Paper: http://ceur-ws.org/Vol-2882/paper58.pdf
Lynn de Rijk : You said it? How mis- and disinformation tweets surrounding the Corona-5G-conspiracy communicate through implying. Proc. of MediaEval 2020, 14-15 December 2020, Online.
A qualitative analysis of 130 mis/disinformation tweets regarding the corona-5G-conspiracy, shows that often meaning is implied, by leaving out coherence markers, putting the words in someone else's mouth through citing and ambiguous phrasing/punctuation.
Presented by: Lynn de Rijk
You said it? How mis- and disinformation tweets surrounding the Corona-5G-conspiracy communicate through implying.
1. You said it? How mis- and disinformation
tweets surrounding the corona-5G-conspiracy
communicate through implying
Lynn de Rijk, Radboud University
Many thanks to Martha Larson, Zhengyu Zhao and John Keates
3. Introduction
What is challenging for automated systems?
> Recognizing implied meaning
What is implied meaning?
- Clear example: sarcasm/irony;
- Also common in everyday communication in less obvious ways.
5. Speech Act Theory
1) Locution (direct SA)
Literal meaning of the utterance.
2) Illocution (indirect SA)
Implied meaning of the utterance.
3) Perlocution
What happens as a result of the utterance.
“Peter, you are standing on
my foot.”
6. Speech Act Theory
1) Locution (direct SA)
Asserting state of affairs.
2) Illocution (indirect SA)
3) Perlocution
“Peter, you are standing on
my foot.”
7. Speech Act Theory
1) Locution (direct SA)
Asserting state of affairs.
2) Illocution (indirect SA)
Requesting Peter to place his foot elsewhere.
3) Perlocution
“Peter, you are standing on
my foot.”
8. Speech Act Theory
1) Locution (direct SA)
Asserting state of affairs.
2) Illocution (indirect SA)
Requesting Peter to place his foot elsewhere.
3) Perlocution
Peter’s action (e.g., moving his foot and/or
apologizing).
“Peter, you are standing on
my foot.”
9. Approach
Research questions
1. How prevalent is implied meaning in mis-/disinformation tweets surrounding the corona-
5G-conspiracy?
2. Are there linguistic cues that signal implied meaning?
Data (n=90)
FakeNews Task Dataset – 5G-corona-conspiracy tweets.
Method (qualitative approach)
Coding
- Speech Acts (direct and indirect);
- Linguistic features.
10. “Do you know what time it is?”
Coding Indirect Speech Acts
11. “Do you know what time it is?”
“Yes, I know.”
Coding Indirect Speech Acts
12. “Do you know what time it is?” WHY
“Yes, I know.”
Coding Indirect Speech Acts
13. Anyone else curious about the majority of
deaths in china seem to be the same areas
they rolled out their stand alone 5G just a
couple months ago. Verry few deaths being
reported in other areas in comparison. #5G
#CoronavirusOutbreak #COVID19
#5gamechanger
EXAMPLE 1
CODES
14. Anyone else curious about the majority of
deaths in china seem to be the same areas
they rolled out their stand alone 5G just a
couple months ago. Verry few deaths being
reported in other areas in comparison. #5G
#CoronavirusOutbreak #COVID19
#5gamechanger
EXAMPLE 1
CODES
• Direct SA: asking
15. Anyone else curious about the majority of
deaths in china seem to be the same areas
they rolled out their stand alone 5G just a
couple months ago. Verry few deaths being
reported in other areas in comparison. #5G
#CoronavirusOutbreak #COVID19
#5gamechanger
EXAMPLE 1
CODES
• Direct SA: asking
• Direct SA: asserting
16. Anyone else curious about the majority of
deaths in china seem to be the same areas
they rolled out their stand alone 5G just a
couple months ago. Verry few deaths being
reported in other areas in comparison. #5G
#CoronavirusOutbreak #COVID19
#5gamechanger
EXAMPLE 1
CODES
• Direct SA: asking
• Direct SA: asserting
• Indirect SA: ?
17. Anyone else curious about the majority of
deaths in china seem to be the same areas
they rolled out their stand alone 5G just a
couple months ago. Verry few deaths being
reported in other areas in comparison. #5G
#CoronavirusOutbreak #COVID19
#5gamechanger
EXAMPLE 1
CODES
• Direct SA: asking
• Direct SA: asserting
• Indirect SA: ?
I’m curious about this, are you as well?
18. Anyone else curious about the majority of
deaths in china seem to be the same areas
they rolled out their stand alone 5G just a
couple months ago. Verry few deaths being
reported in other areas in comparison. #5G
#CoronavirusOutbreak #COVID19
#5gamechanger
EXAMPLE 1
CODES
• Direct SA: asking
• Direct SA: asserting
• Indirect SA: ?
I’m curious about this, are you as well?
In areas with many deaths, 5G was implemented.
19. Anyone else curious about the majority of
deaths in china seem to be the same areas
they rolled out their stand alone 5G just a
couple months ago. Verry few deaths being
reported in other areas in comparison. #5G
#CoronavirusOutbreak #COVID19
#5gamechanger
EXAMPLE 1
CODES
• Direct SA: asking
• Direct SA: asserting
• Indirect SA: ?
I’m curious about this, are you as well?
In areas with many deaths, 5G was implemented.
In other areas with no 5G, there are fewer deaths.
20. Anyone else curious about the majority of
deaths in china seem to be the same areas
they rolled out their stand alone 5G just a
couple months ago. Verry few deaths being
reported in other areas in comparison. #5G
#CoronavirusOutbreak #COVID19
#5gamechanger
EXAMPLE 1
CODES
• Direct SA: asking
• Direct SA: asserting
• Indirect SA: concluding
The coronavirus outbreak and 5G
implementation are related somehow.
I’m curious about this, are you as well?
In areas with many deaths, 5G was implemented.
In other areas with no 5G, there are fewer deaths.
21. Anyone else curious about the majority of
deaths in china seem to be the same areas
they rolled out their stand alone 5G just a
couple months ago. Verry few deaths being
reported in other areas in comparison. #5G
#CoronavirusOutbreak #COVID19
#5gamechanger
EXAMPLE 1
CODES
• Direct SA: asking
• Direct SA: asserting
• Indirect SA: concluding
• No meaningful connectives
(excluding ‘and’ and ‘or’)
22. Results – Implied meaning
Code Number of tweets (n=90) Percentage of the dataset
Indirect SA 59 65.6%
No indirect SA 31 34.4%
23. Results – Direct SA’s
Code (not mutually exclusive) Number of tweets (n = 90) Percentage of the dataset
Direct SA:
- asking
- asserting
- citing
- describing
28
69
24
29
32.2%
76.7%
26.6%
32.2%
24. Results – Direct SA’s
Code (not mutually exclusive) Number of tweets (n = 90) Percentage of the dataset
Direct SA:
- asking
- asserting
- citing
- describing
28
69
24
29
32.2%
76.7%
26.6%
32.2%
Yes, .... Did the 5G rollout in Wuhan damage the innate cellular defense cells
of the population, putting the people at risk of complications and death from
coronavirus? https://t.co/hRf5gfRbV7 #SARSCoV2 #COVID19 #2019nC0V
#CoronavirusOutbreak #Coronaviruswuhan
EXAMPLE 2
25. Results – Direct SA’s
Code (not mutually exclusive) Number of tweets (n = 90) Percentage of the dataset
Direct SA:
- asking
- asserting
- citing
- describing
28
69
24
29
32.2%
76.7%
26.6%
32.2%
Yes, .... Did the 5G rollout in Wuhan damage the innate cellular defense cells
of the population, putting the people at risk of complications and death from
coronavirus? https://t.co/hRf5gfRbV7 #SARSCoV2 #COVID19 #2019nC0V
#CoronavirusOutbreak #Coronaviruswuhan
EXAMPLE 2
26. Results – Direct SA’s
Code (not mutually exclusive) Number of tweets (n = 90) Percentage of the dataset
Direct SA:
- asking
- asserting
- citing
- describing
28
69
24
29
32.2%
76.7%
26.6%
32.2%
Yes, .... Did the 5G rollout in Wuhan damage the innate cellular defense cells
of the population, putting the people at risk of complications and death from
coronavirus? https://t.co/hRf5gfRbV7 #SARSCoV2 #COVID19 #2019nC0V
#CoronavirusOutbreak #Coronaviruswuhan
EXAMPLE 2
27. Results – Linguistic cues
Code (not mutually exclusive) Number of tweets (n = 90) Percentage of the dataset
No meaningful connectives 52 57.7%
Lexical cue phrases:
- Opinion
- Relation
26
19
28.9%
21.1%
Emoji use 12 13.3%
28. Insights
Answering our RQ’s:
Yes, implied meaning is common in mis-/disinformation tweets. Certain linguistic cues
might signal implied meaning, such as the omission of meaningful connectives.
Automated fake news detection
Does implied meaning hinder correct classification of these tweets?
Is implied meaning a signifier of mis-/disinformation?
Hello everyone!
My name is Lynn de Rijk. I am a Linguistics and Communication research master student at the Radboud university in Nijmegen. Today I'll present my paper how mis- and disinformation tweets surrounding the corona 5g conspiracy communicate through implying.
My presentation will take 10 minutes instead of the regular 3, as I’ve taken quite a different approach to other teams.
Therefore I need to explain some theory first.
After the theory, I’ll walk you through the analysis, discuss the results and some relevant insights.
As I said I work in the field of linguistics and communication and my specialty is analyzing patterns in naturally-occurring interaction, especially online interaction.
For this task, I was wondering what might be challenging for automated systems and together with Martha Larson I decided to focus on implied meaning. Implied meaning can be something like sarcasm, where you say one thing, but mean the exact opposite, but our everyday interaction is actually rife with implied meaning as well. So much so that we might not even realize anymore.
To explain what I mean by that, I'll discuss an example, while also immediately introducing the framework I've used to index implied meaning for this study, namely Speech Act Theory. So, here you see an example of a very basic utterance: Peter, you are standing on my foot.
Speech act theory describes three levels to this utterance, the locution or the direct speech act, the illlocution, or the indirect speech act, and the perlocation.
Now, the locution in this utterance would be asserting a state of affairs. So this is phrased as an assertion and it's stating that it's true that Peter is indeed standing on your foot.
Of course, interestingly enough, this is not what this utterance is meant to do. It's not just meant to inform Peter of this state of affairs, what you want is Peter to place his foot his foot elsewhere. So the illocution, or the implied meaning of this utterance, is requesting Peter to place his foot elsewhere
The perlocution is then what actually happens, so in this case, Peter could move his food he could not move his foot, he could apologize. For the purposes of this presentation, I will not discuss the perlocution in future examples, as it’s not relevant to us at this moment.
For the purposes of this study, I have looked at two questions:
How prevalent is implied meaning in mis-/disinformation tweets surrounding the corona-5G-conspiracy? So, is implied meaning possibly something the classifiers you built might have to deal with?
Are there linguistic cues that signal implied meaning? Which could of course be helpful to train automated systems in the future.
So, to find an answer to these questions, I've analyzed 90 tweets from the FakeNews task data set, more specifically, those from the 5G-Corona conspiracy subset.
To analyze the tweets I coded these 90 tweets manually. Coding in this sense simply means labeling each tweet for certain features, looking at what direct and indirect SA’s were present in each tweet and also labeling for possibly relevant linguistic features.
So, I've coded for the direct and the indirect speech acts of each tweet.
But a very relevant question is, how can you code for indirect speech acts, as it always some form of interpretation, right?
How can you reliably code for that?
That’s why I want to look at a short example, which will show that in many cases, we as language users often do interpret an utterance in the exact same way.
So take this question: Do you know what time it is?
This is a possible answer to that question.
And this is a dad joke, right?
And that’s interesting, because “yes, I know” is the answer to the literal question, “do you know X?”. It answers the direct speech act.
But we all know that that is not the answer sought by the speaker, who of course wants to know what time it is.
The reason we understand this question so easily, is because in everyday communication, we tend to try to understand each other. And so when someone says an utterance, what we do as an as a recipient is ask ourselves the question, “Why is this person asking me this question? Why are they saying this?”
Because of course, we are not computers, so we do not only respond to the literal text, but we try to work together to form meaningful communication.
Now with this in the back of our minds, let's look at an example from the dataset. I'll give you a few seconds to read it for yourself.
Very well. I hope I hope you all read it. So, how did I go about coding this? Well, first, I would determine the direct speech acts.
Right, so the first sentence, anyone else curious about the majority of deaths in China is a question.
You'll notice there is no question mark, no punctuation at all actually, but we can see from the word order that this is a question.
The question is followed by two assertions, wherein the user is stating things about the world they see as being true.
Now, what is the indirect speech act? It’s important to know, that I only coded for indirect speech acts, when this was the primary speech act, meaning the tweet is not very meaningful if taken literally.
Think back to the ‘Do you know what time it is?”, this is not very meaningful if taken literally, answered with ‘Yes, I know”.
So, what I did when coding, is first take the literal interpretation and see if that could plausibly be what the author intended to communicate.
In other words, if the literal interpretation could be the answer to our ‘why’-question. So, lets do that now.
Now I don’t know about you, but I find it strange, if someone simply wanted to communicate with me that something happens in one place and somewhere else it doesn’t?
My response to something like this would be: So what? What are you trying to say?
So I would argue that the indirect speech act here is concluding. It’s concluding that: The coronavirus outbreak and 5G implementation are related somehow.
Lastly, a relevant linguistic cue in this tweet is the lack of connectives.
The code is called meaningful connectives, by which I mean connectives other than ‘and’ and ‘or’, which aren't very informative about the relationship the author sees between sentences.
Examples of meaningful connectives are ‘however’, indicating contrast or ‘therefore’, indicating a conclusion being drawn.
So now, you know a little bit about the coding process and the theory behind it, let's look at the results. First, we see that many of these tweets indeed have implied meaning as a primary meaning: in 65,6% of the tweets analyzed this was the case. Communicating through implying is thus common practice in these tweets.
Another interesting find is that citing is a very common practice.
If we follow the link, we find that this part is actually the headline of the linked article.
Note that this leaves little room for the users own opinion and expression.
Lastly, linguistic cues found.
You can see here that meaningful connectives are very often omitted from tweets.
So in 57.7% of the data set, no meaningful connectives were found.
I'm going to skip the lexical cue phrases because we don't have the time, but it's also interesting to point out that emojis are scarcely used.
To answer our RQ’s: we see that implied meaning is common in mis-/disinformation tweets surrounding this conspiracy.
We also see that certain linguistic cues, such as omission of meaningful connectives, might be an indicator of implied meaning.
Now lastly, what does this mean for automated fake news detection?
First, we see a lot of implied meaning in these mis-/disinformation tweets.
This might hinder correctly distinguishing fake news from discussion on the same topic.
If so, the linguistic cues I’ve found might be a way to combat this.
The second question is, is implied meaning a signifier of mis and disinformation? Because if so, training automated systems to recognize implied meaning would aid in classifying these tweets.