TL;DR
This paper introduces a novel hybrid approach to automatic text generation, combining the efficiency of schema-based methods with the flexibility of Rhetorical Structure Theory (RST). By integrating domain-dependent and domain-independent knowledge resources, this system aims to produce more coherent and rhetorically compelling textual structures.
What Problem Does This Paper Address?
Natural language generation (NLG) systems have evolved from generating single sentences to producing coherent multi-sentence paragraphs. This evolution involves two key challenges: deciding what to say (strategic level) and how to say it (tactical level). Early systems relied on "schemata" to define conventional text structures, but these lacked a clear understanding of the rhetorical role each part played. Later, Rhetorical Structure Theory (RST) provided a way to model these relationships, offering more flexibility but proving more difficult to implement. The core problem this paper addresses is how to achieve a balance between the efficiency of schema-based approaches and the flexibility of RST-based methods in planning textual structures.
Methods
The paper proposes a hybrid planning architecture that combines schema-based and RST-based paradigms. The core of this approach is a top-down hierarchical expansion mechanism that leverages multiple knowledge resources. These resources are categorized into:
- Intentional Operators: Representing the writer's intentions, decomposed into subgoals and linked to rhetorical operators or speech acts. These are formalized using a theory of rational interaction.
- Rhetorical Operators: Associated with intentional operators, these define the rhetorical means to achieve specific goals. They have components like prerequisites, constraints, and effects, and can be domain-dependent or domain-independent.
- Intentional and Rhetorical Networks: Hierarchical structures that guide the control flow for using intentional operators and define preferred structures for rhetorical relations, respectively.
The planning process involves expanding intentional operators into subgoals or rhetorical operators into actions. It also incorporates heuristic mechanisms, such as augmented transition network traversal and constructive critics, to manage complexity and make efficient choices.
Datasets
The authors used a selection of English sales letters as their test data. These letters were chosen for their relatively formal structure and were written for a specific audience concerning computer software products. The textual analysis was carried out according to RST, with some modifications. No new benchmarks were introduced.
Key Insights and Ideas
- Bridging the Gap: The paper highlights that schemata and RST plans are not fundamentally opposed but rather exist on a spectrum, with schemata being collections of compiled plans. This insight is crucial for developing a hybrid system.
- Multi-Knowledge Resource Integration: The authors emphasize the necessity of integrating diverse knowledge resources (domain-dependent and domain-independent) to handle the complexity of text generation.
- Intentionality is Key: Explicitly modeling communicative intentions through "intentional operators" is presented as a vital addition to traditional RST, which focuses primarily on rhetorical relations.
- Heuristic Efficiency: The proposed system uses heuristic strategies to selectively apply resources, prioritizing efficiency by avoiding unnecessary computational effort.
Key Results
The paper describes a prototype system designed to demonstrate the hybrid approach. While specific quantitative results are not detailed, the authors claim the system can generate textual structures that are "linguistically motivated and coherent" and "rhetorically compelling." The output is a single structure tree with speech acts and propositions, suitable for input to a functional grammar for sentence generation. The system can produce different textual structures (e.g., lengthy and persuasive vs. short and informative) based on the input goal and desired mode.
Limitations and Future Work
The authors acknowledge that their research is ongoing and they have not identified all necessary knowledge resources or planning mechanisms. Specific limitations mentioned include:
- The need for further evaluation of various "planning critics" in detail.
- The desire to capture richer domain knowledge.
- The need to further develop critic modules and their controlling mechanisms.
Future work will focus on these areas to enhance the system's capabilities and robustness.
Summary updated