Create Testing Framework for NLP applications

[nickumia]

Assuming a purely textual model, how can you guarantee that the NLP agent understands what it knows? I'm substituting understanding with comprehension in this derivation because typically the test of a machine understanding is how well it's able to regurgitate text back to humans. I think there's more to understanding than just comprehension, but for the sake of "knowing what you know", I think comprehension is a solid foundation.

Key ways of testing Reading Comprehension:

• Making connections:
  • Can the system make accurate relationships between elements inside its model?
  • Can the system make accurate relationships between elements in two different sources?
• "Think-Pair-Share":
  • Can the system talk about what it knows and have a discussion about it's knowledge?
  • This one is hard to do in a system. It could be loosely related to adversarial and peer machine learning algorithms... but I feel like that's a stretch..
• Summarization:
  • Can the system summarize what it knows?
  • This one probably needs to be optimized. I honestly don't have a good idea of how this is done in the field currently.
• Visualization:
  • Can the system create a visual of what it knows?
  • For the AIs that do text-to-image and image-to-text, I feel like this has the most value.
  • The problem is, how do you judge the picture? The whole idea of humans judging artwork essentially boils down to "beauty is in the eye of the beholder". To that logic, if the AI thinks the picture represents the text, who are we to tell it it is wrong? There are culturally accepted standards of accuracy, but I wouldn't say any of them are universal.
• Make Predictions:
  • Can the system extrapolate more information than what exists in its knowledge base?
  • This is a really valid one, but just like the discussion one above, it's hard to automate and validate.

References:

• https://www.hmhco.com/blog/reading-comprehension-assessment-strategies
• https://coxcampus.org/the-5-critical-components-of-reading-and-how-to-assess-student-progress/

[nickumia]

Thoughts from talking to a friend with history in the education system:

• Reading comprehension tasks are different at different stages of development
• The key in early stage is to make real world connections.
  • What does this mean for an AI? TBD
• (early stage) The most beneficial way to check for comprehension is through discussions.
  • If we talk to AI enough, is the whole point to convince us that it understands? (i.e. the turing test) I don't think so.
• Can an AI build "personal experience"?
• How to recognize the value proposition of learning? What is important to learn and why.
• AI doesn't have its own human experience, but it learns about human experience by processing data.

Thoughts on reading Reading Comprehension Ability Test-A Turing Test for Reading Comprehension (2019):

• I would fail that year 5 test... those questions are terrible in general haha
• The year 9 test just turned me off visually. It was only looking at it the third time that I realized it was about the BBT (a super awesome show haha) P.S. Is it bad that this questions seem easier?
• The reading CAT is essentially reaffirming that the Turing test is accurate. It sticks to "if it responds like a human, it must be human" logic 😒
• Levels of CAT
  • CAT Level 1 – identifying the facts presented in the article.
  • CAT Level 2 – identifying the facts presented in the article by applying common knowledge of equivalent terms in the context.
  • CAT Level 3 – perform inference with the knowledge expressed in the article or together with the related common knowledge.
  • CAT Level 4 – identifying sentiment (mood, tone, genre), intent and main thought (idea).
• Seemed about how I thought people were thinking about reading comprehension

[nickumia]

Integrating AI in medical education: embracing ethical usage and critical understanding (2023)

• Uses of AI in education:
  • Learner-oriented AI: help students
  • Instructor-oriented AI: help teachers
  • Institution-oriented AI: help administration
• Lack of "knowledge depth" and "critical thinking" in higher level education submissions from AI-generated content
• "Universities should implement strict policies and guidelines for the ethical use of AI applications"
• "Foster a culture of academic integrity"
• "Balancing their use [of AI] with their own unique ideas and voices"
• Turnitin (oh, ol' reliable) does AI-generation scoring along with plagarism.
  • What are the characteristic of AI-generated text?
• "large language model-powered chatbots to answer common medical questions for students and patients. Chatbots can also be used by students to practice their clinical communication skills."
  • This seems counter-intuitive... the AI generates information that lacks knowledge depth, but it is good enough to answer "common medical questions" and "practice communicating information"
• I like the funding statement, conflict of interest statement and publisher's note 😌
• Key references of interest:
  • https://www.nesta.org.uk/report/education-rebooted/
  • https://www.frontiersin.org/articles/10.3389/feduc.2023.1183162/full

[nickumia]

Although outside the scope of this endeavor, getting a qualitative understanding of how good AI-generated text is is also important. Some good reference material for ChatGPT-generated content vs. Experience Teacher-generated content:

• Can ChatGPT make reading comprehension testing items on par with human experts?

[nickumia]

One more article using "Xplainable AI" in COVID-19 prediction:

XAI Integration: To enhance interpretability and explainability, eXplainable Artificial Intelligence (XAI) techniques, specifically the SHapley Additive exPlanations (SHAP)algorithm, are integrated into the model. The SHAP algorithm helps to quantify the contribution of each feature towards the final prediction, providing insights into the importance and influence of different factors

[nickumia]

Current thoughts on the issue:

• Validation is only as good as testing.
• Testing is not a reliable method of determining credibility and knowledge.
  • A better method is required.
  • However, in the meantime, some testing is required
• As I develop new algorithms and programs, I'll re-evaluate this and add things as they appear relevant.