标注方案
用 YAML 定义标注方案,涵盖单选、多选、片段、李克特、滑块、文本等 76 种类型。配置标签、问题和必填项。
标注方案定义标注者要完成的标注任务。Potato 支持 76 种标注类型,可以组合出复杂的标注任务。
标注方案剖析
基本结构
每个方案都定义在 annotation_schemes 数组里:
annotation_schemes:
- annotation_type: radio
name: sentiment
description: "What is the sentiment?"
labels:
- Positive
- Negative
- Neutral必填字段
每个方案都带这三个字段。各个类型还会在此之上追加自己的必填字段,下文介绍每种类型时会一并列出。
| 字段 | 说明 |
|---|---|
annotation_type | 标注类型(radio、multiselect、likert、span、text、number、slider、multirate、select、pairwise、bws、soft_label、confidence、constant_sum、range_slider、semantic_differential、hierarchical_multiselect、card_sort、rubric_eval、extractive_qa、error_span、triage、coreference、span_link 等) |
name | 内部标识符(不含空格,用于输出) |
description | 展示给标注者的说明 |
支持的标注类型
1. 单选(Radio)
从列表中恰好选一项:
- annotation_type: radio
name: sentiment
description: "What is the sentiment of this text?"
labels:
- Positive
- Negative
- Neutral
# Assigns keys 1, 2, 3... in the order the labels are listed
sequential_key_binding: true
# Horizontal layout instead of vertical
horizontal: true顺序绑定把数字 1-9 映射到前九个标签,所以重排标签列表会重新分配按键。若想要一种不随重排变动的绑定,请给每个标签各自指定按键(见键盘快捷键)。
2. 李克特量表(Likert)
带端点标签的评分量表:
- annotation_type: likert
name: agreement
description: "How much do you agree with this statement?"
size: 5 # Number of scale points
min_label: "Strongly Disagree"
max_label: "Strongly Agree"
# Show the numeric value of the selected point
displaying_score: truesize 和两端的端点标签都是必填。五个刻度给标注者一个中立的中点;偶数刻度则强迫他们选边。
3. 多选(Multiselect)
从列表中选多项:
- annotation_type: multiselect
name: topics
description: "Select all relevant topics"
labels:
- Politics
- Technology
- Sports
- Entertainment
- Science
# Require at least one box before the annotator can move on
label_requirement:
required: true
# Allow free text response
has_free_response: truehas_free_response 会加上一个"其他"文本框;在标签集还在修订期间值得打开,因为它收上来的条目正好点出你漏掉了哪些类别。
4. 片段标注(Span)
高亮并标记文本片段:
- annotation_type: span
name: entities
description: "Highlight named entities in the text"
labels:
- PERSON
- ORGANIZATION
- LOCATION
- DATE
# Show the label name on each highlighted span
show_span_labels: true
# Ctrl/Cmd+click to add non-contiguous ranges to one annotation
allow_discontinuous: false
# Keyboard shortcuts for labels
sequential_key_binding: trueallow_discontinuous 管的是由分散区间组成的片段,比如分裂先行词,而不是彼此重叠的片段。除非你要标的现象确实跨越间隔,否则别打开它,因为 Ctrl/Cmd 点击这个动作很容易误触发。
5. 滑块(Slider)
连续数值区间:
- annotation_type: slider
name: confidence
description: "How confident are you in your answer? (0 = not at all, 100 = completely)"
min_value: 0
max_value: 100
starting_value: 50
step: 1
# Shows min_value and max_value at the ends of the track
show_labels: truemin_value、max_value 和 starting_value 三个都是必填。起始值是每位标注者动手之前滑块所在的位置,取中点可以避免把他们锚定到任何一端。滑块不接受 labels 列表。一旦添加,题目会切换为单选(radio)布局,滑块也随之消失,因此量表两端的说明文字请写在 description 中。
6. 文本输入(Text)
自由文本作答:
- annotation_type: text
name: explanation
description: "Explain your reasoning"
# Multi-line input
rows: 4
# Require a minimum length and show a running count
min_chars: 20
show_char_count: true
# Placeholder text
placeholder: "Enter your explanation here..."
# Disable paste (for transcription tasks)
allow_paste: falseallow_paste: false 对转写和自由回忆类任务很要紧,因为在这类任务里粘贴来的答案是从原文抄的,不是标注者产出的。
7. 数字输入(Number)
带约束的数值输入:
- annotation_type: number
name: count
description: "How many entities are mentioned?"
min: 0
max: 100
step: 18. 矩阵评分(Multirate)
用同一套量表给多个条目评分:
- annotation_type: multirate
name: quality_aspects
description: "Rate each aspect of the response"
# Rows: the items being rated
options:
- Accuracy
- Fluency
- Relevance
# Columns: the shared rating scale
labels:
- Poor
- Fair
- Good
- Very Good
- Excellent
# Randomize row order per annotator
option_randomization: true
# Layout options
arrangement: "vertical"
display_config:
num_columns: 2options 和 labels 都是必填,因为矩阵需要行和列。把行的顺序随机化,可以把顺序效应摊到各个条目上,而不是全压在恰好排在最前面的那一项上。
通用选项
键盘快捷键
用键盘绑定加快标注:
# Automatic sequential binding: digits 1, 2, 3... in label order
sequential_key_binding: true若想要一种在标签列表变动时不会跟着移动的绑定,就把标签写成对象,各自给一个 key_value:
labels:
- name: Positive
key_value: "1"
- name: Negative
key_value: "2"
- name: Neutral
key_value: "3"工具提示
给单个标签提供悬停提示。工具提示附着在标签上,不在方案上,因此配置在 labels 里面:
labels:
- name: Positive
tooltip: "Expresses happiness, approval, or satisfaction"
- name: Negative
tooltip: "Expresses sadness, anger, or disappointment"
- name: Neutral
tooltip: "No clear emotional content"若说明需要排版,就把 tooltip_file 指向一个 HTML 文件:
labels:
- name: Economic
tooltip_file: "config/tooltips/economic.html"标签颜色
用于视觉区分的自定义颜色,在 ui 下为整个项目统一定义:
ui:
label_colors:
entities:
PERSON: "#3b82f6"
LOCATION: "#10b981"
ORGANIZATION: "#f59e0b"颜色先按方案名作键,再按标签名。颜色也可以连同工具提示一起写在单个标签上:
labels:
- name: positive
color: "#22C55E"
tooltip: "Positive sentiment"必填项
把某个方案设为提交前必填:
- annotation_type: radio
name: sentiment
description: "What is the sentiment?"
label_requirement:
required: true
labels:
- Positive
- Negative若要求的是某个特定选项而不是随便作答,就用 required_label 点名,它接受单个标签或一组标签。
多个方案
在一个条目上组合多种标注类型:
annotation_schemes:
# Primary classification
- annotation_type: radio
name: sentiment
description: "Overall sentiment"
labels:
- Positive
- Negative
- Neutral
label_requirement:
required: true
sequential_key_binding: true
# Confidence rating
- annotation_type: likert
name: confidence
description: "How confident are you?"
size: 5
min_label: "Guessing"
max_label: "Certain"
# Topic tags
- annotation_type: multiselect
name: topics
description: "Select all relevant topics"
labels:
- Politics
- Technology
- Sports
- Entertainment
has_free_response: true
# Notes
- annotation_type: text
name: notes
description: "Any additional observations"
rows: 4
label_requirement:
required: false这里只有情感标签是必填的。把置信度评分和备注设为可选,能把必填集合收缩到研究真正依赖的那一个字段上,标注者就不会被一个不承担分析重量的框卡住。
高级功能
成对比较(Pairwise)
比较两个条目。两个条目来自实例本身,由 items_key 指定:
- annotation_type: pairwise
name: preference
description: "Which response is better?"
mode: binary
items_key: "responses" # Field holding a list of 2+ items
# Labels for the two sides
show_labels: true
labels:
- "Response A"
- "Response B"
# Tie option (opt-in)
allow_tie: true
tie_label: "No preference"这一对由数据文件提供,所以每一行都带一个至少两项的列表:
{"id": "1", "responses": ["First response text", "Second response text"]}当确实等价的配对很常见时,加一个平局按钮是值得的,因为没有它标注者只能随意断定高下,而这些噪声会以分歧的形式冒出来。
最优最差量表(Best-Worst Scaling)
从一个元组中挑出最优和最差来给条目排序。元组由顶层的 bws_config 块构造,而不是由方案构造;没有它的 bws 方案会在一个空条目列表上渲染出一道问题:
bws_config:
tuple_size: 4 # Items per tuple
min_item_appearances: 8 # Minimum appearances per item across tuples
seed: 42
scoring:
method: counting # counting | bradley_terry | plackett_luce
annotation_schemes:
- annotation_type: bws
name: ranking
description: "Sentiment Intensity"
best_description: "Which sentence expresses the MOST positive sentiment?"
worst_description: "Which sentence expresses the LEAST positive sentiment?"
tuple_size: 4
sequential_key_binding: true数据文件的每一行就是一个条目,Potato 从这个池子里抽取元组。调高 min_item_appearances 会让每个条目的分数更稳,代价是比较次数更多。
下拉选择(Select)
省空间的单项选择:
- annotation_type: select
name: category
description: "Select a category"
labels:
- Category A
- Category B
- Category C
- Category D
- Category E
# Vary the option order between annotators
randomize_order: true下拉框在展开之前会藏起选项,因此适合标签集很长、用单选按钮会把正文挤出屏幕的情形。对于短标签集,单选按钮更快,因为它只要一次点击而不是两次。
数据格式参考
输入
标注方案配合你自己的数据格式工作:
{
"id": "doc_1",
"text": "This is the text to annotate."
}输出
标注以方案名作键保存:
{
"id": "doc_1",
"annotations": {
"sentiment": "Positive",
"confidence": 4,
"topics": ["Technology", "Science"],
"entities": [
{"start": 0, "end": 4, "label": "ORGANIZATION", "text": "This"}
],
"notes": "Clear positive sentiment about technology."
}
}最佳实践
1. 清晰的标签
使用无歧义、互相区分的标签:
下面这些标签在同一个维度上标出了不同的点:
labels:
- Strongly Positive
- Somewhat Positive
- Neutral
- Somewhat Negative
- Strongly Negative下面这些则不然。两个标注者不会在 "OK"、"Fine" 和 "Acceptable" 之间划出同一条线,随之而来的分歧说的是标签集本身,而不是数据:
labels:
- Good
- OK
- Fine
- Acceptable2. 有用的工具提示
为含义微妙的标签加上工具提示:
labels:
- name: Sarcasm
tooltip: "The text says the opposite of what it means"
- name: Irony
tooltip: "A mismatch between expectation and reality"3. 键盘快捷键
为高吞吐量的任务启用快捷键:
sequential_key_binding: true4. 合理的顺序
让标签顺序保持一致:
- 最常见的排在前面
- 按字母顺序
- 按强度(由低到高)
5. 限制选项数量
选项太多会拖慢标注:
- 单选:2-7 个选项
- 多选:5-15 个选项
- 李克特:5-7 个刻度
6. 先自己试标
部署给标注者之前先自己标几个例子,用来发现:
- 有歧义的标签
- 缺失的类别
- 说不清楚的说明