High Level Memory Structures and Text Coherence in Translation

C. J. Yang

Abstract

Effective machine translation requires more than a mapping of lexical items between languages; it necessitates a deep understanding of the situational context and the causal relationships between events. This paper explores the role of high-level memory structures, specifically "scripts," in determining the appropriate use of connectives in Mandarin Chinese translation. By leveraging Conceptual Dependency (CD) and script-based reasoning, the proposed framework identifies when events are "conjoinable" based on world knowledge, allowing for more natural and coherent target-language generation. The research demonstrates that text coherence is fundamentally tied to the system's ability to model human expectations and recognize deviations from stereotypical event sequences.

Introduction

A persistent challenge in machine translation is the "lexical gap"—instances where the target language handles logical connections and transitions differently than the source language. In English, the conjunction "and" serves as a ubiquitous tool for joining events, often regardless of their underlying causal relationship. However, in Mandarin Chinese, the inclusion or omission of connectives is highly sensitive to the perceived relationship between actions.

This paper argues that high-quality translation is impossible without "world knowledge." To produce a translation that sounds natural to a native speaker, a system must understand the "why" behind the sequence of events. By utilizing memory structures to model these relationships, we can bridge the gap between literal word-for-word translation and true situational understanding.

Theoretical Framework: Scripts and Conceptual Dependency

The methodology presented here is rooted in the Symbolic AI tradition, specifically the work of Roger Schank and Robert Abelson. The system relies on two primary components:

  1. Conceptual Dependency (CD): A language-independent representation of meaning. Rather than storing words, the system parses sentences into primitive actions and roles, focusing on the underlying concepts.
  2. Scripts: High-level memory structures that represent stereotypical sequences of events in specific contexts. For example, a "Restaurant Script" includes a predictable chain: entering, being seated, ordering, eating, and paying.

By combining these, the system can interpret a sentence not just as a string of symbols, but as a series of points within a known human activity.

The Conjoinability Criterion

The core insight of this research is the Conjoinability Criterion. In Mandarin Chinese, the grammar is inherently sensitive to the causal structure of the world. If two events are part of the same "causal chain" within a script, they are considered conjoinable.

In such cases, Mandarin typically "forbids" the use of an explicit connective like HEH- (and). Because the listener can easily infer the connection through their own world knowledge, the language treats the sequence as a single, fluid progression. Conversely, if two events are independent or lack an inherent causal link, a connective is required to bridge them.

Case Studies in Translation

To demonstrate this phenomenon, the paper examines several hand-crafted scenarios:

  • Causally Linked Events: Consider the sentence, "John went to a restaurant and ordered a hamburger." In a restaurant script, ordering follows entering. Because these are causally linked, the "and" is omitted in the Mandarin translation to maintain natural flow.
  • Independent Events: Consider the sentence, "John read a book and wrote a letter." While both are actions performed by John, there is no inherent script that links reading a book to writing a letter. In this case, the Mandarin translation must preserve the connective (HEH-) to signal the transition between two distinct activities.

Signaling Script Violations

World knowledge is equally important when events deviate from expectations. If a story describes a violation of a standard script—such as John leaving a restaurant without paying—the translation system must recognize this as an anomaly.

To maintain text coherence, the system must use specific markers (such as GIO or GUT) to signal the violation to the reader. A system without script knowledge might produce a translation that implies a false causality (e.g., "John didn't pay, so he left"), which would be confusing or nonsensical to a native speaker. By understanding the script, the system knows that "leaving" after "not paying" is an unusual sequence that requires explicit emphasis to be understood correctly.

Conclusion and Future Directions

This research posits that translation is not merely a linguistic mapping but a process of understanding a situation and re-describing it within the constraints of the target language's cultural and logical framework. The use of scripts provides a powerful, albeit limited, mechanism for capturing the world knowledge necessary for this task.

However, scripts represent only one layer of human memory. Future work must incorporate more flexible structures, such as Memory Organization Packets (MOPs), plans, and goals, to handle the full complexity of human interaction. While the current scope is limited to specific connectives, it lays the groundwork for a generation of translation systems that prioritize coherence and situational context over simple lexical substitution.

Summary updated

BibTeX
@inproceedings{yang1981high,
  author = {C. J. Yang},
  title = {{High Level Memory Structures and Text Coherence in Translation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure