TL;DR
This paper argues against modular natural language generation systems where conceptual and linguistic decisions are made sequentially. Instead, it proposes that these decisions are deeply intertwined and must be handled through an integrated process, suggesting a "discourse grammar" to manage these interactions. This is crucial for generating more natural and contextually appropriate text.
What Problem Does This Paper Address?
Traditional natural language generation (NLG) systems often operate in a pipeline: first, a "conceptualizer" decides what information to convey and in what order, then a "dictionary" picks words, and finally a "grammar" constructs sentences. The authors challenge this modular approach, arguing that conceptual decisions (like information order) and linguistic decisions (like word choice and sentence structure) are not independent. They highlight cases where the "best" linguistic choice might conflict with the conceptual plan, leading to awkward or unnatural output. The core problem is how to create NLG systems that produce fluent and coherent text by effectively integrating these intertwined decisions.
Methods
The paper's primary method is to demonstrate the interdependence of conceptual and linguistic decisions through detailed examples, particularly focusing on the generation of "result causation" scenarios (e.g., "John shot Mary, causing her death"). They analyze various combinations of information order (result-first vs. act-first), sentence segmentation (single complex sentence vs. multiple simple sentences), and syntactic constructions (active vs. passive voice). This analysis leads to the proposal of a "discourse grammar" that encapsulates the feasible combinations of these decisions, effectively acting as a set of pre-defined discourse structures. The authors argue that this discourse grammar, rather than a purely language-free conceptualizer, is necessary for effective generation.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Interdependence is Key: Conceptual and linguistic decisions are not independent. The choice of words (e.g., "to kill" vs. "to murder") and sentence structures (active/passive, complex/simple sentences) heavily influences and is influenced by the conceptual plan (information order, explicitness).
- "Language-Free" Conceptualizers are Problematic: The assumption that a conceptualizer can operate without linguistic knowledge is flawed. Linguistic constraints can and do impact conceptual choices.
- The Need for a "Discourse Grammar": To handle the complex interactions, a system needs more than just a conceptual planner and a dictionary. It requires a "discourse grammar" that defines valid combinations of conceptual and linguistic choices, essentially pre-defining acceptable discourse structures for specific types of information.
- Priorities are Context-Dependent: There's no universal rule for which decision (conceptual or linguistic) takes priority. The optimal order depends on the specific semantic content and communicative goals (e.g., generating a suicide might prioritize the verb "to commit suicide," while a crime narrative might prioritize chronological order).
- Frozen Phrases Matter: The meaning of fixed expressions or "frozen phrases" cannot be derived from their constituents, meaning they cannot be generated by a modular system that processes components independently.
Key Results
The paper doesn't present quantitative results in terms of accuracy or performance metrics. Instead, its key findings are qualitative:
- It demonstrates through examples that out of 36 possible combinations of information order, sentence segmentation, and syntactic construction for result causation, only 15 are feasible.
- It shows that certain lexical choices (e.g., "to assassinate") are incompatible with specific discourse structures, further illustrating the interdependence.
- It argues that the selection of a discourse structure is a necessary step in generation, and this selection must account for linguistic feasibility.
Limitations and Future Work
- Intuition-Based Choices: The paper acknowledges that current methods for choosing between feasible generation strategies often rely on intuition, which is difficult to operationalize for a generation system.
- "What to Say" vs. "How to Say It": The paper focuses on the "How to say it?" problem. Future work could extend the model to integrate "What to say?" (conceptual content generation) with "How to say it?", potentially considering the hearer's needs and knowledge.
- Operationalizing Discourse Grammar: While the concept of a discourse grammar is proposed, the paper doesn't detail how to automatically construct or learn such a grammar. The current approach relies on pre-defined feasible combinations.
Summary updated