Coping with Data Scarcity: First Steps towards Word Expansion for a Chatbot in the Urban transportation Domain
View/ Open
Date
2020-11-26Author
García Montero, Eneritz
Metadata
Show full item recordAbstract
Hizkuntzaren Prozesamenduan (HP) zenbait arlotan hitzak erabili izan dira tradizionalki
zabaltze-tekniken garapenean, hala nola Informazioaren Berreskurapenean (IB) edota
Galdera-Erantzun (GE) sistemetan. Master tesi honek bi hurbilpen aurkezten ditu
Elkarrizketa-Sistemen (ES) arloan zabaltze-teknikak garatze aldera, zehazkiago
Donostiako (Gipuzkoa) hiri-garraiorako chatbot baten ulertze-modulua garatzera
zuzendurik. Lehenengo hurbilpenak hitz-bektoreak erabiltzen ditu semantikoki antzekoak
diren terminoak erauzteko, kasu honetan FastText-eko aurre-entreinaturiko embedding
sorta espainieraz eta bigarren hurbiltzeak hitzen adiera-desanbiguazioa erabiltzen du
sinonimoak datu-base lexiko baten bidez erauzteko, kasu honetan espainierazko
WordNet-a. Horretarako, ataza kolaboratibo bat diseinatu da, non corpusa osatuko
baitugu balizko-egoera erreal baten sarrerak jasoz. Bestalde, domeinuz kanpo dauden
sarrerak identi katze aldera, bi esperimentu sorta garatu dira. Lehenengo fasean
kali katze sistema bat garatu da, non corpuseko terminoak Term Frequency-Inverse
Document Frequency (TF-IDF) erabiliz ordenatzen baitiren eta ondoren
kali katze-sistema kosinu-antzekotasunaren bidez osatzen da. Bigarren faseak aurreko
kali katze-sistema formalizatuko da, hiru datu-multzo prestatuz eta estrati katuz.
Datu-multzo hauek erregresore lineal bat eta Kernel linealarekin euskarri bektoredun
makina bat entreinatzeko erabili dira. Emaitzen arabera, aurre-entreinaturiko bektoreek
leialtasun handiagoa daukate input errealari dagokionez. Hala ere, datu-base lexikoek
estaldura linguistiko zabalagoa gehituko diote zabalduriko corpus hipotetikoari. Azkenik,
domeinuaren diskriminazioari dagokionez, emaitzek TF-IDF-tik erauzitako termino
gehienen zeukan datu-multzoa hobesten dute. Text expansion techniques have been used in some sub elds of Natural Language
Processing (NLP) such as Information Retrieval or Question-Answering Systems. This
Master's Thesis presents two approaches for expansion within the context of Dialogue
Systems (DS), more precisely for the Natural Language Understanding (NLU) module of
a chatbot for the urban transportation domain in San Sebastian (Gipuzkoa). The rst
approach uses word vectors to obtain semantically similar terms while the second one
involves synonym extraction from a lexical database. For this purpose, a corpus composed
of real case scenario inputs has been exploited. Furthermore, the qualitative analysis of
the implemented expansion techniques revealed a need to lter out-of-domain inputs. In
relation to this problem, two di erent sets of experiments have been carried out. First,
the feasibility of using Term Frequency-Inverse Document Frequency (TF-IDF) and
cosine similarity as discrimination features was explored. Then, linear regression and
Support Vector Machine (SVM) classi ers were trained and tested. Results show that
pre-trained word embedding expansion constitutes a more loyal representation of real case
scenario inputs, whereas lexical database expansion adds a wider linguistic coverage to a
hypothetically expanded version of the corpus. For out-of-domain detection, increasing
the number of features improves both, linear regression and SVM classi cation results.