Literature Introduction. 21/10/2014

Learning Latent Personas of Film Characters

Transcript

1. Literature Introduction Seminar Yamamoto Laboratory Exchange Student Gamar Azuaje
2. Paper Introduction • Learning Latent Personas of Film Characters • David Bamman, Brendan O’Connor, Noah A. Smith • Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Pages: 352—361. Sofia, Bulgaria. August 2013.
3. Summary • Learn personas through characters from different films. • Extract features from text and movie metadata. • Two models, one which contains the aditional non- textual information. • Compare results with a golden cluster.
4. Data • Text source consists of 42,306 movie plot summaries extracted from the English Wikipedia. November 2, 2012. – Median length, approximately 176 words. – Synopsis of events and implicit character descriptions. • Features were extracted from each character using the Stanford CoreNLP Library. – Agent verbs: Verbs for which the entity is the agent. – Patient verbs: Verbs for which the entity was a receptor from the action. – Attributes: adjectives and common nouns used as a modifier of the character. • Movie metadata was extracted from Freebase. November 4, 2012. – Language, country, release date and detailed genre – Characters were associated with actors who played them. – Extra information of gender and age.
5. Data Zombie persona based on three distributions over latent topics .
6. Models • Soft clustering over words to topics – The verb “strangle” is related to topic “assault” • Soft clustering over topics to personas. – “Villian” perform assault actions. • Hard clustering over topics to personas. – “Darth Vader” type Villian.
7. Dirichlet Persona Model
8. Dirichlet Persona Model • Information only from the structured text. – For each character Bags of (r,w). r ∈ {agent, patient, attribute}, w word from text.
9. Persona Regression
10. Persona Regression • Incorporates metadata from the movie genre. – Movie genre, Character age, gender. – Captures likelihood of actor playing a character role. • Latent personas are drawn from the movie metadata using logistic regresion.
11. Evaluation • Character names. – Characters that appear in at least two separate movies were used. – Characters must have at least two tokens in their names. – 970 characters met this criteria and were denoted as golden clusters. • TV Tropes. – User-input information containing common tropes (narrative, character, plot device) – 72 character types were used. 501 characters met this criteria.
12. Evaluation • Variation of information was used to measure the theorical distance between the two clusters. • Lower value means greater similarity.
13. Results Character Names §5.1 TV Tropes §5.2 K Model P = 25 P = 50 P = 100 P = 25 P = 50 P = 100 25 Persona regression 7.73 7.32 6.79 6.26 6.13 5.74 Dirichlet Persona 7.83 7.11 6.44 6.29 6.01 5.57 50 Persona regression 7.59 7.08 6.46 6.30 5.99 5.65 Dirichlet Persona 7.57 7.04 6.35 6.23 5.88 5.60 100 Persona regression 7.58 6.95 6.32 6.11 6.05 5.49 Dirichlet persona 7.64 6.95 6.25 6.24 5.91 5.42 Variation of information between learned personas and gold clusters for different numbers of topics K and personas P. Lower values are better.
14. Results Label Most characteristic words Label Most characteristic words UNITE unite marry woo elope court SWITCH switch confirm escort report instruct PURCHASE purchase sign sell owe buy INFATUATE infatuate obsess acquaint revolve concern SHOOT shoot aim overpower interrogate kill ALIEN alien child governor bandit priest EXPLORE explore investigate uncover deduce CAPTURE capture corner transport imprison trap WOMAN woman friend wife sister husband MAYA maya monster monk goon dragon WITCH witch villager kid boy mom INHERIT inherit live imagine experience share INVADE invade sail travel land explore TESTIFY testify rebuff confess admit deny DEFEAT defeat destroy transform battle inject APPLY apply struggle earn graduate develop CHASE chase scare hit punch eat EXPEL expel inspire humiliate bully grant TALK talk tell reassure assure calm DIG dig take welcome sink revolve POP pop lift crawl laugh shake COMMAND command abduct invade seize surrender SING sing perform cast produce dance RELENT relent refuse agree insist hope APPROVE approve die suffer forbid collapse EMBARK embark befriend enlist recall meet WEREWOLF werewolf mother parent killer father MANIPULATE manipulate conclude investigate conduct DINER diner grandfather brother terrorist ELOPE elope forget succumb pretend like DECAPITATE decapitate bite impale strangle stalk FLEE flee escape swim hide manage REPLY reply say mention answer shout BABY baby sheriff vampire knight spirit DEMON demon narrator mayor duck crime BIND bind select belong refer represent CONGRATULATE congratulate cheer thank recommend REJOIN rejoin fly recruit include disguise INTRODUCE introduce bring mock read hatch DARK dark major henchman warrior sergeant HATCH hatch don exist vow undergo SENTENCE sentence arrest assign convict promote FLIRT flirt reconcile date dance forgive DISTURB disturb frighten confuse tease scare ADOPT adopt raise bear punish feed RIP rip vanish crawl drive smash FAIRY fairy kidnapper soul slave president INFILTRATE infiltrate deduce leap evade obtain BUG bug zombie warden king princess SCREAM scream faint wake clean hear Latent topics learned for K = 50 and P = 100.
15. Results Freq Actions Characters Features 0.109 DARKm, SHOOTa, Jason Bourne (The Bourne Supremacy), Jack Traven Action, Male, War SHOOTp (Speed), Jean-Claude (Taken) film 0.079 CAPTUREp, Aang (The Last Airbender), Carly (Transformers: Dark of Female, Action, INFILTRATEa, FLEEa the Moon), Susan Murphy/Ginormica (Monsters vs. Aliens) Adventure 0.067 DEFEATa, DEFEATp, Glenn Talbot (Hulk), Batman (Batman and Robin), Hector Action, Animation, INFILTRATEa (Troy) Adventure 0.060 COMMANDa, DEFEATp, Zoe Neville (I Am Legend), Ursula (The Little Mermaid), Action, Adventure, CAPTUREp Joker (Batman) Male 0.046 INFILTRATEa, Peter Parker (Spider-Man 3), Ethan Hunt (Mission: Male, Action, Age EXPLOREa, EMBARKa Impossible), Jason Bourne (The Bourne Ultimatum) 34-36 0.036 FLIRTa, FLIRTp, Mark Darcy (Bridget Jones: The Edge of Reason), Jerry Female, Romance TESTIFYa Maguire (Jerry Maguire), Donna (Mamma Mia!) Film, Comedy 0.033 EMBARKa, INFILTRATEa, Perseus (Wrath of the Titans), Maximus Decimus Meridius Male, Chinese INVADEa (Gladiator), Julius (Twins) Movies, Spy 0.027 CONGRATULATEa, Professor Albus Dumbledore (Harry Potter and the Age 58+, Family CONGRATULATEp, Philosopher’s Stone), Magic Mirror (Shrek), Josephine Film, Age 51-57 SWITCHa Anwhistle (Lemony Snicket’s A Series of Unfortunate Events) 0.025 SWITCHa, SWITCHp, Clarice Starling (The Silence of the Lambs), Hannibal Age 58+, Male, MANIPULATEa Lecter (The Silence of the Lambs), Colonel Bagley (The Age 45-50 Last Samurai) 0.022 REPLYa, TALKp, FLIRTp Graham (The Holiday), Abby Richter (The Ugly Truth), Female, Comedy, Anna Scott (Notting Hill) Romance Film 0.020 EXPLOREa, EMBARKa, Harry Potter (Harry Potter and the Philosopher’s Stone), Adventure, Family CAPTUREp Harry Potter (Harry Potter and the Chamber of Secrets), Film, Horror Captain Leo Davidson (Planet of the Apes) 0.018 FAIRYm, COMMANDa, Captain Jack Sparrow (Pirates of the Caribbean: At Action, Family CAPTUREp World’s End), Shrek (Shrek), Shrek (Shrek Forever After) Film, Animation 0.018 DECAPITATEa, Jericho Cane (End of Days), Martin Riggs (Lethal Weapon Horror, Slasher, DECAPITATEp, RIPa 2), Gabriel Van Helsing (Van Helsing) Teen 0.017 APPLYa, EXPELp, Oscar (Shark Tale), Elizabeth Halsey (Bad Teacher), Dre Female, Teen, PURCHASEp Parker (The Karate Kid) Under Age 22 Top 14 inferred latent personas sorted by frequency.
16. Conclusion • Characters personas were inferred automatically from text. • Movie metadata improved the results slightly. • Work could be extended to other domains.