TL;DR
This paper presents a computational model for generating referring expressions (like "the applicant" or "your spouse") in a multilingual system that creates instructions for filling out pension forms. The model extends the Centering Theory to better handle pronoun generation by incorporating discourse structure, and it customizes rules for different languages to ensure natural and accurate text.
What Problem Does This Paper Address?
Generating natural and effective language is a core challenge in Natural Language Generation (NLG). A crucial aspect of this is generating "referring expressions" – phrases that identify specific entities (people, objects, concepts) in a text. The goal is to create expressions that are unambiguous, concise, and help the reader easily follow the discourse. This paper tackles this problem within a specific, complex domain: generating multilingual instructions for pension forms. This domain presents unique challenges:
- Multilinguality: The system needs to produce text in English, German, and Italian, requiring language-specific rules.
- Domain Specificity: Pension forms involve technical terms and specific entity types (applicants, forms, benefits) that need precise identification.
- Discourse Cohesion: Instructions must guide the user logically, and referring expressions play a key role in linking different parts of the text.
- Formality and Politeness: Administrative documents have specific stylistic requirements.
The authors aim to build a system that can generate referring expressions that meet these demands, moving beyond simple entity identification to ensure fluency and reader comprehension.
Methods
The core of the proposed model is based on the theoretical work of J.R. Martin (1992) and extends the Centering Model for pronoun generation.
- Entity Classification: The system first classifies entities into three types:
- Specific Entities: Concrete individuals or collections (e.g., "Gianni Billia").
- Generic Entities: Abstract classes or concepts (e.g., "married women").
- Anchored Entities: Generic entities that become specific in the context of the communicative situation (e.g., "the applicant" when a specific person is reading the form).
- Knowledge Representation: The model uses a knowledge base that represents these entities and their relationships.
- Referring Expression Generation Algorithm: A global algorithm, derived from corpus analysis, guides the selection of referring expressions. This algorithm considers:
- Rhetorical Structure Theory (RST): How clauses are linked semantically to form a coherent text. This helps determine the "focus" of attention.
- Contextual Information: Background knowledge (cultural, social), previously mentioned entities, and the current "focus state" of the reader's attention.
- Style Parameters: Formality and politeness requirements of the output text.
- Pronoun Generation (Centering Model Extension): To handle pronoun generation, the authors extend the Centering Model. The standard Centering Model uses "forward-looking centers" (Cf) and "backward-looking centers" (Cb) to track discourse focus. The proposed extension incorporates information from the rhetorical structure of the discourse to better predict the reader's flow of attention and resolve pronoun ambiguity, especially in languages that allow null pronouns (like Italian). This helps explain exceptions to standard Centering rules.
- Multilingual Customization: At each decision point in the generation process, rules are customized based on the target language (English, German, or Italian) to account for language-specific linguistic realization choices.
Datasets
The paper describes the use of a multilingual corpus of pension form instructions (English, German, Italian) for analysis and to derive the rules for the referring expression component. No specific dataset names or sizes are provided, nor are new benchmarks introduced.
Key Insights and Ideas
- Discourse Structure is Key for Pronouns: The most significant insight is that simply tracking discourse entities (as in the standard Centering Model) is insufficient for pronoun generation. Incorporating the rhetorical structure of the text provides crucial information about how the reader's attention progresses, allowing for more accurate and natural pronoun use, especially in complex cases or languages with null pronouns.
- Multilinguality Requires Rule Customization: The paper emphasizes that a one-size-fits-all approach to referring expressions doesn't work across languages. Language-specific rules are essential for generating natural-sounding text, particularly in how anchored entities are expressed (e.g., possessives in English/German vs. definite articles in Italian).
- Anchored Entities Bridge Generic and Specific: The concept of "anchored entities" is useful for bridging the gap between abstract, generic descriptions and concrete, specific references within a particular communicative context.
- Combining Knowledge Sources: The model effectively integrates various knowledge sources – linguistic, discourse-level, and domain-specific – to make informed decisions about referring expressions.
Key Results
The paper describes a computational model and its implementation within the GIST project. While specific quantitative results (e.g., accuracy scores) are not presented, the key findings are qualitative and algorithmic:
- The proposed model successfully generates referring expressions for a multilingual application domain.
- The extension to the Centering Model effectively addresses pronoun generation by leveraging discourse structure, leading to more natural anaphora.
- The language-specific customization of rules allows the system to handle multilingual variations in referring expression realization.
- The model is described as being "particularly suitable to work in a multilingual domain."
Limitations and Future Work
- Lack of Quantitative Evaluation: The paper focuses on the model's design and theoretical underpinnings rather than presenting empirical evaluation results with concrete metrics.
- User Model Representation: The "Background" knowledge is represented as a list of entities the reader is supposed to know. The paper mentions this is a current representation, implying potential for more sophisticated user modeling.
- Future Work: The authors suggest that the proposed approach, particularly the integration of discourse structure for anaphora, can be used to define "repairing strategies" for exceptions in pronoun generation. The general framework is presented as adaptable to other application domains.
Summary updated