TL;DR
This paper introduces a computational model for "social perlocutions" – the subtle, often implicit, social effects of speech acts beyond their literal meaning. They've implemented this model in a system called LetterGen, which generates socially appropriate email by planning individual speech acts to achieve desired social outcomes like politeness and helpfulness.
What Problem Does This Paper Address?
Traditional speech act theory and computational linguistics have largely focused on the informational or task-oriented functions of language (e.g., informing, requesting). However, communication is also crucial for managing social relationships, maintaining "face," and conveying social nuances like thanks or sympathy. These "social perlocutions" have been largely ignored in computational models, making it difficult to build systems that can generate truly socially appropriate text. This paper aims to bridge this gap by modeling these social effects and integrating them into a practical text generation system.
Methods
The authors propose a computational model of social perlocutions that integrates work from natural language generation (NLG), communication studies, and social psychology. Their core contribution is a taxonomy of social perlocutionary effects, defined in terms of the hearer's mental attitudes. These effects are then organized into a graph representing their causal relationships.
The system, LetterGen, uses this graph-based representation to determine appropriate speech acts. Given a high-level communicative goal (e.g., "decline an invitation politely"), LetterGen traverses the graph to identify speech acts that achieve the goal and mitigate potential negative side effects. It employs a reactive planning approach, where it identifies speech acts, checks for conflicts with user goals (like politeness), and recursively finds new speech acts to resolve these conflicts.
Datasets
Not discussed in the paper. The authors mention analyzing a corpus of letters and email messages for their speech acts and social effects, but no specific dataset is detailed.
Key Insights and Ideas
- Beyond Literal Meaning: Communication is not just about conveying information; it's about achieving social goals and influencing relationships.
- Social Perlocutions are Key: Understanding and modeling the social effects of speech acts (like building affinity, securing favors, or managing impressions) is crucial for natural language generation.
- Interconnectedness of Effects: Social effects are not isolated. They form a web of causal relationships, where one effect can lead to another, and multiple speech acts can contribute to the same effect.
- Graph-Based Reasoning: Representing these complex relationships as a graph allows for a more systematic and computational approach to planning social interactions.
- Mitigation is Essential: Generating socially appropriate text often involves not just achieving the primary goal but also actively mitigating potential negative social side effects.
Key Results
The paper presents a functional prototype, LetterGen, which successfully generates socially appropriate email messages. As an example, Figure 1 shows a politely declined invitation that includes thanking, apologizing, and suggesting an alternative speaker, demonstrating the system's ability to go beyond a simple "no." The authors state that their model has been applied to over 400 effects and constraints, enabling the generation of various message types, including those for initiating/terminating friendships, job applications, congratulations, and apologies.
Limitations and Future Work
The authors acknowledge three major limitations:
- Scope: The model doesn't currently cover all aspects of social interaction, such as relative status, specialized roles (judge, employer), or certain socially-related speech acts (expressing sadness, joy).
- Conditions: The conditions under which effects occur need further elaboration.
- Implementation: The system's ability to query users for necessary background information is limited to avoid making the program too cumbersome.
Future work includes addressing these limitations by exploring techniques for forming detailed user profiles, applying the model to letter understanding (not just generation), and developing more flexible organizational and text templates. They also suggest exploring how to plan the realization of speech acts more dynamically.
Summary updated