notesum.ai

Published at December 5

PoTable: Programming Standardly on Table-based Reasoning Like a Human Analyst

cs.IR

cs.AI

Released Date: December 5, 2024

Authors: Qingyang Mao¹, Qi Liu¹, Zhi Li², Mingyue Cheng, Zheng Zhang, Rui Li

Aff.: ¹State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China; ²Shenzhen International Graduate School, Tsinghua University

Arxiv: http://arxiv.org/pdf/2412.04272v1

Approach

WikiTQ (D)

WikiTQ (T)

TabFact (S)

GPT

LLAMA

GPT

LLAMA

GPT

LLAMA

Binder (ICLR’23)

59.20

50.65

58.86

50.51

84.63

78.16

Dater (SIGIR’23)

56.76

42.78

58.33

43.53

80.98

81.57

Chain-of-Table (ICLR’24)

56.64

62.39

55.60

62.22

84.24

85.62

TabSQLify (NAACL’24)

56.87

55.51

57.02

55.78

78.75

70.70

PoTable (Ours)

63.58

(+4.38)

65.10

(+2.71)

64.73

(+5.87)

65.56

(+3.34)

88.93

(+4.30)

87.06

(+1.44)