=Paper=
{{Paper
|id=Vol-2862/paper1
|storemode=property
|title=Tokamak Elementary: Visual Novel Meets Natural Language Understanding (short paper)
|pdfUrl=https://ceur-ws.org/Vol-2862/paper1.pdf
|volume=Vol-2862
|authors=Daniel Darabos,Gyorgy Lajtai,Andras Nemeth
|dblpUrl=https://dblp.org/rec/conf/aiide/DarabosLN20
}}
==Tokamak Elementary: Visual Novel Meets Natural Language Understanding (short paper)==
Tokamak Elementary: Visual Novel Meets Natural Language Understanding Daniel Darabos, Gyorgy Lajtai, Andras Nemeth Lynx Analytics daniel.darabos@lynxanalytics.com Abstract “Tokamak Elementary” is a game built around a custom con- versational engine. The player can talk with “Toki”, the AI- controlled character as with a chatbot. Toki is a gentle spirit creature summoned from the world of spirits. As such, Toki starts from a mostly clean slate and it’s up to the player to teach them basic concepts of our world (such as what ani- mals are) and how to interact with humans (such as returning greetings). One part of the game is a sandbox environment where the player chats with Toki in freeform natural language. (Fig- ure 1) This part is fun because the player can interact with a goofy AI and teach it things. (Consider (Lionhead Studios 2001).) The other part of the game is a school Toki has to go and pretend to be a human. This part is a non-interactive Figure 1: The free natural language chat interface where the scripted visual novel with a branching dialog tree. (Figure 2) player can chat with Toki. This part is fun because of a complex plot involving school children, magical creatures, and fusion reactors. The novelty of the game comes from the interaction of these matched against the parse trees of all rules stored in Toki’s two parts. At school Toki’s conversational engine gets the di- knowledge base. The best matching combination of rules is alog from other characters as input. Toki answers according executed. The rules query and update the knowledge base to its engine, just like when speaking to the player. The dia- and craft a response. There are rules that allow the player to log branches are selected by simple checks on Toki’s output. add new rules, for example teaching Toki to say ”hello” when (Usually via regular expressions.) greeted. Gameplay alternates between the two parts. Toki goes to The neural network-based parsing approach allows some school, faces challenges and initially fails them. The player flexibility in the input. But we can’t hope to understand all then gets to talk to Toki at home and teach them new facts possible inputs. While providing a challenge and telling an and rules. When the player sends Toki back to school, they interesting story, the visual novel part also serves to demon- will go through the same day (same dialog tree), but Toki strate to the player the kind of grammar structures that Toki will give different responses due to an updated knowledge understands. (Figure 3) base. This will lead to new branches of the dialog tree and Crafting a response is also challenging. The rules give exam- eventually a successful completion of the day. ple responses, and our engine performs substitution according The school setting is an opportunity for real educational con- to the actual rule execution context. We use SpaCy (Honnibal tent. In our video walk-through of the demo (Darabos 2020) and Montani 2017) to tokenize and tag the template sentence we meet Sophie, the music teacher who explains the chro- and Pattern (Smedt and Daelemans 2012) for conjugation to matic scale and a musical interval. Tokamak Elementary does match the target tags. not directly quiz the player, but progress is gated behind Toki Tokamak Elementary is under development. The focus so far achieving good grades. The player is tasked with explaining has been on demonstrating that the basic concept works. We real-world concepts to Toki. Sometimes teaching someone have a playable demo that includes the sandbox mode and else is the best way to learn a lesson ourselves. a single short day at school. (As shown in the video walk- Our conversational engine parses inputs based on a con- through.) We plan to add more days with a gradual ramp up stituency parse tree generated by the Berkeley Neural Parser of difficulty and an unfolding of the story. When we add new (Kitaev and Klein 2018). The parse tree is recursively scenes, we add grammar and knowledge representation fea- tures as needed. (Such as past tense and a concept of time- Copyright c 2020 for this paper by its authors. Use permitted un- lines.) This accumulation of natural language understanding der Creative Commons License Attribution 4.0 International (CC also improves the sandbox chat experience. BY 4.0). Figure 2: Toki goes to a school near the ITER nuclear fusion research site. Could there be a connection between fusion and magic? Figure 4: A journal feature lets the player access the full transcript of everything so far. Here we can see how Toki managed to complete a simple logical reasoning task. Figure 3: A line from the visual novel part of the game. The text that demonstrates grammar that Toki can understand is highlighted in red. This is also an example of educational content in Tokamak Elementary. The player has to teach Toki some music theory to progress. References Darabos, D. 2020. Tokamak elementary demo 2020. https://www.youtube.com/watch?v=5 V5fDM9lrI. Honnibal, M., and Montani, I. 2017. spaCy 2: Natural language understanding with Bloom embeddings, convolu- tional neural networks and incremental parsing. To appear. Kitaev, N., and Klein, D. 2018. Constituency parsing with a self-attentive encoder. CoRR abs/1805.01052. Lionhead Studios. 2001. Black & white. [CD-ROM]. Smedt, T. D., and Daelemans, W. 2012. Pattern for python. Journal of Machine Learning Research 13(66):2063–2067. Figure 5: Once a required magical item is obtained, the player gets the ability to directly inspect Toki’s knowledge base. Here we see the knowledge base entries created upon hearing “a cat has four legs.”