sandy4321/bidding-dataset

Bidding documents for paper "CinBidding: A Dataset for Domain-specific Information Extraction with Limited Data"

Bidding Documents for Information Extraction

Problem

Given a document and a list of pre-defined tags, the system extracts corresponding information from the document to each tag.

For this dataset, the list of pre-defined tags are:

No.	Tag Name (JP)	Tag Name (EN)	Value Type	Description
1	調達年度	Year of Procurement	datetime (year only)
2	都道府県	Prefecture	text
3	入札件名	Bid Subject	text
4	施設名	Facility Name	text
5	需要場所(住所)	Address for Demand	text
6	調達開始日	Start Date of Procurement	datetime
7	調達終了日	End Date of Procurement	datetime
8	公告日	Public Announcement Date	datetime
9	質問票締切日時	Deadline for Questionnaire	datetime
10	資格申請締切日時	Deadline for Applying Qualification	datetime
11	入札書締切日時	Deadline for Bidding	datetime
12	開札日時	Opening Application Date	datetime
13	質問箇所 所属/担当者	PIC for Inquiry of Questions	text
14	質問箇所 TEL/FAX	TEL/FAX for Inquiry of Questions	text
15	資格申請送付先	Address for Submitting Application	text
16	資格申請送付先 部署/担当者名	Department/PIC for Submitting Application	text
17	入札書送付先	Address for Submitting Bid	text
18	入札書送付先 部署/担当者名	Department/PIC for Submitting Bid	text
19	開札場所	Place of Opening Bid	text

Note:
The tag names in correct answer (excel files) may be different from the list above. It is suggested that you use the below line to normalize the tags read from those excel files.

import unicodedata
import re

tag = unicodedata.normalize("NFKC", re.sub('＊|\*|\s+', '', tag))

Dataset

Overview

• Total number of documents: 124
• Train: 82
• Dev (Public Test): 22
• Test (Private Test): 20

Data Collection and Processing

We collected 124 public Japanese bidding documents from the Japan Oil, Gas and Metals National Corporation JOGMEC. Each document is a public notice, which outlines the information about the bidding process, including the dates of the contract, the deadlines for submission, and the contacts of the department or person in charge. These documents are raw texts, so we need to define a set of tags for the annotation process. To do that, we consulted our legal team for the definition. The discussion and definition were internally conducted. Finally, we defined 19 names that represent the categories of extracted information, which we formally refer to as tags. The list of tags covers common important information of a bidding document. The list is unique and remains unchanged in all three train/dev/test sets.

The next step is to parse the collected documents. This is because the collected documents are in the PDF format, so it needs to be converted to the text format for facilitating the annotation process. To do that, we build a document reader library to parse the original documents in the PDF format into structured texts. The document reader bases on pdfplumber, which is a python-based library for extracting text characters, rectangles, lines, and tables from a PDF file. The document reader utilizes the position of characters in the original document to get the exact text of a line or paragraph. However, the original pdfplumber tool can not output texts in a structure, e.g. lines, paragraphs. To address this issue, we defined a handful of rules based on the bullets, numberings, indentation of a text line to construct the hierarchical structure of texts as appeared in the original text. The reason for keeping the hierarchical structure of texts is to facilitate IE models by taking advantage of context information.

Distribution of Tags

Note: the number on x-axis corresponds to tag number in the table above.

Interpretation of Correct Answer files

Column Name	Description
Page No	page number in which the text line appears
Text	text line as appeared in the original document
Index	index of each text line, starting from 1
Parent Index	refer to the index of its immediate title as appeared in the original document
Is Title	whether the current text line is a title or not
Is Table	whether the current text line is in a table or not
Tag	name of tag(s) in the current text line
Value	extracted value from the current text line

Note:
If there are multiple tags and CAs in the same line:

• semicolons ; are used to separate different tags and different CAs
• the number of tags = the number of CAs
  (for cases where the number of tags != the number of CAs, it is certain that the number of tags in that line = 1 or the number of CAs in that line = 1)