A Computational Approach to Topic and Focus in a Production Model

Vincenza PignataroCOLING 1988

TL;DR

This paper proposes a computational model for identifying "topic" and "focus" in natural language, treating them as semantic and pragmatic phenomena rather than purely syntactic ones. This model is then integrated into a natural language generation system, demonstrating its utility in predicting thematic progression and accent placement in simplified discourse.

What Problem Does This Paper Address?

Understanding how humans convey information, particularly what is already known (topic) versus what is new or emphasized (focus), is crucial for building intelligent language systems. Traditional approaches often rely on syntactic cues, which can be insufficient. This paper addresses the challenge of computationally modeling topic and focus by developing a framework that considers semantic, pragmatic, and contextual factors. The authors aim to create an operational definition of these concepts that can be integrated into a natural language generation system, specifically for producing simplified German texts in a "blocks world" scenario. This allows for the simulation of how information is structured and conveyed in a communicative exchange.

Methods

The core of the paper is a computational model for identifying topic and focus. This model operates on semantic representations of utterances and assigns semantic-pragmatic categories based on a defined context. The context is modeled as a pair of sets of statements representing the speaker's assumptions about the world and mutual knowledge.

Key techniques include:

  • Operational Definitions: The authors define "topic" (t), "salient topic" (st), and "focus" (f) through a set of rules that assign labels (e.g., given, chosen, mentioned) to elements within an illocutionary plan based on contextual information.
  • Contextual Categories: A set of labels like g (given), ch (chosen), m (mentioned), ump (mentioned in previous sentence), and their negations, are used to characterize the status of utterance elements with respect to the context.
  • Illocutionary Plans (IP): Utterances are represented as ordered sets of elements (verb and case frame arguments), forming an illocutionary plan.
  • Generation System Integration: The identified topic and focus labels are used to guide the generation process, influencing thematic progression and accent mapping in the output text. The model outlines a four-function generation process (F1-F4) for transforming illocutionary plans and context into surface structures.

Datasets

Not discussed in the paper. The examples are based on a simulated "blocks world" scenario, which serves as a controlled environment for generating and analyzing discourse.

Key Insights and Ideas

  • Topic and Focus as Pragmatic Phenomena: The paper strongly advocates for treating topic and focus as semantic-pragmatic phenomena, heavily influenced by context, rather than solely syntactic features.
  • Contextual Information is Key: The model's effectiveness hinges on its ability to leverage contextual assumptions about mutual knowledge and the state of the world.
  • Predictive Power for Generation: The identified topic and focus labels are not just descriptive but are actively used to guide the generation of natural-sounding text, influencing thematic progression and accent placement.
  • Thematic Progression Rules: The paper proposes specific rules (R1 and R2) that link focused elements in one utterance (e.g., a TAKE action) to topicalized elements in subsequent utterances (e.g., a PUT action), providing a mechanism for discourse coherence.

Key Results

The paper presents several examples demonstrating the application of its rules to generate simplified German sentences. For instance:

  • Example 1: An illocutionary plan for "TAKE, ADRESSEE, OBJECT" results in t(TAKE), t(ADRESSEE), and f(OBJECT), leading to the surface structure "du nimmst einen roten Klotz" (you take a red block), with the accent likely on "roten Klotz" (red block) due to its focus status.
  • Example 2: A "PUT, ADRESSEE, OBJECT, GOAL" plan yields t(PUT), t(ADRESSEE), t(OBJECT), and f(GOAL), resulting in "du stellst ihn auf den Tisch" (put it on the table), with the accent on "Tisch" (table) as the focus.
  • Example 3: Demonstrates how salient topic (st) is assigned to multiple objects in a TAKE action that become the focus of subsequent PUT actions, illustrating how the model handles more complex discourse structures.

Limitations and Future Work

  • Simplified Domain: The model is tested on a highly simplified "blocks world" scenario. Its applicability to more complex, real-world discourse remains to be demonstrated.
  • Context Representation: While a contextual framework is proposed, its full specification and scalability for diverse situations are not detailed.
  • Accent Assignment Nuances: The paper notes that accent assignment to adjectives instead of nouns requires further investigation into lower-level constituent structures.
  • Lexical Choice: The authors suggest extending the semantic categories to better express underdetermined versus determined elements of a set, which is crucial for making appropriate lexical choices (e.g., definite vs. indefinite articles). This points towards future work in refining the semantic representation for generation.

Summary updated

BibTeX
@inproceedings{pignataro1988computational,
  author = {Vincenza Pignataro},
  title = {{A Computational Approach to Topic and Focus in a Production Model}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}
Figure