How AI can learn to manage your Evernote
Evernote is a popular note-taking software which is used by more than 200 million people around the world. However, the app provides only very basic support for organizing data, so users have to expend effort in maintaining its structure. Fortunately, Artificial Intelligence can automate many of these tedious housekeeping activities. This improves the user experience and allows users to devote their attention to more valuable activities.
What is Evernote?
Evernote is the most popular and most powerful note-taking tool on the market. Its tagline is “Remember Everything” – you just drag your document into the app, and Evernote takes care of it. People use Evernote for all types of personal and professional information and use its powerful search function to retrieve it when when they need it.
Evernote stores information in notes. There are only three simple mechanisms for structuring notes called tags, notebooks and stacks. Notes that belong to a topic such as Finances or Projects can be stored in a corresponding notebook, and notebooks can be stacked on top of each other. Alternatively, notes can be given tags to denote the topics that they are associated with. Evernote provides a search function for note contents and for metadata such as author or date created. Thus, a user can search for all the notes with the tag Projects or those that were created by Susan more than a week ago.
Evernote’s strength is also its weakness
Evernote is a universal tool, so it makes no assumptions about the type of information it is given or the application from which it arises: a document might be anything from a recipe for chocolate chip cookies to notes for a PhD thesis.
For this reason, Evernote cannot provide support for organizing information in an application-specific manner. For example, the popular Getting Things Done (GTD) time management system is a common application with Evernote users. It requires future events to be organized according to their due date. Since Evernote does not have this ability, users have to emulate it as best they can by creating and assigning tags and notebooks. Thus, they might decide to create 12 notebooks (or 12 tags) titled January_tasks, February_tasks etc. Each time they create a note describing a new task, they have to remember to move it to the appropriate notebook (or assign the appropriate tag to it). If they forget to do so, or make a mistake, then the note is lost for the GTD system – with potentially serious consequences in the future.
This means that if you want to maintain a particular organizational structure for your information, you will need to perform lots of repetitive actions. For example, every time you scan an invoice, you have to add a tag such as Taxes or Invoices to it. This is tedious and time-consuming, and if you forget to do it, your organizational system immediately breaks down.
Evernote’s general-purposeness is thus both a strength and a weakness: You get maximum flexibility, but at the cost of a lot of manual effort.
Delegating actions to Filterize
As Evernote users ourselves, we were acutely aware of this weakness and undertook to compensate it by building a software tool called Filterize.
Filterize is a web-based SaaS that links to an Evernote account and can emulate certain user actions such as adding a tag to a note or moving a note to a particular notebook. The user can define rules that instruct Filterize what action to take, if a note meets a given condition.
For example, a user might want all notes with the tag Invoice to be stored in the notebook Finances. This is how Filterize displays the corresponding rule in the control panel:
From the moment this rule is activated, whenever the user modifies a note with the tag Invoices, Filterize will move it to the notebook Finances automatically, without the need for any intervention on the part of the user.
Since actions such as adding tags or moving notes between notebooks occur very frequently, rules such as this have the potential of saving the user a large amount of work. In addition, they eliminate the risk of errors or of being forgotten entirely. Some Filterize users have created more than 100 rules, which save them thousands of manual actions every month.
Rules are particularly useful for building dashboards in Evernote, because they are constantly in need of updating. Filterize refreshes some of our users’ dashboards automatically more than 2,000 times per week, without the need for any intervention on their part whatsoever.
However, in order to apply these rules effectively, users must have a well-structured “filing system” in their minds that they can translate into the syntax required by Filterize. This is true for power users, who may have thousands of notes, but is not necessarily the case for a typical user.
Adding Artificial Intelligence to Filterize
We therefore searched for a way to make the benefits of Filterize rules available to mainstream users. These are people who use Evernote in a less structured way and who are also perhaps hesitant about formulating rules.
The key to achieving this was the realization that the users’ intentions were actually already implicitly visible in their Evernote databases: the metadata and locations of notes in notebooks are a mirror image of their “mental filing systems”. Artificial Intelligence algorithms can analyse this information and correlate it with the contents of individual notes in order to reconstruct the actions that users applied to them. These actions can then be converted into rules which Filterize can then execute automatically without the user even having had to formulate them explicitly.
This realization has allowed us to develop our Evernote virtual Assistant (EvA), which we have recently integrated into Filterize. EvA transforms Evernote from a passive repository into a proactive assistant: it learns what you repeatedly do and then offers to do it for you.
The following example occurs immediately after the user has moved a note called Amazon Order – 2019-01-02 to the notebook Finances. This note has the tag Invoice and an attached PDF file. Based on existing notes in the database, EvA has discovered two highly probable rules that might describe this action.
In the first case, EvA has learned that the user has previously moved notes with the tag Invoice and a PDF attachment to the notebook Finances. There are 18 notes in the database with a PDF attachment and the tag Invoice, and 100% of them are located in the notebook Finances.
The second case is similar – except that there was no PDF attachment. The database contains 25 notes with the tag Invoice, and 92% of these are located in the notebook Finances.
The user may now select which (if any) of these two rules Filterize should add to its library and apply to notes automatically in the future.
Using AI to detect patterns
EvA uses an Artificial Intelligence technique known as Frequent Pattern Mining (FPM) to find patterns and correlations in the notes database that reveal how the user structures their information. Frequent pattern mining is a common technique for predicting human behaviour. It is often used in the retail and e-commerce sector to discover facts such as, “people who buy beef patties and tomato ketchup will probably also buy hamburger buns.”
The result of this analysis is a set of rules that might have been used to structure the database. EvA awards each rule a score based on various criteria such as accuracy, complexity and frequency of occurrence. This set is initially very large: for a typical Evernote user, it might contain millions of alternatives. EvA’s second task is therefore to determine which of them are most helpful to the user.
Finding the best rules
In the first elimination step, EvA discards all rules whose accuracy is less than 80%. In other words, there should be no more than 20% exceptions to it in the database. This eliminates the majority of the alternatives, but for an average user might still result in hundreds of thousands of possibilities.
In the next step, EvA computes the Pareto frontier. This is the set of rules whose scores cannot be beaten by another with respect to all evaluation criteria. This step will also eliminate the vast majority of the remaining rules, but might still leave several hundred for a typical Evernote user.
In the third step, EvA selects the rules which it thinks will most closely match the format of rules chosen by the user in the past and which will be the most beneficial if applied.
EvA then displays a ranked list of the top-rated rules to the user, who can then select the most appropriate one. In the image in the previous section, this list contains only two alternatives; in more complex cases, it may contain up to 20. EvA adds the selected rule to a library, and Filterize will apply it automatically any time the user activates it, for example by creating a new note.
Of course, users can view all the rules proposed by EvA at any time and activate/deactivate, modify or delete them as desired.
The number of samples EvA needs in order to detect a plausible rule is surprisingly small. For example, a user may have several notes containing the keywords Cake and Pre-heat oven in the notebook Baking. EvA only needs five such notes for the corresponding rule to be highly plausible.
Typical use cases for EvA
Here are four typical use cases for EvA:
- New users. For most newly registered Filterize users, EvA can immediately suggest up to 15 useful rules on the basis of tags alone. (Analysing the contents of the notes consumes a lot of processing power and takes some time to complete.)
- Archiving scans. One common application of Evernote is to archive scanned documents. Most scanners can be configured to send scanned documents directly to a user’s Evernote account, and Evernote is able to read the content of the scanned images. If a user has filed scanned documents containing the same keywords to the same notebook consistently, EvA can recognize this and offer to do it automatically in the future.
- Setting reminders. In a work environment, one person might assign tasks to another via a shared notebook. EvA can learn to add a reminder to these notes automatically.
- Tagging web clips. Evernote provides a web clipper function for copying web pages into its database. For recurring topics, EvA can learn to read the clipped content and assign an appropriate tag to the note. For example, EvA could learn to automatically add the tag Hobbies to every clip containing the keywords Knitting and Pattern.
The more extensive and consistent an Evernote user’s “mental file system” is, the more he or she can benefit from EvA, since their notes will contain a greater number of identifiable patterns. In the ideal case, EvA will be able to automatically manage every single note that is consistently with previous occurrences.
Benefits and future potential
Filterize rules offer many benefits to the Evernote user. They save time, reduce the risk of misplacing or losing notes and make note management much more convenient. With the addition of EvA, users no longer even need to specify rules explicitly, but just choose which of the various candidates detected by the AI most accurately corresponds to their intent.
We believe that we have only just begun to realize the potential that AI has to offer for Evernote. For example, we are currently developing algorithms that will offer the following features:
- Recover lost notes. Notes that have tags missing are at risk of being lost. EvA will be able to use correlations between note contents and tags to suggest replacing a forgotten tag. For example, a user may have several notes containing the keywords Restaurant and Stars that are all tagged with Reviews, except for one. This note will therefore not show up when the user searches their Evernote for reviews. EvA will be able to suggest repairing that note by adding the missing tag.
- Tidy up the database. EvA will be able to recognize irregularities in the archive and help to repair it. For example, it is possible to find notes that are probably located in the wrong notebook and to suggest the notebook(s) that it is was most likely intended for.
- (Re)structure the archive. Some people use Evernote as an Everything Bucket – a repository into which they throw all their documents, paying little or no attention to organization. In this case, EvA might be able to propose a logical structure for the notes using notebooks and/or tags. The necessary AI tools are available, but it still unclear how effective it could be for an unmanaged Evernote archive.
EvA is now available for beta testing
We have released EvA as a public beta test version. If you already have a Filterize account, you can start using it immediately – simply activate EvA in your Filterize account. Other Evernote users need to register with Filterize first.
Filterize is a cloud service that acts as your personal Evernote assistant. Tell the software how you organize your notes or just let its Artificial Intelligence learn how to do it automatically. Filterize will then manage your notes in the background, eliminating repetitive tasks, avoiding errors and saving you time.