Aprendizaje computacional en grandes volúmenes de texto

Mario Graff (mgraffg@ieee.org, mario.graff@infotec.mx)

Sabino Miranda (sabino.miranda@infotec.mx)

Daniela Moctezuma (dmoctezuma@centrogeo.edu.mx)

Eric S. Tellez (eric.tellez@infotec.mx)

CONACYT, INFOTEC y CentroGEO

https://github.com/ingeotec

  • Aplicaciones
    • Análisis de sentimientos
    • Determinación de autoría
    • Clasificación de noticias
    • Spam
    • Género y edad
  • Conclusiones

Análisis de Sentimientos

  • $\mu$TC
  • Supervisión a distancia (D.S.)
  • Transferencia de conocimiento a otros lenguajes

EvoDAG

EvoDAG

$\mu$TC + EvoDAG

$\mu$TC + EvoDAG

INEGI

Algoritmo Macro $F_1$ Positivo Negativo Neutro
$\mu$TC 0.5061 0.7631 0.5376 0.2174
$\mu$TC + EvoDAG 0.5691 0.7566 0.5849 0.3659
Ord. Voc. Aff. 0.5704 0.7529 0.5843 0.3739
D.S. all tweets 0.4783 0.6200 0.5396 0.2755
D.S. 0.4796 0.6269 0.5422 0.2698

TASS 2015

Algoritmo Macro $F_1$ Positivo Negativo Neutro
$\mu$TC 0.6270 0.7217 0.6403 0.5189
$\mu$TC + EvoDAG 0.6420 0.7288 0.6430 0.5541
Ord. Voc. Aff. 0.6352 0.7139 0.6413 0.5504
D.S. all tweets 0.5439 0.6108 0.6000 0.4209
D.S. 0.5439 0.6108 0.6000 0.4209

SemEval 2015

Algoritmo Macro $F_1$ Positivo Negativo Neutro
$\mu$TC 0.5815 0.6430 0.4333 0.6681
$\mu$TC + EvoDAG 0.5848 0.6322 0.4894 0.6327
Ord. Voc. Aff. 0.5756 0.6292 0.4769 0.6207
D.S. all tweets 0.5194 0.5696 0.4480 0.5407
D.S. 0.5217 0.5622 0.4543 0.5485

SemEval 2016

Algoritmo Macro $F_1$ Positivo Negativo Neutro
$\mu$TC 0.4638 0.5907 0.2980 0.5026
$\mu$TC + EvoDAG 0.5229 0.6223 0.4564 0.4901
Ord. Voc. Aff. 0.5144 0.6154 0.4472 0.4807
D.S. all tweets 0.4664 0.5753 0.4068 0.4172
D.S. 0.4431 0.5577 0.3926 0.3790

Distant Supervision

EvoDAG + Transferir Conocimiento

EvoDAG + Transferir Conocimiento

Model INEGI TASS2015 SemEval2015 SemEval2016
INEGI 0.4796 0.4393 0.4759 0.4312
TASS2015 0.4910 0.5439 0.5065 0.4652
SemEval2015 0.4690 0.4491 0.5217 0.4526
SemEval2016 0.4383 0.4504 0.4684 0.4431

Determinación de autoría

  • (PAN)[http://pan.webis.de/index.html] Eventos de identificación de texto.
  • Identificar la autoría basado en su estilo
  • Tarea
    • genero
    • edad
    • genero - edad
  • Lenguajes
    • Español
    • Inglés
{"text": ["Domingo termina Horario de Verano. A mi la verdad me encanta esta \u00e9poca del a\u00f1o x q los d\u00edas x alguna razon parece q me \"duran mas\".", "@ClaudioXGG a mi me preocupa mucho la distancia con la Sociedad q Reforma Fiscal q vendieron a @EPN caus\u00f3 en forma innecesaria."], "age_group": "35-49", "klass": "male", "gender": "male"}

PAN2013 - Género

Inglés

Approach Accuracy
Meina et al. 0.5921
microTC 0.5867
Santosh et al. 0.5816
Pastor L. et al. 0.5690

PAN2013 - Género

Español

Approach Accuracy
microTC 0.6750
Santosh et al. 0.6473
Pastor L. et al. 0.6299
Cruz et al. 0.6165

PAN2013 - Edad

Inglés

  • Rango de 10, 20, 30 años
Approach Accuracy
microTC 0.6605
Pastor L. et al. 0.6572
Meina et al. 0.6491
Santosh et al. 0.6408

PAN2013 - Edad

Español

Approach Accuracy
microTC 0.6897
Pastor L. et al. 0.6558
Santosh et al. 0.6430
Cruz et al. 0.6219

authorship - CCA

JimGilchrist

Hong Kong Dragon Airlines (Dragonair) is about to buy two new Airbus Industrie consortium aircraft and lease another to cope with increasing demand on its routes into China, industry sources told Reuters.

...

Dragonair recently began new services to the Chinese city of Qingdao and Khaohsiung in Taiwan. It also has full scheduled cargo rights on two of its Chinese routes to Xian and Chengdu and says it has plans for additional Chinese services to Chongqing, Urumqi and Shantou. Other shareholders in Dragonair are the Chinese state-owned China National Aviation Corp with 36 percent, China-backed CITIC Pacific Ltd with 29 percent and Cathay's parent, Swire Pacific Ltd with eight percent. --Air Cargo Newsroom Tel+44 171 542 7706 Fax +44 171 542 5017

authorship - CCA

Approach accuracy macro-f1 micro-f1
microTC 0.7680 0.7544 0.7680
Escalante et al. 2015 0.7372 0.7032 -
Cummins et al. 0.1000 0.0182 -

authorship - NFL

Joe Lapointe

Last season, the consistent strength of the GiantsÕ offense was a combination of three running backs known as Earth, Wind and Fire.

It was meant as a compliment, but those elements of nature can also spell disaster, and this seasonÕs fortunes have bordered on that.

On Tuesday Ahmad Bradshaw (Fire) was on crutches with at least a sprained left ankle, according to the Giants. The foot was in a protective boot and it was uncertain if he could play on Thursday night at Denver, or if he would even make the trip.

ÒWeÕll see,Ó Coach Tom Coughlin said. Coughlin said Bradshaw left SundayÕs 34-31 victory over Atlanta with what the Giants thought was a routine sprain.

ÒBut, evidently, it is more severe than that,Ó Coughlin said. For most of the season, Bradshaw has played with a fractured bone in the small toe of his right foot and has practiced rarely.

...

authorship - NFL

Approach accuracy macro-f1 micro-f1
microTC 0.8444 0.8403 0.8444
Escalante et al. 2015 0.8376 0.7637 -
Cummins et al. 0.7778 0.7654 -

authorship

  • CCA
  • NFL
  • Business
  • Poetry
  • Travel
  • Cricket

OHSUMED

  • Colección de 348,566 (se uso un subset de 20K)
  • MEDLINE,
  • Título
  • Resumen
  • Identificar el sujeto de estudio.
{"text": "Haemophilus influenzae meningitis with prolonged hospital course.\n A retrospective evaluation of Haemophilus influenzae type b meningitis observed over a 2-year period documented 86 cases.\n Eight of these patients demonstrated an unusual clinical course characterized by persistent fever (duration: greater than 10 days), ... had no sequelae.\n", "klass": "C01"}
name accuracy macro-f1 micro-f1
Until 2015 ~0.40 - -
Huynh et al. 2015* 0.5690 - -
microTC 0.4611 0.3959 0.4611

Texto pre-procesado

  • 20 new groups
  • Reuters
    • 8 clases
    • 10 clases
    • 52
  • cade
  • webkb

cade

{"text": "br br email arvores arvores http www apoio mascote natureza vida links foram animais animais brasileira alunos andar sp desenvolvido projeto carlos guarulhos associacao associacao associacao associacao paulista maos homens homens homens uniao mesquita saudavel acao htm coragem diretoria aguia plantas plantas org org universo tradicional tiveram federacao garra penetrar tornaram assasinadas taz sobreviveram cruelmente shu barranco fontagua homenageamos homenageamos homenageamos insanas heroicamente devastadora kuoshu mundomagico kuo paraventi", "klass": "08_cultura"}

Preguntas