Integrating Text Planning and Production in Generation

Eduard H. HovyIJCAI 1985 · OpenReview

Abstract While the task of language generation seems to separate quite naturally into two aspects—text planning and text production—it is necessary to have these stages interact at generator decision points. This interaction ensures that the planner does not require explicit syntactic knowledge and the producer does not require explicit goal-related information. This paper describes the decision points, the types of plans used in making decisions, and a process that performs the task, as embodied in a functional program.

Introduction: The Divide in Language Generation

In the mid-1980s, Natural Language Generation (NLG) systems typically followed one of two architectural paths. They either lacked a planning phase entirely, focusing almost exclusively on syntactic correctness, or they employed a rigid "one-way" pipeline. In the latter, a high-level planner would determine the entire content and structure of a message before handing it off to a generator for realization.

Eduard H. Hovy argues that both approaches are fundamentally flawed when dealing with pragmatics—the social and contextual nuances of language. The "gap" between high-level goals (what we want to achieve) and low-level realization (how we say it) makes it difficult for systems to make the right lexical and structural choices at the right time.

The Pragmatic Gap

The core issue is that pragmatic goals often depend on subtle linguistic choices. For example, if a system aims to evoke sympathy for a "65-year-old beggar," it might be more effective to describe the subject as "the old woman" rather than "the 65-year-old woman." However, a high-level planner should not need to be an expert in English adjectives, and a low-level generator should not need to understand the speaker's emotional or political motivations.

Hovy proposes an integrated approach characterized by two-way communication at specific decision points. Instead of the planner issuing mandatory instructions, it provides "suggestions" on a suggestion list, allowing the generator to query the planner when it encounters a choice in the grammar.

The Five Decision Points of Generation

The system operates through five key decision points where the generator pauses to consult the planner:

  1. Topic Choice: Deciding which piece of information from the knowledge base to address next.
  2. Sentence Content: Determining which pieces of information are relevant enough to be grouped into a single sentence.
  3. Sentence Organization: Choosing the subject and focus (e.g., deciding between "Carter lost the election" or "The margin of defeat was 3 delegates").
  4. Clause Organization: Deciding how to order modifiers and head nouns within a clause.
  5. Word Choice (Lexical Choice): Selecting the specific verb or noun that carries the desired connotation (e.g., choosing between "freedom fighter" and "terrorist").

When the generator reaches one of these points, it presents the available syntactic options to the planner. The planner evaluates these options against active pragmatic goals—such as "be polite," "be dominant," or "convince the hearer"—and returns a preference.

Conflict Resolution and Opportunism

One of the most significant challenges in this architecture is handling conflicting goals. A speaker may simultaneously want to "be honest" and "be polite." Hovy’s system manages this by merging suggestion lists and selecting the most "popular" or highly-ranked option among active plans. This allows for parsimony; planning only occurs when the generator actually needs to make a choice, preventing the system from over-computing details that might never be used.

Furthermore, this interaction allows for opportunism. The generator can take advantage of specific syntactic possibilities that the planner could not have foreseen, such as using a specific adverb to emphasize a small numerical difference that only becomes apparent during the realization phase.

Experimental Evaluation: The 1980 Presidential Primaries

The system was tested using a specialized knowledge base representing the 1980 Carter-Kennedy Presidential primary elections. The goal was to generate different versions of the election results based on varying "sympathies."

For instance, when tasked with convincing a Kennedy supporter that Carter’s performance was not a total failure, the system produced:

"Kennedy only got a small number of delegates in the election on 20 February. Carter just lost by a small number of delegates; he has several delegates more than Kennedy in total."

The system achieved this by choosing the verb "lost" (with a "just" modifier) to minimize the impact of the defeat and selecting the phrase "small number" to frame the delegate gap favorably.

Limitations and Future Work

While the integrated approach marks a significant step forward, Hovy identifies several limitations:

  • Simplistic Conflict Resolution: The current "most popular option" heuristic is a baseline; more sophisticated meta-rules for ranking goals are required.
  • Domain Specificity: The logic is currently tied to a specific property inheritance network. Scaling to open-domain generation would require a significantly larger world-knowledge base.
  • Computational Overhead: The constant back-and-forth between the planner and generator could become a bottleneck in real-time systems as the complexity of the grammar and the number of pragmatic goals increase.

Conclusion

By breaking the rigid pipeline of traditional NLG, Hovy demonstrates that language generation is most effective when planning and production are treated as interleaved processes. This architecture allows for the production of text that is not only syntactically correct but also pragmatically aligned with the speaker's complex social and communicative goals.

Summary updated

Abstract

While the task of language generation seems to separate quite naturally into the two aspects of language generation (text planning and text production), it is necessary to have the planning and the production interact at generator decision points in such a way that the former need not contain explicit syntactic knowledge, and that the latter need not contain explicit goal-related information. This paper describes the decision points, the types of plans that are used in making the decisions, and a process that performs the task. These ideas are embodied in a program.
BibTeX
@inproceedings{hovy1985integrating,
  author = {Eduard H. Hovy},
  title = {{Integrating Text Planning and Production in Generation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
  url = {https://openreview.net/forum?id=B14GZSMubB},
}
Figure