“Unplugged” Semantic Networks and Knowledge Representations

Summary	In this assignment, learners can create their own semantic networks by gluing down printable cards containing concepts and arrows containing relationships. Provided card decks contain concepts related to animals, family, and musical instruments. Learners will be encouraged to reflect on the networks they create and consider the strengths and limitations of the knowledge representation using a provided list of questions that can be used to foster discussion or as a written activity. No technology is required.
Topics	semantic nework, knowledge representations, knowledge-based AI
Audience	K-12. This assignment is suitable for young learners but can also serve as a simple, accessible introduction to knowledge representations or knowledge-based AI for learners of all ages.
Difficulty	No technical skills required, accessible for younger kids (6 and up), takes 15-20 minutes to complete.
Strengths	This is a simple introduction to semantic networks and knowledge representations with no prerequisite technical skills needed. Learners can draw on their prior knowledge and interests when constructing a network. It is also unplugged (i.e. requires no computer), making it useful in a variety of contexts.
Weaknesses	Since this activity does not involve a computer, learners will not get to test out how a computer would use the semantic networks they created.
Dependencies	No dependencies or prerequisite knowledge required.
Variants	This activity was designed for a K-12 environment. Depending on the cards and connections used (which can be customized using the attached Microsoft Word or Adobe InDesign file), building semantic networks could be used to teach about connections in a variety of different subjects including: biology (plants and animals, the human body), foreign language (learning vocabulary and verb conjugations), and learning about the self (family relationships, likes/dislikes).

Activity

Cut out the Semantic Network card deck and have students use the cards to build their own semantic network. The square cards represent objects or ideas. The arrow cards represent relationships. Blank arrows and squares can be used to write-in custom relationships and concepts. Use glue and paper to make your own network to teach an AI agent. An example semantic network built using the cards is shown below. Students can simulate an AI-user interaction using their semantic networks. Two students can trade completed semantic networks and ask their partner questions about the network they created (e.g. What is a cat?). The student’s partner should answer the questions using the semantic network as their only guide (simulating an AI agent whose only knowledge is based on the semantic network). An example interaction is also provided below.

Example simulation:
S1: What does a cat have?
S2: A cat has eyes and fur.
S1: What do a cat and an ant have in common?
S2: A cat and an ant both have eyes.
S1: What is an ant?
S2: An ant is an insect.

Example discussion:
Q: What types of relationships were you able to express using the semantic network? What was harder to capture?
A: I was able to say things like a cat is a mammal, a cat has fur. It was hard to capture certain things like cats don't really dislike birds, they just eat them.
Q: Do you think the computer understood what a cat was based on what you told it?
A: Not really. It knows that a cat has fur and eyes but it doesn't know that cats meow and purr and are sweet and like to climb on things.
Q: Do you think you could capture that information if you built a bigger network with more arrows?
A: I probably could have told it some of those things if I built a really big network. But that would take a lot of effort, and I still don't think it would really understand cats like I do.
Q: Did any of the connections you made surprise you?
A: I was surprised that cats and ants had something in common.
Q: What do you think would happen if you put false information into the network, like telling it that cats have scales?
A: The AI would probably give you the wrong answers then.
Q: What would you do if you wanted to make an AI that could learn new information from people?
A: Maybe I would make it so that the computer could add arrows and squares itself whenever someone told it something new. Like if someone told it cat has teeth, it would add a teeth square and a has arrow connecting it to cat.
Q: Can you explain how the computer would use the network to answer the question what do a snake and a fish have in common?
A: It would look at the snake tile and the fish tile and look at the arrows coming off of them and see if there is any tile that is connected to both.

Key Vocabulary

Knowledge representation: a structure or symbol that a computer uses to store and organize information about the world [1]
Semantic network: a knowledge representation that represents relationships between concepts and ideas in the form of a network. It is generally shown as a graph where concepts/ideas are nodes and relationships are edges or arrows [2]
Knowledge-based artificial intelligence: a computer program that reasons and uses a knowledge base to solve complex problems [3]
Natural language processing: is a subfield of linguistics, computer science, information engineering, and artificial intelligence concerned with the interactions between computers and human (natural) languages [4]

Discussion Questions

What types of relationships were you able to express using the semantic network? What was harder to capture?
Did any of the connections you made surprise you?
What happens if you put false information in the semantic network? If you were building an AI agent, how could you correct for programmer errors?
What does it mean for the computer to understand what a cat is (for example)–if the computer doesn’t know any information outside of what you built into the semantic network?
Do you think the way a computer understands concepts using a semantic network is similar or different to how we as humans understand concepts?
What would you do if the system needed to learn new knowledge (i.e. add to its semantic network over time)?
Compare the semantic network to other ways that computers learn and represent knowledge—for example, a neural network. How are they similar/different? What types of tasks would each knowledge representation be better suited for?

Real-World Connections

WordNet is a natural language processing system that uses semantic networks to help computers understand the connections between different words in the English language. It’s like a dictionary for the computer with many links connecting words to each other so that the computer can understand common-sense relationships. WordNet has been used by search engines like Google to help find relevant pages for a given search term. For example, if you search for something related to a dog, WordNet might help find related pages that use synonyms like puppy or canine.
ConceptNet is similar to WordNet, but connects compound concepts in a semantic network so computers can understand common-sense relationships between entities and ideas. ConceptNet has been used for tasks like summarizing long texts or detecting the emotion of a text (e.g. this message is sad).
Semantic networks can be used to aid in detecting plagiarism by enabling the computer to recognize similarities in word meanings between two texts (even if the specific words differ).

Learning Outcomes

This activity aims to communicate the following learning outcomes:

Learners understand that one way computers store common-sense knowledge is using networks of connected concepts and relationships.
Learners can explain at a high-level how a computer would use the network they built to answer questions (e.g. follow the HAS arrows connected to cat to answer What does a cat have?)
Learners can reflect on the strengths and limitations of semantic networks as a way of representing knowledge.

Assignment in Practice

We have not yet rigorously tested this assignment with a representative audience from multiple grade bands. However, to help instructors determine the appropriate use and context for the assignment, we can share some anecdotal observations/lessons learned from our experience using this lesson in practice. We have previously tested versions of this assignment in informal learning contexts (at-home learning, museums) with ~15 different family groups with children ranging from 4 to 17 years old. Children ages 6 and up have productively participated in the activity. Younger children (6-9) were able to draw on their existing knowledge of animals and families in particular to build networks. Some adult support was provided in the form of verbal instructions and prompting questions during the network building (e.g. What does a cat have?). We have observed that younger children often need some adult support during the latter half of the activity when they are asked to simulate an AI agent using their network. Instructors working with younger kids may want to provide an example demonstration for the class or provide more hands-on support during that portion of the activity. Alternative age-appropriate vocabulary may also lower the barrier to entry for young learners--for instance, referring to the semantic network as a knowledge net or talking about how computers store knowledge/information rather than using the term knowledge representation. In our experience, the activity has also been engaging for older children and adults and allows groups with learners of a variety of ages to work together.