Crescimento (Growth)/Primeiro dia personalizado/Tarefas dos novatos

Grupo:	Growth
Início:	2018-07-24
Equipe:	Gergő Tisza (software engineer), Benoît Evellin (community relations specialist), Elena Tonkovidova (QA engineer), Kosta Harlan (software engineer), Morten Warncke-Wang (data analyst), Rita Ho (senior user experience designer)
Líder:	Roan Kattouw
Gestão:	Marcella Florence (engineering), Marshall Miller (product), Jazmin Tanner (program)
Atualização:	Most updates posted here.

This page is a translated version of the page Growth/Personalized first day/Newcomer tasks and the translation is 12% complete.

Languages:

Rastreado no Phabricator
Tarefa T227728

This page describes the Growth team's work on the "newcomer tasks" project, which is a specific project under the larger "Personalized first day" initiative. This page contains major assets, designs, and decisions. Most incremental updates on progress will be posted on the general Growth team updates page, with some large or detailed updates posted here.

You can quickly see what the team built by looking through these mockups (use arrow keys to navigate):

The design and planning for this project began on 2019-07-24. The first version was deployed in four wikis on 2019-11-20.

In December 2020, we published results showing the positive impact that newcomer tasks have on engagement. See this page for the details.

Situação atual

2019-07-24: first team meeting to discuss newcomer tasks
2019-08-27: team meeting to go over design concepts.
2019-09-09: Phabricator tasks created for engineering work
2019-09-23: desktop user tests complete
2019-09-30: mobile user tests complete
2019-11-20: V1.0 deployed to Czech, Korean, Arabic, and Vietnamese Wikipedias
2019-12-13: first variant test ("initiation") deployed to Czech, Korean, Arabic, and Vietnamese Wikipedias
2020-01-14: testing the addition of topic matching, to be deployed the week of 2020-01-20.
2020-01-21: the option to select topics of interest was added to the suggested edits module
2020-03-05: topic matching upgraded to use ORES models
2020-04-03: results from first variant test
2020-06-08: switched all newcomers to Variant A
2020-06-15: deployed guidance
2020-10-19: deployed Variants C and D
2020-11-20: Published Newcomer tasks experiment analysis
Seguinte: Continue improvements to Newcomer tasks and possibly develop new Structured tasks

Sumário

We think that newcomers should have every opportunity to succeed when they first arrive at the wiki. But frequently, newcomers attempt a task that is too difficult for them, can't find a task they want to do, or can't find ideas for how to remain involved after their first edit. This leads to many of them leaving and not coming back. There have been successful attempts in the past at recommending tasks to editors, and so we believe that the newcomer homepage is a potential place to recommend relevant tasks for newcomers.

We'll need to keep in mind a few things:

Many newcomers arrive with something specific in mind they're trying to accomplish, like add a specific photo to a certain article. We don't want to get in the way of them accomplishing their goal.
Newcomers build up their skills over time by progressing from easier edits to hard ones.
When newcomers are successful early on, they are more motivated to continue editing.

Taking those things into account, we want to recommend tasks to newcomers that arrive at the right place and time for them, teach them skills they need to be successful, and relate to their interests.

A valuable tool we have for helping tasks be relevant to newcomers is the welcome survey, which was originally built specifically for this purpose: personalizing the newcomer's experience. We'll plan to use the optional information newcomers give about their goals and interests to recommend appropriate tasks for them.

One of the largest challenges is going to be figuring out how to gather tasks that are appropriate for newcomers to do. There are many existing sources, such as templates that call for work on articles, recommendations in the Content Translation tool, or suggestions from tools like Citation Hunt. The question will be which of those options help newcomers accomplish their goals.

At first, we'll focus on using the newcomer homepage as the place to recommend tasks, but in the longer term, we can imagine building features that extend into the editing experience to recommend and help newcomers accomplish recommended tasks.

Also in the longer term, we'll be thinking about ways to tie task recommendations into other parts of the newcomer experience, such as the impact module on the homepage, or into the help panel.

The sections below may be changed significantly in the coming weeks, are too technical or less relevant for the understanding of the project. We have decided not to have them translated.

Porque esta idéia é prioridade

We know from research and experience that many newcomers fail early in their editing journey for one of these reasons:

They arrive with a very challenging edit in mind, such as writing a new article or adding an image.

Those tasks are difficult enough that they likely fail and don't return.

They arrive without knowing what to edit, and can't find any edits to make.

We also know that on the newcomer homepage, the most frequently clicked-on module is the "user page" module -- the only thing on the page that encourages users to start editing. This makes us think that many users are looking for a clear way to get started with editing.

And from past Wikimedia endeavors, we've seen that task recommendations can be valuable. SuggestBot is a project that sends personalized recommendations to experienced users, and is a well-received service. The Content Translation tool also serves personalized recommendations based on past translations, and has been shown to increase the volume of editing.

For all these reasons, we think that recommending specific editing tasks for newcomers will give them a clear way to get started. For those newcomers that have an edit in mind that we want to do, we'll encourage them to try some easy edits first to build up their skills. For those newcomers who do not have a specific preference on what to edit, they'll hopefully find some good edits from this feature.

Glossário

There are many terms that sound similar and can be confusing. This section defines each of them.

"Newcomer tasks": The entire workflow that recommends edits for newcomers and guides them through the edits.
"Suggested edits": The name of the specific module that the newcomer tasks workflow adds to the newcomer homepage.
"Task recommendations" or "Task suggestions": Lists of articles that need editing work, suggested automatically to users.
"Personalized": Software that adapts automatically to each user to fit their needs.
"Customized": Software that the user adapts to fit their needs.
"Topic": A content subject, such as "Art", "Music", or "Economics".
"Topic matching": The ability to find tasks for newcomers that match their topics of interest.
"Guidance": Features that help the newcomer complete the suggested task while they are working on it.
"Maintenance template": Templates that are put on articles indicating that work needs to be done on them.

Recommending tasks

The core challenge to this project is: Where will the tasks come from and how will we give the right ones to the right newcomers?

The graphic below shows our priorities when recommending tasks to newcomers.

As shown in the graphic above, we would give newcomers tasks that...

...arrive at the right time and place for a newcomer's journey.
...teach relevant conceptual and technical skills.
...gradually guide users to build up their editing abilities.
...be personalized to their interests.
...show them the value and impact of editing.
...motivate them to participate continually.

For instance, we do not want to give newcomers tasks that are irrelevant to what they hope to accomplish. If a newcomer wants to write a new article, then asking them to add a title description will not teach them skills they need to be successful.

We're splitting this challenge into two parts: the sourcing the tasks and topic matching.

Sourcing the tasks

There are many different places we could find tasks for newcomers to do. Our team listed as many as we could think of and evaluated them for whether they seem to be achievable for the first version of the feature. Below is a table showing the many sources of tasks that we evaluated in coming to the decision to start by using maintenance templates.


Source of task	Explanation	Evaluation
Maintenance templates	Most wikis use templates or categories to indicate articles that need copyediting, references, or other modifications. These are placed manually by experienced users.	Easily accessible. Already used in SuggestBot and GettingStarted .
Work on newest articles	New articles may be good candidates for work because they likely could be improved or expanded. They are also more likely to be about current topics.	Easily accessible, but most new articles are created by experienced users, and may not need help from newcomers.
Add images from Commons	There are articles that have images in some language Wikipedias but not in others. This could be a good task for a newcomer who created their account in order to add an image of their own.	An idea with high potential, but would require a lot of work to build interfaces. There are also questions about how to identify whether an article needs an image, and which one to recommend.
Expand short articles	Many articles are stubs that could be expanded.	This task is probably too open-ended and difficult for a newcomer.
Link to orphan articles	Many articles have no incoming links from any other articles. Users could find articles to link to the orphan articles.	Easy to identify orphans, but may be confusing for a newcomer to have to go find other articles in order to do the task.
Add references	Many articles are in need of additional references or citations.	Probably a challenging task for a newcomer. Frequently covered by maintenance templates.
Add categories	Categories are used for many purposes on the wikis, and adding them to articles that don't have them could be a low-pressure way to contribute.	Newcomers may not have good judgment when it comes to adding categories. This also does not teach editing skills that they need for other tasks.
Content translation	The Content Translation tool could be a good way to structure the editing experience and help newcomers write new articles without having to generate all the content on their own.	An integration here could be great -- we may want to use the welcome survey to distinguish which newcomers are multilingual.
Add sections	There are algorithms in development that can recommend additional section headers based on similar articles.	Writing a new section from scratch may be too challenging a task for a newcomer.
Specific link recommendation	Adding wikilinks is one of the best tasks for newcomers. It would be powerful if we could not only tell a newcomer that an article needs more links, but indicate which specific words or phrases should become an link (internal and/or external, depending on local policies).	Some research has been done on this idea that the team will be looking into, as this idea could be a perfect first edit for a newcomer.
Specific copy edits	Many articles need copyediting, but it would be a better experience for newcomers if we could suggest specific changes to make in article, such as words that are likely misspelled or sentences that likely need to be rephrased.	While this would be an excellent experience for the newcomer, we don't have a way to approach this. Perhaps experienced could flag specific copy edit changes instead of fixing them.

Version 1.0: basic workflow

In version 1.0, we will deploy the basic parts of the newcomer tasks workflow. It will recommend articles to newcomers that require different types of edits, but it will not match the articles to the newcomers' topics of interest (version 1.1), and it will also not guide the newcomers in completing the task (version 1.2).

Maintenance templates

We're going to be starting by using maintenance templates and categories to identify articles that need work. All of our target wikis use some set of maintenance templates or categories on thousands of articles, tagging them as needing copyediting, references, images, links, or expanded sections. And previous task recommendations software, such as SuggestBot, have used them successfully. These are some examples of maintenance categories:

In this Phabricator task, we investigated exactly which templates are present and in what quantities, to get a sense of whether there will be enough tasks for newcomers. There seem to be sufficient numbers for the initial version of this project. We are likely to incorporate other task sources from the table below in future versions.

It's also worth noting that it could be possible to supplement many of these maintenance templates with automation. For instance, it is possible to automatically identify articles that have no internal links, or articles that have no references. This is an area for future exploration.

During the week of October 21, 2019, the members of the Growth team did a hands-on exercise in which we attempted to edit articles with maintenance templates. This helped us understand what challenges we can expect newcomers to face, and gave us ideas for addressing them. Our notes and ideas are published here.

Design

Comparative review

Our team's designer reviewed the way that other platforms (e.g. TripAdvisor, Foursquare, Amazon Mechanical Turk, Google Crowdsource, Reddit) offer task recommendations to newcomers. We also reviewed Wikimedia projects that incorporate task recommendations, such as the Wikipedia Android app and SuggestBot. We think there are best practices we can learn from other software, especially when we see the same patterns across many different types of software. Even as we incorporate ideas from other software, we will still make sure to preserve Wikipedia's unique values of openness, clarity, and transparency. The main takeaways are below, and the full set of takeaways is on this page:

Task types – bucket into 4 types: Rating content, Creating content, Moderating/Verifying content, Translating content
Incentives – Most products offered intangible incentives mainly bucketed into the form of: Awards and ranking (badges), Personal pride and gratification (stats), or Unlocking features (access rights)
Reward incentives – promote badges or attainments of specific milestones (e.g., a badge for adding 50 citations)
Personalization/Customization – Most have at least one facet of personalization/customization. Most common customization is user input on surveys upon account creation or before a task, most common system-based personalization type is geolocalization
Visual design & layout – incentivizing features (stats, leaderboards, etc) and onboarding is visually rich compared to pared back, simple forms to complete short edits.
Guidance – Almost all products reviewed had at least basic guidance prior to task completion, most commonly introductory ‘tours’. In-context help was also provided in the form of instructional copy, tooltips, step-by-step flows, as well as offering feedback mechanisms (ask questions, submit feedback)

Mockups

Our evolving designs can always be found in two sets of interactive mockups (use arrow keys to navigate):

Desktop
Mobile

Those mockups contain explorations of all the difference parts of the user journey, which we have broken down into several parts:

Gathering information from the newcomer: learning what we need in order to recommend relevant tasks.
Feature discovery: the way the newcomer first encounters task recommendations.
Task recommendations: the interface for filtering and choosing tasks.
Guidance during editing: once the newcomer is doing a task, the guidance that helps them understand what to do.
User feedback: ways in which the newcomer can indicate that they are not satisfied with the recommended task.
Next edit: how we continue the user's momentum after the save an edit.

Below are some of the original draft design concepts as the team continues to refine our approach.

Mockup of possible location for newcomer tasks module
Mockup of initial state of module
Mockup of possible design for module
Mockup of guidance while doing a task using the help panel
Mockup of feedback element if a user does not complete a task

User testing

Desktop

During the week of September 16, 2019, we used usertesting.com to conduct six tests of the desktop newcomer tasks prototype with internet users unaffiliated with the Wikimedia movement. In these tests, respondents are compensated for trying out the mockups, speaking aloud on what they observe, and answering questions about the experience. The full results can be found in this Phabricator task. The goals of this testing were:

Gauge the discoverability of the newcomer tasks module
Identify improvements to the usability of the tasks module:
1. Do users understand how to select and review article suggestions?
2. Do users understand how to filter by interests and task difficulty?
3. Do they know how to start editing a suggested article?
Gauge user reactions to the suggestions and expectations about guidance through the task.

Resumo dos resultados

All users thought it made sense and intuitive to get suggestions based on their topics of interest.
Similarly, the different task difficulties was positively received by all participants.
Overall usability of the suggested edits module was extremely high. People knew how to click to view more articles, use the filter to change topics and task levels, and knew to click on the card to open a suggestion for editing.
4/6 Participants did not initially realize that they should click on “See suggested edits” as a way to help them achieve their goal of writing a new article. This seemed to be a common mental model where users separated "Editing" as different from "Creating a new page".
Start module is clearly the starting point for all participants. Moreover many were drawn to “See suggested edits” button as a way to follow the progression of activities in the start module.
Users had a clear understanding and expectation they would be shown suggested articles for editing based on the intro dialogs to add topics and introducing task levels.
Everyone was able to select the popular topics and add their own topic easily.
Everyone understood the purpose of the suggested edits module.
Two people were confused/assumed that they could not create a new article until completing easy and medium tasks.
5 of 6 participants knew to click on the help panel button for guidance once they entered the editor mode.
Four people expected to be able to contact their mentor in the help panel.
Task tips lacked sufficient level of guidance for a couple of participants.

Recomendações

Improve copy and more over user education that creating new content is a form of editing.
Make updates to the Impact module as tested here to aid user understanding of suggested edits.
Provide good in-edit context help. It’s very important for users trying an edit.
- Include a “checklist” for users to revise in the help panel’s task tips.
- Provide short examples of what to do.
- Indicate to users they do not have to copy edit for an entire article.
Including real-time filtering results helps users connect suggestions as article edits and encourage use of the filtering to find matching articles.

Mobile

During the week of September 30, 2019, we used usertesting.com to conduct six tests of the mobile newcomer tasks prototype. The full results can be found in this Phabricator task. The goals of this testing were the same as with desktop, but with the added goal of understanding how the mobile experience should differ from the desktop experience. Mobile user testers were prompted with the scenario of intending to add an image to Wikipedia (whereas desktop respondents were prompted with the scenario of intending to create a new article).

Resumo dos resultados

Overall users found the start module (redesigned) clearly laid out the guided steps to begin.
The extra “Suggested edits” module below, while not especially confusing, was still not where users expected to go to help them with their task to add an image.
Suggested edits was quite intuitive to use, with participants understanding how its different elements (filtering, seeing more articles, etc) worked. However, users do not see the value of doing Suggested edits beyond learning or boredom.
Several people wanted more granular topics to be available than the broad topics listed.
Having the detailed difficulty info was educational, but potentially discouraging. All were surprised “Adding images” was classed as hard, with varying degrees of frustration about this fact.
Filtering by interests is a big selling point.
3 people towards the end of the test assumed there was some “verification” or requirement to do some Easy tasks before Medium/Hard tasks could be achieved
Everyone understood the purpose of the Suggested edits as giving edits that would users learn to edit, and also emphasize that it showed them some edits were harder to do.
All users struggled to use the guidance we offered through the help panel while they were editing. This is a major area we need to think hard about designing before we begin to build it.

Recomendações

Suggested edits call to action is inside start module, not its own card.
Improve copy and user education imagery to better convey that there is real world value in trying suggested edits beyond learning and that task difficulty is a guide only and tasks can be tried out of order.
Add an overlay specifically to introduce personalized introduction to suggested edits.
Including real-time counting of filtered results on both task and topic filters.
Incorporate more granular searching by interest topics by users.
Reiterate when a user opens a suggestion that it is a real, impactful edit.
Update design of the in-task help panel so that all available help content is clearly accessible.

Version 1.1: topic matching

Past research and development shows that users are more likely to do recommended tasks if the tasks match their topical interests. SuggestBot uses an editor's past editing history to find similar articles, and those intelligent results are shown in this paper to be executed on more often than random results. The Content Translation tool also recommends articles based on a user's previous translation history, and those recommendations have increased the translation volume.

In looking at the usage of V1.0 of newcomer tasks, which does not contain topic matching, we see that there are users who navigate through many suggested articles, and end up clicking on none. There are also users who navigate through many, and end up editing only the ones they happen to find that belong to a certain topic, such as medicine. These are also good indicators that topics can be valuable to help newcomers find articles they want to edit.

Our challenge with newcomers is a "cold start problem", in that newcomers do not have any edit history to use when trying to find relevant articles for them to edit. We want to have an algorithm that says what the topic is of each article, and use that to filter the articles that have maintenance templates.

Algoritmos

There are multiple approaches with which we might find articles that match a user's stated topic of interest. While our team identified many, we built prototypes for three methods and tested them:

morelike: assign a seed list of articles that represent each topic area (e.g. "Art" might be represented by the articles for "Painting", "Sculpture", "Dance", and "Weaving".) Use that seed list to find other articles that are similar to those in the seed list by using a similarity algorithm called "morelike".
free text: instead of choosing from a set list of topics, allow newcomers to type in any phrase they want to indicate a topic. Use regular Wikipedia search to surface articles relevant to that phrase.
ORES: ORES is a machine learning service that – among other things – can return a predicted topic for any article. Though this prediction service only works in English Wikipedia, there are ways to translate predictions from English to other wikis.

In this Phabricator task, we evaluated the three methods, and decided to proceed with the ORES model. The Growth team worked with the Scoring team to strengthen the model, and with the Search team to make the model predictions available to the newcomer tasks workflow. During the time that this work was happening, we deployed the somewhat worse-performing morelike algorithm, and switched to the ORES model about a month later.

The ORES model we use now offers 64 topics, and we chose to expose 39 of them to newcomers. The evaluation in four different languages showed that on average, 8.5 out of 10 suggestions for a given topic seem like good matches for that topic.

Design

In designing interfaces that allow newcomers to choose topics of interest, these are some of the considerations:

How to make a long list of about 30 topics not overwhelming to the user?
How to handle multiple layers of topics (e.g. if "Science" has sub-topics of "Biology", "Chemistry", etc.)
Whether users can give feedback when a topic does not match what they selected?

These mockups contain our current designs for this interface. You can navigate with your keyboard's arrow keys. Below are some images of the mockups:

Initial overlay to indicate topics of interest (desktop)
Suggested edits module if user has not yet chosen topics
Dialog for setting topics (desktop)
Dialog for setting topics (mobile)

Version 1.2: guidance

Guidance was deployed on 2020-06-15. For a guide to translating the messages in this feature, see this page.

After newcomers have selected an article from the suggested edits module, they should receive guidance about how to click edit and complete the edit successfully. While it is exciting that some portion of newcomers are completing suggested edits without guidance, we're confident that by adding guidance, we will substantially increase how many newcomers edit.

We decided to repurpose the help panel as the place to deliver this guidance. Reusing the help panel will allow us to build quickly. The guidance contains three phases:

When the user has arrived on the article and before they click edit.
After clicking edit and before saving an edit.
After saving an edit.

Some of the ideas we considered implementing included:

Guidance tailored to each type of edit, varying depending on whether the suggested edit is a copyedit, adding links, adding references, etc.
Reminder that an edit can be small, and that the user does not have to edit the whole article.
Step-by-step walkthrough that is like a checklist for completing the edit.
Highlighting the maintenance templates in the article so that the user can see why the article was suggested.
An indicator that encourages the user to click the edit button.
A place to put videos that demonstrate how to complete the edit.
Suggestions for additional edits after saving the initial edit.
Ability for the user to notify their mentor that they have done an edit, so the mentor can check their work and thank them.

During the last week of December 2019, we user tested desktop and mobile prototypes, which can be found below. We will post the user test results after assembling them.

Below are some images of the prototype:

Help panel displaying guidance before clicking edit on desktop
Overlay after saving an edit on desktop
Help panel "peeking" from bottom of page on mobile
Help panel contents on mobile

Variant testing

After deploying the first version of newcomer tasks, we want to start testing different variants of the feature, so that we can improve it iteratively. Rather than just having one design of newcomer tasks, and seeing if newcomers are more productive with it than without it, we plan to test more than variant of newcomer tasks at a time, and compare them. We have compiled an exhaustive list of all the ideas of variants to test -- but we will only end up testing perhaps 10 per year, because of the effort and time it takes to build, test, and analyze.

In March, April, and May 2020, we'll be testing variants that aim to get more users into the newcomer tasks flow.

See this page for the list of variant tests and their results.

Measurement and results

Controlled experiment

In December 2020, we published the results of a controlled experiment showing that newcomer tasks have a positive impact on engagement. These are our most important results, and give us confidence that these features should expand to more wikis. See this page for the details.

Usage

Starting in December 2019, we have been tracking several key metrics from newcomers tasks. The graphs shown in this section are our main charts of those metrics as of 2020-08-17.

Summary

Since deploying newcomer tasks in November 2019, we have seen steady increases in both the number of edits from the feature and the number of editors using the feature. These increases are due to two elements: (a) improvements to the feature, and (b) expanding the feature to more wikis.

Specific charts

Conversion funnel: the first graph is the most important to our team. Each line shows how many newcomers arrive at each stage of our "conversion funnel", meaning how far they progress into the newcomer tasks workflow, as a percentage of newcomers who visit their homepage. We want the users to move through the stages of (1) interacting with the module (blue), (2) selecting an article (red), (3) clicking edit on the article (orange), (4) saving an edit (green). In general, we want to see all the lines go up.

Since the early days of the feature, the percentage of users who have clicked edit and who have saved edits have steadily gone up. In January 2020, something like 2% of newcomers who visited their homepage saved a suggested edit. In August 2020, that has grown to 5.3%, which is more than double.
In August, almost all users who selected a task clicked edit, which can be seen by the closeness of the red and orange lines.
We think that these improvements are due to the two major features we deployed between January and August: topic matching (which allows newcomers to find more interesting articles) and guidance (which encourages them to click edit and explains how to complete the edit).

Edits: the second graph shows the number of newcomer task edits completed each week, with a separate line for each wiki and a "total" line in black. From December to August 17, there have been 15,126 edits completed through newcomer tasks. It is clear that this has grown over time, which is certainly to be expected because we have gone from 4 wikis to 12 between January and August.

But looking at the individual wikis' lines, it is possible to see growth over time.

It is common for the number of suggested edits completed each week on a wiki to vary a lot. One of the reasons is that a small number of enthusiastic newcomers can create dozens or hundreds of edits in a short time, but then may not be on wiki on other weeks.
Arabic Wikipedia, being one of the largest wikis that has the feature, consistently creates the most edits.

Editors: in addition to tracking the number of edits, we also want to make sure that increasing numbers of newcomers are participating. The third graph shows the number of users completing newcomer tasks each week, broken out by wiki.

Similarly to the graph of edits, this number also has increased steadily, and the addition of new wikis (such as French Wikipedia on week 21 and Persian Wikipedia on week 32) are clearly visible.
We believe that the effect of "guidance" is visible. This was released before week 25. There have been over 100 users of newcomer tasks every week since its release, whereas only three weeks had previously reached that level.

Edit quality

The Growth team's ambassadors have gone through over 300 edits saved by newcomers and marked whether or not each edit was productive (meaning that it improved the article). We are happy to see that about 75% of the edits are productive. This is similar to the baseline rate for newcomer edits, and we're glad that this feature has not encouraged vandalism. Most of the edits are copyedits, with many also adding links, and some even adding content and references. About a third of users who make one suggested edit go on to make additional suggested edits. Many also go on to make edits that are not suggested by the feature, which is behavior we are happy to see.

The high-quality edits we're seeing encourage us to improve the feature so that more newcomers begin and complete its workflow.