Leveraging new technologies
The term artificial intelligence brings to mind images of popular science fiction computers or voices, like the HAL 9000, from "2001: A Space Odyssey." Computer programs or systems designed to perform or emulate tasks traditionally only possible by humans is broadly known by this name. It encompasses speech recognition, decision-making, visual perception, and translation between languages. It is not a new concept, but it is still emerging as a widely available and immediately relevant technology. The chatbot proposed in Strand C is a kind of artificial intelligence-driven software that understands natural language input and attempts to respond is if it is a real person. (Reshmi & Balakrishnan, 2016) We see the automatic chat greetings in the bottom right corner of countless web pages that include shopping, customer support, and personal assistants, as just a small sampling.
Exploration of AI in education has just barely begun. Intelligent Tutoring Systems can now be found in online textbooks and mobile apps. Amazon’s Echo devices are entering classrooms to be queried by eager students. Data miners are searching for ways to access and use the vast stores of data collected. Woolf, et al., identified five areas for future research at the intersection of artificial intelligence and education. One of these was the idea that each learner could have their own mentor or guide. (Woolf, Lane, Chaudhri, & Kolodner, 2013) Science education is just one of the environments where the concept of AI-as-mentor can be put to the test.
The HAL 9000 didn't originally have a calm, even-tempered voice. It was originally recorded with human-like emotions that set a standard for AI responses today. (Flahive, 2018)
open source: software that is free and publicly published such that any developer may copy, alter, and customize the software to create a new iteration of it.
Natural Language Understanding
In 2016, a biology professor at Georgia Tech designed an intelligent teacher assistant chatbot application that he used to automate responses to frequently asked questions from his students. Students could largely could not distinguish between the computer response system and a living teacher's assistant. (McFarland, 2016) Also known as natural language processing, the term natural language understanding encompasses computer programs or algorithms designed to allow automatic interpretation of written or spoken language. The development of software that performs well in language understanding and manipulation can be utilized in a vast array of practical applications. We use voice-mediated versions of this software when we talk to Siri, Alexa, Google, and Cortana. Natural language is the pillar of language translation services and the ability of an end-user to speak or write to a device and be able to receive a response as a result of that input. Out of many smaller natural language efforts, there have grown a handful of major platforms that allow developers to use the services. Most have a free pricing tier, along with premium services at a cost.
Braun, Hernandez-Mendez, Matthes, & Langen (2017) developed techniques to measure each of the major platforms against one another. They compared how well each engine performed in identifying structured information from unstructured, everyday language. Behind the user-interface, natural language understanding or processing involves categorizing the components of language and recognizing patterns. The quality of these programs hinges on how well the code is able to represent and distinguish the nuances of human speech and writing. They found that Microsoft's LUIS performed very well, but that in some cases RASA was sufficiently comparable in most areas, while being free and open-source. Each of the platforms represent impressive achievements of computer programming.
This is also a broad, widely applicable term. It refers to algorithms that rely on a training model in which humans "teach" the algorithm to identify examples of a given entity. Facial and image recognition can fall into this category, along with complex data analysis techniques.
As part of a project this past summer, a group of students examined how we could put together natural language tools to evaluate interactions in an online LMS. Specifically, we used machine learning tools to train a model of sentiments and thoughts to look for in discussion board interactions.
intelligent tutoring system: computer programs designed to guide students through an interactive learning process.
Intelligent Tutoring Systems
Other technologies can support science learning in relatively new ways, made possible by more recent advances in artificial intelligence and better computer hardware. Intelligent tutoring systems are one such technology. They are pieces of software designed to guide students from where their understanding lies in in a content area, toward a given objective. Effective tutoring serves as a bridge between prior knowledge and the objectives of a new learning activity or skill. It will provide sufficient instruction to help the student avoid frustration without over-distilling the needed information either. (Wood & Wood, 1996) Intelligent tutoring systems have advanced as viable classroom technologies with the development of artificial intelligence and more ready access to powerful computing processing. They attempt to capture effective teaching techniques employed by experienced teachers and tutors. Sometimes they are embedded in a game format. Other platforms may be more explicit. Below are three commercial examples of intelligent tutoring systems, along with one non-profit offering from the Concord Consortium.
Lexia Reading Software
Lexia is a type of intelligent tutoring system that adapts to student reading levels and responses. It guides teachers toward resources to help students based on their responses within an interesting game environment.
McGraw Hill Learnsmart and SmartBook
The Smartbook program is designed to continually assess student understanding and make recommendations based on learning as students work through written material and illustrations.
The Concord Consortium has long worked to develop rich interactive simulations and materials for science education. The latest version of a dragon-themed genetics game is an intelligent tutoring system that tracks competencies and guides students toward increased understanding of genetics and heredity.
ALEKS Math Tutoring System
ALEKS has been widely adopted in post-secondary math programs. It utilizes artificial intelligence to assess and provide targeted instruction, practice, and assessment based on student knowledge and progress.
Previous Work with Conversational Agents
Graesser, Jeon, & Dufty (2008) developed with an education chatbot agent that they called AutoTutor. It began primarily as a vehicle for researching a conversational method of tutoring students, particularly within the sciences. They examined dialogue tailored toward addressing misconceptions that had been compiled in advance. The AutoTutor asked questions about student expectations or preconceptions, and conversation is moved forward as the student answers the questions. They noted that while the computer didn't understand the student's statements per se, the students were still able to learn from the conversations. It has been a decade since the initial design of AutoTutor and the work of Graesser et al. Since that time, the natural language understanding components of the agents like AutoTutor have advanced considerably, as discussed in Strand D.
In a more recent work, Graesser et al., discuss further insights gained by continued work with the AutoTutor platform and described methods used to score assessments within the AutoTutor system. They discussed a variety of ways that a task can be evaluated and a score computed. They note that the conversations embedded in technology acted as stealth assessments, in which student understanding could be evaluated without it feeling like a test. (Graesser, Cai, Morgan, & Wang, 2017) The evaluation can be partly a function of the forward progression of the conversation and the student's ability to express ideas in a way that emulates expectations. This is different than a simple grading scheme that results in the judgement on the quality of a student's answers. AutoTutor also categorizes student speech in ways that inform whether or not a student understands a particular concept as well as analyzes the speech for markers of frustration, confusion, anger and more. Another component of possible assessment is to look for matches between expected student intents or entities and the actual intents or entities presented.
intent: in natural language processing, an intent is the motivation or direction of a specific piece of a conversation. Associating particular these motivations with labels allows if-then scenarios to be designed based on which intents are identified by the intelligent agent.
entity: a specific piece of data in a chatbot conversation. Entities can be situated as options within a specific intent.
My favorite color is red.
My favorite color is blue.
The intent is favorite color. The entity is either red or blue.
Imagining an Intelligent Science Mentor
Shown below is a second educational technology project that I completed that involved an imagining of an artificially intelligent science project mentor that could guide students through the phases of developing a science fair project or smaller scale inquiry project. This mentor was envisioned as a way to emulate student interactions during the inquiry process by helping students to identify a research question, formulate a testable hypothesis, refine a procedure, and analyze data. Several different platforms were examined as tools with which such a mentor could be built. The mentor’s function as a consultative partner during inquiry science activities positions it as an intelligent tutoring system, in addition to an assessment engine and mentor. Google's Dialogflow, IBM's Watson, and Microsoft's LUIS all feature graphical interfaces that allow non-programmers the ability to construct conversations that utilize natural language understanding.
The development of such a tool is an interesting thought exercise, and it is not outside the boundaries of what artifical intelligence is capable of. Systems that support these types of tools are quickly becoming more powerful and more readily available. In 2014, Nye, Graesser, and Hu documented their work with a platform that utilized natural language understanding. AutoTutor gave rise to several projects related to the science mentor vision in which students give written responses to prompts and a computer understands and responds based on the input it receives. (2014) The remarkable work of this team developed several key ideas in relation to the potential of conversation-based assessment. They argue that by using rich conversations, feedback is timely and relevant, discourse taps into higher-level reasoning skills, and students are actively engaged in conversations. The AutoTutor performed well in Turing-style reviews of transcripts when compared to human tutors. The popularization and accessibility of natural language understanding tools makes further development of these ideas accessible to non-programmers.
natural language understanding: computer algorithms or programs that can interpret input in human language, either written or spoken.
the Turing test: an imitation test invented to evaluate the ability of an artificial agent to seem human. The test has long been held as a standard, albeit a controversial one, in AI development.