Difference between revisions of "Winter Milestone 5 Templates"

From crowdresearch
Jump to: navigation, search
(System)
(Undo revision 17774 by Vrindabhatia (talk))
 
(9 intermediate revisions by 2 users not shown)
Line 1: Line 1:
 
Please use the following template to write up your introduction section this week.  
 
Please use the following template to write up your introduction section this week.  
  
== System ==
+
== System (for task feed and open gov write up) ==
  
We're going to borrow introduction section from this paper as an example: [[:Media:Twitch Crowdsourcing (private).pdf | Vaish R, Wyngarden K, Chen J, et al. Twitch crowdsourcing: crowd contributions in short bursts of time. Proceedings of the 32nd annual ACM conference on Human factors in computing systems. ACM, 2014: 3645-3654.]] Please note how this section was divided into different parts. Please follow the same template.  
+
We're going to borrow systems section from this paper as an example: [[:Media:Twitch Crowdsourcing (private).pdf | Vaish R, Wyngarden K, Chen J, et al. Twitch crowdsourcing: crowd contributions in short bursts of time. Proceedings of the 32nd annual ACM conference on Human factors in computing systems. ACM, 2014: 3645-3654.]] Please note how this section was divided into different parts. Please follow the same template.  
  
 
=== Brief introduction of the system ===  
 
=== Brief introduction of the system ===  
Line 42: Line 42:
 
=== Module 1: Census ===  
 
=== Module 1: Census ===  
  
==== Problem ====
+
==== Problem/Limitations ====
 
Despite progress in producing effective understanding of
 
Despite progress in producing effective understanding of
 
static elements of our physical world — routes, businesses
 
static elements of our physical world — routes, businesses
Line 197: Line 197:
 
wider range of factual queries.
 
wider range of factual queries.
  
== Science ==
+
== Methods (for task authorship write up) ==
 
+
We're going to borrow introduction section from this paper as an example: [[:Media:2015 eta (private).pdf | Cheng, J., Teevan, J. & Bernstein, M.S. (2015). Measuring Crowdsourcing Effort with Error-Time Curves. CHI 2015.]]. Please note how this section was divided into different parts. Please follow the same template.
+
  
 +
We're going to borrow methods section from this paper as an example: [[:Media:2015 eta (private).pdf | Cheng, J., Teevan, J. & Bernstein, M.S. (2015). Measuring Crowdsourcing Effort with Error-Time Curves. CHI 2015.]]. Please note how this section was divided into different parts. Please follow the same template.
  
=== Phenomenon you're interested in ===
+
=== Study introduction ===  
Imagine that a requester wants to use Amazon Mechanical
+
STUDY 1: ETA VS. OTHER MEASURES OF EFFORT
Turk to label 10,000 images with a fixed set of tags. How
+
We begin by comparing ETA and other measures of difficulty
much should workers be paid to label each image? Would labeling
+
(including time and subjective difficulty) across a number of
an image with twice as many tags result in a task that
+
common crowdsourcing tasks. After describing the experimental
is twice as much effort? Should the tags be provided in a drop
+
setup, designed to elicit the necessary data to generate
down list or with radio buttons? Answering these questions
+
error-time curves and other measures for each task, we show
requires a fine-grained understanding of the amount of effort
+
how closely the different measures matched.
the task requires. This process today involves trial and error:
+
requesters observe the wait time and quality on test tasks,
+
guess what might have been causing any problems, tweak the
+
task, and repeat. An accurate measure of the effort required
+
to complete a crowdsourced task would enable requesters to
+
  
compare different approaches to their tasks, iterate toward a
+
=== Study method ===
better design, and price their tasks objectively. It could also
+
Method: Study 1 and all subsequent experiments reported in this paper
help workers decide whether to accept a task, or even allow
+
were conducted using a proprietary microtasking platform
systems to offer tasks based on difficulty or time availability.
+
that outsources crowd work to workers on the Clickworker
However, despite its potential value, task effort is challenging
+
microtask market. The platform interface is similar to that
to estimate. Workers face cognitive biases in assessing diffi-
+
of Amazon Mechanical Turk; users upload HTML task files,
culty [21], while requesters cannot easily observe the process
+
workers choose from a marketplace listing of tasks, and data
and, as experts, categorically underestimate completion times
+
is collected in CSV files. We restricted workers to those residing
[12]. These limits suggest the need for a behavioral approach
+
in the United States. Across all studies, 470 unique workers
to measure effort. One approach might be to let the market
+
completed over 44 thousand tasks. A followup survey
identify hard tasks by reacting to the posted price [30].
+
revealed that approximately 66% were female. We replicated
 +
Study 1 on Amazon Mechanical Turk and found empirically
  
=== The puzzle (observations we can't account for yet) ===
+
similar results, so we only report results using Clickworker in
However, prices cannot easily make fine distinctions in an inelastic
+
this paper.
market such as Mechanical Turk [14]. Another approach
+
might be to use task duration as a signal of difficulty, but this
+
is unreliable because workers regularly accept multiple tasks
+
simultaneously and interleave work [29]. Measures such as
+
reaction time [32] are not easy to apply to typical crowd tasks:
+
reaction time metrics tend to use simplistic tasks (e.g., shape
+
or color recognition), while others may be too involved for
+
crowd work (e.g., [9]).
+
  
=== Experimental design ===  
+
=== Method specifics and details ===  
In this paper, we propose a data-driven behavioral measure
+
Primitive Crowdsourcing Task Types
of effort that can be easily and cheaply calculated using the
+
We began by populating our evaluation tasks with common
crowd. Our metric, the error time area (ETA), draws on
+
crowdsourcing task types, or primitives, that appear commonly
cognitive psychology literature on speed-accuracy tradeoff
+
as microtasks or parts of microtasks. To do this, we
curves [32], and represents the effort required for a worker to
+
looked at the types of tasks with the most available HITs
accurately complete a task. To create it, we first recruit workers
+
on Amazon Mechanical Turk, at reports on common crowdsourcing
to complete the task under different time limits. Next, we
+
task types [15], and at crowdsourcing systems described
fit a curve to the collected data relating the error rate and time
+
in the literature (e.g., [4]). After several iterations
limit (Figure 1). Last, we compute ETA by taking the area
+
we identified a list of ten primitives that are present in most
under this error-time curve. Because ETA is calculated using
+
crowdsourcing workflows (Table 1, Figure 2). For example,
a data-driven approach, task difficulty can be determined
+
the Find-Fix-Verify workflow [4] could be expressed using
with minimal effort and without analytical modeling. Rather
+
a combination of the FIND (identify sentences which need
than measuring average duration independent of work quality,
+
shortening), FIX (shortening these sentences), and BINARY
ETA computes quality as a function of duration and thus
+
primitives (verifying the shortening is an improvement). In
can be used to estimate a wage for a task. ETA also allows
+
many cases, the primitives themselves (or repetitions of the
requesters to compare multiple task designs; for example, we
+
same primitive) make up the entire task, and map directly to
find that tagging an image with an open textbox is less effort
+
common Mechanical Turk tasks (e.g., finding facts such as
than choosing between a fixed list of 16 options, but more
+
phone numbers about individuals (SEARCH)).
effort than choosing between a fixed list of 8 options.
+
We instantiated these primitives using a dataset of images of
 +
people performing different actions (e.g., waving, cooking)
 +
[34] and a corpus of translated Wikipedia articles selected because
 +
they tend to contain errors [1].
  
=== Evaluation methods ===  
+
=== Experimental Design for the study ===  
 +
We presented workers with a mixed series of tasks from the
 +
ten primitives and manipulated two factors: the time limit
 +
and the primitive. Each primitive had seven different possible
 +
time limits, and one untimed condition. The exact time limits
 +
were initialized using how long workers took when not under
 +
time pressure. The result was a sampled, not fully-crossed,
 +
design. For each worker we randomly selected five primitives
 +
for them to perform; for each primitive, three questions of that
 +
type were shown with each of the specified time limits. The
 +
images or text used in these questions were randomly sampled
 +
and shuffled for each worker. To minimize practice effects,
 +
workers completed three timed practice questions prior
 +
to seeing any of these conditions. The tasks were presented
 +
in randomized order, and within each primitive the time conditions
 +
were presented in randomized order. Workers were
 +
compensated $2.00 and repeat participation was disallowed.
 +
A single task was presented on each page, allowing us to
 +
record how long workers took to submit a response. Under
 +
timed conditions, a timer started as soon as the worker advanced
 +
to the next page. Input was disabled as soon as the
 +
timer expired, regardless of what the worker was doing (e.g.,
 +
typing, clicking). An example task is shown in Figure 3.
  
After describing ETA, we explore the metric via four studies:
+
=== Measures from the study ===
Study 1: ETA vs. other measures of effort. For ten common
+
The information we logged allowed us to calculate behavioral
microtasking primitives (e.g., multiple choice questions,
+
measures for each primitive:
long-form text entry), we show that the ETA metric
+
– ETA. The ETA is the area under the error-time curve.
represents effort better than existing measures.
+
– Time@10. We also calculated the time it takes to achieve
Study 2: ETA vs. market price. We then compare ETA as
+
an error rate at the 10th percentile.
well as other measures to the market prices of these primitives
+
– Error. We measured the error rate against ground truth
on a crowdsourcing platform.
+
for each primitive. If there were many possible correct
– Study 3: Modeling perceptual costs. By augmenting ETA
+
responses, we manually judged responses while blind to
with measures of perceptual effort, we find we can better
+
condition. Automatically computing distance metrics (e.g.,
model a worker’s perceived difficulty of a task.
+
edit distance) resulted in empirically similar findings.
– Study 4: Tasks without ground truth. In order to capture
+
– Time. We measured how long workers took to complete the
how well people do a task, ETA requires ground truth. We
+
primitive without any time limit.
extend the metric to also work for subjective tasks.
+
After each task block was complete, we additionally asked
 +
workers to record several subjective reflections:
 +
Estimated time. We asked workers to report how long they
 +
thought they spent on a primitive absent time pressure.
 +
Time estimation has previously been used as an implicit
 +
signal of task difficulty [5].
 +
– Relative subjective duration (RSD). RSD, a measure of
 +
how much task time is over- or underestimated [5], is obtained
 +
by dividing the difference between estimated and
 +
actual time spent by the actual time spent.
 +
– Task load index (TLX). The NASA TLX [10] is a validated
 +
metric of mental workload commonly used in human factors
 +
research to assess task performance. It consists of a
 +
survey that sums six subjective dimensions (e.g., mental
 +
demand).
 +
A separate experimental design that contained all ten primitives,
 +
where each worker completed three untimed practice
 +
questions followed by three untimed questions for each primtive
 +
(with the primitives presented in random order), was used
 +
to obtain the
 +
– Subjective rank. Workers considered all of the primitives
 +
they completed and ranked them in order of effort required.
 +
As rankings produce sharper distinctions than individual ratings
 +
[2], we consider subjective rank to represent our ground
 +
truth ranking of the primitives. However, rank would not be a
 +
deployable solution for requesters. Ranking means that workers
 +
would need to test the new task against at least log(n)
 +
of the primitives, incurring a large fixed overhead. Further,
 +
ranking is ordinal, and cannot quantify small changes in effort.
 +
In contrast, ETA is an absolute ranking, can measure
 +
small changes in effort, and only needs to be measured for
 +
the target task to compare it with other tasks.
  
=== The Result ===  
+
=== What do we want to analyze? ===
We then demonstrate how ETA can be used for rapidly prototyping
+
Analysis
tasks. ETA makes it possible to characterize tasks in
+
60 workers completed Study 1, with 30 performing each
terms of their monetary cost and human effort, and paves the
+
primitive. We averaged our dependent measures across all
way for better task design, payment, and allocation.
+
30 workers, and compared the ranking of primitives induced
 +
by each measure to the average subjective ranking (subjective
 +
rank was obtained by having 40 other workers rank all
 +
ten primitives). We used the Kendall rank correlation coeffi-
 +
cient to capture how closely each measure approximated the workers’ ranks, with Holm-corrected p-values calculated under
 +
the null hypothesis of no association. A rank correlation
 +
of 1 indicates perfect correlation; 0 indicates no correlation.
 +
Measures that capture the subjective ranking accurately can
 +
analyze new tasks types without comparing them against multiple
 +
benchmark tasks.

Latest revision as of 06:21, 14 February 2016

Please use the following template to write up your introduction section this week.

System (for task feed and open gov write up)

We're going to borrow systems section from this paper as an example: Vaish R, Wyngarden K, Chen J, et al. Twitch crowdsourcing: crowd contributions in short bursts of time. Proceedings of the 32nd annual ACM conference on Human factors in computing systems. ACM, 2014: 3645-3654. Please note how this section was divided into different parts. Please follow the same template.

Brief introduction of the system

Twitch is an Android application that appears when the user presses the phone’s power/lock button (Figures 1 and 3). When the user completes the twitch crowdsourcing task, the phone unlocks normally. Each task involves a choice between two to six options through a single motion such as a tap or swipe.

How is the system solving critical problems

To motivate continued participation, Twitch provides both instant and aggregated feedback to the user. An instant feedback display shows how many other users agreed via a fadeout as the lock screen disappears (Figure 4) or how the user’s contributions apply to the whole (Figure 5). Aggregated data is also available via a web application, allowing the user to explore all data that the system has collected. For example, Figure 2 shows a human generated map from the Census application. To address security concerns, users are allowed to either disable or keep their existing Android passcode while using Twitch. If users do not wish to answer a question, they may skip Twitch by selecting ‘Exit’ via the options menu. This design decision has been made to encourage the user to give Twitch an answer, which is usually faster than exiting. Future designs could make it easier to skip a task, for example through a swipe-up.

Introducing modules of the system

Below, we introduce the three main crowdsourcing applications that Twitch supports. The first, Census, attempts to capture local knowledge. The following two, Image Voting and Structuring the Web, draw on creative and topical expertise. These three applications are bundled into one Android package, and each can be accessed interchangeably through Twitch's settings menu.

Module 1: Census

Problem/Limitations

Despite progress in producing effective understanding of static elements of our physical world — routes, businesses and points of interest — we lack an understanding of human activity. How busy is the corner cafe at 2pm on Fridays? What time of day do businesspeople clear out of the downtown district and get replaced by socializers? Which neighborhoods keep high-energy activities going until 11pm, and which ones become sleepy by 6pm? Users could take advantage of this information to plan their commutes, their social lives and their work.

Module preview

Existing crowdsourced techniques such as Foursquare are too sparse to answer these kinds of questions: the answers require at-the-moment, distributed human knowledge. We envision that twitch crowdsourcing can help create a human-centered equivalent of Google Street View, where a user could browse typical crowd activity in an area. To do so, we ask users to answer one of several questions about the world around them each time they unlock their phone. Users can then browse the map they are helping create.

System details

Census is the default crowdsourcing task in Twitch. It collects structured information about what people experience around them. Each Census unlock screen consists of four to six tiles (Figures 1 and 3), each task centered around questions such as: • How many people are around you? • What kinds of attire are nearby people wearing? • What are you currently doing? • How much energy do you have right now? While not exhaustive, these questions cover several types of information that a local census might seek to provide. Two of the four questions ask users about the people around them, while the other two ask about users themselves; both of which they are uniquely equipped to answer. Each answer is represented graphically; for example, in case of activities, users have icons for working, at home, eating, travelling, socializing, or exercising. To motivate continued engagement, Census provides two modes of feedback. Instant feedback (Figure 4) is a brief Android popup message that appears immediately after the user makes a selection. It reports the percentage of responses in the current time bin and location that agreed with the user, then fades out within two seconds. It is transparent to user input, so the user can begin interacting with the phone even while it is visible. Aggregated report allows Twitch users to see the cumulative effect of all users’ behavior. The data is bucketed and visualized on a map (Figure 2) on the Twitch homepage. Users can filter the data based on activity type or time of day.


Module 2: Photo Ranking

Problem/Limitations

Beyond harnessing local observations via Census, we wanted to demonstrate that twitch crowdsourcing could support traditional crowdsourcing tasks such as image ranking (e.g., Matchin [17]). Needfinding interviews and prototyping sessions with ten product design students at Stanford University indicated that product designers not only need photographs for their design mockups, but they also enjoy looking at the photographs. Twitch harnesses this interest to help rank photos and encourage contribution of new photos.

Module details

Photo Ranking crowdsources a ranking of stock photos for themes from a Creative Commons-licensed image library. The Twitch task displays two images related to a theme (e.g., Nature Panorama) per unlock and asks the user to slide to select the one they prefer (Figure 1). Pairwise ranking is considered faster and more accurate than rating [17]. The application regularly updates with new photos. Users can optionally contribute new photos to the database by taking a photo instead of rating one. Contributed photos must be relevant to the day’s photo theme, such as Nature Panorama, Soccer, or Beautiful Trash. Contributing a photo takes longer than the average Twitch task, but provides an opportunity for motivated individuals to enter the competition and get their photos rated. Like with Census, users receive instant feedback through a popup message to display how many other users agreed with their selection. We envision a web interface where all uploaded images can be browsed, downloaded and ranked. This data can also connect to computer vision research by providing high-quality images of object categories and scenes to create better classifiers.

Module 3: Structuring the Web

Problem/Limitations

Search engines no longer only return documents — they now aim to return direct answers [6,9]. However, despite massive undertakings such as the Google Knowledge Graph [36], Bing Satori [37] and Freebase [7], much of the knowledge on the web remains unstructured and unavailable for interactive applications. For example, searching for ‘Weird Al Yankovic born’ in a search engine such as Google returns a direct result ‘1959’ drawn from the knowledge base; however, searching for the equally relevant ‘Weird Al Yankovic first song’, ‘Weird Al Yankovic band members’, or ‘Weird Al Yankovic bestselling album’ returns a long string of documents but no direct answer, even though the answers are readily available on the performer’s Wikipedia page.

Module preview

To enable direct answers, we need structured data that is computer-readable. While crowdsourced undertakings such as Freebase and dbPedia have captured much structured data, they tend to only acquire high-level information and do not have enough contributors to achieve significant depth on any single entity. Likewise, while information extraction systems such as ReVerb [14] automatically draw such information from the text of the Wikipedia page, their error rates are currently too high to trust. Crowdsourcing can help such systems identify errors to improve future accuracy [18]. Therefore, we apply twitch crowdsourcing to produce both structured data for interactive applications and training data for information extraction systems.

Module details

Contributors to online efforts are drawn to goals that allow them to exhibit their unique expertise [2]. Thus, we allow users to help create structured data for topics of interest. The user can specify any topic on Wikipedia that they are interested in or want to learn about, for example HCI, the Godfather films, or their local city. To do so within a oneto-two second time limit, we draw on mixed-initiative information extraction systems (e.g., [18]) and ask users to help vet automatic extractions. When a user unlocks his or her phone, Structuring the Web displays a high-confidence extraction generated using ReVerb, and its source statement from the selected Wikipedia page (Figure 1). The user indicates with one swipe whether the extraction is correct with respect to the statement. ReVerb produces an extraction in SubjectRelationship-Object format: for example, if the source statement is “Stanford University was founded in 1885 by Leland Stanford as a memorial to their son”, ReVerb returns {Stanford University}, {was founded in}, {1885} and Twitch displays this structure. To minimize cognitive load and time requirements, the application filters only include short source sentences and uses color coding to match extractions with the source text. In Structuring the Web, the instant feedback upon accepting an extraction shows the user their progress growing a knowledge tree of verified facts (Figure 5). Rejecting an extraction instead scrolls the user down the article as far as their most recent extraction source, demonstrating the user’s progress in processing the article. In the future, we envision that search engines can utilize this data to answer a wider range of factual queries.

Methods (for task authorship write up)

We're going to borrow methods section from this paper as an example: Cheng, J., Teevan, J. & Bernstein, M.S. (2015). Measuring Crowdsourcing Effort with Error-Time Curves. CHI 2015.. Please note how this section was divided into different parts. Please follow the same template.

Study introduction

STUDY 1: ETA VS. OTHER MEASURES OF EFFORT We begin by comparing ETA and other measures of difficulty (including time and subjective difficulty) across a number of common crowdsourcing tasks. After describing the experimental setup, designed to elicit the necessary data to generate error-time curves and other measures for each task, we show how closely the different measures matched.

Study method

Method: Study 1 and all subsequent experiments reported in this paper were conducted using a proprietary microtasking platform that outsources crowd work to workers on the Clickworker microtask market. The platform interface is similar to that of Amazon Mechanical Turk; users upload HTML task files, workers choose from a marketplace listing of tasks, and data is collected in CSV files. We restricted workers to those residing in the United States. Across all studies, 470 unique workers completed over 44 thousand tasks. A followup survey revealed that approximately 66% were female. We replicated Study 1 on Amazon Mechanical Turk and found empirically

similar results, so we only report results using Clickworker in this paper.

Method specifics and details

Primitive Crowdsourcing Task Types We began by populating our evaluation tasks with common crowdsourcing task types, or primitives, that appear commonly as microtasks or parts of microtasks. To do this, we looked at the types of tasks with the most available HITs on Amazon Mechanical Turk, at reports on common crowdsourcing task types [15], and at crowdsourcing systems described in the literature (e.g., [4]). After several iterations we identified a list of ten primitives that are present in most crowdsourcing workflows (Table 1, Figure 2). For example, the Find-Fix-Verify workflow [4] could be expressed using a combination of the FIND (identify sentences which need shortening), FIX (shortening these sentences), and BINARY primitives (verifying the shortening is an improvement). In many cases, the primitives themselves (or repetitions of the same primitive) make up the entire task, and map directly to common Mechanical Turk tasks (e.g., finding facts such as phone numbers about individuals (SEARCH)). We instantiated these primitives using a dataset of images of people performing different actions (e.g., waving, cooking) [34] and a corpus of translated Wikipedia articles selected because they tend to contain errors [1].

Experimental Design for the study

We presented workers with a mixed series of tasks from the ten primitives and manipulated two factors: the time limit and the primitive. Each primitive had seven different possible time limits, and one untimed condition. The exact time limits were initialized using how long workers took when not under time pressure. The result was a sampled, not fully-crossed, design. For each worker we randomly selected five primitives for them to perform; for each primitive, three questions of that type were shown with each of the specified time limits. The images or text used in these questions were randomly sampled and shuffled for each worker. To minimize practice effects, workers completed three timed practice questions prior to seeing any of these conditions. The tasks were presented in randomized order, and within each primitive the time conditions were presented in randomized order. Workers were compensated $2.00 and repeat participation was disallowed. A single task was presented on each page, allowing us to record how long workers took to submit a response. Under timed conditions, a timer started as soon as the worker advanced to the next page. Input was disabled as soon as the timer expired, regardless of what the worker was doing (e.g., typing, clicking). An example task is shown in Figure 3.

Measures from the study

The information we logged allowed us to calculate behavioral measures for each primitive: – ETA. The ETA is the area under the error-time curve. – Time@10. We also calculated the time it takes to achieve an error rate at the 10th percentile. – Error. We measured the error rate against ground truth for each primitive. If there were many possible correct responses, we manually judged responses while blind to condition. Automatically computing distance metrics (e.g., edit distance) resulted in empirically similar findings. – Time. We measured how long workers took to complete the primitive without any time limit. After each task block was complete, we additionally asked workers to record several subjective reflections: – Estimated time. We asked workers to report how long they thought they spent on a primitive absent time pressure. Time estimation has previously been used as an implicit signal of task difficulty [5]. – Relative subjective duration (RSD). RSD, a measure of how much task time is over- or underestimated [5], is obtained by dividing the difference between estimated and actual time spent by the actual time spent. – Task load index (TLX). The NASA TLX [10] is a validated metric of mental workload commonly used in human factors research to assess task performance. It consists of a survey that sums six subjective dimensions (e.g., mental demand). A separate experimental design that contained all ten primitives, where each worker completed three untimed practice questions followed by three untimed questions for each primtive (with the primitives presented in random order), was used to obtain the – Subjective rank. Workers considered all of the primitives they completed and ranked them in order of effort required. As rankings produce sharper distinctions than individual ratings [2], we consider subjective rank to represent our ground truth ranking of the primitives. However, rank would not be a deployable solution for requesters. Ranking means that workers would need to test the new task against at least log(n) of the primitives, incurring a large fixed overhead. Further, ranking is ordinal, and cannot quantify small changes in effort. In contrast, ETA is an absolute ranking, can measure small changes in effort, and only needs to be measured for the target task to compare it with other tasks.

What do we want to analyze?

Analysis 60 workers completed Study 1, with 30 performing each primitive. We averaged our dependent measures across all 30 workers, and compared the ranking of primitives induced by each measure to the average subjective ranking (subjective rank was obtained by having 40 other workers rank all ten primitives). We used the Kendall rank correlation coeffi- cient to capture how closely each measure approximated the workers’ ranks, with Holm-corrected p-values calculated under the null hypothesis of no association. A rank correlation of 1 indicates perfect correlation; 0 indicates no correlation. Measures that capture the subjective ranking accurately can analyze new tasks types without comparing them against multiple benchmark tasks.