Table Retrieval

Task Overview

The goal of the Table Retrieval (TR) task is to search a table that contains the data for an answer, given a single document (annual securities report) and the question.

Task Definition

• Input: Question (question), annual security report HTML (doc_id)

• Output: Table (table_id)

• Evaluation: Accuracy

Dataset

An example from questions_tr.json:

{

"question_train1": {

  "question": "大和ハウス工業の2019年の個別のShareholdersEquityにおける「自己株式の処分」を含む表は？",

  "doc_id": "S100ITAZ",

  "table_id": "S100ITAZ-0000000-tab1"

  },

}

The template is structured as follows.

• question_id (object)

  • question_{data type}{sequence number}

  • The data type is either train, valid or test.

  • The sequence number is not continuous for each task to prevent duplication of question sentences between tasks.

• question: Question sentence (str)

  • This is the question sentence used as the query.

  • All questions include the company name, period, and item name (search target).

  • Some questions include the individual/consolidated element and member element when available in the XBRL.

• doc_id (str)

  • S100{4 capital letters}

  • This is the document number assigned by EDINET.

• table_id (str)

  • {doc_id}-{HTML file number (7 digits)}-tab{consecutive number}

  • This is an ID assigned to the HTML annual securities report by the Task Organizer.

  • It is provided as the "table-id" attribute of the table tag in the HTML.