-
Notifications
You must be signed in to change notification settings - Fork 2
/
2019_qb.tex
executable file
·1315 lines (947 loc) · 36.5 KB
/
2019_qb.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
\documentclass[xcolor=dvipsnames]{beamer}
\usetheme[
showdate=true, % show the date on the title page
alternativetitlepage=true, % Use the fancy title page.
titlepagelogo=general_figures/shell, % Logo for the fir\
st page.
]{UMD}
%\usetheme{Rochester}
%\usepackage{beamerthemesplit}
\usepackage{xmpmulti}
\usepackage{booktabs}
\usepackage{graphicx,float,wrapfig, bbm}
\usepackage{amsfonts, bbold, comment}
\usepackage{mdwlist}
\usepackage{tikz}
\usepackage{subfigure}
\usepackage{colortbl}
\usepackage{multirow}
\usetikzlibrary{shapes.geometric}
\definecolor{xred}{HTML}{DB4437}
\definecolor{xyellow}{HTML}{F4B400}
\definecolor{xgreen}{HTML}{0F9D58}
\newcommand*{\tcircle}[1]{\tikz[anchor=base,baseline=-2.5pt] \node[circle,fill=#1,scale=0.9] (X) {};}
\newcommand*{\tsquare}[1]{\tikz[anchor=base,baseline=-2.5pt] \node[fill=#1,scale=1.2] (X) {};}
\newcommand*{\tdiamond}[1]{\tikz[anchor=base,baseline=-2.5pt] \node[diamond,fill=#1,scale=0.7] (X) {};}
\newcommand*{\ttriangle}[1]{\tikz[anchor=base,baseline=-1.5pt] \node[regular polygon,regular polygon sides=3,fill=#1,scale=0.6] (X) {};}
\newcommand{\fsi}[2]{
\begin{frame}[plain]
\vspace*{-1pt}
\makebox[\linewidth]{\includegraphics[width=\paperwidth]{#1}}
\begin{center}
#2
\end{center}
\end{frame}
}
\newcommand{\abr}[1]{\textsc{#1} }
\newcommand{\pos}[1]{{\texttt{#1}}}
\newcommand{\e}[2]{\mathbb{E}_{#1}\left[ #2 \right] }
\newcommand{\ind}[1]{\mathbb{I}\left[ #1 \right] }
\newcommand{\ex}[1]{\mbox{exp}\left\{ #1\right\} }
\newcommand{\g}{\, | \,}
\newcommand{\citename}[1]{#1 }
\newcommand{\gfxs}[2]{
\begin{center}
\includegraphics[width=#2\linewidth]{simtrans/#1}
\end{center}
}
\newcommand{\gfxq}[2]{
\begin{center}
\includegraphics[width=#2\linewidth]{qb/#1}
\end{center}
}
%\usecolortheme{ucdblack}
\title[HCQA]{Cooperative and Competitive Machine Learning through Question Answering}
\author{ Jordan Boyd-Graber et al.}
\date{2018}
\institute[Maryland] % (optional, but mostly needed)
{University of Maryland}
\begin{document}
\frame{
\titlepage
\tiny
}
\fsi{qb/turing}{Turing Test: Definition of AI (Image from Wall Street
International)}
\fsi{qb/mimno-tweet}{Models pretty of do bag well word}
\fsi{qb/mimno-tweet}{Bag of word models do pretty well! \\ \pause
How do we separate interesting from mechanistic?}
\begin{frame}{Overview}
\begin{itemize}
\item How to do QA right
\item How to compete against humans
\item How to explain QA (and measure it)
\item How to make QA tasks difficult
\item Connections to other NLP tasks
\end{itemize}
\end{frame}
\fsi{qb/human_reading}{SQuAD: Ignore Knowledge}
\fsi{qb/jeopardy}{IBM Watson: QA Solved!}
\fsi{qb/asms}{Arkansas School for Mathematics and Sciences}
\fsi{qb/caltech}{California Institute of Technology}
\fsi{qb/princeton}{Princeton University}
\fsi{qb/pub_quiz}{Princeton Pub Quiz}
\fsi{qb/naqt_jbg_1}{Affiliations, Courtesy of National Academic Quiz Tournaments}
%\fsi{qb/naqt_jbg_2}{Statistics}
\fsi{qb/boulder}{University of Colorado Boulder}
\fsi{qb/maqt}{University of Maryland}
% Add in some trivia / QB image
\fsi{qb/jeopardy}{IBM Watson: QA Solved?}
\fsi{qb/klingon}{ Not just an academic question for me \dots }
\begin{frame}{Is Jeopardy! the right venue?}
\begin{center}
\href{https://www.youtube.com/embed/fgEoiGQmbgk?start=86&autoplay=1}{ghoBE'!}
\end{center}
\gfxq{final_jeopardy}{1.0}
\end{frame}
\begin{frame}{What is a good question?}
\begin{columns}
\column{.5\linewidth}
\gfxq{siri}{.6}
\column{.5\linewidth}
\begin{itemize}
\item Reason across sentences
\item Written by experts
\item Compare/compete with humans
\item Test uncertainty
\item Advance science
\end{itemize}
\end{columns}
\end{frame}
\fsi{qb/squad}{SQuAD: Rajpurkar et al., 2016 (Reason)}
\fsi{qb/triviaqa}{TriviaQA: Joshi et al., 2017 (Experts)}
\fsi{qb/squad_2}{Test uncertainty (Rajpurkar et al., 2018)}
\begin{frame}
\frametitle{This isn't new!}
\begin{columns}
\column{.5\linewidth}
\begin{itemize}
\item Game called ``quiz bowl''
\item Two teams play each other
\begin{itemize}
\item Moderator reads a question
\item When a team knows the answer, they signal (``buzz'' in)
\item If right, they get points; otherwise, rest of the question is read to the other team
\end{itemize}
\item Hundreds of teams in the US alone
\only<2>{\item Example \dots}
\end{itemize}
\column{.5\linewidth}
\includegraphics{qb/quizbowl}
\end{columns}
\end{frame}
\begin{frame}[t]
\frametitle{Sample Question}
\small
\only<2->{This man wrote a ballet in which 60 white nymphs perform the
Rainbow Dance, so named because different colored lights are focused
on them---\only<3->{that ballet Alethes and Iris was never produced
because the theater manager was convinced it would cause a
fire. He invented the first \only<4->{cow catcher and campaigned
for passage of an act against "street nuisances" like outdoor
musicians, who he claimed wasted a quarter of his life. He
responded to \only<5->{William Whewell's statement that no
scientific explanation of the universe was possible in his
\only<6->{"Ninth Bridgewater Treatise," which attacked the
eight previous treatises funded by the Earl of
Bridgewater. He also wrote about the influence of
aristocrats on science in his \only<7->{Reflections on the
Decline of Science in England . This author of the memoir
\only<8->{Passages From the Life of a Philosopher improved
on \only<9->{Jacquard's invention of punched cards by
proposing a device which incorporated an
\only<10->{input, storage system, processor, control
unit, and output device. FTP, name this inventor of
the \only<11->{Analytic Engine and the Difference
Engine.}}}}}}}}}}
\only<12->{ {\bf Charles Babbage}}
\end{frame}
\begin{frame}[t]
\frametitle{Sample Question}
The Swiss-Italian architect Pietro Antonio Solari
\only<2->{built several fortified towers in this city, which
often vied for power with its northern rival Tver. A ruler
of this city prevailed in the} \only<3->{Great Stand on the
Ugra River. A prince from this city was nicknamed for
winning a battle on the} \only<4->{Don river. Partly because
a ruler of this city married} \only<5->{Sophia Palaiologina,
the niece of the last Byzantine Emperor, this city styled
itself the} \only<6->{``Third Rome'' after the fall of
Constantinople. Another prince of this city stopped paying
tribute to the} \only<7->{Mongols in 1476, ending the
``Tatar yoke.''} \only<8->{The Grand Duchy headquartered in
this city came to an end in 1547 with the ascension of}
\only<9->{ Ivan IV, who made it his capital. For 10 points,
name this city where Ivan III renovated the
Kremlin,} \only<10->{the capital of Russia.}\\
\vspace{.5cm} \only<11->{ {\bf Moscow} (Moskva / Muscovy)}
\end{frame}
\begin{frame}
\frametitle{Question Structure Enables Compeition}
\begin{columns}
\column{.5\linewidth}
\includegraphics[width=1.0\linewidth]{qb/jeopardy}
\column{.5\linewidth}
\begin{itemize}
\item Watson must decide to answer {\bf once}, after
complete question
\item Quiz Bowl: decide after each word
\item Obscure clues at start, easy at end
\item ``Gold standard'' in trivia community
\end{itemize}
\end{columns}
\end{frame}
\begin{frame}{Checklist}
\begin{itemize}
\item Reason across sentences\only<2->{: question and source}
\item Written by experts\only<3->{: trivia community}
\item Compare/compete with humans\only<4->{: by design}
\item Test uncertainty\only<5->{: by design}
\item \alert<6>{Advance science} \only<6>{(More later)}
\end{itemize}
\end{frame}
\begin{frame}{How to approach this problem \dots}
\only<1>{
\begin{columns}
\column{.5\linewidth}
\gfxq{guess}{0.8}
\column{.5\linewidth}
\gfxq{buzzer}{0.8}
\end{columns}
}
\only<2>{
\gfxq{guess}{0.5}
}
\end{frame}
\begin{frame}{}
\begin{columns}
\column{.4\linewidth}
\begin{center}
\includegraphics[width=0.8\linewidth]{general_figures/mohit}
\end{center}
\column{.6\linewidth}
\begin{block}{ {\bf \href{http://cs.colorado.edu/~jbg//docs/2014_emnlp_qb_rnn.pdf}{A Neural Network for Factoid Question Answering over Paragraphs}}}
\underline{\href{http://cs.umd.edu/~miyyer/}{Mohit Iyyer}}, {\bf Jordan Boyd-Graber}, Leonardo Claudino, Richard Socher, and Hal {Daum\'{e} III}. \emph{Empirical Methods in Natural Language Processing}, 2014
\end{block}
\begin{block}{ {\bf \href{file:///Users/jbg/public_html/docs/2015_acl_dan.pdf}{Deep Unordered Composition Rivals Syntactic Methods for Text Classification}}}
\underline{\href{http://cs.umd.edu/~miyyer/}{Mohit Iyyer}}, Varun
Manjunatha, {\bf Jordan Boyd-Graber} and Hal {Daum\'{e} III}. \emph{Empirical Methods in Natural Language Processing}, 2014
\end{block}
\end{columns}
\end{frame}
\begin{frame}{Vector Space Model}
\only<1>{\gfxq{unigram_models_0}{.8}}
\only<2>{\gfxq{unigram_models_1}{.8}}
\only<3>{\gfxq{unigram_models_2}{.8}}
\only<4>{\gfxq{unigram_models_3}{.8}}
\only<5>{\gfxq{unigram_models_4}{.8}}
\only<6>{\gfxq{unigram_models_5}{.8}}
\only<7>{\gfxq{unigram_models_6}{.8}}
\only<8>{\gfxq{unigram_models_7}{.8}}
\only<9>{\gfxq{unigram_models_8}{.8}}
\end{frame}
% \begin{frame}{Training}
% \begin{columns}
% \column{.5\linewidth}
% \begin{itemize}
% \item Initialize embeddings from \textsc{word2vec}
% \item Randomly initialize composition matrices
% \item Update using \textsc{warp}
% \begin{itemize}
% \item Randomly choose an instance
% \only<2->{\item Look where it lands}
% \only<4->{\item Has a correct answer}
% \only<5->{\item Wrong answers may be closer}
% \only<6->{\item Push away wrong answers
% \item Bring correct answers closer}
% \end{itemize}
% \end{itemize}
% \column{.5\linewidth}
% \only<1>{\gfxq{warp_training_5}{.8}}
% \only<2>{\gfxq{warp_training_4}{.8}}
% \only<3>{\gfxq{warp_training_3}{.8}}
% \only<4>{\gfxq{warp_training_2}{.8}}
% \only<5>{\gfxq{warp_training_1}{.8}}
% \only<6>{\gfxq{warp_training_0}{.8}}
% \end{columns}
% \end{frame}
\begin{frame}{More complicated representations}
\gfxq{embedding}{1.0}
\end{frame}
\begin{frame}{How to approach this problem \dots}
\only<1>{
\begin{columns}
\column{.5\linewidth}
\gfxq{guess}{0.8}
\column{.5\linewidth}
\gfxq{buzzer}{0.8}
\end{columns}
}
\only<2>{
\gfxq{buzzer}{0.5}
}
\end{frame}
\begin{frame}{}
\begin{columns}
\column{.5\linewidth}
\includegraphics[width=0.7\linewidth]{general_figures/hehe}
\column{.5\linewidth}
\begin{block}{{\bf
\href{http://cs.colorado.edu/~jbg//docs/qb_emnlp_2012.pdf}{Besting
the Quiz Master: Crowdsourcing Incremental
Classification Games}}}
{\bf Jordan Boyd-Graber}, He He, and Hal {Daum\'{e} III}. \emph{Empirical Methods in Natural Language Processing}, 2012
\end{block}
\begin{block}{{\bf
\href{http://www.cs.colorado.edu/~jbg/docs/2016_icml_opponent.pdf}{Opponent Modeling in Deep Reinforcement Learning}}}
He He, {\bf Jordan Boyd-Graber}, Kevin Kwok, and Hal
{Daum\'{e} III}. \emph{International Conference of Machine Learning}, 2016
\end{block}
\end{columns}
\end{frame}
\begin{frame}
\frametitle{Interface}
\begin{columns}
\column{0.5\linewidth}
\begin{center}
\includegraphics[width=0.8\linewidth]{qb/screenshot}
\end{center}
\column{0.5\linewidth}
\only<2>{
\begin{itemize}
\item 7000 questions: first day
\item 43000 questions: two weeks
\item 461 unique users
\item Imitated \dots
\end{itemize}
\gfxq{protobowl}{.8}
}
\end{columns}
\end{frame}
\begin{frame}{Not all opponents are equal}
\gfxq{player_profile}{.9}
\pause
Varies by category!
\end{frame}
\begin{frame}{Many models, choose which: DRON-MoE}
\only<1>{\gfxq{dron-moe}{.8}}
\only<2>{\gfxq{dron-moe2}{.8}}
\end{frame}
\begin{frame}{Reward}
\gfxq{dqn_results}{.5}
\end{frame}
\begin{frame}{Reward: Closer Look}
\only<1>{\gfxq{reward1}{.8}}
\only<2>{\gfxq{reward2}{.8}}
\only<3>{\gfxq{reward3}{.8}}
\only<4>{\gfxq{reward4}{.8}}
\end{frame}
\begin{frame}{Experiment 1}
\begin{columns}
\column{.25\linewidth}
\gfxq{colby_jeo}{1.0}
Colby Burnett:
\$375,000
\column{.25\linewidth}
\gfxq{ben_jeo}{1.0}
Ben Ingram:
\$427,534
\column{.25\linewidth}
\gfxq{alex_jeo}{1.0}
Alex Jacobs: \$151,802
\column{.25\linewidth}
\gfxq{kristin_jeo}{1.0}
Kristin Sausville: \$95,201
\end{columns}
\pause
\begin{center}
End result: 200-200 tie!
\end{center}
\end{frame}
\fsi{qb/hsnct1}{}
\fsi{qb/jennings}{23. October 2015, Seattle}
\fsi{qb/jennings_handshake}{300-160}
% \fsi{qb/hsnct_2016}{330-60} % Not sure of exact score, double check
\fsi{qb/nasat}{Humans 345-145 (vs. OUSIA)}
\fsi{qb/hsnct_2017}{Best Human Trivia Players {\bf Barely} Defeated,
June 2017}
\fsi{qb/hcqa}{Best Human Trivia Players {\bf Soundly} Defeated,
December 2017 \\ (OUSIA)}
\fsi{qb/seattle_crowd}{Question Answering is Not a Trivial Activity}
\fsi{qb/chicago_crowd}{http://qanta.org}
\fsi{qb/boring_dot_products}{Boring Dot Products}
\fsi{general_figures/blackbox}{}
\fsi{interpretability/lime_explanation}{LIME: Ribeiro et al. (2016)}
\fsi{interpretability/lime_image_explain}{LIME: Ribeiro et al. (2016)}
\fsi{interpretability/mt_task}{LIME: Ribeiro et al. (2016)}
\fsi{interpretability/mt_results}{What can humans do for ML?}
\fsi{general_figures/kill_all_humans}{}
\fsi{general_figures/bennet_robot}{}
\fsi{general_figures/tng_poker}{}
\fsi{simtrans/centaur-chess}{Centaur Chess}
\begin{frame}{}
\begin{columns}
\column{.4\linewidth}
\begin{center}
\includegraphics[width=0.8\linewidth]{general_figures/shi}
\end{center}
\column{.6\linewidth}
\begin{block}{{\bf What can AI do for me: Evaluating Machine Learning Interpretations in Cooperative Play}} \underline{\href{http://users.umiacs.umd.edu/~shifeng/}{Shi Feng}} and {\bf Jordan Boyd-Graber}. \emph{Intelligent User Interfaces}, 2019
\end{block}
\end{columns}
\end{frame}
\begin{frame}{Team-Based Interpretability}
\only<1>{\gfxq{qb_centaur_1}{.9}}
\only<2>{\gfxq{qb_centaur_2}{.9}}
\only<3>{\gfxq{qb_centaur_3}{.9}}
\only<4>{\gfxq{qb_centaur_6}{.9}}
\end{frame}
\fsi{qb/augment/screenshot_all}{Interface}
\fsi{qb/augment/screenshot_guesses}{}
\fsi{qb/augment/screenshot_highlight}{{\bf Highlighting}}
\fsi{qb/augment/screenshot_evidence}{}
\begin{frame}{Experts vs. Novices}
\begin{block}{Experts}
Trivia experts, familiar with task, enjoy the task
\end{block}
\begin{block}{Mechanical Turkers}
Mechanical Turkers: easily overwhelmed, need the help
\end{block}
\end{frame}
\fsi{qb/augment/tools_acc}{Evidence helps novices, experts are expert}
\fsi{qb/augment/tools_buzz}{Hights help experts}
\begin{frame}{Regression Analysis}
For each triple (player, question, interpretations), we predict the outcome
(correct answer or not) with a logistic regression. The features include:
\begin{itemize}
\item player ID
\item question ID
\item buzzing position
\item enabled interpretations: individual and combinations
\end{itemize}
\pause
\begin{block}{Coefficients tell story!}
\begin{itemize}
\item {\bf Big, Positive}: Help
\item {\bf Big, Negative}: Hurt
\item {\bf Small}: Neutral
\end{itemize}
\end{block}
\end{frame}
\fsi{qb/augment/coefs_0}{Everything helps: Evidence for novies,
Highlight for experts}
\fsi{qb/augment/coefs_1}{Synergistic effects}
\fsi{qb/augment/coefs_2}{Highlight and evidence help experts most}
\fsi{qb/augment/coefs_3}{For novices, less synergy}
\begin{frame}{Improvement through Reinforcement Learning}
\only<1>{\gfxq{rl_centaur_2}{.9}}
\only<2>{\gfxq{rl_centaur_3}{.9}}
\only<3>{\gfxq{rl_centaur_4}{.9}}
\only<4>{\gfxq{rl_centaur_5}{.9}}
\only<5>{\gfxq{rl_centaur_6}{.9}}
\end{frame}
\begin{frame}{Can we improve QA systems?}
\begin{columns}
\column{.6\linewidth}
\gfxq{trick/pyramid}{.9}
\column{.4\linewidth}
\begin{itemize}
\item Questions should be pyramidal
\item But for whom?
\begin{itemize}
\item Quotes
\item Reusing clues
\end{itemize}
\item Adversarial writing
\item Improve questions
\end{itemize}
\end{columns}
\end{frame}
\begin{frame}{What do we mean by ``adversarial''?}
\gfxq{trick/flow_chart_horizontal_label}{1.0}
\begin{itemize}
\item Round 1: Only IR interpretations
\item Round 2: IR and RNN (influence functions) interpretations
\pause
\item Another reason we need to have good explanations of QA
\end{itemize}
\end{frame}
\fsi{qb/trick/brahms_0}{\href{http://write.qanta.org}{http://write.qanta.org}}
\fsi{qb/trick/brahms_1}{}
\fsi{qb/trick/brahms_2}{}
\fsi{qb/trick/brahms_3}{}
\fsi{qb/trick/brahms_4}{}
\fsi{qb/trick/brahms_5}{}
\fsi{qb/trick/round_one}{Round 1: Only IR-based QA system}
\fsi{qb/trick/round_two}{Round 2: RNN-based QA system}
\begin{frame}{Competition}
\gfxq{trick/pace}{.8}
\begin{itemize}
\item December 15: Seven top human teams, fourteen computer teams
\item Top four teams from each ``division'' faced off against each
other
\pause
\item All computer teams lost to human teams
\pause
\item But two games were really close; strongest system was based on BERT
\item YouTube video series soon! \dots
\end{itemize}
\end{frame}
% Hard but not really a good question: F3 Q24
\begin{frame}{Impossible Until the End}
\alert<3>{Ritchie Watson commended this play's historical accuracy for
getting the price for a dozen eggs right---ten cents---to defend
against Elizabeth Hardwick’s contention that it was a sentimental
history.} \alert<4>{At the end of this play, a man wonders why a wheelchair is
at the top of a staircase, and} \alert<5>{Alexandra announces that she is leaving
her mother. Leo is pressured into stealing a set of valuable railroad
bonds in this drama. In this play, which takes its title from the Song
of Solomon,} Regina Hubbard schemes \tdiamond{xgreen}
\tcircle{xgreen} to obtain a majority share in a cotton mill. For 10
points, \tsquare{xgreen} \ttriangle{xgreen} name this play by
Lillian Hellman. \\
\pause
\textbf{Answer}: The\ Little\ Foxes\\
\only<3>{Academic literature}
\only<4>{Vague plot summary}
\only<5>{Avoid last names}
\end{frame}
% Tricky and hard: F3 Question 25
\begin{frame}{Tricky and impossible for current systems}
In Our Town, a character \ttriangle{xred} with this given name explains Grover’s Corners’ place in the universe. In The Crucible, a character \tcircle{xred} with this first name contends that the girls’ actions are part of their “silly seasons” and is the wife of Francis Nurse. A novel with this name, which conducts hidden messages to Rommel in The English Patient, is titled for a character who is killed in a boating accident at Manderley. For 10 points, give this name of a Daphne du Maurier \tdiamond{xyellow} \tsquare{xgreen} gothic novel which is also the first name of Miss Sharp, the protagonist of William Thackeray's Vanity Fair. \tcircle{gray} \ttriangle{gray}
\ttriangle{xred} Thornton Wilder
\tcircle{xred} Richard
\tcircle{gray} Richard
\ttriangle{gray} Thornton Wilder \\
\textbf{Answer}: Rebecca\\
\end{frame}
% Tricky and hard: F3 Question 29
\begin{frame}{Close, but \dots}
An army that took its name from this geographical feature had a
British doctor, James Paroissien, as its Surgeon General, and recent
scholarship by Peter Blanchard revealed its use of slaves. That army
crossed this geographical feature according \ttriangle{xred} to
Thomas Maitland’s \tcircle{xred} plans at Uspallata and Los
Patos. Spanish forces under Rafael Maroto were defeated in the
foothills of these mountains by an army led by José de San Martín and
Bernardo O'Higgins. \tsquare{xred} For 10 points, name this mountain
range of South America that played a role \ttriangle{xyellow} in the
independence of Chile. \tdiamond{gray} \tcircle{gray} \tsquare{gray}
\ttriangle{xred} Angel Falls
\tcircle{xred} Mountain
\tsquare{xred} Battle of Chacabuco
\tdiamond{gray} Battle of Chacabuco
\tcircle{gray} Mountain
\tsquare{gray} Battle of Chacabuco \\
\pause
\textbf{Answer}: Andes\\
\end{frame}
% Showing strength of QB format: Packet F3 Question 30
\begin{frame}{Showing the strength of quiz bowl format}
A painting by this artist \tsquare{xred} presents a tree as a cupboard with two doors, behind which are a white sphere and a miniature lighted house. A series of three paintings by this man shows a daytime sky with cumulus clouds above a dark house illuminated by a lamppost. This artist of Blood Will Tell and The Empire of Lights depicted a clock set to 12:43 \tcircle{xgreen} as a locomotive speeds out of a fireplace and a man in \tsquare{xyellow} a bowler hat with an \ttriangle{xgreen} apple in front of his face. For 10 points, name this Belgian surrealist painter \tdiamond{xgreen} of the Son of Man and Time Transfixed.
\tsquare{xred} Argon \\
\pause
\textbf{Answer}: René\ Magritte\\
\end{frame}
% Why it's a good learning framework: Packet F3 Question 31
\begin{frame}{Our QA Systems are Crazy}
\only<1-2>{
\begin{block}{Out of Touch}
On "The Critic", this man was replaced by a drunken animatronic bear from the Country Bear Jamboree but nobody seemed to notice. This man was edited to appear in the film "Contact", which prompted an angry statement from Mike McCurry. He was portrayed by Dennis Quaid in "The Special Relationship", who ate McDonald's every day to prepare for the role. A fictionalized version of this man named Henry Burton, a charismatic Southern governor running for the Democratic nomination, is portrayed by John Travolta in "Primary Colors". For ten points, name this American president who played the saxophone on an appearance on the Arsenio Hall Show.
\end{block} }
\only<2>{{\bf Samuel L. Jackson?}}
\end{frame}
\begin{frame}{Matching Entites Across Sentences}
\begin{block}{\only<2->{Magic Flute}}
At its premiere, \alert<3>{the librettist of this opera} portrayed
\alert<4>{a character who asks for a glass of wine with his dying wish}. \alert<4>{That
character} in this opera is instructed to ring some bells to summon
his love. At its beginning, \alert<5>{a man} who claims to have killed a (*)
serpent has a padlock put on \alert<5>{his} mouth because of \alert<5>{his} lying. The
plot of this opera concerns a series of tests that \alert<5>{Tamino} must
undergo to rescue Tamina from Sorastro. For 10 points, name this
Wolfgang Mozart opera titled for \alert<6>{an enchanted woodwind instrument}.
\end{block}
\only<3-4>{{\bf Not all references are named (\alert<3>{Emanuel
Schikaneder}, \alert<4>{Papageno})}}
\only<5>{Need to be able to match pronouns across sentences (or have
deep world knowledge)}
\only<6>{Requires semantic knowledge}
\end{frame}
\begin{frame}{}
\begin{columns}
\column{.4\linewidth}
\includegraphics[width=0.8\linewidth]{general_figures/hehe} \\
\includegraphics[width=0.8\linewidth]{general_figures/alvin}
\column{.6\linewidth}
\begin{block}{ {\bf \href{http://umiacs.umd.edu/~jbg//docs/2015_emnlp_rewrite.pdf}{Syntax-based Rewriting for Simultaneous Machine Translation}}}
He He, Alvin Grissom II, {\bf Jordan Boyd-Graber}, and Hal {Daum\'{e} III}. \emph{Empirical Methods in Natural Language Processing}, 2015
\end{block}
\begin{block}{ {\bf \href{http://umiacs.umd.edu/~jbg/docs/2016_naacl_interpretese.pdf}{Interpretese vs. Translationese: The Uniqueness of Human Strategies in Simultaneous Interpretation}}}
He He, {\bf Jordan Boyd-Graber}, and Hal {Daum\'{e} III}.
\emph{North American Association for Computational Linguistics}, 2016
\end{block}
\end{columns}
\end{frame}
\begin{frame}{Simultaneous Interpretation is Hard!}
\begin{columns}
\column{.5\linewidth}
\begin{itemize}
\item Exhausting for humans
\item Computers not trusted
\item Differential strengths
\item Same word-by-word characteristic
\end{itemize}
\column{.5\linewidth}
\gfxs{computer-interpreter}{1.0}
\end{columns}
\end{frame}
\begin{frame}{How we could translate a sentence}
\only<1>{\gfxs{example_3}{.9}}
\only<2>{\gfxs{example_4}{.9}}
\only<3>{\gfxs{example_5}{.9}}
\only<4>{\gfxs{example_6}{.9}}
\only<5>{\gfxs{example_7}{.9}}
\only<6>{\gfxs{example_8}{.9}}
\only<7>{\gfxs{example_9}{.9}}
\only<8>{\gfxs{example_10}{.9}}
\only<9>{\gfxs{example_11}{.9}}
\only<10>{\gfxs{example_12}{.9}}
\only<11>{\gfxs{example_13}{.9}}
\only<12>{\gfxs{example_14}{.9}}
\only<13>{\gfxs{example_15}{.9}}
\only<14>{\gfxs{example_16}{.9}}
\only<15>{\gfxs{example_17}{.9}}
\only<16>{\gfxs{example_18}{.9}}
\only<17>{\gfxs{example_19}{.9}}
\end{frame}
\fsi{simtrans/interpreter-screenshot}{Same human-computer cooperation}
\begin{frame}{Human-Computer Question Answering}
\begin{itemize}
\item Pit machine learning algorithms against humans
\item Fun task for human participants
\item Good system for discriminating depth and confidence of
algorithms
\item Knowledge and language comprehension needed to do well
\item Simple approaches work well if you have the whole question
\item Good competition for QA systems
\item Can get much harder
\begin{itemize}
\item Speech
\item Specialized knowledge
\item Adversarial authoring
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{Future \dots}
\begin{itemize}
\item Computers dominate on ``normal'' questions
\item Not so much on adversarial questions
\item Stronger QA systems (e.g., Google's!)
\begin{itemize}
\item Need more thought on how to explain
\item Expose QA system to more users
\item Increase the breadth of resources
\item Multilingual
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{Find out More!}
\begin{itemize}
\item Code: \url{http://github.com/Pinafore/qb}
\item Shared Task \url{http://qanta.org}
\end{itemize}
\end{frame}
% \begin{frame}{Come to UMD}
% \begin{columns}
% \column{.5\linewidth}
% \only<1>{
% \begin{center}
% \includegraphics[width=.9\linewidth]{umd/umd} \\
% \includegraphics[width=.9\linewidth]{umd/qb_team}
% \end{center}
% }
% \column{.5\linewidth}
% \begin{itemize}
% \item Looking for undergrads/grads/interns
% \item A great place for natural language
% processing and machine learning
% \item Not too shabby at quiz bowl either
% \end{itemize}
% \end{columns}
% \end{frame}
\frame{
\frametitle{But wait, there's more!}
\vspace{-.5cm}
\begin{columns}
\column{.5\linewidth}
\begin{block}{Computational Social Science}
\centering
\includegraphics[width=0.9\linewidth]{teaparty/figures/framing} \\
\cite{nguyen-13b,nguyen-15}
\end{block}
\begin{block}{Interactive Machine Learning}
\centering
\includegraphics[width=0.4\linewidth]{interactive_topic_models/new_interface} \\
\cite{Smith-17,Poursabzi-16}
\end{block}
\column{.5\linewidth}
\begin{block}{Multilingual Topic Models}
\begin{center}
\begin{large}
$p_{\mbox{topic}}(e | f)$ \\
\end{large}
\cite{eidelman-12,hu-14}
\end{center}
\vspace{-.3cm}
\end{block}
\begin{block}{Sentiment / Internal State}
\centering
\includegraphics[width=0.4\linewidth]{general_figures/diplomacy} \\
\cite{niculae-15,sayeed-12,boyd-graber-10}
\end{block}
\end{columns}
}
\frame{
\frametitle{Thanks}
\begin{block}{Collaborators (not listed on previous slides)}
Anupam Guha (Maryland), Manjhunath Ravi (Colorado), Danny Bouman (UMD UG),
Stephanie Hwa (UMD UG), Yogarshi Vyas (UMD), Larry Davis
(UMD), Naho Orita (Tohoku), Snigdha Chaturvedi (UMD), Varun
Manjunatha (UMD), Srijan Kumar (UMD), Vlad Niculae
(Cornell), Cristian Danescu-Niculescu-Mizil (Cornell),
Richard Socher (Salesforce), Leonardo Claudino (UMD)
\end{block}
\begin{columns}