Artificial Conversational Companions

Sviatlana DanilavaAAAI 2013 · DOI

Introduction

The evolution of conversational agents has largely focused on transactional efficiency—providing weather updates, setting reminders, or answering factual queries. However, in the domain of Second Language Acquisition (SLA), the requirement for an agent shifts from a simple tool to a "companion." This paper explores the development of Artificial Conversational Companions (ACCs), agents designed to sustain meaningful, multi-week relationships with language learners by prioritizing the social and temporal nuances of human interaction.

The Interactional Gap in Chatbot Design

Current chatbot architectures excel at processing content—the "what" of a conversation—but frequently stumble on the "how." In human communication, meaning is co-constructed through interactional rules, including turn-taking, response timing, and social cues. For a language learner, "meaningful interaction" involves more than just grammatical accuracy; it requires a partner who can navigate errors gracefully and recognize signals of frustration or closeness.

The author identifies a significant gap: most systems fail to account for the social co-construction of dialogue. To address this, the research moves beyond simple content-based responses toward "interaction profiles," which track the social and temporal connections between participants over time.

Methodology: Data-Driven Conversation Analysis

The proposed framework utilizes a data-driven approach rooted in conversation analysis. Rather than relying solely on pre-programmed linguistic rules, the research models computational behavior on actual human-to-human Instant Messaging (IM) interactions.

Key components of this methodology include:

  1. Interaction Profiles: Based on linguistic theories (Spranz-Fogasy 2002), these profiles monitor the phenomena of talk and the evolving relationship between the user and the agent.
  2. Responsiveness Modeling: This involves analyzing "responsiveness values"—the specific timing and sequence of message turns. By tracking these values, the system can recognize behaviors such as self-repairs or moments of non-understanding.
  3. AIML Integration: These social and error-correction layers are integrated into an Artificial Intelligence Markup Language (AIML) architecture, extending standard chatbot capabilities with a sophisticated interactional layer.

Empirical Foundation: The IM Corpus

To build these models, the study utilized a custom-built corpus of naturalistic interactions. The dataset consists of 72 IM dialogues, totaling approximately 4,800 messages and 52,000 tokens. The participants included nine advanced learners of German and four native German speakers.

The interactions took place over a period of four to eight weeks, with individual sessions lasting between 20 and 90 minutes. A unique "forwarding chatbot" setup was employed to strip away status indicators like "typing..." animations. This ensured that the "joint activity" of the conversation was constructed purely through the text messages and their timing, providing a clean environment to study how social presence is established through text alone.

Key Insights: IM as a Hybrid Medium

A central finding of the research is the status of Instant Messaging as a hybrid medium: "conceptually oral" but "medially written." This duality makes IM an ideal laboratory for SLA. It captures the spontaneity and social dynamics of speech while providing the persistence of text, which is crucial for language reflection and error correction.

The research highlights several critical factors in digital companionship:

  • Social Closeness via Metadata: Social distance is negotiated through non-verbal cues such as emoticons, language complexity, and politeness levels.
  • Error Correction as Strategy: Instead of merely flagging errors, the framework looks at how experts negotiate meaning during a breakdown, allowing the agent to employ similar strategies.
  • Responsiveness as a Feature: The timing of a reply is often as informative as the text itself. Responsiveness values can automatically detect when a user is struggling or attempting to correct themselves.

Conclusion and Future Work

The research establishes the foundational requirements for an agent to be perceived as a companion rather than a tool. By developing a specialized annotation scheme for "conceptually oral" learner language and outlining an AIML-based system architecture, the paper provides a roadmap for more human-centric SLA agents.

Future work will focus on the quantitative evaluation of specific components, such as the accuracy of self-correction recognition, and qualitative user interviews to better understand the subjective experience of "companionship" in human-agent relationships. While currently focused on German, the interaction profile methodology holds potential for broader applications in health, fitness, and other long-term digital companionship contexts.

Summary updated

BibTeX
@inproceedings{danilava2013artificial,
  author = {Sviatlana Danilava},
  title = {{Artificial Conversational Companions}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2013},
  doi = {10.1609/aaai.v27i1.8495},
}
Figure