Detecting Deception Through Linguistic Analysis

							Detecting Deception
Through Linguistic
      Analysis
      Judee K. Burgoon
           J. P. Blair
         Tiantian Qin
    Jay F. Nunamaker, Jr.
                  Introduction
   Intelligence analysts are required to sift through
    mountains of information
   Humans are generally bad at detecting deception
   Mounting evidence suggests that CMC makes
    humans even less accurate at detecting
    deception
   There is a need to develop tools to help humans
    detect deception
                   Background
   Desert Survival
     2 (Deception – Truthful) X 2 (FtF or CMC)
     All discussions were transcribed

     Linguistic Analysis conducted on 27 possible
      indicators of deception
   Mock Theft Pilot
     2 (Deception – Truthful) X 2 (FtF or Text)
     Transcribed and Analyzed
                  Hypotheses
   Deceivers will display higher
     Quantity
     Expressiveness
                     Hypotheses
   Deceivers will display less
       Complexity
          Vocabulary
          Grammatical
                     Method
   Students recruited from a multi-sectioned
    communications class
   Half were randomly assigned to be thieves and
    half to be innocents
   Thieves “stole” a wallet that was left in their
    classroom
   Innocents were told that a “theft” would occur
    on a given day
                     Method
   Subjects were motivated to do well by telling
    them that they could earn $10 if they convinced
    the interviewer that they were innocent
   An additional $50 was to be awarded to the
    person who was the most successful at
    convincing the interviewer.
                     Method
   Participants were interviewed by trained
    interviewers using a standardized BAI format in
    one of three modalities
     FtF
     Text/Chat

     Audio Conference

   All interviews were recorded and transcribed
                          Method
   Analysis was conducted using shallow parsers (Grok
    and Iskim) or look-up dictionaries
   Classes of Cues
       Quantity (Words, Syllables, Sentences)
       Vocabulary Complexity (Big words and Syllables per word)
       Grammatical Complexity (Short and Long Sentences, Flesch-
        Kincaid, and others)
       Expressiveness (Rate of Adj and Adv, emotiveness index,
        affective terms)
     MOCK THEFT RESULTS
Decision
Tree                                ASL



Analysis                            15.75


                   Sen&Rm                       ASW


                     1                          1.4




                                            1
              2             ASL                       ASL


                            12.52                     18




                    2                  1        1           2




Note: Sample Tree from text modality with no duplicated cues
     MOCK THEFT RESULTS
Decision
Tree
Analysis                          ASL



                                  15.75


                    Sent-                         1
                    Comp


                     22




              1              2




Note: Sample tree from text modality, significant cues only
                DECISION TREE
                IMPROVEMENT
78.58%


75%




62.5%


60.42%




58.33%




         Original
                                   Significant                      Only txt; Significant
                    No Duplicate
                                                 Only txt; No duplicate
                  Conclusions
   We were able to identify some linguistic
    indicators of deception
   Modality also appears to affect several indicators
   These indicators could be subjected to a pruned
    tree algorithm to classify subjects as truthful or
    deceptive
   Future research will serve to further improve
    modeling
               Future Research
   The linguistic model will be improved by
     Adding more data
     Improving dictionaries

     Focusing on different models for different
      communication contexts
     Adding subjective operator evaluations