McRae, K. Nuances of knowing: Brain potentials reveal implicit effects of domain knowledge on word processing in the absence of sentence-level knowledge. Psychophysiology e14422 (2023). [DOI] [PubMed]
  • 8.Duan, Y., Zhou, C., Wang, Z., Wang, Y.-K. & Lin, C.-t. DeWave: Discrete Encoding of EEG Waves for EEG to Text Translation. In Thirty-seventh Conference on Neural Information Processing Systems (2023).
  • 9.Menon, R. S. & Kim, S.-G. Spatial and temporal limits in cognitive neuroimaging with fMRI. Trends in cognitive sciences3, 207–216 (1999). [DOI] [PubMed] [Google Scholar]
  • 10.Loued-Khenissi, L., Döll, O. & Preuschoff, K. An overview of functional magnetic resonance imaging techniques for organizational research. Organizational Research Methods22, 17–45 (2019). [Google Scholar]
  • 11.Baillet, S. Magnetoencephalography for brain electrophysiology and imaging. Nature neuroscience20, 327–339 (2017). [DOI] [PubMed] [Google Scholar]
  • 12.Kutas, M. & Hillyard, S. A. Brain potentials during reading reflect word expectancy and semantic association. Nature307, 161–163 (1984). [DOI] [PubMed] [Google Scholar]
  • 13.Van Petten, C. & Kutas, M. Interactions between sentence context and word frequency in event-related brain potentials. Memory & cognition18, 380–393 (1990). [DOI] [PubMed] [Google Scholar]
  • 14.Kutas, M. In the company of other words: Electrophysiological evidence for single-word and sentence context effects. Language and cognitive processes8, 533–572 (1993). [Google Scholar]
  • 15.Laszlo, S. & Federmeier, K. D. A beautiful day in the neighborhood: An event-related potential study of lexical relationships and prediction in context. Journal of Memory and Language61, 326–338 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Laszlo, S. & Federmeier, K. D. The N400 as a snapshot of interactive processing: Evidence from regression analyses of orthographic neighbor and lexical associate effects. Psychophysiology48, 176–186 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Kutas, M. & Federmeier, K. D. Thirty years and counting: finding meaning in the N400 component of the event-related brain potential (ERP). Annual review of psychology62, 621–647 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Aurnhammer, C., Delogu, F., Schulz, M., Brouwer, H. & Crocker, M. W. Retrieval (N400) and integration (P600) in expectation-based comprehension. Plos one16, e0257430 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Heilbron, M., Armeni, K., Schoffelen, J.-M., Hagoort, P. & De Lange, F. P. A hierarchy of linguistic predictions during natural language comprehension. Proceedings of the National Academy of Sciences119, e2201968119 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.He, T., Boudewyn, M. A., Kiat, J. E., Sagae, K. & Luck, S. J. Neural correlates of word representation vectors in natural language processing models: Evidence from representational similarity analysis of event-related brain potentials. Psychophysiology59, e13976 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Goldstein, A. et al. Shared computational principles for language processing in humans and deep language models. Nature neuroscience25, 369–380 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Guo, Y., Liu, T., Zhang, X., Wang, A. & Wang, W. End-to-end translation of human neural activity to speech with a dual–dual generative adversarial network. Knowledge-based systems277, 110837 (2023). [Google Scholar]
  • 23.Hollenstein, N. et al. Decoding EEG brain activity for multi-modal natural language processing. Frontiers in Human Neuroscience 378 (2021). [DOI] [PMC free article] [PubMed]
  • 24.Schwartz, D. & Mitchell, T. Understanding language-elicited EEG data by predicting it from a fine-tuned language model. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 43–57 (2019).
  • 25.Sulpizio, S., Arcara, G., Lago, S., Marelli, M. & Amenta, S. Very early and late form-to-meaning computations during visual word recognition as revealed by electrophysiology. Cortex157, 167–193 (2022). [DOI] [PubMed] [Google Scholar]
  • 26.MNE Developers: Kiloword Dataset. https://mne.tools/stable/documentation/datasets.html#kiloword-dataset.
  • 27.Hollenstein, N. et al. ZuCo, a simultaneous EEG and eye-tracking resource for natural sentence reading. Scientific data5, 1–13 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Hollenstein, N., Troendle, M., Zhang, C. & Langer, N. ZuCo 2.0: A Dataset of Physiological Recordings During Natural Reading and Annotation. In Proceedings of the 12th Language Resources and Evaluation Conference, 138–146 (2020).
  • 29.Hollenstein, N., Pirovano, F., Zhang, C., Jäger, L. & Beinborn, L. Multilingual Language Models Predict Human Reading Behavior. In 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, 106–123 (Association for Computational Linguistics (ACL), 2021).
  • 30.Wang, Z. & Ji, H. Open vocabulary electroencephalography-to-text decoding and zero-shot sentiment classification. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, 5350–5358 (2022).
  • 31.Luck, S. J. & Kappenman, E. S.The Oxford handbook of event-related potential components (Oxford university press, 2013).
  • 32.Potter, M. C. Rapid serial visual presentation (RSVP): A method for studying language processing. In New methods in reading comprehension research, 91–118 (Routledge, 2018).
  • 33.Dambacher, M. et al. Stimulus onset asynchrony and the timeline of word recognition: Event-related potentials during sentence reading. Neuropsychologia50, 1852–1870 (2012). [DOI] [PubMed] [Google Scholar]
  • 34.Barber, H. A., van der Meij, M. & Kutas, M. An electrophysiological analysis of contextual and temporal constraints on parafoveal word processing. Psychophysiology50, 48–59 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Kosch, T., Schmidt, A., Thanheiser, S. & Chuang, L. L. One does not simply RSVP: mental workload to select speed reading parameters using electroencephalography. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1–13 (2020).
  • 36.Plöchl, M., Ossandón, J. P. & König, P. Combining EEG and eye tracking: identification, characterization, and correction of eye movement artifacts in electroencephalographic data. Frontiers in human neuroscience6, 278 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Rayner, K. Eye movements in reading and information processing: 20 years of research. Psychological bulletin124, 372 (1998). [DOI] [PubMed] [Google Scholar]
  • 38.Zhang, W., Li, N., Wang, X. & Wang, S. Integration of sentence-level semantic information in parafovea: Evidence from the RSVP-flanker paradigm. PloS one10, e0139016 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Stites, M. C., Payne, B. R. & Federmeier, K. D. Getting ahead of yourself: Parafoveal word expectancy modulates the N400 during sentence reading. Cognitive, Affective, & Behavioral Neuroscience17, 475–490 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Payne, B. R., Stites, M. C. & Federmeier, K. D. Event-related brain potentials reveal how multiple aspects of semantic processing unfold across parafoveal and foveal vision during sentence reading. Psychophysiology56, e13432 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.Li, N., Dimigen, O., Sommer, W. & Wang, S. Parafoveal words can modulate sentence meaning: Electrophysiological evidence from an RSVP-with-flanker task. Psychophysiology59, e14053 (2022). [DOI] [PubMed] [Google Scholar]
  • 42.Amazon Mechanical Turk. https://www.mturk.com/.
  • 43.Bianchi, B. et al. Human and computer estimations of Predictability of words in written language. Scientific reports10, 4396 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Murray, H. & El-Leithy, S. Behavioural experiments in cognitive therapy for posttraumatic stress disorder: why, when, and how? Verhaltenstherapie31, 50–60 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45.Gagné, N. & Franzen, L. How to run behavioural experiments online: Best practice suggestions for cognitive psychology and neuroscience. Swiss Psychology Open: the official journal of the Swiss Psychological Society3 (2023).
  • 46.Bhattasali, S., Brennan, J., Luh, W.-M., Franzluebbers, B. & Hale, J. The Alice Datasets: fMRI & EEG observations of natural language comprehension. In Proceedings of the Twelfth Language Resources and Evaluation Conference, 120–125 (2020).
  • 47.Li, J. et al. Le Petit Prince multilingual naturalistic fMRI corpus. Scientific data9, 530 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48.Troyer, M. & Kutas, M. Harry Potter and the Chamber of What?: The impact of what individuals know on word processing during reading. Language, cognition and neuroscience35, 641–657 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49.Wu, S., Ramdas, A. & Wehbe, L. Brainprints: identifying individuals from magnetoencephalograms. Communications Biology5, 852 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50.Armeni, K., Güçlü, U., van Gerven, M. & Schoffelen, J.-M. A 10-hour within-participant magnetoencephalography narrative dataset to test models of language comprehension. Scientific Data9, 278 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51.De Leeuw, J. R. jsPsych: A JavaScript library for creating behavioral experiments in a Web browser. Behavior research methods47, 1–12 (2015). [DOI] [PubMed] [Google Scholar]
  • 52.Taylor, W. L. "Cloze procedure”: A new tool for measuring readability. Journalism quarterly30, 415–433 (1953). [Google Scholar]
  • 53.de Varda, A. G., Marelli, M. & Amenta, S. Cloze probability, predictability ratings, and computational estimates for 205 English sentences, aligned with existing EEG and reading time data. Behavior Research Methods 1–24 (2023). [DOI] [PMC free article] [PubMed]
  • 54.Frank, S. L., Fernandez Monsalve, I., Thompson, R. L. & Vigliocco, G. Reading time data for evaluating broad-coverage models of English sentence processing. Behavior research methods45, 1182–1190 (2013). [DOI] [PubMed] [Google Scholar]
  • 55.Frank, S. L., Otten, L. J., Galli, G. & Vigliocco, G. The ERP response to the amount of information conveyed by words in sentences. Brain and language140, 1–11 (2015). [DOI] [PubMed] [Google Scholar]
  • 56.Dambacher, M. & Kliegl, R. Synchronizing timelines: Relations between fixation durations and N400 amplitudes during sentence reading. Brain research1155, 147–162 (2007). [DOI] [PubMed] [Google Scholar]
  • 57.Brysbaert, M. How many words do we read per minute? A review and meta-analysis of reading rate. Journal of memory and language109, 104047 (2019). [Google Scholar]
  • 58.PsychoPy: The Open-Source Software for Running Psychology Experiments. https://www.psychopy.org/.
  • 59.Amazon Web Services. Amazon Simple Storage Service (S3). https://aws.amazon.com/s3/.
  • 60.Brain Products. ActiCHamp. https://www.brainproducts.com/solutions/actichamp/.
  • 61.Klem, G. H. The ten-twenty electrode system of the international federation. The international federation of clinical neurophysiology. Electroencephalogr. Clin. Neurophysiol. Suppl.52, 3–6 (1999). [PubMed] [Google Scholar]
  • 62.Wang, Z., Healy, G., Smeaton, A. F. & Ward, T. E. An investigation of triggering approaches for the rapid serial visual presentation paradigm in brain computer interfacing. In 2016 27th Irish Signals and Systems Conference (ISSC), 1–6 (IEEE, 2016).
  • 63.Altmann, G. T. & Mirković, J. Incrementality and prediction in human sentence processing. Cognitive science33, 583–609 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 64.Kutas, M., DeLong, K. A. & Smith, N. J. A look around at what lies ahead: Prediction and predictability in language processing. Predictions in the brain: Using our past to generate a future190207 (2011).
  • 65.Quach, B. M. DERCo: A Dataset for Human Behaviour in Reading Comprehension Using EEG. OSF 10.17605/OSF.IO/RKQBU (2024). [DOI] [PMC free article] [PubMed]
  • 66.MNE Developers. MNE-Python: Software for processing MEG and EEG data. https://mne.tools/1.4/index.html.
  • 67.Nolan, H., Whelan, R. & Reilly, R. B. FASTER: fully automated statistical thresholding for EEG artifact rejection. Journal of neuroscience methods192, 152–162 (2010). [DOI] [PubMed] [Google Scholar]
  • 68.Jung, T.-P. et al. Removing electroencephalographic artifacts by blind source separation. Psychophysiology37, 163–178 (2000). [PubMed] [Google Scholar]
  • 69.Hyvärinen, A. & Oja, E. Independent component analysis: algorithms and applications. Neural networks13, 411–430 (2000). [DOI] [PubMed] [Google Scholar]
  • 70.Ablin, P., Cardoso, J.-F. & Gramfort, A. Faster independent component analysis by preconditioning with Hessian approximations. IEEE Transactions on Signal Processing66, 4040–4049 (2018). [Google Scholar]
  • 71.Pion-Tonachini, L., Kreutz-Delgado, K. & Makeig, S. ICLabel: An automated electroencephalographic independent component classifier, dataset, and website. NeuroImage198, 181–197 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 72.Jas, M., Engemann, D. A., Bekhti, Y., Raimondo, F. & Gramfort, A. Autoreject: Automated artifact rejection for MEG and EEG data. NeuroImage159, 417–429 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 73.Smith, F. The role of prediction in reading. Elementary English52, 305–311 (1975). [Google Scholar]
  • 74.Luke, S. G. & Christianson, K. Limits on lexical prediction during reading. Cognitive psychology88, 22–60 (2016). [DOI] [PubMed] [Google Scholar]
  • 75.Szewczyk, J. M. & Federmeier, K. D. Context-based facilitation of semantic access follows both logarithmic and linear functions of stimulus probability. Journal of memory and language123, 104311 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 76.Orlando, R. IPA: Incremental Parsing and Alignment. https://github.com/Riccorl/ipa.
  • 77.Luck, S.An introduction to the event-related potential technique (MIT press, 2014).
  • 78.Dimigen, O., Sommer, W., Hohlfeld, A., Jacobs, A. M. & Kliegl, R. Coregistration of eye movements and EEG in natural reading: analyses and review. Journal of experimental psychology: General140, 552 (2011). [DOI] [PubMed] [Google Scholar]
  • 79.Mecarelli, O. Electrode placement systems and montages. Clinical Electroencephalography 35–52 (2019).
  • 80.Staub, A. The effect of lexical predictability on eye movements in reading: Critical review and theoretical interpretation. Language and Linguistics Compass9, 311–327 (2015). [Google Scholar]
  • 81.Drieghe, D., Brysbaert, M., Desmet, T. & De Baecke, C. Word skipping in reading: On the interplay of linguistic and visual factors. European Journal of Cognitive Psychology16, 79–103 (2004). [Google Scholar]
  • 82.Dimigen, O. & Ehinger, B. V. Regression-based analysis of combined EEG and eye-tracking data: Theory and applications. Journal of Vision21, 3–3 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 83.Dickinson, E. R., Adelson, J. L. & Owen, J. Gender balance, representativeness, and statistical power in sexuality research using undergraduate student samples. Archives of Sexual Behavior41, 325–327 (2012). [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Questionnaire for the five articles (52.5KB, docx)
Transcripts of five articles (20.1KB, docx)
41597_2024_3915_MOESM3_ESM.jpg (536.1KB, jpg)

Online Consent Form for Behavioural Word-Prediction Experiment

Human next-word prediction performance (9.1KB, xlsx)

Data Availability Statement

All the scripts for preprocessing can be found at https://github.com/Tayerquach/DERCo.

Articles from Scientific Data are provided here courtesy of Nature Publishing Group

RESOURCES