When selecting a Digital Asset Management system (DAM), today’s marketing leaders are looking for Artificial Intelligence (AI)-enabled features. Auto tagging, sometimes known as smart tagging, is at the very top of their list.
Tagging based on AI enables assets to be tagged automatically upon ingestion. This smart tagging service creates tags or metadata without human intervention, saving marketers time while reducing the potential for human error. Most importantly, tagging assets or adding metadata to assets using AI improve the searchability and findability of those assets in the DAM, which promote asset reuse, and increase the ROI (return on investment) of a DAM system.
Today, most DAM systems are equipped with smart tagging services, which are based on standard API services, such as Microsoft, Google, Amazon, Clarifai and others. These tagging services work on visual recognition, meaning they can describe images or videos based on what you can see using a generic vocabulary.
More advanced DAMs also have the ability to plug in learned AI using machine-learning models that can customize tagging to align with a specific company, market, industry, etc. Additionally, these models teach the DAM system to use company and business-specific vocabularies for tagging.
Let’s look at an example:
(Credits to Pixabay nastya_gepp)
A standard tagging service will use generic tags to describe this image, such as: person, human, woman, face, smile, cellphone, clothing, indoor, coffee, cup, photography, sitting, drink, restaurant, cup, etc.
A trained smart tagging service, like Aprimo AI, can provide tags that are more specific to a line of business. Business-specific tagging models can be taught to identify brands, products, objects and people. For example, if you are in the business of selling jewelry, you may want tags specifying the exact watch or bracelet the woman is wearing in the picture. On the other hand, if you’re a clothing retailer, you may want the smart tagging to identify that the woman is wearing a specific striped shirt and that the model is Helena.
Trained smart tagging will undoubtably improve your content discoverability. The bigger question is: “Is it worth the investment for you”?
Should I invest in trained AI?
To answer this, you should evaluate the following four questions:
1: Do you need business-specific vocabulary in your tags?
Machine-learned models are valuable if you need the ability to tag and search assets using business-specific vocabulary. If you’re happy with the generic tags the API services are offering today, then you likely don’t need anything more.
2: Do you have the scale to make AI worth it?
To train a learned AI model, you need between 5 and 100 images per tag. If your tags are visually closer, you’ll need more training data. If you have fewer tags and they are visually dispersed, you’ll need less training data.
Because the automated tagger competes against real human effort attributing the tags manually, you also need to identify how much time and completeness of tags you’ll gain by introducing learned AI. If you only have 1000 images in your database or only upload 100 new images per month, then you probably don’t have enough mass because training for 20 tags would take more time, effort, and content than you have.
The number of users who need to upload and tag assets in the DAM is another key factor to determining if a learned AI model makes sense for your business. If you have numerous users uploading assets and tagging them, then the benefit from an automated tagging service is greater than if you only have a couple of disciplined content administrators in charge of your tagging. A trained AI service ensures that the quality and consistency of tags are always the same, which can’t be said when many different people are uploading content.
3: Do you have a controlled vocabulary?
Do you have a controlled vocabulary (i.e. a standard set of terms that everyone in your organization uses), or are you willing to work with one? If so, trained AI may be a great fit for your business. When training AI models, having a controlled vocabulary is critical because the correctness and accuracy of tags:
- Enable a high-performing content model
- Reduce ambiguous tags
- Ensure consistency in tagging and more accurate search results
4: Are your tags visual?
Can you visually identify the characteristics you would like to see tagged through the AI service? For instance, I can teach the model to recognize a brand and model of car because there is a visual difference between car brands and models. In contrast, I cannot teach the model to add tags based on what’s under the car (e.g. the axle, brakes, transmission, etc.) because it has no impact on how the car looks visually. Ultimately, you can train your AI model to be aware of only those things you can see with your eye.
Considerations for your tagging strategy
It’s important to note that this blog only covers part of what you need from a tagging or metadata attribution strategy. Other things you should consider:
- What about non-visual tags? In our example, the picture could have been taken in “Nashville“ or at a “Starbucks”. For these non-visual tags, we need to depend on other tagging or metadata attribution mechanisms.
- Not all automated asset enrichment is driven through tags. DAMs store a lot of searchable metadata on your assets (filename, fields and classifications, embedded metadata, textual content, etc.) that can enrich assets in other ways.
- You can improve findability of assets by adding metadata or tags in non-AI ways, like logic or metadata templates, which can greatly reduce manual effort without having to invest the time or money in AI.
At Aprimo, we believe learned smart tagging can have big benefits with only a little bit of effort. But it may not be for everyone. What are your thoughts on AI and DAM? We’d love to hear your feedback on social or in the comments section!