Introduction
As Intelligent Virtual Agents (IVAs) like Siri and Alexa become more sophisticated, human users increasingly treat them as social entities. This shift leads to the inclusion of "relational language"—elements like small talk, personal backstories, and justifications—that are not strictly necessary for the task at hand. While these strategies build rapport in human-to-human interaction, they often act as "social noise" for IVAs, degrading intent detection and response accuracy. This paper presents a novel corpus of annotated relational language and demonstrates that filtering this noise significantly enhances agent performance.
Methodology and Categorization
The researchers developed a multi-stage annotation process to identify and categorize non-task-oriented text. Annotators first isolated segments unnecessary for business intent and then classified them into six primary relational strategies:
- Greetings: Standard social openers (e.g., "Hello").
- Backstory: Self-disclosure used to create intimacy or context (e.g., "My son is graduating").
- Justification: Explaining the reason for a request to establish credibility.
- Gratitude: Polite closings (e.g., "Thanks in advance").
- Ranting: Negative narratives or excessive complaining.
- Expressing Emotion: Statements of feeling (e.g., "I love that!").
To validate the dataset, a character-level alignment score was used to ensure high agreement on the boundaries of these social segments.
Dataset and Prevalence
The study analyzed 8,000 initial conversation turns across four domains: three commercial IVAs (Telecom, Airline, Train Travel) and one human-to-human forum (TripAdvisor).
Analysis revealed that relational language is pervasive. While it appeared in 94% of human-to-human TripAdvisor posts, it was also present in up to 77% of Telecom IVA requests. "Backstory" emerged as the dominant strategy across all domains, indicating a deep-seated human need to explain the "why" behind a request, even to a machine.
Impact on AI Performance
The core finding is that relational language actively obfuscates primary intent. By feeding both original and "cleaned" (relational text removed) requests into three commercial IVAs, the researchers observed a clear "confidence boost" in intent detection.
Human A-B testing further confirmed these results: users significantly preferred responses generated from cleaned requests. The AI, no longer "distracted" by social fluff, provided more accurate and relevant answers. The study argues that many "multi-intent" inputs are actually a combination of one task intent and one relational intent.
Conclusion and Future Directions
This research highlights the importance of distinguishing between task intent and relational intent. While current systems benefit from filtering out social noise, the authors suggest that future IVAs should move beyond mere filtering. The ultimate goal is for agents to recognize and reciprocate these strategies—acknowledging a user's backstory while still fulfilling the task—to build greater trust and believability.
Summary updated