Talking NPCs in a Virtual Game World

Tina Klüwer, Peter Adolphs, Feiyu Xu, Hans Uszkoreit, Xiwen ChengACL 2010

TL;DR

This paper introduces KomParse, a natural language dialog system for non-player characters (NPCs) in the virtual world Twinity. It combines knowledge-intensive question answering with task-specific dialog to create more immersive and interactive virtual experiences, addressing the limitations of current NPC communication capabilities.

What Problem Does This Paper Address?

Multi-user online games and virtual worlds like Twinity are becoming increasingly popular, offering new social and economic platforms. A crucial element for immersion and engagement in these worlds is the interaction with non-player characters (NPCs). However, current NPCs often have limited communication abilities, restricting their roles and the overall user experience. This paper tackles the challenge of creating more sophisticated and versatile NPCs that can engage in natural language conversations, fulfilling various roles from task-oriented assistants to conversational companions. This advancement is key to populating virtual worlds and making them more dynamic and believable.

Methods

The KomParse system employs a hybrid approach to NPC dialog, integrating several key techniques:

  • Semantic Web Technologies: The system leverages RDF (Resource Description Format) for knowledge representation and OWL (Web Ontology Language) for defining domain-specific ontologies. This allows for structured, queryable knowledge bases that can support semantic search and inference.
  • Hybrid Dialog Strategy: KomParse combines different dialog strategies based on the NPC's role. For task-oriented NPCs (like a furniture sales agent), it uses a complex, ontology-driven task model based on dialog graphs. For more open-ended conversational NPCs (like a bartender), it relies heavily on question-answering capabilities and a broader "gossip domain" knowledge base.
  • Modular Architecture: The system is designed with a modular architecture, separating the Twinity client (handling avatar representation and game interaction) from the KomParse server (housing the conversational agent logic). This allows for flexibility and easier integration with different platforms.
  • Input Analysis and Dialog Management: The conversational agent includes an input analyzer for linguistic processing (parsing, semantic interpretation) and a dialog manager that interprets user utterances within the dialog context, determines the next system action, and controls the conversation flow.
  • Robust Chitchat Handling: To ensure robustness, a fallback A.L.I.C.E. chatbot is integrated to handle chitchat utterances that the primary dialog system cannot manage.
  • Frame-Based Task Modeling: For task-driven scenarios, a frame-based sub-component ("form") is used to store underspecified object descriptions and incrementally build up specifications as the conversation progresses, increasing flexibility.

Datasets

The paper does not explicitly mention datasets used for training the core NLP models. However, it highlights the creation and use of:

  • Extensive Furniture Ontology: Developed with a project partner, containing 95,258 triples, 123 furniture classes, and numerous color and style instances.
  • Cocktail Ontology: Defining 13 cocktail classes with ingredients and tastes, comprising 21,880 triples.
  • Biographical Ontology ("Gossip Ontology"): Defining biographical and career concepts, accompanied by a large database of celebrities (nearly 600,000 individuals) and their relations, automatically acquired from the web.
  • Wizard-of-Oz Corpus: A corpus of 18 dialogs (3,171 turns, 4,313 utterances, 23,015 words) from a furniture sales scenario, used to inspire the task model for the sales agent.

Key Insights and Ideas

  • Hybrid Approach is Key: The success of KomParse hinges on its ability to blend different dialog strategies. A purely task-driven or purely conversational approach would limit NPC utility. The hybrid model allows NPCs to be both functional assistants and engaging characters.
  • Semantic Web for Richer NPCs: Using RDF and OWL for knowledge representation allows NPCs to access and reason over complex domain knowledge, enabling more nuanced and informed interactions. This moves beyond simple keyword matching.
  • Ontology-Driven Task Modeling: The use of ontologies to structure domain knowledge and inform dialog graphs provides a flexible and scalable way to manage task-oriented conversations, especially in dynamic environments like virtual worlds.
  • Separation of Concerns for Scalability: The clear separation between the game client and the conversational agent server makes the system more adaptable to different virtual world platforms, reducing the need for complete reimplementation.
  • The "Gossip Domain" for Engagement: The bartender NPC's access to a broad "gossip domain" demonstrates how NPCs can be made more entertaining and engaging by providing trivia and social information, going beyond strict task fulfillment.

Key Results

The paper presents a system demonstration rather than quantitative results from extensive evaluations. However, it highlights:

  • Successful Integration: The KomParse system is successfully integrated into the Twinity virtual world, enabling NPCs to converse with users.
  • Two Distinct NPC Personalities: The system supports two distinct NPC types: a task-oriented furniture sales agent and a more conversational bartender, showcasing the flexibility of the hybrid approach.
  • Robustness through Fallback: The integration of an A.L.I.C.E. chatbot provides a fallback mechanism for handling unexpected or out-of-domain utterances, contributing to a more robust user experience.
  • Demonstrated Functionality: The paper provides examples of conversations, illustrating how the system can handle requests for furniture and engage in trivia-based chat.

Limitations and Future Work

  • Limitations:
    • Speech Recognition: The system currently relies on text-based chat and Voice-over-IP, but the paper notes that imperfect speech recognition is a potential hurdle for future voice-enabled interactions.
  • Evaluation: The system is presented as a work in progress, with a need for more comprehensive evaluation steps.
  • Future Work:
    • Component Evaluation: Focus on evaluating individual components like parsing and dialog interpretation for precision and recall.
  • End-User Evaluation: Conduct experiments with end-users to assess the usability of the two distinct NPC scenarios.
  • Opinion Mining and Sentiment Analysis: Integrate these functionalities to better understand user preferences, particularly for the furniture sales agent.
  • Platform Extension: Adapt the avatar interfaces to extend the system to other virtual world platforms.

Summary updated

BibTeX
@inproceedings{kluwer2010talking,
  author = {Tina Klüwer and Peter Adolphs and Feiyu Xu and Hans Uszkoreit and Xiwen Cheng},
  title = {{Talking NPCs in a Virtual Game World}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2010},
}
Figure