This document discusses detecting patterns of child grooming behavior on social media. It presents an approach for identifying the stages of online child grooming (trust development, grooming, physical approach) using chat conversation data annotated with grooming stages. Features like n-grams, part-of-speech tags, sentiment, psycholinguistic dimensions and discourse patterns are extracted from sentences and used to build classifiers. Analysis finds discourse and psycholinguistic features are effective for automatic classification of grooming stages. Future work involves adding temporal features and distinguishing between teen and predatory sexual content.
Detecting child grooming behaviour patterns on social media
1. DETECTING CHILD GROOMING
BEHAVIOUR PATTERNS ON SOCIAL MEDIA
Miriam Fernandez
m.fernandez@open.ac.uk
Harith Alani
h.alani@open.ac.uk
A. Elizabeth Cano
amparo.cano@open.ac.uk
2. INTRODUCTION
SOCIAL MEDIA PLATFORMS
• Widely spread across the Internet.
• Quick and inexpensive tools for
personal and group communications.
• No age, geographical and cultural
boundaries.
• Anonymity of the users is not
compromised.
http://i.huffpost.com/gen/1539078/thumbs/o-SOCIAL-MEDIA-facebook.jpg
3. INTRODUCTION
ONLINE CHILDREN EXPOSURE TO
PAEDOPHILES
* Source: NSPCC
• 12% of 11- 16 year olds in
the UK received unwanted
sexual messages*.
• 8% of 11-16 year olds in
the UK received requests to
send or respond to a sexual
message*.
http://news.bbcimg.co.uk/media/images/58277000/jpg/_58277172_138045933.jpg
4. INTRODUCTION
ONLINE CHILD GROOMING
Premeditated behaviour intending to secure the trust of a minor as a first
step towards future engagement in sexual conduct.
http://www.saferinternet.at/uploads/pics/442916_web_R_K_by_Christian_Seidel_pixelio.de.jpg
5. INTRODUCTION
ONLINE CHILD GROOMING
Pseudo-victims
posing as minors
Predators seeking to
groom minors
Chat-conversations
6. INTRODUCTION
ONLINE CHILD GROOMING
Pseudo-victims
posing as minors
Predators seeking to
groom minors
Chat-conversations
Archive chat conversation
Convicted Predators
7. INTRODUCTION
IDENTIFYING GROOMING STAGES
PERVERTED JUSTICE
Perverted Justice
!
• 530 chat-room conversations.
• Involving:
A. PJ volunteers posing as minors.
B. Adults seeking to begin a
sexual relationship with a
minor.
Perverted Justice, http://www.perverted-justice.com/Perverted Justice, http://www.perverted-justice.com/
8. INTRODUCTION
ONLINE CHILD GROOMING
OLSON’S THEORY OF LURING COMMUNICATION (LTC)
Gain
Access
age, gender, likes,
dislikes, family..
[1] Towards a Theory of Child Sexual Predator’s Luring Communication. L.N. Olson et al
Time
Approach
predator: where are you from
victim: where I from or
where I am now?
what’s your asl?
Deceptive Trust
Development
9. INTRODUCTION
ONLINE CHILD GROOMING
OLSON’S THEORY OF LURING COMMUNICATION (LTC)
Gain
Access
[1] Towards a Theory of Child Sexual Predator’s Luring Communication. L.N. Olson et al
Time
Grooming
Approach
predatoCrO: sRoE do you masturbate?
victim: not really that borin
predator: what do you like in sex?
raise victim’s
curiosity, .
10. INTRODUCTION
ONLINE CHILD GROOMING
OLSON’S THEORY OF LURING COMMUNICATION (LTC)
Gain
Access
Deceptive Trust
Development
[1] Towards a Theory of Child Sexual Predator’s Luring Communication. L.N. Olson et al
Time
Cycle of Entrapment
Grooming
Approach Isolation
Physical
Approach
Sexual
Conduct
predator: do you like to meet sometime?
victim: maybe you seem cool
…
predator: but i'm sorry your parents
home all the time
victim: no
11. MOTIVATION
IDENTIFYING GROOMING STAGES
PERVERTED JUSTICE
• Sexually abused children driven to voluntarily
agree to physically approach the predator [36].
• Understanding predator’s manipulative strategies
could help in educating children on how to
react when expose to such situations.
12. RELATED WORK
IDENTIFYING GROOMING STAGES
Predators/Victims chat-room conversations
• Predator detection (Kontostathis et al. [15], Michalopoulos et al.
[17], Escalante et al. [6])
Online Child Grooming
• Annotation Tools (Kontostathis et al. [15])
• Empirical Analysis of child grooming stages (Gupta et al.[9])
• Discriminate child grooming from adults cyber-sex (Bogdanova
et al.[2])
13. OBJECTIVES
DETECTING ONLINE CHILD GROOMING
IN ONLINE CHAT-ROOMS
• Create classification models to identify online
child grooming stages:
1. Trust development
2. Grooming
3. Physical approach
!
• Analyse discriminative features characterising
these stages.
14. APPROACH
IDENTIFYING GROOMING STAGES
PERVERTED JUSTICE DATA SET
Annotated Chat-room Conversations (Kontostathis et al[15])
• 50 conversations transcripts
• Conversations with 83 to 12K lines.
• Predator’s sentences manually labelled by two
annotators.
• Annotations labels: 1)Trust development,
2)Grooming, 3)Seek for physical approach, 4)
Other.
16. APPROACH
IDENTIFYING GROOMING STAGES
DATA PREPROCESSING
Challenges in processing chat-room conversations
• Irregular and ill-formed words.
• Chat slang and teen-lingo
• Emoticons.
! Generated a list of over 1K terms and definitions:
17. APPROACH
IDENTIFYING GROOMING STAGES
FEATURE EXTRACTION
Feature Description
n-grams n-grams (n=1,2,3) BoW extracted from a sentence.
Syntactic (POS) POS tags extracted from a sentence.
Sentiment Polarity Average sentiment polarity of terms contained in a
sentence
Content Complexity, Readability, Length.
Psycho-linguistic LIWC dimensions, based on cosine similarity of a
sentence to a dictionary
Discourse Semantic Frames, describing lexical use of English in texts.
18. APPROACH
IDENTIFYING GROOMING STAGES
PSYCHO-LINGUISTIC FEATURES
• Authorship profiling [12] shown that different groups of people
writing about a particular genre use language differently.
• E.g. frequency in the use of certain words.
• LIWC dataset [26] covers 60 dimensions of language.
STYLE PATTERNS
prepositions e.g., for, beside
conjunctions e.g., however, whereas
cause e.g., cuz, hence
PSYCHOLOGICAL
swearing e.g., damn, bloody
affect e.g., agree, dislike
sexual e.g., naked, porn
19. APPROACH
IDENTIFYING GROOMING STAGES
DISCOURSE FEATURES
• Qualitative analysis [5] of PJ’s predators transcripts revealed
frequent use of fixated discourse (i.e predator unwillingness
to change a topic).
• FrameNet [1], incorporate semantic generalisations of a
discourse.
• Covers 1K patterns used in English (e.g.,Intentionality act,
Causality, Grant Permission)
20. APPROACH
IDENTIFYING GROOMING STAGES
DISCOURSE FEATURES
SEMAFOR [4] to extract semantic frames from sentences.
Sentence: Your mom will let you stay home?, I’m happy
FRAME SEMANTIC ROLE LABEL
Grant Permission Target
Action
you
stay home
Grantee
Grantor
Action
you
your mom
stay home
Emotion Directed Target
Experiencer
happy
I
23. EXPERIMENTS
IDENTIFYING GROOMING STAGES
FEATURE ANALYSIS
Top discriminative features
predator: lots of luck right like
your pictures i see you keep it in
place and look
24. EXPERIMENTS
IDENTIFYING GROOMING STAGES
FEATURE ANALYSIS
Top discriminative featuprreesdator: you ever had anyone
run there fingers real litely over
your body
25. EXPERIMENTS
IDENTIFYING GROOMING STAGES
FEATURE ANALYSIS
Top discriminative features
predator: do you want me to
come ther ? good
26. EXPERIMENTS
CONCLUSIONS
• Psycho-linguistic and discourse features provide an insight of
the mindset of predators in online grooming stages.
• Discourse patterns are effective features for the automatic
classification of sentences into online grooming stages.
!
28. EXPERIMENTS
FUTURE WORK
• Some stages of online grooming are not sequential (e.g.
predator convincing child to meet in person during trust-development).
- Adding temporal features to the analysis could aid in
characterising such back-forth changes on these stages.
• Characterise sexual content between teens and between
adults (challenging since both involve the use of sexual
content).