|
require 'rturk'
module Agents
class HumanTaskAgent < Agent
default_schedule "every_10m"
description <<-MD
You can use a HumanTaskAgent to create Human Intelligence Tasks (HITs) on Mechanical Turk.
HITs can be created in response to events, or on a schedule. Set `trigger_on` to either `schedule` or `event`.
The schedule of this Agent is how often it should check for completed HITs, __NOT__ how often to submit one. To configure how often a new HIT
should be submitted when in `schedule` mode, set `submission_period` to a number of hours.
If created with an event, all HIT fields can contain interpolated values via [JSONPaths](http://goessner.net/articles/JsonPath/) placed between < and > characters.
For example, if the incoming event was a Twitter event, you could make a HITT to rate its sentiment like this:
{
"expected_receive_period_in_days": 2,
"trigger_on": "event",
"hit": {
"max_assignments": 1,
"title": "Sentiment evaluation",
"description": "Please rate the sentiment of this message: '<$.message>'",
"reward": 0.05,
"questions": [
{
"type": "selection",
"key": "sentiment",
"name": "Sentiment",
"required": "true",
"question": "Please select the best sentiment value:",
"selections": [
{ "key": "happy", "text": "Happy" },
{ "key": "sad", "text": "Sad" },
{ "key": "neutral", "text": "Neutral" }
]
},
{
"type": "free_text",
"key": "feedback",
"name": "Have any feedback for us?",
"required": "false",
"question": "Feedback",
"default": "Type here...",
"min_length": "2",
"max_length": "2000"
}
]
}
}
As you can see, you configure the created HIT with the `hit` option. Required fields are `title`, which is the
title of the created HIT, `description`, which is the description of the HIT, and `questions` which is an array of
questions. Questions can be of `type` _selection_ or _free\\_text_. Both types require the `key`, `name`, `required`,
`type`, and `question` configuration options. Additionally, _selection_ requires a `selections` array of options, each of
which contain `key` and `text`. For _free\\_text_, the special configuration options are all optional, and are
`default`, `min_length`, and `max_length`.
If all of the `questions` are of `type` _selection_, you can set `take_majority` to _true_ at the top level to
automatically select the majority vote for each question across all `max_assignments`.
As with most Agents, `expected_receive_period_in_days` is required if `trigger_on` is set to `event`.
MD
event_description <<-MD
Events look like:
{
}
MD
def validate_options
errors.add(:base, "'trigger_on' must be one of 'schedule' or 'event'") unless %w[schedule event].include?(options[:trigger_on])
if options[:trigger_on] == "event"
errors.add(:base, "'expected_receive_period_in_days' is required when 'trigger_on' is set to 'event'") unless options[:expected_receive_period_in_days].present?
elsif options[:trigger_on] == "schedule"
errors.add(:base, "'submission_period' must be set to a positive number of hours when 'trigger_on' is set to 'schedule'") unless options[:submission_period].present? && options[:submission_period].to_i > 0
end
if options[:take_majority] == "true" && options[:hit][:questions].any? { |question| question[:type] != "selection" }
errors.add(:base, "all questions must be of type 'selection' to use the 'take_majority' option")
end
end
def default_options
{
:expected_receive_period_in_days => 2,
:trigger_on => "event",
:hit =>
{
:max_assignments => 1,
:title => "Sentiment evaluation",
:description => "Please rate the sentiment of this message: '<$.message>'",
:reward => 0.05,
:questions =>
[
{
:type => "selection",
:key => "sentiment",
:name => "Sentiment",
:required => "true",
:question => "Please select the best sentiment value:",
:selections =>
[
{ :key => "happy", :text => "Happy" },
{ :key => "sad", :text => "Sad" },
{ :key => "neutral", :text => "Neutral" }
]
},
{
:type => "free_text",
:key => "feedback",
:name => "Have any feedback for us?",
:required => "false",
:question => "Feedback",
:default => "Type here...",
:min_length => "2",
:max_length => "2000"
}
]
}
}
end
def working?
last_receive_at && last_receive_at > options[:expected_receive_period_in_days].to_i.days.ago && !recent_error_logs?
end
def check
setup!
review_hits
if options[:trigger_on] == "schedule" && (memory[:last_schedule] || 0) <= Time.now.to_i - options[:submission_period].to_i * 60 * 60
memory[:last_schedule] = Time.now.to_i
create_hit
end
end
def receive(incoming_events)
if options[:trigger_on] == "event"
setup!
incoming_events.each do |event|
create_hit event
end
end
end
# To be moved either into an initilizer or a per-agent setting.
def setup!
RTurk::logger.level = Logger::DEBUG
RTurk.setup(ENV['AWS_ACCESS_KEY_ID'], ENV['AWS_ACCESS_KEY'], :sandbox => ENV['AWS_SANDBOX'] == "true") unless Rails.env.test?
end
protected
def review_hits
reviewable_hit_ids = RTurk::GetReviewableHITs.create.hit_ids
my_reviewed_hit_ids = reviewable_hit_ids & (memory[:hits] || {}).keys.map(&:to_s)
log "MTurk reports the following HITs [#{reviewable_hit_ids.to_sentence}], of which I own [#{my_reviewed_hit_ids.to_sentence}]"
my_reviewed_hit_ids.each do |hit_id|
hit = RTurk::Hit.new(hit_id)
assignments = hit.assignments
log "Looking at HIT #{hit_id}. I found #{assignments.length} assignments#{" with the statuses: #{assignments.map(&:status).to_sentence}" if assignments.length > 0}"
if assignments.length == hit.max_assignments && assignments.all? { |assignment| assignment.status == "Submitted" }
if options[:take_majority] == "true"
options[:hit][:questions].each do |question|
counts = question[:selections].inject({}) { |memo, selection| memo[selection[:key]] = 0; memo }
assignments.each do |assignment|
answers = ActiveSupport::HashWithIndifferentAccess.new(assignment.answers)
answer = answers[question[:key]]
counts[answer] += 1
end
end
else
event = create_event :payload => { :answers => assignments.map(&:answers) }
log "Event emitted with answer(s)", :outbound_event => event, :inbound_event => Event.find_by_id(memory[:hits][hit_id.to_sym])
end
assignments.each(&:approve!)
memory[:hits].delete(hit_id.to_sym)
end
end
end
def create_hit(event = nil)
payload = event ? event.payload : {}
title = Utils.interpolate_jsonpaths(options[:hit][:title], payload).strip
description = Utils.interpolate_jsonpaths(options[:hit][:description], payload).strip
questions = Utils.recursively_interpolate_jsonpaths(options[:hit][:questions], payload)
hit = RTurk::Hit.create(:title => title) do |hit|
hit.max_assignments = (options[:hit][:max_assignments] || 1).to_i
hit.description = description
hit.question_form AgentQuestionForm.new(:title => title, :description => description, :questions => questions)
hit.reward = (options[:hit][:reward] || 0.05).to_f
#hit.qualifications.add :approval_rate, { :gt => 80 }
end
memory[:hits] ||= {}
memory[:hits][hit.id] = event && event.id
log "HIT created with ID #{hit.id} and URL #{hit.url}", :inbound_event => event
end
# RTurk Question Form
class AgentQuestionForm < RTurk::QuestionForm
needs :title, :description, :questions
def question_form_content
Overview do
Title do
text @title
end
Text do
text @description
end
end
@questions.each.with_index do |question, index|
Question do
QuestionIdentifier do
text question[:key] || "question_#{index}"
end
DisplayName do
text question[:name] || "Question ##{index}"
end
IsRequired do
text question[:required] || 'true'
end
QuestionContent do
Text do
text question[:question]
end
end
AnswerSpecification do
if question[:type] == "selection"
SelectionAnswer do
StyleSuggestion do
text 'radiobutton'
end
Selections do
question[:selections].each do |selection|
Selection do
SelectionIdentifier do
text selection[:key]
end
Text do
text selection[:text]
end
end
end
end
end
else
FreeTextAnswer do
if question[:min_length].present? || question[:max_length].present?
Constraints do
lengths = {}
lengths[:minLength] = question[:min_length].to_s if question[:min_length].present?
lengths[:maxLength] = question[:max_length].to_s if question[:max_length].present?
Length lengths
end
end
if question[:default].present?
DefaultText do
text question[:default]
end
end
end
end
end
end
end
end
end
end
end
|
