Conference PaperPDF Available

Detecting Deception through Linguistic Analysis

Authors:

Abstract and Figures

Tools to detect deceit from language use pose a promising avenue for increasing the ability to distinguish truthful transmissions, transcripts, intercepted messages, informant reports and the like from deceptive ones. This investigation presents preliminary tests of 16 linguistic features that can be automated to return assessments of the likely truthful or deceptiveness of a piece of text. Results from a mock theft experiment demonstrate that deceivers do utilize language differently than truth tellers and that combinations of cues can improve the ability to predict which texts may contain deception.
Content may be subject to copyright.
Detecting Deception through Linguistic Analysis
Judee K. Burgoon
1
, J. P. Blair
2
, Tiantian Qin
1
, Jay F. Nunamaker, Jr
1
,
1
Center for the Management of Information, University of Arizona
2
Department of Criminal Justice, Michigan State University
Submitted to the Intelligence and Security Informatics Symposium, Tucson, June 2003.
Portions of this research were supported by funding from the U. S. Air Force Office of Scientific
Research under the U. S. Department of Defense University Research Initiative (Grant #F49620-01-1-
0394). The views, opinions, and/or findings in this report are those of the authors and should not be
construed as an official Department of Defense position, policy, or decision.
Abstract.
Tools to detect deceit from language use pose a promising avenue for increasing the ability to
distinguish truthful transmissions, transcripts, intercepted messages, informant reports and the like from deceptive ones.
This investigation presents preliminary tests of 16 linguistic features that can be automated to return assessments of the
likely truthful or deceptiveness of a piece of text. Results from a mock theft experiment demonstrate that deceivers do
utilize language differently than truth tellers and that combinations of cues can improve the ability to predict which
texts may contain deception.
1 Introduction
One of the most daunting challenges facing intelligence and law enforcement communities in the wake
of 9/11 is anticipating and preventing terrorist attacks. Doing so requires sifting through a daily avalanche
of information and communications to discriminate good information from bad, and deceptive
communications from truthful ones. Much of this task falls to human analysts. Aside from the enormity of
the task, the empirical evidence is compelling that humans, even professionally trained ones, are typically
very poor at detecting deception and fallacious information [1, 2, 3]. Moreover, evidence is building that
detection abilities are even more faulty when communication is computer-mediated, especially if is text-
based such as with e-mail [4, 5]. Increasing reliance worldwide on electronic communications thus carries
with it the very real likelihood of further decrements in detection accuracy and heightened vulnerabilities to
national and international security. Human information processors are in serious need of tools that can
assist them in filtering information and alerting them to suspicious information. Aside from the security
implications, such tools would also be of great benefit to law enforcement in dealing with a wide array of
criminal investigations.
Our program of research, funded by the Department of Defense, has begun to develop such tools
by examining linguistic features and content characteristics of texts for reliable indicators of deceit. These
indicators can then be incorporated into software tools that automate their detection and subject them to
statistical analysis for probabilistic estimates of truthfulness or deceptiveness. Decades of research have
confirmed that there are few indicators of deceit that remain invariant across genres of communication,
situations, communicators, cultures, and other features of communication contexts. Yet combinations of
indicators have shown to have good predictive ability in specified contexts [6]. The research to be reported
here was guided by the objectives of identifying those indicators that are (1) the least context-sensitive and
(2) the most amenable to automation. We present preliminary results from a mock theft experiment that is
still in progress, our purposes being to illustrate the promise of examining text-based linguistic indicators
and of examining such indicators using a particular statistical approach that examines combinations of cues.
2 Background
Two experiments from our research program predate the one to be reported here. One was
modeled after the familiar Desert Survival Problem, in which pairs of participants were given a scenario in
which their jeep had crashed in the Kuwaiti desert, read material we developed from an Army field manual
that we entitled, “Imperative Information for Surviving in the Desert,” and then were asked to arrive at a
consensus on the rank-ordering of salvageable items in terms of their importance to survival. The task was
conducted via email over the course of several days. In half of the pairs, one person was asked to deceive
the partner by advocating choices opposite of what the experts recommend (e.g., discarding bulky clothing
and protective materials so as to make walking more manageable). Partners discussed the rankings and
their recommendations either face-to-face or using a computer-mediated form of communication such as
text chat, audioconferencing, or videoconferencing. All discussions were recorded and transcribed then
subjected to linguistic analysis of such features as number of words, number of sentences, number of
unique words (lexical diversity), emotiveness, and pronoun usage. Of the 27 indicators that were examined,
several proved to reliably distinguish truth tellers from deceivers. Deceivers were more likely to use longer
messages but with less diversity and complexity, and greater uncertainty and “distancing” in language use
than truth tellers. These results revealed that systematic differences in language use could help predict
which messages originated from deceivers and which, from those telling the truth.
The second experiment was designed as a pilot effort for the experiment to be reported below. In
this experiment, participants staged a mock theft and were subsequently interviewed by untrained and
trained interviewers via text chat or face-to-face (FtF) interaction [4, 5]. The FtF interactions were later
transcribed, and the transcripts and chats were submitted to linguistic analysis on the same features as noted
above, plus several others that are available in the Grammatik tool within WordPerfect. Due to the small
sample size, none of the differences between innocents (truth tellers) and thieves (deceivers) were
statistically significant, but patterns were suggestive of deceivers tending toward briefer messages (fewer
syllables, words, and sentences; shorter and simpler sentences) of greater complexity (e.g., greater
vocabulary and sentence complexity, lower readability scores) than truth tellers (higher Flesch-Kincaid,
sentence complexity, vocabulary complexity, syllables per word).
The patterns found in these first efforts suggested that we should expect to find many linguistic
differences between deceivers and truth tellers with a larger, and well-designed experiment. We therefore
hypothesized that deceptive senders display higher (a) quantity, (b) non
immediacy, (c) expressiveness,
(d) informality, and (e) affect; and less (f)
complexity, (g) diversity, and (h) specificity of language
in their messages than truthful senders.
3 Method
Students were recruited from a multi-sectioned communication class by offering them credit for
participation and the chance to win money if they were successful at their task. Half of the students were
randomly assigned to be “thieves,” i.e., those who would be deceiving about a theft, and the other half
became “innocents,” i.e., those who would be telling the truth.
Interviewees in the deceptive condition were assigned to “steal” a wallet that was left in a
classroom. In the truthful condition, interviewees were told that a “theft” would occur in class on an
assigned day. All of the interviewees and interviewers then appeared for interviews according to a pre-
assigned schedule. We attempted to motivate serious engagement in the task by offering interviewers $10
if they could successfully detect whether their interviewee was innocent or guilty and successfully detect
whether they were deceiving or telling the truth on a series of the interview questions. In turn, we offered
interviewees $10 if they convinced a trained interviewer that they were innocent and that their answers to
several questions were truthful. An additional incentive was a $50 prize to be awarded to the most
successful interviewee.
Interviewees were then interviewed by one of three trained interviewers under one of three
modalities— Face to Face (FtF), text chat, or audioconferencing. The interviews followed a standardized
Behavioral Analysis Interview format that is taught to criminal investigators [7]. Interviews were
subsequently transcribed and submitted to linguistic analysis. Clusters of potential indicators, all of which
could be automatically calculated with a shallow parser (Grok or Iskim) or could use a look-up dictionary,
were included. The specific classes of cues and respective indicators were as follows:
1. Quantity (number of syllables, number of words, number of sentences)
2. Vocabulary Complexity (number of big words, number of syllables per word)
3. Grammatical Complexity (number of short sentences, number of long sentences, Flesh-Kincaid grade
level, average number of words per sentence, sentence complexity, number of conjunctions)
4. Specificity and Expressiveness (emotiveness index, rate of adjectives and adverbs, number of affective
terms)
4 Results
In the following two subsections, we investigate data from two perspectives: analysis of individual cues
and cluster analysis. To analyze how well individual cues distinguish messages of deceivers from those of
truth tellers, we conducted multivariate analyses of related groups of cues followed by directional t-tests on
individual cues to identify which ones contribute most to differentiating deceivers and truthful tellers. The
cluster analysis answers the question of whether combinations of cues (in a hierarchy structure) can
improve overall ability to differentiate deceivers from truth tellers. Furthermore, unlike the traditional
statistical cluster analysis, we used a data-mining algorithm -- C4.5 [8] -- to cluster the cues and obtain a
hierarchical tree structure. In this way, we fulfilled the “automatic” requirement of automating deception
detection.
4.1 Individual cue analysis
Results were based on data of 41 subjects whose modality was text chat (txt) or audio. (In the future,
data for face-to-face (FtF) will be included after those sessions have been completed and all video files are
transcribed). Among 41 subjects, 29 interacted via txt and 20, via audio; 26 were “thieves” (i.e., deceivers)
and 23 were “innocents (i.e., truth tellers). Table 1 presents descriptive statistics for the 16 cues that were
analyzed.
Table 1. Means (Deviations) for 16 cues
Deceiver Truthful-teller Cues
Txt Audio Txt Audio
Syllables 121.06(82.723) 146.44(75.583) 144.58(86.732) 189.27(86.383)
Words 86.06(58.297) 106.00(53.708) 106.25(64.435) 140.36(62.094)
Sentences 5.88(3.621) 7.11(4.372) 6.00(5.117) 6.91(2.948)
Short sentences 2.82(2.298) 3.11(3.140) 2.58(3.872) 2.09(1.57)
Long sentences 0.24(0.437) 0.44(0.520) 0.92(1.16) 1.09(1.22)
Simple sentences 1.29(1.21) 2.11(2.205) 2.20(2.64) 1.36(0.920)
Big words 6.29(6.64) 6.78(4.32) 6.55(3.77) 8.28(5.71)
Average syllables per word 1.42(0.107) 1.36(0.089 1.36(0.068) 1.34(.096)
Average words per sentence 15.45(6.53) 16.93(9.82) 21.02(7.76) 22.94(13.36)
Flesch-Kincaid grade level 7.28(2.839) 7.32(3.986) 8.91(3.223) 8.82(3.087)
Sentence complexity 41.12(16.86) 43.33(18.66) 52.91(26.44) 49.73(24.08)
Vocabulary complexity 9.88(7.54) 7.11(4.48) 7.27(4.60) 7.27(4.22)
Conjunctions 4.47(3.48) 5.30(4.03) 5.00(3.82) 9.27(7.08)
Rate of Adjectives & Adverbs 0.12(0.014) 0.12(0.016) 0.12(0.014) 0.13(0.014)
Emotiveness 0.29(0.100) 0.29(0.141) 0.29(.073) 0.27(.160)
Affect 0.24(0.562) 0(0) 0.55(0.820) 0.33(0.651)
Results of the multivariate tests and t-tests are shown in Table 2. (For plots of means by deception
condition and modality, see the figures in the appendix.) The multivariate analysis on indicators of quantity
of language produced a significant multivariate effect for deception (p = .033) and no modality by
deception interaction. Deceivers said or wrote less than truth tellers.
The multivariate analysis of indicators of complexity at both the sentence level (simple sentences,
long sentences, short sentences, sentence complexity, Flesch-Kincaid grade level, number of conjunctions,
average-words-per-sentence (AWS)) and vocabulary level (vocabulary complexity, number of big words,
average-syllables-per-word (ASW)) did not produce overall multivariate effects, but several individual
variables did show the effects of deception condition. Deceivers had significantly fewer long sentences,
AWS and sentence complexity than truth tellers; and a lower Flesch-Kincaid grade level than truth tellers.
This meant their language was less complex and easier to comprehend. The t-tests also provided weak
support for deceivers having fewer ASW (p = .102) and conjunctions (p = .149) in messages than truth
tellers. Thus, deceivers used less complex language at both the lexical (vocabulary) and grammatical
(sentence and phrase) levels. Modality effects also showed that subjects in text chat used fewer
conjunctions than in audio chat, indicating that that modality was less likely to exhibit compound and
complex sentences.
For the analyses of message specificity and expressiveness (adjectives and adverbs, emotiveness,
and affect), the multivariate test showed a trend toward a main effect for the deception condition (p = .101).
There was a significant univariate difference on affect, such that deceivers used less language referring to
emotions and feelings than did truth tellers.
Table 2. Univariate F-tests (p-values) and t-tests for individual cue analysis
Test of between-subject effects
Cues
Modality Condition Modality*Condition
Independent Samples t-Test
Syllables 2.054(.159) 1.842(.182) .156(.695) 1.502(.140)
Words 2.363(.131) 2.407(.128) .162(.689) 1.702(.096)*
Sentences .810(.373) .001(.972) .018(.894) .111(.912)
Short sentences .122(.016)* .588(.447) .225(.637) -.725(.472)
Long sentences .547(.464) 6.566(.014)* .005(.947) 2.781(.008)*
Simple sentences .029(.886) .002(.969) 1.874(.178) .061(.951)
Big words .462(.500) .288(.594) .146(.704) .616(.541)
Average syllables per word 1.949(.17) 1.703(.199) .413(.524) -1.668(.102)
Average words per sentence .374(.544) 4.368(.042)* .006(.936) 2.414(.021)*
Flesch-Kincaid grade level .001(.979) 2.690(.108) .005(.943) 1.958(.056)*
Sentence complexity .006(.940) 2.055(.159) .181(.673) 1.779(.082)*
Vocabulary complexity .657(.422) .512(.478) .657(.422) -.997(.324)
# of Conjunctions 3.393(.072)* 2.569(.116) 1.496(.228) 1.426(.163)
Rate Adjectives and Adverbs 0.150(.700) .329(.569) .301(.586) -.596(.554)
Emotiveness 0.020(.889) .054(.818) .060(.808) -.233(.817)
Affect 1.591(.214) 3.291(.214) .004(.948) 1.630(.110)
* p < .05, one-tailed.
4.2 Cluster analysis by C4.5
Although many linguistic cues were not significant as shown in section 1, they can form a
hierarchy tree that performs relatively well in discriminating deceptive communicators from truthful ones.
Among many data-mining algorithms, we chose C4.5 because it provides a clear cluster structure
(compared with neural network), as well as satisfactory precision [9]. C4.5 used a pruned tree to cluster the
cues. This algorithm cuts off redundant branches while constraining error rates. We used software of Weka
(University of Waikato in New Zealand; Witten and Frank, 2000) to implement the C4.5.
Figure 1 is the
output of a pruned tree. “1” stands for truthful condition, “2” stands for deception condition.
The correct prediction rate using 15-fold cross-validation is 60.72%, which is reasonably
satisfactory given the small size of the data set.
As shown in the Figure 1, a combination of linguistic cues can well categorize deception
behaviors. For example, sentence level complexity combined with vocabulary or affect acted as good
classifier. Those significant linguistic cues in section 1 also played important roles in the cluster
classification: number of conjunctions, FK grade level, AWS, affect. On the other hand, the cluster
structure also showed consistency with multivariate tests: not all linguistic cues contribute in identifying
deceptions. There were “unhelpful” cues, such as emotiveness, which showed no significance in both the
single level structure and cluster analysis (hierarchy structure). However, it is premature to conclude the
ineffectiveness of any linguistic cues at this point. Further investigations with larger data sets will give us
deeper insight into the intra-relations of cues.
The confusion matrix shows the number of misclassifications: 10 out of 37 true conditions are
misclassified as deceptive, and 19 out of 35 deceptive conditions are misclassified. The tree mentioned
above produced less misclassifications in the truth condition than in the deceptive condition, which implies
a “truth-biased” judgment. In other words, the cluster method is cautious in designating messages as
untrustworthy.
#conjunction <= 10
| #conjunction <= 3
| | #bigwds <= 7
| | | FK_grade <= 8.383333
| | | | #bigwds <= 3
| | | | | Voc_comp <= 12
| | | | | | number_of_sentences <= 2
| | | | | | | Average_words_per_sentence <= 7.625: 1 (2.0)
| | | | | | | Average_words_per_sentence > 7.625: 2 (3.0)
| | | | | | number_of_sentences > 2: 2 (7.0)
| | | | | Voc_comp > 12: 1 (2.0)
| | | | #bigwds > 3: 2 (9.0)
| | | FK_grade > 8.383333: 1 (7.0/1.0)
| | #bigwds > 7: 1 (5.0/1.0)
| #conjunction > 3
| | Average_words_per_sentence <= 31.25
| | | Affect <= 1: 1 (23.0/4.0)
| | | Affect > 1
| | | | Average_words_per_sentence <= 14.166667: 2 (3.0)
| | | | Average_words_per_sentence > 14.166667: 1 (4.0)
| | Average_words_per_sentence > 31.25: 2 (3.0)
#conjunction > 10: 2 (4.0)
=== Confusion Matrix ===
a b <-- classified as
27 10 | a = truthful
19 16 | b = deceptive
Fig. 1. Pruned tree output from C4.5
5 Discussion
This investigation was undertaken largely to demonstrate the efficacy of utilizing linguistic cues,
especially ones that can be automated, to flag potentially deceptive discourse, and to use statistical
clustering techniques to select the best set of cues to reliably distinguish truthful from deceptive
communication. This investigation demonstrates the potential of both the general focus on language
indicators and the use of hierarchical clustering techniques to improve the ability to predict what texts
might be deceptive.
As for the specific indicators that might prove promising, these results provide some evidence for
the hypothesis that deceivers behave differently than truth tellers in communications via text chat and/or
audio chat. Although many tests were not significant due to the small sample size, there was a trend shown
in the profile plots demonstrating that: deceivers’ messages were briefer (i.e., lower on quantity of
language), were less complex in their choice of vocabulary and sentence structure, and lack specificity or
expressiveness in their text-based chats. This is consistent with profiles found in nonverbal deception
research showing deceivers tend to adopt, at least initially, a fairly inexpressive, rigid communication style
with “flat” affect. It appears that their linguistic behavior follows suit and also demonstrates their inability
to create messages rich with the details and complexities that characterize truthful discourse. Over time,
deceivers may alter these patterns, more closely approximating normal speech in many respects. But it is
possible that language choice and complexity may fail to show changes because deceivers are not accessing
real memories and real details, and thus will not have the same resources in memory upon which to draw.
Unlike asynchronous experiments such as the Desert Survival experiment (DSP), subjects did not
have sufficient time to provide detailed lies that contained more quantity and complexities [10]. The
differences in synchronicity in these two tasks points to time for planning, rehearsal, and editing as a major
factor that may alter the linguistic patterns of deceivers and truth tellers. As a consequence, no single
profile of deception language across tasks is likely to emerge. Rather, it is likely that different cue models
will be required for different tasks. Consistent with interpersonal deception theory [11], deceivers may
adapt their language style deliberately according to the task at hand and their interpersonal goals. If the
situation does not afford adequate time for more elaborate deceits, one should expect deceivers to say less.
But if time permits elaboration, and/or the situation is one in which persuasive efforts may prove beneficial,
deceivers may actually produce longer messages. What may not change, however, is their ability to draw
upon more complex representations of reality because they are not accessing reality. In this respect,
complexity measures may prove less variant across tasks and other contextual features. The issue of context
invariance thus becomes an extremely important one to investigate as this line of work proceeds.
Modality also plays a role in communication. Subjects talked more than they wrote, but message
complexity did not seem to be much different between the text and audio modalities. Future research will
explore the effect of different communication modalities on the characteristics of truthful and deceptive
messages.
Although clustering analysis did not consider modality effects, it provided a hierarchy tree
structure to capture the combined characteristics of the cues. It also provided an exploratory threshold
value to separate deceptive and true messages. It should also be noted that the analysis in this study used
the absolute values of linguistic characteristics to classify statements as truthful or deceptive. Because
people vary greatly in their usage of the language (e.g. some people naturally use more or less complex
language than others), the use of cue values that are relative to the sender of the message may result in
greater classification accuracy. This would require building a model of the baseline speech patterns of an
individual and then comparing an individual message to this model.
Future research will consider intra-connections among linguistic cues, tasks, and modalities. More
data will also enhance reliability of current results, but it is clear from these results alone that linguistic
cues that are amenable to automation may prove valuable in the arsenal of tools to detect deceit.
References
1. Burgoon, J. K., Buller, D. B., Ebesu, A., Rockwell, P.: Interpersonal Deception: V: Accuracy in
Deception Detection. Communication Monographs 61 (1994) 303-325
2. Levine, T., McCornack, S.: Linking Love and Lies: A Formal Test of the McCornack and Parks
Model of Deception Detection. J. of Social and Personal Relationships 9 (1992) 143-154
3. Zuckerman, M., DePaulo, B., Rosenthal, R.: Verbal and Nonverbal Communication of Deception.
In: Berkowitz, L. (ed.): Advances in Experimental Social Psychology, Vol. 14. Academic Press,
New York (1981) 1-59
4. Burgoon, J., Blair, J. P., Moyer, E.: Effects of Communication Modality on Arousal, Cognitive
Complexity, Behavioral Control and Deception Detection during Deceptive Episodes. Paper
submitted to the Annual Meeting of the National Communication Association, Miami. (2003,
November)
5. Burgoon, J., Marett, K., Blair, J. P.: Detecting Deception in Computer-Mediated Communication.
In: George, J. F. (ed.):
Computers in Society: Privacy, Ethics & the Internet. Prentice-Hall,
Upper Saddle River, NJ
(in press)
6. Vrij, A.: Detecting Lies and Deceit. John Wiley and Sons, New York (2000)
7. Inbau, F. E., Reid, J. E., Buckley, J. P., Jayne, B. C.: Criminal Interrogations and Confessions. 4th
edn. Aspen, Gaithersburg, MD (2001)
8. Quinlan, J. R.: C4.5. Morgan Kaufmann Publishers, San Mateo, CA (1993)
9. Spangler, W., May, J., Vargas, L.: Choosing Data-Mining Methods for Multiple Classification:
Representational and Performance Measurement Implications for Decision Support. J. Management
Information Systems 16 (1999) 37-62
10. Zhou, L. Twitchell, D., Qin, T., Burgoon, J. K., Nunamaker, J. F., Jr.: An Exploratory Study into
Deception Detection in Text-based Computer-Mediated Communication. In: Proceedings of the 36
th
Annual Hawaii International Conference of System Sciences. Big Island, Los Alamitos, CA (2003)
11. Buller, D. B., Burgoon, J. K.: Interpersonal Deception Theory. Communication Theory
6 (1996)
203-242
Appendix: Individual cue comparisons by modality and deception condition*
Estimated Marginal Means of total number of syllables
Modality
21
Estimated Marginal Means
200
180
160
140
120
100
GUILT
guilty
innocent
Estimated Marginal Means of total number of words
Modality
21
Estimated Marginal Means
150
140
130
120
110
100
90
80
GUILT
guilty
innocent
Estimated Marginal Means of number of sentences
Modality
21
Estimated Marginal Means
7.2
7.0
6.8
6.6
6.4
6.2
6.0
5.8
GUILT
guilty
innocent
Estimated Marginal Means of short sentences
Modality
21
Estimated Marginal Means
3.2
3.0
2.8
2.6
2.4
2.2
2.0
GUILT
guilty
innocent
Estimated Marginal Means of long sentences
Modality
21
Estimated Marginal Means
1.2
1.0
.8
.6
.4
.2
0.0
GUILT
guilty
innocent
Estimated Marginal Means of simple sentences
GUILT
innocentguilty
Estimated Marginal Means
2.2
2.0
1.8
1.6
1.4
1.2
Modality
1
2
Estimated Marginal Means of big words
GUILT
innocentguilty
Estimated Marginal Means
8.5
8.0
7.5
7.0
6.5
6.0
Modality
1
2
Estimated Marginal Means of average syllables per word
GUILT
innocentguilty
Estimated Marginal Means
1.44
1.42
1.40
1.38
1.36
1.34
1.32
Modality
1
2
Estimated Marginal Means of average words per sentence
GUILT
innocentguilty
Estimated Marginal Means
24
22
20
18
16
14
Modality
1
2
Estimated Marginal Means of Flesch-Kincaid grade level
GUILT
innocentguilty
Estimated Marginal Means
9.5
9.0
8.5
8.0
7.5
7.0
Modality
1
2
Estimated Marginal Means of sentence complexity
GUILT
innocentguilty
54
52
50
48
46
44
42
40
Modality
1
2
Estimated Marginal Means of vocabulary complexity
GUILT
innocentguilty
Estimated Marginal Means
10.5
10.0
9.5
9.0
8.5
8.0
7.5
7.0
6.5
Modality
1
2
Estimated Marginal Means of #conjunction
GUILT
innocentguilty
Estimated Marginal Means
10
9
8
7
6
5
4
Modality
1
2
Estimated Marginal Means of RateAdAdj
GUILT
innocentguilty
Estimated Marginal Means
.13
.12
.11
.10
Modality
1
2
Estimated Marginal Means of Emotive
GUILT
innocentguilty
Estimated Marginal Means
.30
.29
.28
.27
Modality
1
2
Estimated Marginal Means of Affect
GUILT
innocentguilty
Estimated Marginal Means
.6
.5
.4
.3
.2
.1
0.0
Modality
1
2
* Modality 1 = Text, Modality 2 = Audio
... Linguistic features have been used in both supervised and unsupervised learning approaches. It is essential to recognise that prospective deceivers use short sentences, certain phrasal verbs as well as certain verb tenses as part of their language, which has been revealed by some cooperative experiments between psychologists, linguists and computer scientists [22,[29][30][31][32]. For example, 16 language variables were investigated by Burgoon et al. [29] in order to see if they may help distinguish between deceptive and truthful communications. ...
... It is essential to recognise that prospective deceivers use short sentences, certain phrasal verbs as well as certain verb tenses as part of their language, which has been revealed by some cooperative experiments between psychologists, linguists and computer scientists [22,[29][30][31][32]. For example, 16 language variables were investigated by Burgoon et al. [29] in order to see if they may help distinguish between deceptive and truthful communications. They conducted two experiments in order to construct a database, in which they set up face-to-face or computer-based discussions with one volunteer acting as the deceiver and the other acting genuinely. ...
Article
Full-text available
Efforts have been dedicated by researchers in the field of natural language processing (NLP) to detecting and combating fake news using an assortment of machine learning (ML) and deep learning (DL) techniques. In this paper, a review of the existing studies is conducted to understand and curtail the dissemination of fake news. Specifically, we conducted a benchmark study using a wide range of (1) classical ML algorithms such as logistic regression (LR), support vector machines (SVM), decision tree (DT), naive Bayes (NB), random forest (RF), XGBoost (XGB) and an ensemble learning method of such algorithms, (2) advanced ML algorithms such as convolutional neural networks (CNNs), bidirectional long short-term memory (BiLSTM), bidirectional gated recurrent units (BiGRU), CNN-BiLSTM, CNN-BiGRU and a hybrid approach of such techniques and (3) DL transformer-based models such as BERTbase and RoBERTabase. The experiments are carried out using different pretrained word embedding methods across four well-known real-world fake news datasets—LIAR, PolitiFact, GossipCop and COVID-19—to examine the performance of different techniques across various datasets. Furthermore, a comparison is made between context-independent embedding methods (e.g., GloVe) and the effectiveness of BERTbase—contextualised representations in detecting fake news. Compared with the state of the art’s results across the used datasets, we achieve better results by solely relying on news text. We hope this study can provide useful insights for researchers working on fake news detection.
... Deception involves the manipulation of language and careful construction of messages or stories that appear truthful to others. There are many studies focusing on identifying the characteristics of liars, fraudsters, and deceivers by analyzing frauds in financial statements, deceiving emails, fake profiles, and deceptive conversations on dating sites, using different techniques [37][38][39][40]. It has been widely assumed by those deception models that deceivers and deception leave their footprints [41]. ...
... Linguistic cues were used in text-based computer mediated systems by Zhou et al. [44] and were proven very useful in detecting deception. Linguistic cues such as (1) word count, (2) pronouns, (3) expression words, and (4) exclusive words which turned out to be associated with deception [38,41]. Deceivers face a constant struggle in writing due to their lack of familiarity with what they are explaining; and to avoid conflicts with their own statements, therefore, they provide fewer details [41,45,46]. ...
Article
Full-text available
Crowdfunding has seen an enormous rise, becoming a new alternative funding source for emerging companies or new startups in recent years. As crowdfunding prevails, it is also under substantial risk of the occurrence of fraud. Though a growing number of articles indicate that crowdfunding scams are a new imminent threat to investors, little is known about them primarily due to the lack of measurement data collected from real scam cases. This paper fills the gap by collecting, labeling, and analyzing publicly available data of a hundred fraudulent campaigns on a crowdfunding platform. In order to find and understand distinguishing characteristics of crowdfunding scams, we propose to use a broad range of traits including project-based traits, project creator-based ones, and content-based ones such as linguistic cues and Named Entity Recognition features, etc. We then propose to use the feature selection method called Forward Stepwise Logistic Regression, through which 17 key discriminating features (including six original and hitherto unused ones) of scam campaigns are discovered. Based on the selected 17 key features, we present and discuss our findings and insights on distinguishing characteristics of crowdfunding scams, and build our scam detection model with 87.3% accuracy. We also explore the feasibility of early scam detection, building a model with 70.2% of classification accuracy right at the time of project launch. We discuss what features from which sections are more helpful for early scam detection on day 0 and thereafter.
... On the one hand, deceivers can intentionally write longer and more elaborated passages to appear more comprehensive if they perceive persuasive effort as beneficial in the particular circumstance (Burgoon et al., 2003). In fact, in examining deception in an instant messaging context, Hancock et al. (2007) found that the deceivers use more words overall (Hancock et al., 2007), while Anderson and Simester (2013) found they use more explanations to appear more persuasive. ...
... Previous studies have found a negative correlation between sentence complexity and deception due to the cognitive load necessary to maintain a fake story that contradicts or does not result from a real experience, and such a process requires psychological effort (Tausczik & Pennebaker, 2010). Other studies have suggested that deception is associated not only with lower sentence complexity but also lower language diversity (Burgoon et al., 2003;Zhou et al., 2004). Hence, deceivers have less mental capacity to produce more logically flowing complex texts compared to truth tellers. ...
... All true utterances obtained in our study were significantly longer than false ones. As in other studies [7,30,31] in which the true statements turned out to be longer than false statements, we believe it is related to the topic on which participants expressed their opinion. These were issues on which they had an established opinion, and they were therefore able to provide more arguments, facts, details, and finally their own beliefs. ...
Article
Full-text available
Lying appears in everyday oral and written communication. As a consequence, detecting it on the basis of linguistic analysis is particularly important. Our study aimed to verify whether the differences between true and false statements in terms of complexity and sentiment that were reported in previous studies can be confirmed using tools dedicated to measuring those factors. Further, we investigated whether linguistic features that differentiate true and false utterances in English-namely utterance length, concreteness, and particular parts-of-speech-are also present in the Polish language. We analyzed nearly 1,500 true and false statements, half of which were transcripts while the other half were written statements. Our results show that false statements are less complex in terms of vocabulary, are more concise and concrete, and have more positive words and fewer negative words. We found no significant differences between spoken and written lies. Using this data, we built classifiers to automatically distinguish true from false utterances, achieving an accuracy of 60%. Our results provide a significant contribution to previous conclusions regarding linguistic deception indicators.
... The many and varied taxonomies of mis-and disinformation creates a complex landscape and contribute to the tapestry of definitions for the two phenomena [24]. Wardle [50] identified six different types of misinformation [51] which laid the foundation for the mis/disinformation taxonomy, furthered by the work of Farrell and colleagues [16], and Burgoon and colleagues [7] who discussed misinformation in terms of deceptive language and false context. It is important to note this multitude of taxonomies, as it contributes to showing that awareness of policies for determining mis-and disinformation is both necessary and advantageous when devising a user-centred approach to anti-disinformation tools. ...
Article
Full-text available
With the increasing amounts of mis- and disinformation circulating online, the demand for tools to combat and contain the phenomenon has also increased. The multifaceted nature of the phenomenon requires a set of tools that can respond effectively, and can deal with the different ways in which disinformation can present itself, In this paper, after consulting independent fact-checkers to create a list, we map the landscape of tools available to combat different typologies of mis and disinformation on the basis of three levels of analysis: the employment of policy-regulated strategies, the use of co-creation, and the preference for manual or automated processes of detection. We then create a model in which we position the different tools across three axes of analysis, and show how the tools distribute across different market positions.
... Gravanis et al. [22] first reviewed features listed by other authors. Following up, Burgoon et al. [23] suggested four main groups of linguistic features: ...
Article
Full-text available
The unprecedented scale of disinformation on the Internet for more than a decade represents a serious challenge for democratic societies. When this process is focused on a well-established subject such as climate change, it can subvert measures and policies that various governmental bodies have taken to mitigate the phenomenon. It is therefore essential to effectively identify and counteract fake news on climate change. To do this, our main contribution represents a novel dataset with more than 2300 articles written in French, gathered using web scraping from all types of media dealing with climate change. Manual labeling was performed by two annotators with three classes: “fake”, “biased”, and “true”. Machine Learning models ranging from bag-of-words representations used by an SVM to Transformer-based architectures built on top of CamemBERT were built to automatically classify the articles. Our results, with an F1-score of 84.75% using the BERT-based model at the article level coupled with hand-crafted features specifically tailored for this task, represent a strong baseline. At the same time, we highlight perceptual properties as text sequences (i.e., fake, biased, and irrelevant text fragments) at the sentence level, with a macro F1 of 45.01% and a micro F1 of 78.11%. Based on these results, our proposed method facilitates the identification of fake news, and thus contributes to better education of the public.
Article
The global economic disruption brought by COVID‐19 crisis can set a stage for the prevalence of financial statement frauds, which jeopardize the efficient functioning of capital markets. In this paper, we propose a nuanced method to detect frauds by tracking granular changes in disclosures over time. Specifically, we first align paragraphs between consecutive disclosures by their similarities. This alignment can be solved as an optimization‐based matching problem. Then we identify three types of changed contents: recurrent, newly added, and deleted contents. For each type, we measure the changes in terms of fraud‐relevant linguistics features, such as sentiment and uncertainties. Further, we formulate a firm's MD&A change trajectory over years as a multivariate time series composed of these granular metrics. We implement a deep learning model to predict frauds using the change trajectory as an input. Extensive experiments demonstrate that our model significantly outperforms benchmark models, and its performance increases with the length of the change trajectory. Moreover, we found specific types of changes are strongly associated with frauds, including weak modal or reward words in newly added or deleted contents. Our work provides an optimization‐based method to define change trajectories and trace information mutation in narratives. Finally, our study contributes to the fraud detection literature with a new predictive signal – disclosure change trajectories with an effective deep learning architecture. This article is protected by copyright. All rights reserved
Article
Full-text available
Background Little research has been done on nonverbal deception cues in computer-mediated communication (CMC). However, deception is a daily occurrence and since much communication is shifting towards CMC, it is important to understand the difference between truthful and deceptive messages. Objective This research obtained more insight in the use of emoji in deceptive messages by answering the question: Are the frequency and type of emoji different in deceptive compared to truthful online messages? Methods Participants sent three screenshots of deceptive, and truthful messages to WhatsApp. The used emoji were counted and sorted into levels of valence (positive, negative, and neutral) and intensity (strong versus weak). Results The results indicated that participants used more negative, weak emoji in deceptive compared with truthful messages and more positive, weak, and strong emoji in truth compared with deceit. No difference was found for the emoji frequency. Discussion The results are discussed in the light of earlier research. However, this is the first study investigating the use of emoji in the context of computer-mediated deception. Conclusion The type of emoji can be indicative of used as a nonverbal deception cue in online messages.
Article
Full-text available
This paper reports a replication and extension of the McCornack & Parks model of relational deception detection which argued that the association between relational involvement and accuracy in detecting deception is mediated by judgmental confidence and truth-bias; and that relational involvement, confidence, truth-bias and accuracy form a causal chain. Recent research testing some of these links has yielded results which cast doubt upon the validity of the model. Moreover, research examining suspicion has raised questions concerning the generalizability of the model across various levels of aroused suspicion. The present study represents the first rigorous path-analytic test of the model, as well as a test of the model's generalizability. Testing the model in a sample of ninety romantically involved couples, this study found unqualified support for the model as it was originally specified. In addition, the model seems to be generalizable across levels of aroused suspicion. The implications of the model for research on relationships are discussed.
Article
Full-text available
Previous research on accuracy in deception detection has typically occurred in a noninteractive context, which has resulted in many potentially salient influences being ignored. Guided by interpersonal deception theory, the current experiment examined the influences of suspicion, deception type, question type, relational familiarity, and expertise on accuracy in detecting truth and deceit. An adult sample of novices and a second sample of experts (military intelligence instructors and related military personnel) participated in interviews with strangers or acquaintances during which interviewees gave some truthful answers and some deceptive answers, the latter being one of three types. Interviewers, half of whom were induced to be suspicious, followed a standard interview protocol that introduced different question strategies. Results showed that (a) accuracy was much higher on truth than deception, (b) novices were more accurate than experts, (c) accuracy depended on type of deception being perpetrated and whether suspicion was present or absent, (d) suspicion impaired accuracy for experts, (e) truth‐biases intensified with familiar others, especially when interviewers were suspicious, and (f) question strategy ameliorated or aggravated inaccuracy.
Article
Full-text available
Interpersonal deception theory (IDT) represents a merger of interpersonal communication and deception principles designed to better account for deception in interactive contexts. At the same time, it bas the potential to enlighten theories related to (a) credibility and truthful communication and (b) interpersonal communication. Presented here are key definitions, assumptions related to the critical attributes and key features of interpersonal communication and deception, and 18 general propositions from which specific testable hypotheses can be derived. Research findings relevant to the propositions are also summarized.
Article
Full-text available
Data-mining techniques are designed for classification problems in which each observation is a member of one and only one category. We formulate ten data representations that could be used to extend those methods to problems in which observations may be full members of multiple categories. We propose an audit matrix methodology for evaluating the performance of three popular data-mining techniques--linear discriminant analysis, neural networks, and decision tree induction-- using the representations that each technique can accommodate. We then empirically test our approach on an actual surgical data set. Tree induction gives the lowest rate of false positive predictions, and a version of discriminant analysis yields the lowest rate of false negatives for multiple category problems, but neural networks give the best overall results for the largest multiple classification cases. There is substantial room for improvement in overall performance for all techniques.
Conference Paper
Full-text available
Deception is an everyday occurrence across all communication media. The expansion of the Internet has significantly increased the amount of textual communication received and stored by individuals and organizations. Inundated with massive amounts of textual information transmitted through computer-mediated communication, CMC, people remain largely unsuccessful and inefficient in detecting those messages that may be deceptive. Creating an automated tool that could help people flag the possible deceptive messages in CMC is desirable, but first it is necessary to understand cues used to deceive in textual instances. This study focuses on the identification of deceptive cues deceivers use in a textual CMC environment. 30 dyads (n =14 truthful, n = 16 deceptive) were able to complete the desert survival problem. Findings have demonstrated significant differences between the content within truthful and deceptive messages. Several cues were also found to be significantly more present when deceivers write messages.
Article
Full-text available
Deception is an everyday occurrence across all communication media. The expansion of the Internet has significantly increased the amount of textual communication received and stored by individuals and organizations. Inundated with massive amounts of textual information transmitted through Computer-mediated Communication, CMC, people remain largely unsuccessful and inefficient in detecting those messages that may be deceptive. Creating an automated tool that could help people flag the possible deceptive messages in CMC is desirable, but first it is necessary to understand cues used to deceive in textual instances. This study focuses on the identification of deceptive cues deceivers use in a textual CMC environment. 30 dyads (n = 14 truthful, n = 16 deceptive) were able to complete the Desert Survival Problem. Findings have demonstrated significant differences between the content within truthful and deceptive messages. Several cues were also found to be significantly more present when deceivers write messages.
Article
Data-mining techniques are designed for classification problems in which each observation is a member of one and only one category. We formulate ten data representations that could be used to extend those methods to problems in which observations may be full members of multiple categories. We propose an audit matrix methodology for evaluating the performance of three popular data-mining techniques - linear discriminant analysis, neural networks, and decision tree induction -using the representations that each technique can accommodate. We then empirically test our approach on an actual surgical data set. Tree induction gives the lowest rate of false positive predictions, and a version of discriminant analysis yields the lowest rate of false negatives for multiple category problems, but neural networks give the best overall results for the largest multiple classification cases. There is substantial room for improvement in overall performance for all techniques.
Article
Lying and lie detection are the two components that, together, make up the exchange called as the “communication of deception.” Deception is an act that is intended to foster in another person a belief or understanding that the deceiver considers false. This chapter presents a primarily psychological point of view and a relatively microanalysis of the verbal and nonverbal exchange between the deceiver and the lie detector. The chapter discusses the definition of deception. It describes the deceiver's perspective in lie-detection, including the strategies of deception and behaviors associated with lie-telling. The lie-detector's perspective is also discussed in the chapter, and it has described behaviors associated with the judgments of deception and strategies of lie detection. The chapter discusses the outcomes of the deceptive communication process—that is, the accuracy of lie detection—and explores methodological issues, channel effects in the detection of deception, and other factors affecting the accuracy of lie detection.
Criminal Interrogations and Confes-sions
  • F E Inbau
  • J E Reid
  • J P Buckley
  • B C Jayne
Inbau, F. E., Reid, J. E., Buckley, J. P., Jayne, B. C.: Criminal Interrogations and Confes-sions. 4th edn. Aspen, Gaithersburg, MD (2001)
Detecting Lies and Deceit
  • A Vrij
  • A. Vrij
Vrij, A.: Detecting Lies and Deceit. John Wiley and Sons, New York (2000)