Understanding Attribution Explanations

Attribution explanations highlight specific parts of a table—such as rows, columns, or cells—that are most relevant to the answer predicted by a TableQA model. These explanations help you understand which information of the input the system considered important when predicting the answer.

Table caption: 1947 Kentucky Wildcats Football Team

Statement to verify: "The Wildcats kept the opposing team scoreless in 4 games."

Game	Date	Opponent	Result	Wildcats Points	Opponents	Record
1	9999-09-20	Ole Miss	Loss	7	14	0 - 1
2	9999-09-27	Cincinnati	Win	20	0	1 - 1
4	9999-10-11	9 Georgia	Win	26	0	3 - 1 , 20
5	9999-10-18	10 Vanderbilt	Win	14	0	4 - 1 , 14
9	9999-11-15	Evansville	Win	36	0	7 - 2

In this example, the TableQA model has highlighted specific rows and cells to explain its prediction:

The entire rows for games 2, 4, 5, and 9 are highlighted in yellow.
Within these rows, the "Opponents" column cells containing "0" or "scoreless" are highlighted in green.

These highlights indicate that the system identified four games where the opposing team did not score, verifying the statement as CORRECT. The yellow highlighting shows the relevant rows, while the green highlighting represents the cells containing fine-grained information needed to verify the statement.

By using different colors for highlighting, the system provides a more nuanced explanation:

Yellow highlights (rows): Show the overall context of the relevant games.
Green highlights (cells): Pinpoint the exact information (opposing team's score of 0) that directly answer the question.

During the experiment, you will use explanations to debug model prediction. Your task will be to look at the provided explanations and determine if model will output the wrong prediction.