Coupling Retrieval and Meta-Learning for Context-Dependent Semantic Parsing
39878Authors
nan
Abstract
AbstractIn this paper, we present an approach to incorporate retrieved datapoints as supporting evidence for context-dependent semantic parsing, such as generating source code conditioned on the class environment. Our approach naturally combines a retrieval model and a meta-learner, where the former learns to find similar datapoints from the training data, and the latter considers retrieved datapoints as a pseudo task for fast adaptation. Specifically, our retriever is a context-aware encoder-decoder model with a latent variable which takes context environment into consideration, and our meta-learner learns to utilize retrieved datapoints in a model-agnostic meta-learning paradigm for fast adaptation. We conduct experiments on CONCODE and CSQA datasets, where the context refers to class environment in JAVA codes and conversational history, respectively. We use sequence-to-action model as the base semantic parser, which performs the state-of-the-art accuracy on both datasets. Results show that both the context-aware retriever and the meta-learning strategy improve accuracy, and our approach performs better than retrieve-and-edit baselines.
Keywords
Full Content
Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 855–866 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 855 � � ! "# $ % &'()*'++,+-./0//123+4560(7)*8(1*(9)61:/0;<(1=18>(7<80?@ A6/1B2+1BC(?D/E+7/0+7?+-F8B./0/G1/,?<8</12H7+*(<<81B9A6/1BI'+69H@J@3'81/ K8*7+<+-0J(<(/7*'G<8/9F(8L81B93'81/ MNOPQRSTUVWXYZW[[\RW]TUVWX^_[R[O_`QO_a] MQObV]NZ]V]QOV]ZUW]NcdPO^TUWaeP[Pfb_aPU gh i10'8<5/5(79j(57(<(10/1/557+/*'0+81*+7; 5+7/0(7(078(>(22/0/5+810</<<655+7081B(>8; 2(1*(-+7*+10(k0;2(5(12(10<(4/108*5/7<81B9 <6*'/<B(1(7/081B<+67*(*+2(*+12808+1(2+1 0'(*,/<<(1>87+14(10@l67/557+/*'1/067/,,? *+4E81(</7(078(>/,4+2(,/12/4(0/;,(/71(79 j'(7(0'(-+74(7,(/71<0+m12<848,/72/0; /5+810<-7+4 0'(07/8181B2/0/9/120'(,/00(7 *+1<82(7<7(078(>(22/0/5+810</</5<(62+0/<n -+7-/<0/2/50/08+1@)5(*8m*/,,?9+677(078(>; (78</*+10(k0;/j/7((1*+2(7;2(*+2(74+2(, j80'/,/0(10>/78/E,(j'8*'0/n(<*+10(k0(1; >87+14(10810+*+1<82(7/08+19/12+674(0/; ,(/71(7,(/71<0+608,8I(7(078(>(22/0/5+810< 81/4+2(,;/B1+<08*4(0/;,(/7181B5/7/28B4 -+7-/<0/2/50/08+1@o(*+126*0(k5(784(10< +13lp3l.q/123)rG 2/0/<(0<9j'(7( 0'(*+10(k07(-(7<0+*,/<<(1>87+14(1081sG; tG*+2(</12*+1>(7</08+1/,'8<0+7?97(<5(*; 08>(,?@o(6<(<(u6(1*(;0+;/*08+14+2(,/< 0'(E/<(<(4/108*5/7<(79j'8*'5(7-+74<0'( <0/0(;+-;0'(;/70/**67/*?+1E+0'2/0/<(0<@J(; <6,0<<'+j0'/0E+0'0'(*+10(k0;/j/7(7(078(>; (7/120'(4(0/;,(/7181B<07/0(B?8457+>(/*; *67/*?9/12+67/557+/*'5(7-+74<E(00(70'/1 7(078(>(;/12;(280E/<(,81(<@ v w 3+10(k0;2(5(12(10<(4/108*5/7<81B/84<0+4/5 /1/067/,,/1B6/B(600(7/1*(0+/<076*067/,,+B8; */,-+74x(@B@<+67*(*+2(y*+12808+1(2+1/B8>; (1*+10(k0x(@B@*,/<<(1>87+14(10yxD81B(0/,@9 z{|}~D+1B(0/,@9z{|}~i??(7(0/,@9z{|~i?; (7(0/,@9z{|~)6'7(0/,@9z{|~)6'7/12G70I89 z{|y@)0/12/72/557+/*'(<0?58*/,,?,(/71/+1(; <8I(;m0<;/,,4+2(,+10'((1087(07/8181B2/0/<(09 j'8*'8<-(2j80'(/*'(k/45,(8128>826/,,?810'( 07/8181B5'/<(/124/n(<57(28*08+1<-+7(/*'0(<0 (k/45,(810'(81-(7(1*(5'/<(@+j(>(790/n81B o+7n2+1(j'8,(0'8</60'+7j/</1810(71/0K8*7+<+-0 J(<(/7*'@ *+2(B(1(7/08+1/</1(k/45,(957+B7/44(7<6<6; /,,?2+1+0j780(*+2(<-7+4 <*7/0*'810'(7(/, j+7,2@o'(10'(?j780(/58(*(+-*+2(81/5/7; 08*6,/7(1>87+14(1090'(?0?58*/,,?,(>(7/B(5/<0 (k5(78(1*(+1j78081B+77(/281B*+2(<810'(<848; ,/7<806/08+1/</B682/1*(@K(/1j'8,(92/0/5+810< -+7/0/<n4/?>/7?j82(,?x6/1B(0/,@9z{|/y9 0'6<808<2(<87/E,(0+,(/71/5(7<+1/,8I(24+2; (,-+70'(0/7B(02/0/5+810@i10'8<j+7n9j(<062? '+j0+/60+4/08*/,,?7(078(>(<848,/72/0/5+810<81 /*+10(k0;2(5(12(10<*(1/78+/126<(0'(4/<0'( <655+7081B(>82(1*(0+-/*8,80/0(<(4/108*5/7<81B@ &'(7(/7(7(*(10/00(450</0(k5,+8081B7(; 078(>(2(k/45,(<0+8457+>(0'(B(1(7/08+1+-,+B; 8*/,-+74/120(k0@J(078(>(;/12;(280/557+/*'(< x/<'84+0+(0/,@9z{|~6/1B(0/,@9z{|E~o6 (0/,@9z{|~A6(0/,@9z{|y0?58*/,,?m7<06<(/ *+10(k0;812(5(12(107(078(>(70+m120'(4+<07(,; (>/102/0/5+8109/120'(16<(80/</1/22808+1/, 81560+-0'((28081B4+2(,@+j(>(79/*+10(k0; /j/7(7(078(>(78<>(7?845+70/10-+70'(0/<n+- *+10(k0;2(5(12(10<(4/108*5/7<81B@ +7(k/4; 5,(<9/<<'+j181 8B67(|9*,/<<(1>87+14(10*/1 '(,50'(7(078(>(72(*82(j'(0'(70'(2(<87(2*+2( +-8<B(1(7/0(2E?287(*0,? */,,81B+780(7/081B0'( /77/? 0+81*7(4(10(/*'(,(4(10@ 670'(74+7(97(078(>(; /12;(280/557+/*'(<0?58*/,,?*+1<82(7+1,?+1( <848,/7(k/45,(0+(280@i1<(4/108*5/7<81B90'( 5/00(71+-/<076*067/,+605604/?*+4(-7+428-; -(7(107(078(>(2(k/45,(<@&'(7(/,<+(k8<0j+7n< 0+608,8I(46,085,((k/45,(<0+B682(0'(<(4/108* 5/7<(7x/?/08(0/,@9z{|~6/1B(0/,@9z{|/y9 '+j(>(790'(<(/557+/*'(<(80'(76<(/'(678<08* j/?0+(k5,+800'(7(078(>(2,+B8*/,-+74<6*'/< 81*7(/<81B0'(57+E/E8,80?+-/*08+1<x/?/08(0/,@9 z{|y+76<(/7(,(>/1*(-61*08+12(<8B1(2/12 ,(/71(2E/<(2+1(k5(708<(/E+600'(0/7B(0,+B8; */,-+74x6/1B(0/,@9z{|/y@o'(1j(*+1<82(7 0'(*+10(k0(1>87+14(10980¡<1+1078>8/,0+2(<8B1 856 ��� � � � !"� ���$%&'()�)'*++�,(-$'./.)012�(-$'.-.30+�,.2(*'(4*&'.�5� ���������71%&'.89�:.);'.-.30+<� ���������71%&'.89�=.(>?0+<� ���������@AB�C77+�*�+)*'*2�01�0?(+�:.)012�(3�$'*).D� ���������$%&'()�:1(7�*77E71%&'.�*2>FG�� ��������H� IJ"�KL !�!M � L! � �NL� ��E*G� ���$%&'()�:1(7�(3)EG�O�0?(+D*77EPG<Q�������������������� E&G� ���$%&'()�:1(7�(3)EG�O� ������S12�E(30�(�T�F<�(U�:.);'.-.30+D'.3>0?<�(VVGO� ���������������:.);'.-.30+8(9VTP<�Q��Q� WX YZ[X �\� � ��� � �X �X �]^ _`]�`��a X �\�a a X � bccdef ]\` � ghijkhlhmno� � � p qfa a � �a \�qX� f� � � X r �a X X � � X� s ��� ��q �ta� f \� � �� � � q]uX�� v XfwxYyz{r� fwxY|`f �� � �X �X � � � � �\ ��a X� v�a� � X � q]W� fwxY}` �� X � �a� ��~ q ��a �]a `�q � � Xf a �a � �� a a �a �qs �aa �� � �q a � �a X �a� �q v � � �aa �[s^[s]p � fwxY`�[t ]t��� fwxY` ��f �q� X �X � �� X �� ~ �q XX �a �� { � �aa �� � � a �\�� v� a X � �� � X a \ � �\� W �� �\ �� �� � X �X a a � [a�a� X��a �� � �X �X� � X� p f �q f� X �� �� ~ � Xf��� WX Y��a X � �� � �X �X]^_` af X� �X �� � iW � f � a q Z]Y` � � �\ ga � �\ �� ��a]X c kh¡¢ghijkhlhmno`f �]w` �lf X� a]Xg£cbccde`v a [s^[sY��]p � fwxY`¤��\ …
Content truncated to 500 words
Labeled Text
AbstractIn this paper, we present an approach to incorporate retrieved datapoints as supporting evidence for context-dependent semantic parsing, such as generating source code conditioned on the class environment. Our approach naturally combines a retrieval model and a meta-learner, where the former learns to find similar datapoints from the training data, and the latter considers retrieved datapoints as a pseudo task for fast adaptation. Specifically, our retriever is a context-aware encoder-decoder model with a latent variable which takes context environment into consideration, and our meta-learner learns to utilize retrieved datapoints in a model-agnostic meta-learning paradigm for fast adaptation. We conduct experiments on CONCODE and CSQA datasets, where the context refers to class environment in JAVA codes and conversational history, respectively. We use sequence-to-action model as the base semantic parser, which performs the state-of-the-art accuracy on both datasets. Results show that both the context-aware retriever and the meta-learning strategy improve accuracy, and our approach performs better than retrieve-and-edit baselines.
Publication Details
- Paper ID:
- 39878
- Year:
- 2019
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan