Skip to content

Latest commit

 

History

History
201 lines (100 loc) · 5.9 KB

dataset-creation-guidelines.md

File metadata and controls

201 lines (100 loc) · 5.9 KB
Raw

How to Create Datasets for Different NLP Problems

Table of contents

  1. Named Entity Recognition Dataset Guide
  2. Translation Dataset Guide

Named Entity Recognition

Named Entity Recognition (NER) is a classification task that identifies words in a text that refer to entities or predefined categories (such as dates, person, organization and location names).

Example of NER: John [PERSON] is going to Kenya [LOCATION] on Monday [DATE]

For annotation, we will be following the Message Understanding Conference (MUC-6) named entity definition guideline for identifying Persons, organizations, locations, dates and times, see https://cs.nyu.edu/faculty/grishman/NEtask20.book_1.html

Person Annotation [PER]

  1. Personal names (including firstnames, middle names and last names) should be tagged excluding titles e.g Mr, Mrs, President, Professor, e.t.c. Examples:

President Donald [PER] Trump [PER]

Donald [PER] Trump [PER] Jr. [PER]

  1. Family names should be tagged. Example:

The Kennedy [PER] family

Abiy [PER] Ahmed's [PER] wife.

Some tokenization may separate "Ahmed" and "'s", in that case, we have, For example in CoNLL format, where 'O' is no label.

Abiy, PER

Ahmed, PER

's, O

wife, O

  1. Personal names that refers to an organization, location, events, law, disease, prizes should not be tagged. Example,

St. Michael

Nobel Prize

George [PER] Washington [PER] was born in 1732.

George Washington University was established in 1821.

The White House is located in Washington.

Organization Annotation [ORG]

  1. Corporate designators such as "Co." are organizations. Example:

The Bridgestone [ORG] Sports [ORG] Co. [ORG]

  1. Proper names that includes sports teams, stock exchanges, multinational organizations, political parties, unions, government parastatal. Many of them comes in abbreviations e.g WHO, NCDC, NASDAQ, EU, AU

The World [ORG] Health [ORG] Organization [ORG]

AU [ORG]

  1. Proper names referring to facilities e.g factories, hotels, universities, airports, hospitals, churches except (a) the name is clearly the name of the structure/place and not of an organization, or (b) the name is known to be the name of an organization but is used only in reference to the facility as a structure/place. Example:

The Frankfurt [ORG] Airport [ORG] Management have new rules for Aircraft.

I am traveling from the Frankfurt Airport .....will not be tagged

I will stay at the IBIS [ORG] Hotel [ORG]

The Supreme [ORG] Court [ORG] of [ORG] Nigeria [ORG]

The convict will be taken to court.

The patient will be transferred to the General [ORG] Hospital [ORG] Ikoyi [ORG].

The patient is in the hospital.

Location Annotation [LOC]

  1. All country names,region names, state names and city names.

United [LOC] States [LOC] of [LOC] America [LOC]

USA [LOC]

Cape [LOC] Town [LOC]

Niger [LOC] Delta [LOC] Region [LOC]

  1. Multiword expressions in which place names are separated by a comma are to be tagged as separate instances of LOCATION. Example:

Cairo [LOC], Egypt [LOC]

Washington [LOC] , D.C. [LOC]

  1. Do not label Nationalities and street addresses as locations.

An American is traveling to Abuja [LOC]

A Nigerian hospital

53140 Gatchell Road

  1. Place names can sometimes be part of an organization. The annotation will follow these guidelines: (a) If there is a corporate designator, it marks the end of the organization name; (b) if there is no corporate designator, the "of " is part of the organization name. Example:

Hyundai [ORG] of [ORG] Korea [ORG] , [ORG] Inc. [ORG]

Hyundai [ORG], [ORG] Inc.[ORG] of Korea [LOC]

John [PER] is working in BOSCH [ORG], Germany [LOC].

University [ORG] of [ORG] California [ORG] in Los [LOC] Angeles [LOC]

DATE Annotation [DATE]

Tag all absolute and relative dates or periods, including days, months, years. We could combine DATE and TIME annotation as [DATETIME]

  1. Absolute date expressions are to be tagged. Absolute date expression must indicate a specific segment of date i.e the particular day, season, financial quarters, years, decade or a particular century. Example:

Monday [DATE]

10th [DATE] of [DATE] October [DATE]

April [DATE] 6[DATE],[DATE] 2020 [DATE]

Summer [DATE]

the Autumn [DATE] report

Winter [DATE] 2020 [DATE]

fourth [DATE] quarter [DATE]

third [DATE] quarter [DATE] of [DATE] 1991 [DATE]

first [DATE] half [DATE] of [DATE] the [DATE] year [DATE]

1995 [DATE]

1980s [DATE]

19th [DATE] century [DATE]

  1. Relative time expressions should also be tagged. Example:

July [DATE] last [DATE] year [DATE]

this [DATE] June [DATE]

next [DATE] summer [DATE]

thirty [DATE] days [DATE] before [DATE] the [DATE] end [DATE] of [DATE] the [DATE] year [DATE]

  1. Special days, such as holidays, that are referenced by name should be tagged. Example:

because of the observance of All [DATE] Saints' [DATE] Day [DATE]

The Christmas [DATE] day [DATE]

  1. Expression indicating periods between two dates should be tagged. Example:

We are on vacation between [DATE] July [DATE] 1 [DATE] and [DATE] July [DATE] 8 [DATE]

Her visit is from [DATE] July [DATE] 1 [DATE] and [DATE] July [DATE] 8 [DATE]

TIME Annotation [TIME]

This refers to times smaller than a day.

  1. Absolute time expressions are to be tagged. Absolute time expression must indicate a specific segment of time i.e a particular minute and hour. Example:

20 [TIME] minutes [TIME] after [TIME] 10 [TIME]

midnight [TIME]

twelve [TIME] o'clock [TIME] noon [TIME]

noon [TIME]

5 [TIME] p.m. [TIME] EST [TIME]

5:40 [TIME]

  1. Time expression including the city time zone should be tagged. Example:

1:30 [TIME] p.m. [TIME] Chicago [TIME] time [TIME]

  1. Expression indicating periods between two time should be tagged. Example:

The election is from 2 [TIME] p.m.[TIME] to 4 [TIME] p.m. [TIME]

Translation

TODO