Skip to content

kachok/hitman

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Web - template and web app for creating and running Mechanical Turk HITs as External Questions 


Backend - command line tools for interacting with MTurk API and creating HITs in MTurk

RUN ONCE:
	load_data_to_db.py
		populates langugages table in database
		loads data from wikilanguages-pipeline into database to vocabulary and dictionary tables respectively

    prepare_images_files.py
		generates images descriptor files for all vocabulary and dictionary items
	
	java  -Djava.awt.headless=true generateimages sentences.txt segments.txt images
	#java  -Djava.awt.headless=true generateimages test_sentences.txt test_segments.txt ../test_images
	#java  -Djava.awt.headless=true generateimages pilot_sentences.txt pilot_segments.txt ../pilot_images
		renders PNG files
		10MB of images per 1000 words+sentences
		
	obama4australia.py
		generates images for Barack Obama and Australia in all languages

	#java  -Djava.awt.headless=true Str2ImgObama support_sentences.txt support_sequence.txt ../pilot_images/support
		

	generate_voc_hits.py
		creates hittypes in MTurk and populates hittypes table in database
		generates batches of vocabulary HITs in database based on vocabulary data
		populates hits table and voc_hits_data tables in database

	add_voc_hits_to_mturk.py
		generates HITs in MTurk and reference them to batches of words in database
		creates HITs in MTurk for every HIT in datbase with empty mturk_hit_id column
		updates hits table with non-empty mturk_hit_id value
		
	generate_synonyms_table.py
		generates list of synonyms to be used in Synonyms HIT as controls

	generate_non_synonyms_table.py
		generates list of non-synonyms to be used in Synonyms HIT as controls
		
RUN CONTINUOUSLY:
		
	
	get_assignments.py
		tasks retreives all completed assignments from all HITs and loads results into proper hits_results tables
		(works both for vocabulary and synonyms HITs)	

	generate_syn_hits.py
		takes processed results from Voc HIT (from vochitsresults table) and populates syn hits and syn hits data table 
		task can be run multiple times without creating duplicates
		
	add_syn_hits_to_mturk.py
		generates HITs in MTurk and reference them to batches in database
		creates HITs in MTurk for every HIT in datbase with empty mturk_hit_id column
		updates hits table with non-empty mturk_hit_id value	
	
	get_assignments.py
	
	-- run again to get results of new syn hits
		
	revew_step2.py
		reviews results of step2 (synonyms HITs)
		grade syn hits results (based on controls)
		select all assignments pending update in MTurk
		for each pending assignments
			grade assignments one-by-one
			update status in Mturk and create extra assignment for related hit if assignments data was bad (data_quality<=50%)
			close assignment in database
			
	
	review_step1.py
		select all syn hits that have 3 assignments completed with data_status>50% each

		CREATE OR REPLACE VIEW syn_hits_completed AS 
 		-- select all syn HITs that have 3 assignments with good quality of data. Those HITs considered completed
		 -- and data can be pushed to Step 1 to Vocabulary HITs/Assignments		
		
		
		for each hit
			get 3 assignment above (syn_hits_completed x assignments (by hit_id))
			weight results for each word translation and multiply it by performance of worker
			
			-- weight+result per hit per assignment
			SYN_HITS_RESULTS_WIGHTED_PAIRS
			
select voc_assignment_id, pair_id, are_synonyms, max(weight)
from
(
select voc_assignment_id, pair_id, are_synonyms, avg(worker_performance) as weight from (

	select shd.voc_assignment_id, a.hit_id, a.id as assignment_id,shr.quality,shr.pair_id, shwp.quality as worker_performance, shr.are_synonyms from assignments a, syn_hits_completed shc, syn_hits_results shr, syn_hits_data shd, syn_hits_workers_performance shwp
	where  
	shd.hit_id=a.hit_id
	and
	shd.id=shr.pair_id
	and
	a.hit_id=shc.id
	and
	shr.assignment_id=a.id
	and 
	shr.is_control=0 -- remove controls (don't need them)
	and
	a.worker_id=shwp.id
	and 
	a.data_status>0.5 -- cut off nonperforming assignments
	
) as r

group by voc_assignment_id, pair_id, are_synonyms
) as rr

group by voc_assignment_id, pair_id, are_synonyms

			
			
			sort result by weight?
				
-- backtracking database from step2 to step1

select * from hits

select * from assignments

select * from vocabulary

select * from dictionary

select * from voc_hits_data
where hit_id in (2,3)

select * from voc_hits_results
where assignment_id=3

select * from voc_hits_results
where assignment_id=3
and is_control=0


select * from assignments
where id in (11,12)

select * from syn_hits
-- hit_id 41/42

select * from syn_hits_data
where id in (11,12)
-- hit_id=42

select * from syn_hits_data
where is_control=0

select * from syn_hits_results


-- voc assignments that are completed
select * from assignments
where id in (8,4)
-- hit_id 36/29 for assg_id 8/4


select * from voc_hits_data
where id in (36,29)


select * from syn_hits sh, assignments a, workers w, syn_hits_data shd
where sh.id=a.hit_id 
and a.worker_id=w.id
and
shd.is_control=0
and
shd.hit_id=sh.id


-- trace from voc_assignment_id to 
select * from assignments a, hits h
where
a.hit_id=h.id 
and
a.id=6

select * from voc_hits_results

select * from voc_hits_data

select * from assignments

select * from syn_hits_data




		
		
		
	*review-step2.py
		review results of step2

		HOW IT WORKS:
			for all assignments:
				+review if passed/failed based on control
				+mark assignments as passed/failed
				+go to MTurk and accept/reject assignments as needed
				+add extra assignment for each rejected assignment to corresponding HITs

			for all HITs:
				+select all HITS if # of accepted assignments=3
				
				-calculate passed/failed translations for each HIT for each word_id
				
						select * from synonymshits h, synonymshitassignments a, synonymshitresults r, synonymshitsdata d
						where a.mturk_hit_id=h.mturk_hit_id
						and
						r.assignment_id=a.id
						and 
						r.pair_id=d.id
						
						
						
						
						-- select words and 1/0 for correct/incorrect (need to be grouped)

						select word_id, d.assignment_id,
							case WHEN are_synonyms='no' THEN 0
							else 1 
							end as correct

						from synonymshits h, synonymshitassignments a, synonymshitresults r, synonymshitsdata d
						where a.mturk_hit_id=h.mturk_hit_id
						and
						r.assignment_id=a.id
						and 
						r.pair_id=d.id
						and 
						d.is_control=0
						
				-push fail/pass status to step 1
				
				
				 do something with all the QA data and pass it to step 1

	*review-step1.py 
		review results of step1
		
		HOW IT WORKS:
			for all assignments:
				-review if passed/failed based on results pushed from step2
				-mark assignments as passed/failed
				-go to MTurk and accept/reject assignments as needed
				-add extra assignment for each rejected assignment to corresponding HITs

			for all HITs:
				-select all HITS if # of accepted assignments=3
				
				-calculate passed/failed translations for each HIT for each word_id
				
				-build dictionary

				-tada!
	
	
	
---
tasks estimates

100 langs
10,000 words

1250 HITS per language

125,000 1 step HITs

375,000 1 step Assignments


375,000*2 synonym pair

750,000 pairs

100,000 step 2 HITS? (4 times less then step 1)

300,000 assignments for step 2


----


	
Rules for grading tasks:


First 10 assignments performed by worker are auto-approved

After that assignments are rated based on scale:
100%-70% approved, results good
70%-50% approved, results are questionable
50%-0% rejected, results are thrown away


give ? for questionable controls, see for majority vote

	
	
	







About

Template for creating and running Mechanical Turk HITs as External Questions

Resources

Stars

Watchers

Forks

Packages

No packages published