Abstract
Generating natural language in domains where entities undergo physical transformation presents a significant challenge for traditional NLP systems. This paper examines EPICURE, a system introduced by Robert Dale in 1989, which addresses the generation of referring expressions (REs) in the context of cooking recipes. By employing a specialized ontology for masses and sets, a state-based model of ingredient transformation, and a two-level semantic architecture, EPICURE produces efficient and unambiguous references. The system demonstrates mastery over complex linguistic phenomena such as pronominalization, one-anaphora, and elision, guided by a mathematical metric of discriminatory power.
Introduction: Beyond Singular Individuals
In the late 1980s, most natural language generation research assumed a world populated by static, singular individuals. However, real-world tasks like following a recipe involve entities that are far more complex. Ingredients often exist as masses (e.g., "two ounces of salt") or sets (e.g., "three pounds of carrots"). More importantly, these entities are dynamic; they change state throughout a discourse. A set of whole carrots may be peeled, then grated into a mass.
Traditional systems struggled to maintain identity across these shifts. If a recipe instructs a cook to "Drain and rinse them," the system must correctly identify "them" as the specific beans previously mentioned, despite their change in state. EPICURE was designed to navigate this "messy" domain by generating REs that are both adequate for identification and linguistically efficient.
Generalized Physical Objects and State-Based Modeling
To handle the complexity of the kitchen, EPICURE moves beyond simple individual identifiers. It utilizes an ontology of Generalized Physical Objects (physobj) represented as Knowledge Base Entities (KBEs).
As shown in Figure 2, a KBE captures the structure (set vs. mass), quantity, and physical properties of an ingredient. Crucially, EPICURE models the execution of the recipe. Every operation—such as chopping, mixing, or grating—triggers the creation of a new KBE.
This state-based modeling (Figure 1) allows the system to track the ingredient's transformation and refer to it accurately at any point in the process, ensuring that "the carrots" correctly identifies the grated mass rather than the original set.
Two-Level Semantics: From Deep to Surface
EPICURE bridges the gap between internal data and English prose using a two-level semantic framework:
- Deep Semantic Structure (DSS): This level represents the "recoverable" content—the essential information the hearer needs to identify the referent.
- Surface Semantic Structure (SSS): This level maps more directly to the final syntactic form. It handles linguistic nuances like elision (omitting redundant words) and one-anaphora (e.g., "the red one").
The transition from DSS to SSS is governed by the principle of efficiency. Following Gricean maxims, the system avoids over-specification. If a single bowl of carrots is the only one present, the system will generate "the carrots" rather than "the three pounds of orange carrots."
Algorithmic Precision: Discriminatory Power
A key innovation in EPICURE is the Discriminatory Power Algorithm. When choosing adjectives (e.g., "the small red capsicum"), the system calculates a value $F$ for each potential attribute:
$$F(<a,v>, U) = \frac{N-n}{N-1}$$
Here, $N$ is the number of entities in the current context ($U$), and $n$ is the number of entities that possess the attribute-value pair $<a,v>$. This formula measures how effectively an attribute filters out "distractor" entities. The system iteratively selects attributes with the highest discriminatory power until the target entity is uniquely identified.
Results and Linguistic Capabilities
The final SSS is realized into English using a PATR-like unification grammar. EPICURE successfully generates a variety of sophisticated forms:
- Pronominalization: Using "it" or "them" by tracking the "discourse centre" (the result of the most recent operation).
- One-Anaphora: Generating "the small red one" by comparing the current entity's category with previously mentioned ones.
- Elision: Omitting redundant information common in "recipe-speak," such as "Add the garlic [$\phi$]" when the onions are the current focus.
Limitations and Future Directions
While pioneering, EPICURE has limitations. Its discriminatory power metric is currently optimized for "absolute" adjectives (color, shape) and struggles with relative terms like "large." Additionally, its model for managing the "context" of distractors is relatively simple, benefiting from the small number of ingredients typical in recipes. Future work would require more robust focus space management and formal quantitative evaluation of the generated text's effectiveness compared to human-authored instructions.
Summary updated