Today we’re publishing the fourth edition of our Community Standards Enforcement Report, detailing our work for Q2 and Q3 2019. We are now including metrics across ten policies on Facebook and metrics across four policies on Instagram.
These metrics include:
- Prevalence: how often content that violates our policies was viewed
- Content Actioned: how much content we took action on because it was found to violate our policies
- Proactive Rate: of the content we took action on, how much was detected before someone reported it to us
- Appealed Content: how much content people appealed after we took action
- Restored Content: how much content was restored after we initially took action
We also launched a new page today so people can view examples of how our Community Standards apply to different types of content and see where we draw the line.
Adding Instagram to the Report
For the first time, we are sharing data on how we are doing at enforcing our policies on Instagram. In this first report for Instagram, we are providing data on four policy areas: child nudity and child sexual exploitation; regulated goods — specifically, illicit firearm and drug sales; suicide and self-injury; and terrorist propaganda. The report does not include appeals and restores metrics for Instagram, as appeals on Instagram were only launched in Q2 of this year, but these will be included in future reports.
While we use the same proactive detection systems to find and remove harmful content across both Instagram and Facebook, the metrics may be different across the two services. There are many reasons for this, including: the differences in the apps’ functionalities and how they’re used – for example, Instagram doesn’t have links, re-shares in feed, Pages or Groups; the differing sizes of our communities; where people in the world use one app more than another; and where we’ve had greater ability to use our proactive detection technology to date. When comparing metrics in order to see where progress has been made and where more improvements are needed, we encourage people to see how metrics change, quarter-over-quarter, for individual policy areas within an app.
What Else Is New in the Fourth Edition of the Report
- Data on suicide and self-injury: We are now detailing how we’re taking action on suicide and self-injury content. This area is both sensitive and complex, and we work with experts to ensure everyone’s safety is considered. We remove content that depicts or encourages suicide or self-injury, including certain graphic imagery and real-time depictions that experts tell us might lead others to engage in similar behavior. We place a sensitivity screen over content that doesn’t violate our policies but that may be upsetting to some, including things like healed cuts or other non-graphic self-injury imagery in a context of recovery. We also recently strengthened our policies around self-harm and made improvements to our technology to find and remove more violating content.
- On Facebook, we took action on about 2 million pieces of content in Q2 2019, of which 96.1% we detected proactively, and we saw further progress in Q3 when we removed 2.5 million pieces of content, of which 97.3% we detected proactively.
- On Instagram, we saw similar progress and removed about 835,000 pieces of content in Q2 2019, of which 77.8% we detected proactively, and we removed about 845,000 pieces of content in Q3 2019, of which 79.1% we detected proactively.
- Expanded data on terrorist propaganda: Our Dangerous Individuals and Organizations policy bans all terrorist organizations from having a presence on our services. To date, we have identified a wide range of groups, based on their behavior, as terrorist organizations. Previous reports only included our efforts specifically against al Qaeda, ISIS and their affiliates as we focused our measurement efforts on the groups understood to pose the broadest global threat. Now, we’ve expanded the report to include the actions we’re taking against all terrorist organizations. While the rate at which we detect and remove content associated with Al Qaeda, ISIS and their affiliates on Facebook has remained above 99%, the rate at which we proactively detect content affiliated with any terrorist organization on Facebook is 98.5% and on Instagram is 92.2%. We will continue to invest in automated techniques to combat terrorist content and iterate on our tactics because we know bad actors will continue to change theirs.
- Estimating prevalence for suicide and self-injury and regulated goods: In this report, we are adding prevalence metrics for content that violates our suicide and self-injury and regulated goods (illicit sales of firearms and drugs) policies for the first time. Because we care most about how often people may see content that violates our policies, we measure prevalence, or the frequency at which people may see this content on our services. For the policy areas addressing the most severe safety concerns — child nudity and sexual exploitation of children, regulated goods, suicide and self-injury, and terrorist propaganda — the likelihood that people view content that violates these policies is very low, and we remove much of it before people see it. As a result, when we sample views of content in order to measure prevalence for these policy areas, many times we do not find enough, or sometimes any, violating samples to reliably estimate a metric. Instead, we can estimate an upper limit of how often someone would see content that violates these policies. In Q3 2019, this upper limit was 0.04%. Meaning that for each of these policies, out of every 10,000 views on Facebook or Instagram in Q3 2019, we estimate that no more than 4 of those views contained content that violated that policy.
- It’s also important to note that when the prevalence is so low that we can only provide upper limits, this limit may change by a few hundredths of a percentage point between reporting periods, but changes that small do not mean there is a real difference in the prevalence of this content on the platform.
Progress to Help Keep People Safe
Across the most harmful types of content we work to combat, we’ve continued to strengthen our efforts to enforce our policies and bring greater transparency to our work. In addition to suicide and self-injury content and terrorist propaganda, the metrics for child nudity and sexual exploitation of children, as well as regulated goods, demonstrate this progress. The investments we’ve made in AI over the last five years continue to be a key factor in tackling these issues. In fact, recent advancements in this technology have helped with rate of detection and removal of violating content.
For child nudity and sexual exploitation of children, we made improvements to our processes for adding violations to our internal database in order to detect and remove additional instances of the same content shared on both Facebook and Instagram, enabling us to identify and remove more violating content.
- In Q3 2019, we removed about 11.6 million pieces of content, up from Q1 2019 when we removed about 5.8 million. Over the last four quarters, we proactively detected over 99% of the content we remove for violating this policy.
While we are including data for Instagram for the first time, we have made progress increasing content actioned and the proactive rate in this area within the last two quarters:
- In Q2 2019, we removed about 512,000 pieces of content, of which 92.5% we detected proactively.
- In Q3, we saw greater progress and removed 754,000 pieces of content, of which 94.6% we detected proactively.
For our regulated goods policy prohibiting illicit firearm and drug sales, continued investments in our proactive detection systems and advancements in our enforcement techniques have allowed us to build on the progress from the last report.
- In Q3 2019, we removed about 4.4 million pieces of drug sale content, of which 97.6% we detected proactively — an increase from Q1 2019 when we removed about 841,000 pieces of drug sale content, of which 84.4% we detected proactively.
- Also in Q3 2019, we removed about 2.3 million pieces of firearm sales content, of which 93.8% we detected proactively — an increase from Q1 2019 when we removed about 609,000 pieces of firearm sale content, of which 69.9% we detected proactively.
- In Q3 2019, we removed about 1.5 million pieces of drug sale content, of which 95.3% we detected proactively.
- In Q3 2019, we removed about 58,600 pieces of firearm sales content, of which 91.3% we detected proactively.
New Tactics in Combating Hate Speech
Over the last two years, we’ve invested in proactive detection of hate speech so that we can detect this harmful content before people report it to us and sometimes before anyone sees it. Our detection techniques include text and image matching, which means we’re identifying images and identical strings of text that have already been removed as hate speech, and machine-learning classifiers that look at things like language, as well as the reactions and comments to a post, to assess how closely it matches common phrases, patterns and attacks that we’ve seen previously in content that violates our policies against hate.
Initially, we’ve used these systems to proactively detect potential hate speech violations and send them to our content review teams since people can better assess context where AI cannot. Starting in Q2 2019, thanks to continued progress in our systems’ abilities to correctly detect violations, we began removing some posts automatically, but only when content is either identical or near-identical to text or images previously removed by our content review team as violating our policy, or where content very closely matches common attacks that violate our policy. We only do this in select instances, and it has only been possible because our automated systems have been trained on hundreds of thousands, if not millions, of different examples of violating content and common attacks. In all other cases when our systems proactively detect potential hate speech, the content is still sent to our review teams to make a final determination. With these evolutions in our detection systems, our proactive rate has climbed to 80%, from 68% in our last report, and we’ve increased the volume of content we find and remove for violating our hate speech policy.
While we are pleased with this progress, these technologies are not perfect and we know that mistakes can still happen. That’s why we continue to invest in systems that enable us to improve our accuracy in removing content that violates our policies while safeguarding content that discusses or condemns hate speech. Similar to how we review decisions made by our content review team in order to monitor the accuracy of our decisions, our teams routinely review removals by our automated systems to make sure we are enforcing our policies correctly. We also continue to review content again when people appeal and tell us we made a mistake in removing their post.
Updating our Metrics
Since our last report, we have improved the ways we measure how much content we take action on after identifying an issue in our accounting this summer. In this report, we are updating metrics we previously shared for content actioned, proactive rate, content appealed and content restored for the periods Q3 2018 through Q1 2019.
During those quarters, the issue with our accounting processes did not impact how we enforced our policies or how we informed people about those actions; it only impacted how we counted the actions we took. For example, if we find that a post containing one photo violates our policies, we want our metric to reflect that we took action on one piece of content — not two separate actions for removing the photo and the post. However, in July 2019, we found that the systems logging and counting these actions did not correctly log the actions taken. This was largely due to needing to count multiple actions that take place within a few milliseconds and not miss, or overstate, any of the individual actions taken.
We’ll continue to refine the processes we use to measure our actions and build a robust system to ensure the metrics we provide are accurate. We share more details about these processes here.
The post Community Standards Enforcement Report, November 2019 Edition appeared first on About Facebook.
Messenger API to support Instagram
Today, we are announcing updates to the Messenger API to support Instagram messaging, giving businesses new tools to manage their customer communications on Instagram at scale. The new API features enable businesses to integrate Instagram messaging with their preferred business applications and workflows; helping drive more meaningful conversations, increase customer satisfaction and grow sales. The updated API is currently in beta with a limited number of developer partners and businesses.
Instagram is a place for emerging culture and trend creation and discovering new brands is a valuable part of this experience. Messaging plays a central role in helping people connect with brands in personal ways through story replies, direct messages, and mentions. Over the last year, total daily conversations between people and businesses on Messenger and Instagram grew over 40 percent. For businesses, the opportunity to drive sales and improve customer satisfaction by having meaningful interactions with people on Instagram messaging is huge.
“Instagram is a platform for community building, and we’ve long approached it as a way for us to connect with our customers in a place where they are already spending a lot of their time. With the newly launched Messenger API support for Instagram, we are now able to increase efficiency, drive even stronger user engagement, and easily maintain a two-way dialogue with our followers. This technology has helped us create a new pipeline for best-in-class service and allows for a direct line of communication that’s fast and easy for both customers and our internal team.” – Michael Kors Marketing
Works with your tools and workflows
Businesses want to use a single platform to respond to messages on multiple channels. The Messenger API now allows businesses to manage messages initiated by people throughout their Instagram presence, including Profile, Shops, and Stories. It will be possible for businesses to use information from core business systems right alongside Instagram messaging, enabling more personal conversations that drive better business outcomes. For example, businesses integrating with a CRM system can give agents a holistic view of customer loyalty. Furthermore, existing investments in people, tools, and workflows to manage other communication channels can be leveraged and extended to support customers on Instagram. This update will also bring Facebook Shops messaging features to the Messenger API so businesses can create more engaging and connected customer experiences.
To get started, businesses can easily work with developers to integrate Instagram messaging with their existing tools and systems.
Increases responsiveness and customer satisfaction
Customers value responsiveness when they have questions or need help from businesses. For the first time on Instagram, we’re introducing new features that will allow businesses to respond immediately to common inquiries using automation, while ensuring people are seamlessly connected to live support for more complex inquiries. One of our alpha partners, Clarabridge, reported their client brands had improved their response rate by up to 55% since being able to manage Instagram DMs through their platform.
The updates to the Messenger API are part of our overall effort to make it easier for businesses to reach their customers across our family of apps.
Messenger API support for Instagram is currently in beta with a focus on providing high quality, personalized messaging experiences on Instagram while increasing business efficiency. Adidas, Amaro, Glossier, H&M, MagazineLuiza, Michael Kors, Nars, Sephora and TechStyle Fashion Group and other consumer brands are already participating in the beta program. We are excited about early results some businesses saw during alpha testing, including higher response rates, reduced resolution times, and deeper customer insights as a result of integrations. We’re also testing with a limited number of developer partners; and are delighted at the initial response.
“On average, brands have saved at least four hours per agent per week by streamlining social community management within the Khoros platform, plus shortened response rates during business hours — which is crucial to meet as customers who message brands on social media expect a quick reply.” – Khoros
Required migration to token-based access for User Picture and oEmbed endpoints
As part of our Graph API 8.0 release, we announced breaking changes to how developers can access certain permissions and APIs. Starting October 24, 2020, developers need to migrate to token-based access in order to access User Picture and oEmbed endpoints for Facebook and Instagram.
This post outlines these changes and the necessary steps developers need to take to avoid disruption to their app.
Facebook will now require client or app access tokens to access a user’s profile picture. Beginning on October 24, 2020 queries for profile pictures made against user IDs without an access token will return a generic silhouette rather than a profile picture. This is a breaking change for partners. While client or app tokens will be required for user ID queries, they will continue to be a best practice (and not required) for ASID queries for the time being.
Facebook and Instagram oEmbed
We are also deprecating the existing Legacy API oEmbed endpoints for Facebook and Instagram on October 24, 2020, which will be replaced with new Graph API endpoints. If developers don’t make this change and continue to attempt to call the existing oEmbed API, their requests will fail and developers will receive an error message instead. These new endpoints will require client or app access tokens or ASID queries.
Ready to make the switch? You can read more about these changes in our developer documentation for User Picture and also visit our changelog for Facebook OEmbed and IG OEmbed for details on how to start calling these Graph API endpoints.
PyTorch Tutorials Refresh – Behind the Scenes
Hi, I’m Jessica, a Developer Advocate on the Facebook Open Source team. In this blog, I’ll take you behind the scenes to show you how Facebook supports and sustains our open source products – specifically PyTorch, an open source deep learning library. With every new release version, PyTorch pushes out new features, updates existing ones, and adds documentation and tutorials that cover how to implement these new changes.
On May 5, 2020, PyTorch released improvements to Tutorials homepage with new content and a fresh usability experience out into the world (see the Twitter thread) for the community. We introduced keyword based search tags and a new recipes format (bite-sized, ready-to-deploy examples) and more clearly highlighted helpful resources, which resulted in the fresh homepage style you see today.
Something Went Wrong
We’re having trouble playing this video.
As the framework grows with each release, we’re continuously collaborating with our community to not only create more learning content, but also make learning the content easier.
The tutorials refresh project focused on re-envisioning the learning experience by updating the UX and updating the learning content itself.
Our 3 major goals for the refresh were:
- Reduce blocks of text and make it easy for users to find important resources (e.g. PyTorch Cheat Sheet, New to PyTorch tutorials)
- Improve discoverability of relevant tutorials and surface more information for users to know about the available tutorial content
- Create content that allows users to quickly learn and deploy commonly used code snippets
And we addressed these goals by:
- Adding callout blocks with direct links to highlight important resource such as the beginner tutorial, the PyTorch Cheat Sheet and new recipes
- Adding filterable tags to help users easily find relevant tutorials; and formatting the tutorials cards with summaries so users know what to expect without having to click in
- Creating a new learning format, Recipes, and 15 brand new recipes covering some of the most popular PyTorch topics such as interpretability and quantization as well as basics such as how to load data in PyTorch
- In summary:
Add callouts with direct links to highlight important resources
Improve discoverability of relevant tutorials and surface more information for users to know about the available tutorial content
Add filterable tags to help users easily find relevant tutorials. Reformat tutorial cards with summaries so users know what to expect
Create content that allow users to quickly learn and deploy commonly used code snippets
Create a new learning format – Recipes. These are bite-sized, actionable examples of how to use specific Pytorch features, different from our previous full-length tutorials
Why We Made These Changes
Now, what drove these changes? These efforts were driven by feedback from the community; two sources of feedback were the UX research study and direct community interactions:
- UX Research study – Earlier in 2020, we conducted a UX research study of our website in collaboration with the Facebook UX Research team to understand how our developer community is using the website and evaluate ways we can improve it to better meet their needs
- In-person events and online feedback – The team gathered community feedback about existing tutorials to help identify learning gaps.
We used these channels of input to fuel revisioning our learning experience.
Rethinking the Learning Experience
Given the feedback from the UX Research study and the in-person workshop, we went back and rethought the current learning experience.
- 3 levels
- Level 1: API docs. Currently these exist and they contain code snippets that provide an easily understandable (and reproducible) example that allows a user to implement that particular API
- Level 2: The missing puzzle piece
- Level 3: Tutorials ideally provide an end-to-end experience that provides users an understanding of how to take data, train a model and deploy it into a production setting using PyTorch. These exist, but needed to be pruned of outdated content and cleaned up to better fit this model
- Realized we were missing something in between, content that was short, informative, and actionable. That’s how recipes were born. Level 2: Recipes are bite-sized, actionable examples of how to use specific PyTorch features, different from our tutorials
What Was the Process
Just as it took a large team effort, this was more of a marathon as opposed to a sprint. Let’s look at the process:
Timeline of the process:
Overall, the project took about 6 months, not including the UX research and prior feedback collection time. It started off with the kickoff discussion to align on the changes. We assessed the existing tutorials, pruned outdated content and decided on new recipe topics and assigned authors. In the meantime, marketing and documentation engineers collaborated with our web design team on the upcoming UI needs, created mocks to preview with the rest of the team and built out the infrastructure.
For logistics, we created a roadmap and set milestones for the team of authors. We held weekly standup meetings, and the team bounced ideas in chat. The changes were all made in a staging branch in GitHub, which allowed us to create previews of the final build. Next, the build process. Many of the recipe authors were first time content creators, so we held a live onboarding session where we discussed teaching mindset, writing with an active voice, outlining, code standards and requirements; and this was all captured in a new set of content creation documentation.
The bulk of the process was spent in building out the content, copy editing and implementing the UI experience.
With the product out the door, we took some time to perform a team retrospective – asking what went well? What went poorly? What can we do better next time? In addition, we continue to gather ongoing feedback from the community through GitHub issues.
Moving forward, we are brainstorming and forming a longer-term plan for the PyTorch learning experience as it relates to docs and tutorials.
Ways to Improve
Looking back on ways we could have improved:
- Timeline – Our timeline ended up taking longer than anticipated because it had been coupled with a previous version release and team members were serving double-duty working on release content, as well as tutorials refresh content. As version release approached, we took a step back and realized we needed more time to put out a more polished product.
- Testing – In software development, if there is an impending deadline, typically testing is the first thing to get squished; however, more focused testing will always save time in the bigger picture. For us, we would always welcome more time for more CI tests of the tutorial build, as well as beta tests of the user experience. Both of these are ongoing works in progress, as we continue to improve the tutorials experience overall.
So what’s next? We understand that this was just one change in a larger landscape of the overall PyTorch learning experience, but we are excited to keep improving this experience for you, our dedicated PyTorch user.
We would like to hear from you about your experience in the new tutorials. Found a tutorial you loved? Tweet about it and tag us (@PyTorch). Ran into an issue you can help fix? File an issue in https://github.com/pytorch/tutorials. We are excited to continue building the future of machine learning with you!