12 ways Skillsoft Percipio Uses AI to Enhance Learning
Article authored by:
- Potoula Chresomales, SVP Product
- Murali Sastry, SVP Engineering
- Troy Collinsworth, Chief Architect
We often get the question: how does Skillsoft’s Percipio platform use AI to improve the user experience?
At Skillsoft, we have incorporated AI into our Percipio platform in meaningful ways. We continually experiment and test so that we can innovate, learn, and share those insights with our community. We’re proud to be at the forefront of providing AI-driven, transformative learning experiences – efforts of which were recognized by Business Intelligence Group’s 2023 Artificial Intelligence Excellence Awards.
We use AI to do four things that benefit users of the platform and users that access Skillsoft content through learning management system (LMS) integrations:
- Personalize learning for each user, helping them to progress from Point A to Point B quickly and efficiently;
- Improve search and discovery, making it easier for learners to find what they need;
- Link skills, roles, and learning together to guide career paths; and
- Generate new content and curation, enabling users to assess their skills as they build confidence and understanding and to curate learning paths for trending topics automatically.
Let’s go over the 12 ways we are using AI today. And for those that like to look under the hood, we’ll share the AI model we used and any lessons learned from the experiments.
AI is used to personalize learning for each user based on their profile, search behavior, learning activity, skill assessments, and role. In addition to personalized learning paths generated from Skill Benchmark Assessments and interests, Percipio uses AI to:
Recommend content based on recent activity (in production) – Percipio learns a user’s behavior, finds other similar users, and recommends content based on “most similar” users' behaviors. This is modeled after how Amazon recommends products (you might have noticed the “people also bought” section). These recommendations personalize the homepage and the automated re-engagement notifications.
AI model used: collaborative filtering, feedforward
Lesson learned: These recommendations are used most often and users who access them spend 59% more time learning indicating value and relevance.
Recommend content in specific patterns and sequences (in QA testing) – Our AI specialists have spent the last quarter testing and training a new collaborative filtering model that recommends content based on common sequences inherent in the way learners have consumed content.
AI model used: collaborative filtering, feedforward, trained on consumption sequences
Subscribe to the Skillsoft Blog
We will email when we make a new post in your interest area.
Thanks for signing up!
Improve Search and Discovery
Organizations like Google and Meta have made big investments in AI models to improve search relevance and have made those available for others to use. We have adopted those models and deployed them in Percipio, and also used AI to improve search in LMS integrations:
Provide targeted search results for users (in production) – Percipio uses Elasticsearch as the search engine and has incorporated Google’s BERT model into search to deliver more targeted and relevant results especially for learners using multiple word search phrases to find something specific.
AI model used: Google’s BERT
Search and recommendation improvements (in progress) – The search and recommendations team is currently upgrading Percipio search from Google's BERT to Meta's (Facebook's) DistilRoBERTa. Our test results showed that the RoBERTa model, which is trained differently, produces more relevant results. This was found to be true for shorter phases and overlapping terms like micro-services vs micro-behaviors. It was also found to be less sensitive to improper cased acronyms like url vs URL.
AI model used: Meta’s DistilRoBERTa
Video descriptions to improve Search in an LMS (in production) – When Skillsoft content is delivered in a learning management system (LMS), it can be more difficult to surface relevant content because our advanced search capabilities, described above, are typically not in use. LMSs often rely on titles and content descriptions to drive relevancy in search results. Skillsoft uses AI to automatically create detailed descriptions for each of the 40,000+ videos to improve search results.
AI model used: GPT-3
Lessons learned: Quality and yield was strongest when both video transcripts and course objectives were used in the model to generate the video descriptions. Next, the team will experiment with ChatGPT.
Quality checks on course descriptions for LMS integrations (in production) – Learning management systems typically rely on course titles and course descriptions to drive search results. Skillsoft content is authored by subject matter experts who create the course descriptions as well. The quality and completeness of these course descriptions varies by author. Occasionally less relevant search and recommendations results were traced to description quality. A GPT model was fine-tuned to rank description quality so poor-quality descriptions could be remediated.
AI model used: GPT-3, fine-tuned
Lessons learned: 10% were identified as inadequate quality. The GPT-3 description generator described previously will be used to remediate lower quality descriptions with higher quality descriptions.
Link Skills, Roles, Learning Together
Skillsoft has a proprietary skills taxonomy, and all content and Skill Benchmarks Assessments are mapped to it. In addition, Skillsoft maps content and Skill Benchmarks to clients’ own taxonomies or industry wide taxonomies (e.g., DDAT, SFIA). The next step in this evolution:
Role-skill framework (evaluation) – GPT and ChatGPT, in conjunction with other services, are being evaluated for effectiveness in managing our stock and custom skills dictionaries. Specific areas of evaluation include normalizing, de-duping, and mapping roles and skills, categorizing the sea-of-skills into meaningful clusters, and auto-curating customer-specific taxonomies and role-skill profiles from job descriptions.
Generate New Content and Curation
AI models are being used to generate content and images and to curate content to meet specific customer needs.
Customer-specific curation of content (in production) – Customers often have specific curation needs that may include aligning content to their own taxonomy or identifying content for inclusion or exclusion from their programs so they can customize the experience to their learners and culture. . An AI model scans and ranks content based on customers’ needs. The customer reviews the content before deploying.
AI model used: GPT-3, fine-tuned
Generate assessment questions (experiment) – Percipio currently assesses skills and knowledge using many different methods including Skill Benchmarks, course assessments, final exams for Aspire Journeys, and mobile flash cards for learning reinforcement. These require a very deep pool of assessment questions to enable each user to assess and reassess without seeing the same questions. Skillsoft is using AI to generate assessment questions to be used for these different platform features.
AI model used: GPT-3 fine-tuned and GPT-3.5 zero-shot
Lessons learned: The GPT-3 quality was mediocre: 50% were usable, 40% needed edits and 10% had to be discarded. The GPT-3.5 zero-shot performed worse, yielding only ~21% acceptable questions. We are exploring ChatGPT and post generation reflection (PGR) techniques which are showing promise in improving yield.
Curation of learning paths (experiment) – Curation of learning paths is typically done by expert curators in each domain and based on sound instructional design. We've collected a tremendous amount of data about how users actually learn and all of this data will be used to automatically curate learning paths for skills and roles.
AI model used: Collaborative-Filtering, GPT
Chatbot for learners (in production) – Percipio has a chatbot currently deployed to select customers that provides Level 1 customer support to learners answering the most common questions. This will be available to all customers in 2023.
AI model: Salesforce Einstein Chatbot
Lesson Learned: For participating customers, 45% of the level 1 common questions were answered by the chatbot.
Create unique images for blog posts and course thumbnails (experiment) – We used DALL-E 2 to create the image for this blog post and plan to experiment with creating eye-catching thumbnail images that will engage and energize the learner. These will appear in Percipio and through LMS integrations.
AI model: DALL·E 2
ChatGPT early use
Skillsoft’s AI team is already using ChatGPT internally to improve our productivity and accelerate our own AI efforts. The AI team has been utilizing GPT-3 and ChatGPT to generate code to implement Percipio features. These features include enhancements to the AI collaborative filtering model and improved search features (auto-suggest and type-ahead). It has been found to be quite effective. The team is exercising great care to ensure we are not violating copyrighted code.
Using these models, the possibilities are endless.
AI model: ChatGPT and GPT-3