This research shows that function words can be useful as features for machine learning models tasked with detecting conspiratorial content in COVID-19 related Twitter posts. A significance test exposes that the distribution of function words between fake and legitimate content varies greatly. Further, a support vector machine classifier is demonstrated to perform above chance when using function word-only features, achieving a Matthews correlation coeficient of 0.139 on unseen test data.
Previous research into detecting conspiratorial online content
indicates that it can be distinguished by the author’s writing style
[
Fortunately, recent years have seen a growing body of research
on the importance of stylistic features in misinforming and
conspiratorial content. For instance, Posadas-Durán et al. [
For social media and especially Twitter data, the literature is
rather sparse. Del Tredici and Fernández [
Word my is the they used am during their he she
P-Value # No Consp. # Consp. stems from a Russian troll account, which while related, is still not focused on fake news detection specifically.
This research thus aims to expand on the available literature by analysing how function word usage is distributed between authors of conspiratorial versus legitimate content, and quantifying whether function words alone can act as suficient features in the classification of fake news content. 2
The data were provided through the MediaEval 2021 Conference,
for the task ‘FakeNews: Corona Virus and Conspiracies Multimedia
Analysis’ [
To investigate a possible diference in distribution between conspiratorial and non-conspiratorial content, most of this research thus focused on classes 1 and 3. First, function words were extracted from all data using spaCy.2 To test for significance, a 2 test was performed on the distribution of function words between the classes 1 and 3.
Next, the usefulness of the extracted function words was tested. As a representative model, a support vector machine (SVM) with non-linear kernels was chosen, and evaluated using classicfiation 1All posts are written in English and were collected between January 17, 2020 and June 30, 2021. 2For the full list of function words, see https://github.com/explosion/spaCy/blob/ master/spacy/lang/en/stop_words.py accuracy and Matthews correlation coeficient (MCC). Two scenarios were investigated, one with data containing all three classes, and one for data containing the classes 1 and 3 only. To account for random efects, 100 runs were computed for both scenarios, each with a random 10% validation split. As a final evaluation, a single SVM model was fitted on all available data, and evaluated on an unseen test set. Due to organisational means, this evaluation is only available for the 3-class case using MCC. Preprocessing of function words was done using the TfidfVectorizer of SciKit Learn.3 3
The results of the 2 test can be seen in Table 1, showing the 10 function words with lowest p-value, all of which are below 1%. 5 are pronouns (50%), which is higher than the overall frequency of pronouns in the function words (≈11.48%). Only 2 out of the 10 words occur more often in class 3 (conspiracy), while the remaining 8 occur more often in class 1 (no conspiracy).
The following are all sub 5% function words, 40 in total, sorted from lowest to highest p-value: my, is, the, they, used, am, during, their, he, she, by, this, him, serious, doing, might, his, if, us, but, be, these, all, seem, about, part, her, along, could, your, due, have, are, here, using, at, per, when, would, now
Of these words, 12 are pronouns (30%), marked in italic. 7 belong to class 3, marked in bold, while 33 belong to class 1.
Table 2 shows the results of the SVM classifiers. 4
The results of the 2 test show that a significant diference in the distribution of function words can be observed between conspiratorial and non-conspiratorial content, confirming the idea that such words are indeed important for distinguishing fake from legitimate content. This is further supported by the classification results, where an above chance performance on unseen test data was achieved. Unsurprisingly, the classification performance in the binary case was higher than in the full case, since an overlap in style between non-conspiratorial content as well as content which does not actively support conspiracies, but merely discusses them, is to be expected.
Interestingly, the category of function words most common in the
list of sub 5% p-value words were pronouns, for which the relative
frequency was greatly increased compared to their frequency in
all function words. Further, all of these pronouns except their and
these, occurred more often in the non-conspiracy category. Most
prominent are third person singular pronouns (he, she, him, his), all
featured more in class 1, which indicates that conspiracy authors
are less likely to talk about a person at length. This could be because
giving extensive detail (e.g. ‘She said X ’) rather than implying makes
their claims falsifiable, which they might be interested to avoid [
Apart from pronouns, the overall majority of function words with p-values below 5% belong to the non-conspiratorial class. This indicates that authors of fake content use a more simplistic style, as the complexity of a text correlates with the number of diferent function words used.
An earlier analysis on a smaller subset of the data showed different patterns in the function word distributions, most notably the presence of ‘hedging’ words like quite, rather and somehow. However, these patterns disappeared when the larger data set was released. Therefore, it is important to note that the data set at hand, with only 1554 total posts, is a very limited subset of all COVID-19 related data found on Twitter. Thus, it cannot be ruled out that the patterns found in this research, although powerful in predicting on the chosen dataset, may not generalise.
This limitation is extended by the fact that the data analysed in this report does not contain author information. As stylistic information correlates very strongly with the author of a text, the patterns found could, in theory, be caused by a few authors having a disproportionately high representation in the data. This efect unfortunately could not be accounted for due to the missing authorship information.
In conclusion, this research has shown that function words are a strong proxy for detecting conspiratorial content in the context of COVID-19 related fake news on Twitter. To address the limitations of this research, future work should explore in how far these results generalise to larger corpora of Twitter data, diferent domains of conspiracy, as well as other social media platforms.
I thank Martha Larson for her critical input during the development of the research question as well as the analysis process, and Lynn de Rijk for her input on function word usage and interesting lexical patterns in fake news content.