data_index_by_user
int32 0
87.9k
| question
stringlengths 5
77
| answer
sequence |
---|---|---|
0 | κ·Έλ€μ μ¨μ μμ‘° μκ° κΈ°κ³λ₯Ό μ΄λμμ 촬μνμ΅λκΉ? | [
"νΌλ μνμΈ λ¦¬μ‘°νΈ"
] |
1 | λκ° κ΅μ ν΄μμμ ν΅νκΆμ κ°μ§κ³ μμ΅λκΉ? | [
"μ΄λ μ λ°λ"
] |
2 | λκ° νμ΄ν 곡격μ μν΄ μΌν©λκΉ? | [
"λ§λ¦¬."
] |
3 | μΈμ μΆμ
κ΅κ΄λ¦¬ κ°ν λ° ν΅μ μ κ΄ν λ²λ₯ μ΄ ν΅κ³Όλμλκ°? | [
"1986λ
11μ 6μΌ"
] |
4 | μΈμ νΈμλ₯΄ν 리μ½κ° λ―Έκ΅μ 첨κ°λμμ΅λκΉ? | [
"1950"
] |
5 | 64κ°μ κ΅κ° μνμμμ μ΅κ³ μ μ§μ μ¬λ°°μ°μ μ μ λ μ¬λ | [
"μμ΄λΌ μμ¬."
] |
6 | ν° μ§μ μ΄λ μͺ½μ΄ μ λ©΄μ
λκΉ? | [
"λΆ."
] |
7 | λ¨μν리카 곡νκ΅μ κ΄μ μμΉ λ¨μ²΄μ μ΄λ¦ | [
"κ°λ¦μ§ν μ",
"λ¬μ¨ λ§λΈλΌ λ² μ΄ λ©νΈλ‘ν΄λ¦¬ν μμ μ΄",
"μν
ν€λ κ΄μμμΉλ¨μ²΄",
"μΈ μλ€ μ(μ)",
"μνλ€μ€λ²κ·Έ μ",
"λ²νλ‘ μ κ΄μμ",
"μμΏ λ₯΄νλ λ μ(City ofEkurhuleni Metropolitan City)"
] |
8 | λκ° 2019λ
μνΌλ³Όμ μ£Όμ΅ν©λκΉ? | [
"μ νλν, μ‘°μ§μ"
] |
9 | μ체 μΈμμ μΈλμμ 첫 λ²μ§Έ μ νλ₯Ό μΆμν ν΄μ
λλ€. | [
"2014λ
"
] |
10 | μ±κ²½μ λμ€λ λ©λ¦¬ λ§κ·Έλ¬λ μ λν΄μλ μ΄λμ μ΄μΌκΈ°νλκ°? | [
"μ μ½"
] |
11 | λκ° λν΅λ Ήμ μν΄ ν΅ μΆκ΅¬λ₯Ό μ΄λ°ν©λκΉ? | [
"aide-de-camp"
] |
12 | μλμλΌλ μ΄λ¦μ μΆμ²λ 무μμ
λκΉ? | [
"그리μ€"
] |
13 | λμ€λ μ±λμ λͺ©μ리λ₯Ό λΈ μ¬λμ λꡬμ
λκΉ? | [
"\"λ²μ¦\" λΈλ μ΄λλ",
"μΊλ¨Έλ°."
] |
14 | λ΄μμ λ²μ κ²°νΌ μ°λ Ήμ μ΄λ»κ² λλμ? | [
"18"
] |
15 | λ λ ν« μΉ λ¦¬ ννΌ ν¬μ΄λ μΈμ νμλμ? | [
"2018λ
3μ"
] |
16 | λκ° μν νΈν
νΈλμ€λ°λμμμ λ§λΉμ€λ₯Ό μ°κΈ°νλ€. | [
"μ¬λ μλ€λ¬",
"μ
λ λ κ³ λ©μ¦."
] |
17 | μ€ννΈλΌμμ λ§ν λ€νΈμν¬μ μ±λ λ²νΈλ 무μμ
λκΉ? | [
"768(HD)",
"68(SD)"
] |
18 | FA μ»΅ μΈλ―Έ κ²°μΉμ μ μΈμ μΉλ¬μ§λμ? | [
"μλ‘μ΄ λΈλ¦¬ μ€νλμ"
] |
19 | μΈμ μμ΄ν ν°μΉ 6μ μ΄ λμλμ? | [
"2015λ
7μ 15μΌ"
] |
20 | 1966λ
μλμ»΅μμ μ¨μ€νΈ ν μ μλ€ | [
"λ§ν΄ νΌν°μ€",
"μ§μ€ν νμ€νΈ.",
"λ°λΉ 무μ΄"
] |
21 | λκ° λ―Έλ―Έ λ°μ΄μ€μ μ£Όμ κ°λ₯Ό λΆλ₯Έλ€. | [
"μ ν΄λ¨Έ."
] |
22 | λκ° μκΉλ‘ λ΄€μ΄μΌ λΌ. | [
"μ¬ν€ μ‘΄μ¨."
] |
23 | μν λ₯ μν° νΈλΌμ΄μ¦μ μΆμ°ν μ¬λ | [
"μ»€νΈ λ¬μ
",
"λλ° μ€λΈλΌμ΄μΈ.",
"μ‘΄ λ§μ½λΉμΉ.",
"μΌμ΄νΈ νλμ¨",
"μ§λ λ‘λ리κ²μ€",
"λ§ν¬ μνλ²κ·Έ"
] |
24 | 맀λμ¨ μ€νμ΄ κ°λ μμ λΉν΄ λΉλ²κ° 곡μ°νμ λ | [
"2010λ
8μ 31μΌ"
] |
25 | μμλλ¬Ό λΆλ₯λ κ·Έ μλ¨μ κΈ°λ°μΌλ‘ νμλ€. | [
"locomotion"
] |
26 | λ§λλλ‘ μ΄μ£Όν μ΅μ΄μ μ¬λμ λꡬμμ΅λκΉ? | [
"μ΄μ¬λ μμΈμ 무ν¨λ§λ"
] |
27 | λΉνκ³Ό ν°λλΌλ μ νμ μΈ μ½ν κΈ°ν λ²μ£Όμ΄λ€. | [
"κ·Ήμ±"
] |
28 | μ μΌκ°νμ κ°μ ν¬κΈ°λ μΌλ§μΈκ°? | [
"60Β°."
] |
29 | λΉν°μ μΌμλ μΈμ λ°μν©λκΉ? | [
"λ‘μ½μ½ μλ"
] |
30 | λκ° λλ₯Ό μν΄ μ°°λ¦¬ νΈνΈμ ν¨κ» λ
Έλλ₯Ό λΆλ₯Έλ€. | [
"λ―Έκ΅ κ°μ μΌμΌλΌλ."
] |
31 | κ°μ μλ μ 6 μν ꡬμμλ μ΄λ€ μΉ΄μ΄ν°κ° μλ€. | [
"νν΄.",
"λμΉΌλΈ.",
"μ½₯."
] |
32 | 1971λ
λ°©κΈλΌλ°μ νμ μ μ λΉμ μΈλκ΅°μΌλ‘ ν₯νλ μ¬λ | [
"μΌ λ§λ₯μΌ."
] |
33 | λκ° μ΄λκ°μ μμ
μ μΌλμ? | [
"쑴 배리."
] |
34 | λκ° μ€ ν볡μμ νΌκ°λ‘μ κ²°νΌμ λ
Έλνλ€. | [
"μλ λ§ν°μ€",
"Gundula Janowitz"
] |
35 | λκ° μλ
μ μ€λ«λμ μΆκ΅¬λ₯Ό νμ΅λκΉ? | [
"μ΄κΈμ€."
] |
36 | ν리 λμκ°μ μ£½μμ ννΈκ° μΈμ λμμκΉ? | [
"2007λ
7μ 21μΌ"
] |
37 | λ΄ μμ
μ λͺ ν μ°κ·Ή μμ¦ 2 | [
"25."
] |
38 | 곡μ ν λ‘ μ λ°λ° μΉμ
μμ μ΄λ€ νμ΄ λ¨Όμ λ§νλκ°. | [
"κΈμ ν."
] |
39 | λ―Έκ΅ νλ²μ λ§μ§λ§ μμ μ μΈμ μμ΅λκΉ? | [
"1992λ
5μ 5μΌ"
] |
40 | νλΌμ€ν±μΌλ‘ λ§λ€ μ λΆ κΈ°λ° μν | [
"μ₯μμ λ
Ήλ§"
] |
41 | λΆνμ΄ μΈμ 첫 λ―Έμ¬μΌμ λ°μ¬νμ΅λκΉ? | [
"1984λ
"
] |
42 | 4κ°μ κ° μ¬μμ μλ κ΄λ¬Όμ μ§κ΅¬μ λμ΄λ₯Ό κ²°μ νκΈ° μν΄ μ°λλ₯Ό μΈ‘μ νλ€. | [
"μ§λ₯΄μ½."
] |
43 | μ°λ¦¬μκ² λ§μ§λ§μΌλ‘ κ°κΈ°μΌμμ΄ μΌμ΄λ κ²μ μΈμ μ
λκΉ? | [
"2017λ
8μ 21μΌ"
] |
44 | μΈλκ° μ²μμΌλ‘ μ¬λ¦Όν½μ μ°Έμ¬νμ λ | [
"1900λ
"
] |
45 | 루ν¬μ 리μλ νμ λΌλ μ¬μ€μ΄ μΈμ λ°νμ§λκ°? | [
"μμ§μ κ·νμμ"
] |
46 | λμ΄μ΄ λ΄λ©μμ½μ μΈκ΅¬λ μΌλ§μ
λκΉ? | [
"41,629"
] |
47 | κ·Έλ€μ΄ gmc μ¬μ μ λ§λ λ§μ§λ§ ν΄λ μΈμ μμ΅λκΉ? | [
"2008λ
"
] |
48 | μλ‘μ΄ μμ€ν
λ₯΄λ΄μ λ€λλλ μλ―Όμ§μ μ£Όμ μ μ°©μ§μλ€. | [
"New Netherland"
] |
49 | λΆμ ν μ©κ΄λ‘μ λμ Έμ§ μ±κ²½μμ | [
"λ©μ€ν",
"μλ² λλ€κ³ ",
"μ€λλΌν¬."
] |
50 | ꡬκ°μ μν μμ€ν
μ μ΄λ»κ² λμμ΄ λ©λκΉ? | [
"μ€ν μ μ‘μ μμ±ν©λλ€."
] |
51 | μ§μμ κ²°νΌνκ³ λ λ μ¬λ | [
"μμΌμ¬ μ€νμ΄νΈ"
] |
52 | λ―Έκ΅μ΄ μΈμ λ§μ§λ§μΌλ‘ λ―Έμ€ μ°μ£Όμ μΉλ¦¬λ₯Ό λ΄€λμ? | [
"2012λ
"
] |
53 | νμ ν΄λΆνμμ 촬μμ΄ μΈμ λ°μνμ΅λκΉ? | [
"2010λ
5μ 20μΌ"
] |
54 | λ©λ¦¬μ λν΄ λκ° λ
Έλλ₯Ό λΆλ₯΄λμ§. | [
"μ‘°λμ 리μΉλ¨Ό."
] |
55 | μλ‘μ΄ ννzee λΈλ¦¬μ§λ μΈμ μμ±λ κΉμ? | [
"2018λ
λ§"
] |
56 | μλ©μ΄μΉ΄μ μμΉν λ¦¬μ€ κ·Έλλ° κ°μ μ΄λμ μμ΅λκΉ? | [
"ν¬νλλ κ΅κ΅¬"
] |
57 | μ°λ¦¬λ μ΄λμ μμ μ μ¬μ μμ μ»μλκ°? | [
"νλμ€ μ¬λλ€"
] |
58 | μλ‘ κ±΄μ€ μ€μΈ ν νΌ κ²½κΈ°μ₯μ μ΄λμ μμ΅λκΉ? | [
"λ°λ νλ§κΈ° μμΉκ΅¬"
] |
59 | λκ° icc t 20 μλμ»΅ 2016μμ μ°μΉνμ΅λκΉ? | [
"μμΈλ μ λ"
] |
60 | λκ° μ μμΉ΄λ₯Ό λ² λμ€ μμΈμκ² κ²°νΌμν€λμ? | [
"λ‘λ μ΄."
] |
61 | μ¬κΈ°μ νμ΄λ μΈμλ₯μ΄λ μΈμ μ°λ¦¬μ
λκΉ? | [
"1980"
] |
62 | λΉμ μ μκ³Όλνμμ μ΄λ μ λ μ‘Έμ
ν©λκΉ? | [
"μλ°μ¬(MD)"
] |
63 | 루ν΄νΌλμ μ€νμμ μνμ
μμ μμ²μ 무μμ΄μλκ°? | [
"ν¬λ₯¨ ν΅"
] |
64 | λκ° μ’μ μμ¬ μνΌμλ 7μμ κΈλ μμ νλ€. | [
"ν°ν¬μ€ μ°λ¦¬λ²."
] |
65 | λκ° μ€ν νΈλ νΉμμ ν¨νΈλ¦ μ€νμ΄νΈλ₯Ό μ°μ£Όν©λκΉ? | [
"μ₯- νΌμΉ΄λ₯΄ μ μ₯."
] |
66 | ν΄λ¦¬ λμκ°μ μ§ λ μ¨μ¬λ¦¬μ μ€λͺ
| [
"보λ νλ체μ€μΉ΄ λΌμ΄νΈ"
] |
67 | μν λλ¬Όμμμ M. ν° λλ¬Όμ 무μμ
λκΉ? | [
"λΆκ·Ήλ§κ΄λμ΄."
] |
68 | λκ° λ² κ°μ€μμ μΌμ΄λλ μΌμ΄λΌλ 문ꡬλ₯Ό λ§λ€μλκ°? | [
"R&R ννΈλ",
"κ΄κ³ λνμ¬"
] |
69 | μν¬μ μ½λ νΌμ€νΈ μμ΄μ μ€μνΈ λΌμ΄νλ μΈμ μμ΅λκΉ? | [
"2005λ
3μ 18μΌ"
] |
70 | λκ° 5μΈκΈ° μ μ μ±κ²½μ λλ μΈμλ₯Ό μμνμ΅λκΉ? | [
"μνλ€μ€ ꡬν
λ² λ₯΄ν¬"
] |
71 | λν΅λ Ήμ΄ μ£½μΌλ©΄ λκ° λ€μ μ€μ μ λ€. | [
"λΆμ¬μ₯λ."
] |
72 | νλμ€ ν리μ λ
ΈνΈλ₯΄λ΄ λμ±λΉμ μλ μ λͺ
ν μ°½λ¬Έμ 무μμ΄μλμ? | [
"μ₯λ―Έμ°½."
] |
73 | μλ‘μ΄ μλ
μμ λ μ΄κ±΄μ μ§μ
μ 무μμ
λκΉ? | [
"μμ½ν ν맀 rep"
] |
74 | λκ° μ€νλΌμ μ λ Ήμμ λΌμΈμ μ°κΈ°νλκ°. | [
"ν¨νΈλ¦ μμ¨."
] |
75 | νλμ€μ μλ μ λΉμ 무μμΈκ°? | [
"곡νλΉμ",
"μ¬νλΉ"
] |
76 | ν©κΈκΈ°μ¬λ€μ μΈμ nhlμ ν©λ₯νμ΅λκΉ? | [
"2017"
] |
77 | ocμ μμ¦ 1μ λͺ νΈμ΄ μμ΅λκΉ? | [
"27"
] |
78 | 골격근μ μ‘°μ νλ μλΆ μ΄λ λ΄λ°μ μ΄λμμ λ°κ²¬λ©λκΉ? | [
"μ μ€μν."
] |
79 | μ°λ¦¬λ μ΄λ€ μ νμ μ λΆ κ΅¬μ‘°λ₯Ό κ°μ§κ³ μμ΅λκΉ? | [
"νλ² κ³΅νκ΅"
] |
80 | μΈμ μ¬κ³Ό μ»΄ν¨ν°κ° μ¬κ³Ό μΈν¬λ‘ λ³κ²½λμμ΅λκΉ? | [
"2007λ
1μ 9μΌ"
] |
81 | μν κ²½κΈ°μμ λ¨μ λμ λκ° μ΄κ²Όλμ? | [
"λΈλΌμ΄ μ€ν μ΄λ§¨."
] |
82 | νλμ κΉμ΄λ λλ΅μ μΌλ‘ λ€μκ³Ό κ°λ€. | [
"νμ₯μ μ λ°"
] |
83 | μμ±ν΄ μ£Όμ λμλ€μ oλ‘ μμνλ€. | [
"μ€μ
νλ¦°μΈμ€.",
"μ€ν¬ νλ²",
"μ€λ₯΄ν
.",
"μ€μ
λ‘.",
"μ¬λ¦ΌνΌμ.",
"μ€λ‘λΉ",
"μ€μΉ΄λ
Έκ°.",
"μ€λ§.",
"μ€ν¬λΉ"
] |
84 | λ―Έκ΅μ ν λ€μ΄λ μμ¦μ μΈμ μ
λκΉ? | [
"3μλΆν° 6μκΉμ§"
] |
85 | μ¬λ¦Όν½μ 2λ
λ§λ€ μΈμ μμλμλμ? | [
"1994λ
"
] |
86 | λκ° λ²κ³Ό μ§μμ κ°λ°©μ λ§νλκ°? | [
"μ€ν°λΈ M. μ§ν¨ν€"
] |
87 | μ νλΈ κ΅¬λ
μ μκ° κ°μ₯ λ§μ μ¬λ | [
"ν¨λνμ΄."
] |
88 | κ·Έλ€μ μΈμ κ΅μ μ°μ£Ό μ κ±°μ₯μ λ°μ¬νμ΅λκΉ? | [
"1998λ
"
] |
89 | μμ μνμμ μ리 μμΉ΄λ₯Ό μ°κΈ°ν μ¬λ | [
"μ§ μμΌλ."
] |
90 | λκ° λ΄κ° λΉμ μ νΈλμ μ΄μ νλ λ
Έλλ₯Ό μΌλμ? | [
"μ§λ―Έ μ°λ¦¬.",
"μ μ μλ μ°λ.",
"μ½λ ν΄λ§ν΄."
] |
91 | λ©μμλΌλ μκ°μ μ΄λμμ μλμ? | [
"μ λκ΅"
] |
92 | λꡬμ λν μν λΆνλ μ μ°¨μμ΅λκΉ? | [
"μλ¦ λ¦¬λλΈ.",
"ν΄λ‘€λ μλΈλΌν¨μ€"
] |
93 | μ λ½ μ°ν© κ°μ
μ μν΄μλ κ΅κ°λ€μ΄ μ΄λ ν μ νμ μ λΆλ₯Ό κ°μ§κ³ μμ΄μΌ νλ€. | [
"λ―Όμ£Όμ£Όμ"
] |
94 | νλΈλ‘μ μΈμμ μΈμ λμμκΉ? | [
"2016λ
2μ 14μΌ"
] |
95 | 2010λ
elm streetμ μ
λͺ½μμ νλ λ ν¬λ£¨κ±°λ₯Ό μ°κΈ°ν μ¬λμ΄λ€. | [
"μ¬ν€ μΌ ν€μΌλ¦¬."
] |
96 | μ¬μμΉκ΅¬κ° λμ²λΌ λ¨κ±°μ μΌλ©΄ λκ° λ
Έλλ₯Ό ν΄. | [
"νΈμ€ μΈν."
] |
97 | λκ° ν΄λ¦¬λΈλλ κΈ°λ³λ μ μ λ λΈλ‘ μ μμ€λ₯Ό νλ μ΄νμ΅λκΉ? | [
"λ§μ΄μ λ―Έ ννΈ."
] |
98 | 리λλ μμν° μ΄λμ μλμ? | [
"λ§μ°μ€νΌμ€"
] |
99 | μνμμ λκ° λμ μ¬μ΄ λΌμ²Όμ μ£Όμ°μ 맑μλκ°? | [
"λ μ΄μ²Ό λ°μ΄μΈ ",
"ν리λ°μ΄ κ·Έλ μΈμ .",
"νΌμνλμΈμ½ νλΉλ
Έ.",
"μμ΄μΈ κΈλ ",
"μ ν΄λΌνλ¦°"
] |
End of preview. Expand
in Dataset Viewer.
Dataset Card for "kor_nq_open"
Source Data Citation Information
@article{doi:10.1162/tacl\_a\_00276,
author = {Kwiatkowski, Tom and Palomaki, Jennimaria and Redfield, Olivia and Collins, Michael and Parikh, Ankur and Alberti, Chris and Epstein, Danielle and Polosukhin, Illia and Devlin, Jacob and Lee, Kenton and Toutanova, Kristina and Jones, Llion and Kelcey, Matthew and Chang, Ming-Wei and Dai, Andrew M. and Uszkoreit, Jakob and Le, Quoc and Petrov, Slav},
title = {Natural Questions: A Benchmark for Question Answering Research},
journal = {Transactions of the Association for Computational Linguistics},
volume = {7},
number = {},
pages = {453-466},
year = {2019},
doi = {10.1162/tacl\_a\_00276},
URL = {
https://doi.org/10.1162/tacl_a_00276
},
eprint = {
https://doi.org/10.1162/tacl_a_00276
},
abstract = { We present the Natural Questions corpus, a question answering data set. Questions consist of real anonymized, aggregated queries issued to the Google search engine. An annotator is presented with a question along with a Wikipedia page from the top 5 search results, and annotates a long answer (typically a paragraph) and a short answer (one or more entities) if present on the page, or marks null if no long/short answer is present. The public release consists of 307,373 training examples with single annotations; 7,830 examples with 5-way annotations for development data; and a further 7,842 examples with 5-way annotated sequestered as test data. We present experiments validating quality of the data. We also describe analysis of 25-way annotations on 302 examples, giving insights into human variability on the annotation task. We introduce robust metrics for the purposes of evaluating question answering systems; demonstrate high human upper bounds on these metrics; and establish baseline results using competitive methods drawn from related literature. }
}
@inproceedings{lee-etal-2019-latent,
title = "Latent Retrieval for Weakly Supervised Open Domain Question Answering",
author = "Lee, Kenton and
Chang, Ming-Wei and
Toutanova, Kristina",
booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
month = jul,
year = "2019",
address = "Florence, Italy",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/P19-1612",
doi = "10.18653/v1/P19-1612",
pages = "6086--6096",
abstract = "Recent work on open domain question answering (QA) assumes strong supervision of the supporting evidence and/or assumes a blackbox information retrieval (IR) system to retrieve evidence candidates. We argue that both are suboptimal, since gold evidence is not always available, and QA is fundamentally different from IR. We show for the first time that it is possible to jointly learn the retriever and reader from question-answer string pairs and without any IR system. In this setting, evidence retrieval from all of Wikipedia is treated as a latent variable. Since this is impractical to learn from scratch, we pre-train the retriever with an Inverse Cloze Task. We evaluate on open versions of five QA datasets. On datasets where the questioner already knows the answer, a traditional IR system such as BM25 is sufficient. On datasets where a user is genuinely seeking an answer, we show that learned retrieval is crucial, outperforming BM25 by up to 19 points in exact match.",
}
- Downloads last month
- 62