Introduction
In human-robot interaction (HRI), communication is rarely strictly literal. Social norms, such as politeness, lead humans to frequently employ Indirect Speech Acts (ISAs). For instance, an utterance like "I could use some water" is rarely a mere statement of fact; in most contexts, it is a request for action. For a robot to function effectively as a teammate, it must perform pragmatic analysis—interpreting the intended meaning behind literal words based on social context.
Historically, robots relied on handcrafted "idiomatic" rules to map utterances to intentions. However, these rules are difficult to scale and struggle with the inherent "gray areas" of social norms. This research proposes a method for robots to autonomously learn these sociocultural norms and quantify the associated uncertainty using Dempster-Shafer Theory (DST).
Methodology: Modeling Uncertainty and Ignorance
The core of this approach is the application of Dempster-Shafer Theory, a framework for modeling uncertainty that offers more flexibility than traditional Bayesian probability. While standard probability must assign a value to every possible outcome, DST explicitly allows for "ignorance"—the state of having insufficient evidence to make a determination.
The Learning Pipeline
The researchers represent norms as logical rules: $u \wedge C \Rightarrow_{[\alpha, \beta]} i$, where $u$ is the utterance, $C$ represents contextual conditions, and $i$ is the inferred intention. The uncertainty is captured by the interval $[\alpha, \beta]$, where $\alpha$ represents the "belief" (evidence in favor) and $\beta$ represents the "plausibility" (lack of evidence against).
The learning process involves an iterative algorithm that updates these intervals based on human endorsements. To decide which rules to use, a threshold ($\tau$) is applied to the center of a rule's uncertainty interval.
Experimental Design
To train and validate the system, the authors conducted two human-subject experiments:
- Experiment 1 (Generation): 163 participants identified potential meanings for 19 ISA sentences, resulting in a pool of 25 unique candidate intentions.
- Experiment 2 (Detection): 37 participants evaluated these intentions across varied scenarios (e.g., Restaurant vs. Private Home) and roles (e.g., Waiter vs. Friend). This provided the ground-truth data for training the uncertainty intervals.
Results and Discussion
The learning algorithm successfully demonstrated convergence, narrowing uncertainty intervals as more data was processed. For example, the rule "If I am a waiter and a guest says 'I need noodles,' they want me to get them noodles" converged to a high-belief interval of $[0.68, 0.95]$.
A key finding was the sensitivity of the system to the selection of the threshold $\tau$. As shown in the results, a moderate threshold ($\tau = 0.5$) allowed the system to adopt 115 norms, whereas a stricter threshold ($\tau = 0.81$) filtered the set down to the 6 most robustly supported norms. This highlights a fundamental "clarification trade-off": a conservative robot (high $\tau$) avoids social blunders but requires frequent clarification, while a decisive robot (low $\tau$) acts quickly but risks misinterpretation.
Conclusion
By utilizing DST, robots can distinguish between certainty and ignorance, allowing them to seek clarification when evidence is lacking rather than making incorrect assumptions. While currently limited to specific domains like food service, this framework provides a foundation for lifelong learning in robots, enabling them to refine their understanding of complex social norms through continuous interaction.
Summary updated