Youtaqa : système de questions-réponses intelligent basé sur le deep learning et la recherche d’information

Abstract

Users' need for comfort and the demand to have accurate answers to their questions are present nowadays which has given a new purpose to artificial intelligence. The best known search engines such as Google tend to offer brief answers to so-called »factoid» questions. This task is considered difficult in terms of the complexity of the queries and even their answers which can be the combination of several passages.

For this, in this thesis, our goal is based on the design and implementation of a Question-Answering system that can overcome the difficulties mentioned above and that is able to answer questions in several areas accurately and precisely using the Wikipedia knowledge base. Our system named YouTAQA starts by collecting the passages that can answer the query entered by the user and ends by extracting the start and end of the exact answer using Deep Learning. That being said, our system is capable of doing the complete pipeline, from collecting the relevant passages, to extracting the final answer requiring only the question as input. The Deep Learning modules of our system were implemented using the pre-trained BERT model which has been designed to perform various NLP tasks.

Experiments on the dataset have demonstrated the effectiveness of the proposed method and the results of the comparison have shown that our architecture improved the Question-Answering domain.

Keywords: Information Retrieval, Deep Learning, Natural Language Processing, Transfer Learning.

P~.t

~:'_w1 ·
· _{w?w oeh}tl·F#177; Tqy' ûA,A_>) _Yl_{à w}O.....1A_~ T_bAW_m1 T~I_rl·yd~ts_mIT_~A>

TM~ T
·_r`m1 .1..2-_b1 ûA
·_r~ TMS
·i TMym.3 ._·àAnW}1 °A
·@l A^3/4d_§d. A^3/4AR_r ûW_ài _·t1 _rRA.2,1

_·tl A_h3A4_?) _Yt~ Tl·F#177;1 dyq`~ "<sub>.</sub>"<sub>4</sub>->
· · T<sub>b</sub>`} T_mhm1 £@ _rbt`3 .z_~w 'L'A_tA_?) _oe§dq~ _Yl TM_?w?

^3/4

.T-._'w-,4 d_à · AA-_?§z  _wk3  i ·k_m§

_~lOEt( centsnk_m§ Tl·F#177;l _Yl_à T.0._>2
_·A\ú @yfn3 _oey_mO.3 _Yl_à An
·d d_mt`_§ , _r
·@_mI £@ ·
· , I@_h

· 1di.tFA_t T'd, ûA_s._m1 · d_§d`1 ·
· Tl·F#177;l <sub>Y</sub>l<sub>à</sub> T.0.,31 <sub>Y</sub>l<sub>à C</sub> AqI £®<sub>à</sub>i <sub>Cw</sub>
·@<sub>m</sub>I ûA,<sub>w</sub>`OI _Yl_à 1s1 _Yl_à T_iA_>3l A_hnkm_{§ ·}t1 _%oVAqm1_%oym,.t_u YouTAQA _Ymsm1 AnA\ú Lt_b§ .A_§dy_byk_§ T_àwF_w .Deep Learning I Tynq~
_·1d,tFA, Tqy'd1 T-_?A. 31 T_§A_hú T_§ld_i e_r~tFA._{s ·h}tn_§
·d~ts_mI centsli i _(c)@1

_&I_r"tFl _Y) , Tl} ûlÐ T,_w...#177;1 _%om? · , 4:u,_b1 û1_wW-> TMk._?
·Ayq1 _Yl_{à C} A' AnA\ú  4
· , -Ð %o

«_bs BERT _~Ð_wmú
_·Idi.tFA_~ AnA\ú _&._§Cd3 _oe.3 .TM.idm
· 1s1 Xq
· 4.4Wt~ _·t1 TyA_hn1 T_?A_~3~

.Tflt~m1 T_§wOEl1 T_~rb1
_·A_h °I #177; centsmymO~ _oe.3 (c)@1 ,.._§Cdt1

Antyn.  i Tú_CAq_m1 _acAtú _rh\3 , Ts_rtq_mlTq_§rW1 TyA`
· ûAúAy_b1 T_àwm-_{14 Y}l_{à L}.._tCA-₁₄t1 _c.R_w3

.Tl·F#177;~ _Yl_à T_t,A_>2 _r::
·i A^3/4A14..4 _·W`~

._oel`tI TMqú ,Ty`y_bW1 TOEll T,A` ,«y<sub>mà oe</sub>l`~ ,ûA_wl`_m1 _Yl_à ........_b1 :a1t6._o.J~ .it.o..t_,Jt