The Small World of Words project (SWOW) project is a scientific project to map word meaning in various languages. In contrast to dictionaries, it focusses on the aspects of word meaning that are shared between people without imposing restrictions on what aspects of meaning should be considered. The methodology is based on a continued word association task, in which participants see a cue word and are asked to give three associated responses to this cue word. The current version includes over 3 million responses obtained from over 90,000 participants, for more than 12,000 cues.
In this repository you will find a basic analysis pipeline for the English SWOW project which allows you to import an preprocessing the data as well as compute some basic statistics.
Suggestions are always appreciated, and do not hesitate to get in touch if you any questions.
Obtaining the data
In addition to the scripts, you will need to retrieve the word association data.
Currently word association and participant data is available for 12,292 cues. The data consists of over 3 million reponses collected between 2011 and 2018. They are currently submitted for publication. Note that the final version is subject to change.
If you want to use these data for your own research, you can obtain them from the Small World of Words research page. Download the English data (the Dutch data still need to be updated to be used with an R pipeline) and put the file in the
processed folder (depending on the file you got).
Please note that data themselves are licensed under Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License. They cannot be redistributed or used for commercial purposes.
To cite these data: De Deyne, S., Navarro, D., Perfors, A., Brysbaert, M. & Storms, G. (2018). Measuring the associative structure of English: The “Small World of Words” norms for word association.Manuscript submitted for publication.
If you find any of this useful, please consider sharing the word association study.
The datafile consists of participant information about age, gender, native language and location. For a subset of the data, we also provided information about education (we only started collecting this later on).
participantID: unique identifier for the participant
created_at: time and date of participation
age: age of the participant
nativeLanguage: native language from a short list of common languages
gender: gender of the participant (Female / Male / X)
education: Highest level of education: 1 = None, 2 = Elementary school, 3 = High School, 4 = College or University Bachelor, 5 = College or University Master
city: city (city location when tested, might be an approximation)
country: country (country location when tested)
Word association data
The raw data consist of the original responses and spell checked responses. The spell-checked was performed at the server side and for now this script is not included in the current repository.
However, you can find a list spelling corrections and English capitalized words in the
section: identifier for the snowball iteration (e.g. set2017)
cue: cue word
R1Raw: raw primary associative response
R2Raw: raw secondary associative response
R3Raw: raw tertiary associative response
R1: corrected primary associative response
R2: corrected secondary associative response
R3: corrected tertiary associative response
The preprocessed data consist of normalizations of cues and responses by spell-checking them, correcting capitalization and Americanizing. This file is generated by the preprocessData.R script.
In addition to normalizing cues and responses, this script will also extract a balanced dataset, in which each cue is judged by exactly 100 participants. Because each participant generated three responses, this means each cue has 300 associations. The participants were selected to favor native speakers.
In many cases, this preprocessed data is used to derive the associative strengths (i.e. the conditional probability of a response given a cue). These data can be derived using the createAssoStrengthTable.R script.
Use createSWOWENGraph.R to extract the largest strongly connected component for graphs based on the first response (R1) or all responses (R123). The results are written to output/adjacencyMatrices and consist of a file with labels and a sparse file consisting of three values corresponding to row- and column-indices followed by the association frequencies.
In most cases, associative frequencies will need to be converted to associative strengths by dividing with the sum of all strengths for a particular cue.
Vertices that are not part of the largest connected component are listed in a report in the
Use createResponseStats.R to calculate a number of response statistics. Currently the script calculates the number of types, tokens and hapax legomena responses (responses that only occur once). The results can be found in the
Use createCueStats.R to calculate cue statistics. Only words that are part of the strongly connected component are considered. Results are provided for the R1 graph and the graph with all responses (R123). The file includes the following:
coverage: (how many of the responses are retained in the graph after removing those words that aren't a cue or aren't part of the strongest largest component).
H: Shannon entropy of response distributions for each cue
unknown: the number of unknown responses
x.R2: the number of missing R2 responses
x.R3: the number of missing R3 responses
R1 - R2 response chaining
Later responses can be affected by the previous response a participant gave. In general, this is quite rare, but for some cues this effect can be more pronounced. To investigate response chaining, we compare the conditional probabilities of the second response when preceeded with a mediated R1 response with conditional probabilities when R2 is not preceeded by this mediator. An example of this analysis is available in calculateR12ResponseChaining.R.
Spelling and lexica
We tried to check the spelling of the most common responses (those occurring at least two times in the data), but it's quite likely that some corrections can be improved and some misspellings are missed. This is where git can make our lives a bit easier. If you find errors, please check the correction file and submit a pull request with additional or ammended corrections.
Two files are of importance:
- EnglishProperNames.txt: List of proper names that should not be corrected when found as response
- EnglishCustomDict.txt responses that are manually checked. The data in these files take priority over automated (and sometimes faulty) spell-checking. As such, exceptions that should not be touched can be easily included by given including the original response and a correction that is identical. The current file consists of over 11,000 corrections.
The spelling list is merged with SUBTLEX-US and words from the VARCON list to obtain an English lexicon file. This file is used check responses at the individual level and remove participants who provide predominantly non-English responses in the preprocessData.R script.