Markup is an open-source annotation tool for transforming unstructured documents into a structured format that can be used for ML, NLP, etc.
Markup learns as you annotate in order to speed up the process by suggesting complex annotations to you.
There are also a few different in-built tools, including:
- A data generator that helps you to produce synthetic data for training the suggestion model
- An annotator diff tool that helps you to compare annotations produced by multiple annotators
It's still very much a work in progress (and the documentation is severely lacking), but the ultimate goal is to make a tool that's as useful as https://prodi.gy/, without the $400 price tag.
This looks incredible! I’ve been following doccano for awhile but they were still working on active learning. Will you be adding an open source license like MIT?
Beautiful. So many annotation tools focus on "text classification" which assumes you've already got segmented samples. In the real world of documents that's a whole challenge in itself.
Another challenge is that sometimes you're working with PDFs and that means not only ingesting but also displaying. The difficulty is in keeping track of annotations and predictions across the PDF<->text string boundary, both ways.
There are understandably even fewer solutions to that problem because it's a harder UI to build.
Much appreciated! That's true, and lots of the tools that do feature text annotation can be quite restrictive in that they don't allow you to add attributes / repeatedly annotate the same span of text.
Support for PDFs and other doc types is definitely on the backlog, but I keep holding off due to the challenges you mentioned.
Really nice tool - thanks for making this! What is your plan for this? Is this a side-project that you'll potentially turn into a business, or is this just a hobby on the side of your full-time job?
Just asking because I think many folks would be happy to pay to support a small ISV to ensure it's long-term sustainability. Not via donations, but actual pricing.
Thanks for your kind words! It's just a hobby project that I work on alongside my full-time job right now, and to be honest, I'm still trying to figure out a plan. My intention is to keep the core functionality free forever, but I could definitely see a future where there are premium collaborative features or some cost for training custom suggestion models, for example.
If you're one of those folks who would consider supporting a tool such as this, do you have an idea in mind as to what sort of features you'd be willing to pay for?
I'm in the market for a tool like this. At the moment I'm using Prodigy but interested in other options. Features that I'd be willing to pay for (or rather my employer):
1 team functionality with multiple user accounts
2 easy to use workflow for double annotation where each text is annotated by exactly two annotators. The software should make sure that a text is never shown to more than 2 annotators and never shown to the same annotator twice
3 make it easy to review the 2 versions and solve conflicts
4 smarter alternative to review would be a warning system that identifies annotations that may have errors (because a model trained on the other data predicts a different result) and automatically flags it for review by another annotator
5 stats on the annotators: speed, accuracy, statistics on how frequently they assign different labels to detect potential misunderstandings of the annotation schema
6 GUI with overview of all annotation datasets, with stats like % finished annotating (with stages for double annotation and review), the types of annotation done, frequencies of labels to detect imbalances
7 functions to mass-edit the annotations, like renaming or removing an entity type
Another thing I'd be interested in is some integration with a third party annotation provider. There are companies that offer annotation as a service and it's also available on Google Cloud and AWS. Having that integrated into an annotation tool would make it very easy to get large amounts of well annotated training material.
But finally, and much more importantly: The workflow for annotators has to be perfected first, so they can work as efficiently and consistently as possible. Getting this right is more important to me than any of the other features I listed.
Mind if I ask what sort of team features you make use of with Prodigy? Are there any aspects you feel are lacking? Initial thoughts are that it'd be helpful for teams to be able to set group annotation goals, share docs / annotations / configs, view ongoing sessions, assign annotators to sessions, and view stats on each annotator (as per point 5).
> The software should make sure that a text is never shown to more than 2 annotators and never shown to the same annotator twice
For this I plan to let teams set the threshold for the number of documents that should overlap and the number of annotators a text should be shown to. In some situations it could be useful for there to be some % of overlap for all annotators to help determine the inter-annotator agreement across the entire team.
> The workflow for annotators has to be perfected first
Totally agree. My biggest concern is building out the above on top of an inefficient workflow. That's one of the primary driving forces behind the current re-write of the tool.
Love the smart flagging, mass-edit, and integrated provider ideas!
I use these team features in Prodigy: I start annotation sessions with different session_id and with the feed_overlap flag. I run Prodigy from an EC2 instance that annotators connect to.
The Prodigy team is working on a new version called Prodigy Scale with more team features. I'm looking forward to that release! For now it feels like a hack to use Prodigy in a team.
Inter-annotator agreement is key! You could consider making that highly visible in your tool. It's something that every team should measure and strive to maximize.
For developers who use spaCy in production (like me), I imagine it would be very hard for your tool to come out on top of Prodigy. But there could be an opportunity with price-sensitive hobby users or devs who use a different NLP library.
I'm a hobbyist in this space, but my sense is that there is definitely a Github-like platform you could eventually build towards (in the distant future of course). Dataset annotation and workflow is a big area, covering text (which you have the basics of), images, audio, and so many others even within those broader categories (eg: MRI images vs. other types of images, PDF text vs. HTML text, etc).
As psimm mentioned on this same thread branch, there are all sorts of workflows and features that the relevant business users need, and the best way is to actually get 5 of them on a separate screenshare call to better understand their specific workflows. You can then abstract those examples given your own knowledge of the space into entry-tier and premium-tier features, and price accordingly.
Do beware of feature explosion, because many of those discussions will be a list of desired features which, while nice in the long run, don't help you get you a stable next stage in a sustainable manner.
At this stage, find one thing your tool can do exceedingly well for a number of users (ideally who will pay for that).
Interestingly, the Github guys also started out as a side-business while working full-time jobs.
I think you're off to the great start, nice and intuitive tool for text-annotating, so get that flow right before expanding into others. You definitely have something pretty cool here. Congrats again on the launch!
Looks like an interesting project. Would you have some kind of a summary of the methodology you're using for the annotation suggestions? What kind of learning, and which kinds of features?
Just to preface this summary, it's all a bit hacked together at the moment, and I'm in the process of rewriting the tool from scratch so this description is likely to change.
To generate the suggestions there's an active learner with an underlying random forest classifier, that has been fed ~60 seed sentences [1], to classify positive sentences (e.g. contains a prescription) and negative sentences (e.g. doesn't contain a prescription).
All positive sentences are fed into a sequence-to-sequence RNN, that has been trained on ~50k synthetic rows of data [2] which maps unstructured sentences (e.g. patient is on pheneturide 250mg twice a day) to a structured output with the desired features (e.g. name: pheneturide; dose: 250; unit: mg; frequency: 2). The synthetic data was generated using Markup's in-built data generator [3].
The outputs of the RNN are validated to ensure they meet the expected structure and are valid for the sentence (e.g. the predicted drug name must exist somewhere within the original sentence).
All non-junk predictions are shown to the user who can accept, edit, or reject each. Based on the users' response, the active learner is refined (currently nothing is fed back into the RNN).
> Document to annotate - The document you intend to annotate (must be .txt file)
Any thoughts on supporting additional file formats? I'm actually interested in annotating HTML files / web pages. It would be great if I could browse for a local HTML file or enter in a URL and the HTML content would be rendered for it to be annotated using the entities.
For sure! I've just tested it locally and HTML annotation is possible with only a few minor changes. I've just been overly restrictive by limiting it to text files (primarily to avoid PDFs and MS docs that have some additional challenges). Will deploy the updates later today.
Love the idea of being able to enter a URL to retrieve the HTML for annotation.
The primary competitor would be Prodigy IMO, although that's not FOSS. Several open-source alternatives look interesting, such as Docanno, Brat, and Webanno.
In terms of what makes Markup better, I'd say it's mainly that:
- It's quick and easy to annotate with because it suggests complex annotations to you and will learn based on your feedback. Plus it allows for integrated querying of external ontologies, so if you highlight a span of text then it'll suggest matches from the ontology, ordered by similarity.
- It's flexible and configurable. For example, you'll soon be able to provide a small number of example sentences which will generate synthetic data to train a model for you (without any code), and those model weights can be saved for future usage.
- It includes a growing selection of in-built tools, such as a tool for statistically comparing two sets of annotations / annotators.
- It's possible to capture both simple and complex annotations. Most existing tools focus on classification or named-entity recognition, without giving users' the option of capturing attributes nor allowing them to annotate the same span of text repeatedly. Being able to do so enables the production of datasets for more complex use cases.
That's fantastic. I was about to start a project in October building something that's almost completely there already, for a specific use case (annotation of therapy sessions).
26 comments
[ 3.3 ms ] story [ 62.9 ms ] threadMarkup is an open-source annotation tool for transforming unstructured documents into a structured format that can be used for ML, NLP, etc.
Markup learns as you annotate in order to speed up the process by suggesting complex annotations to you.
There are also a few different in-built tools, including:
- A data generator that helps you to produce synthetic data for training the suggestion model
- An annotator diff tool that helps you to compare annotations produced by multiple annotators
It's still very much a work in progress (and the documentation is severely lacking), but the ultimate goal is to make a tool that's as useful as https://prodi.gy/, without the $400 price tag.
Another challenge is that sometimes you're working with PDFs and that means not only ingesting but also displaying. The difficulty is in keeping track of annotations and predictions across the PDF<->text string boundary, both ways.
There are understandably even fewer solutions to that problem because it's a harder UI to build.
https://github.com/allenai/pawls
Support for PDFs and other doc types is definitely on the backlog, but I keep holding off due to the challenges you mentioned.
Just asking because I think many folks would be happy to pay to support a small ISV to ensure it's long-term sustainability. Not via donations, but actual pricing.
If you're one of those folks who would consider supporting a tool such as this, do you have an idea in mind as to what sort of features you'd be willing to pay for?
But finally, and much more importantly: The workflow for annotators has to be perfected first, so they can work as efficiently and consistently as possible. Getting this right is more important to me than any of the other features I listed.
> team functionality with multiple user accounts
Mind if I ask what sort of team features you make use of with Prodigy? Are there any aspects you feel are lacking? Initial thoughts are that it'd be helpful for teams to be able to set group annotation goals, share docs / annotations / configs, view ongoing sessions, assign annotators to sessions, and view stats on each annotator (as per point 5).
> The software should make sure that a text is never shown to more than 2 annotators and never shown to the same annotator twice
For this I plan to let teams set the threshold for the number of documents that should overlap and the number of annotators a text should be shown to. In some situations it could be useful for there to be some % of overlap for all annotators to help determine the inter-annotator agreement across the entire team.
> The workflow for annotators has to be perfected first
Totally agree. My biggest concern is building out the above on top of an inefficient workflow. That's one of the primary driving forces behind the current re-write of the tool.
Love the smart flagging, mass-edit, and integrated provider ideas!
The Prodigy team is working on a new version called Prodigy Scale with more team features. I'm looking forward to that release! For now it feels like a hack to use Prodigy in a team.
Inter-annotator agreement is key! You could consider making that highly visible in your tool. It's something that every team should measure and strive to maximize.
For developers who use spaCy in production (like me), I imagine it would be very hard for your tool to come out on top of Prodigy. But there could be an opportunity with price-sensitive hobby users or devs who use a different NLP library.
As psimm mentioned on this same thread branch, there are all sorts of workflows and features that the relevant business users need, and the best way is to actually get 5 of them on a separate screenshare call to better understand their specific workflows. You can then abstract those examples given your own knowledge of the space into entry-tier and premium-tier features, and price accordingly.
Do beware of feature explosion, because many of those discussions will be a list of desired features which, while nice in the long run, don't help you get you a stable next stage in a sustainable manner.
At this stage, find one thing your tool can do exceedingly well for a number of users (ideally who will pay for that).
Interestingly, the Github guys also started out as a side-business while working full-time jobs.
I think you're off to the great start, nice and intuitive tool for text-annotating, so get that flow right before expanding into others. You definitely have something pretty cool here. Congrats again on the launch!
To generate the suggestions there's an active learner with an underlying random forest classifier, that has been fed ~60 seed sentences [1], to classify positive sentences (e.g. contains a prescription) and negative sentences (e.g. doesn't contain a prescription).
All positive sentences are fed into a sequence-to-sequence RNN, that has been trained on ~50k synthetic rows of data [2] which maps unstructured sentences (e.g. patient is on pheneturide 250mg twice a day) to a structured output with the desired features (e.g. name: pheneturide; dose: 250; unit: mg; frequency: 2). The synthetic data was generated using Markup's in-built data generator [3].
The outputs of the RNN are validated to ensure they meet the expected structure and are valid for the sentence (e.g. the predicted drug name must exist somewhere within the original sentence).
All non-junk predictions are shown to the user who can accept, edit, or reject each. Based on the users' response, the active learner is refined (currently nothing is fed back into the RNN).
[1] https://github.com/samueldobbie/markup/blob/master/data/text...
[2] https://raw.githubusercontent.com/samueldobbie/markup/master...
[3] https://www.getmarkup.com/tools/data-generator/
Any thoughts on supporting additional file formats? I'm actually interested in annotating HTML files / web pages. It would be great if I could browse for a local HTML file or enter in a URL and the HTML content would be rendered for it to be annotated using the entities.
Love the idea of being able to enter a URL to retrieve the HTML for annotation.
In terms of what makes Markup better, I'd say it's mainly that:
- It's quick and easy to annotate with because it suggests complex annotations to you and will learn based on your feedback. Plus it allows for integrated querying of external ontologies, so if you highlight a span of text then it'll suggest matches from the ontology, ordered by similarity.
- It's flexible and configurable. For example, you'll soon be able to provide a small number of example sentences which will generate synthetic data to train a model for you (without any code), and those model weights can be saved for future usage.
- It includes a growing selection of in-built tools, such as a tool for statistically comparing two sets of annotations / annotators.
- It's possible to capture both simple and complex annotations. Most existing tools focus on classification or named-entity recognition, without giving users' the option of capturing attributes nor allowing them to annotate the same span of text repeatedly. Being able to do so enables the production of datasets for more complex use cases.