High Level Memory Structures and Text Coherence in Translation

TL;DR

This paper explores how high-level memory structures, specifically "scripts," are crucial for accurate translation, particularly when dealing with sentences that describe events with causal relationships. It argues that understanding these underlying causal connections, often implicit in the source language, is essential for generating grammatically correct and semantically meaningful translations in the target language (Chinese, in this case).

What Problem Does This Paper Address?

The paper tackles the challenge of translating sentences where lexical items lack direct equivalents in the target language, and more importantly, where the meaning of the sentence relies on understanding implicit causal relationships between events. Traditional translation methods might struggle with this because they focus on word-for-word or phrase-level translation without grasping the underlying narrative or event structure. The authors highlight that simply translating individual events can lead to nonsensical or grammatically incorrect sentences if the causal links between them are ignored, especially in languages like Chinese where explicit connectives might be omitted when causality is clear from context.

Methods

The core of the paper's approach is to leverage "scripts" – a concept from Artificial Intelligence that represents stereotypical sequences of events in a particular domain (e.g., a restaurant script). The authors propose that when two events in a sentence are part of the same script and are causally related, they can be conjoined in the target language (Chinese) without explicit connectives like "and" (HEH-). Conversely, if events are not causally related, even if they belong to the same script, a connective is necessary. This "conjoinability criterion" is presented as a key mechanism for determining when and how to translate conjunctions. The paper also touches upon the generation of specific words or markers (like "GIO" in the Chinese translation) that signal violations of expected script-based event sequences.

Datasets

Not discussed in the paper. The examples provided are hypothetical sentences and their translations, not derived from a formal dataset.

Key Insights and Ideas

  • Scripts as Translation Guides: High-level memory structures like scripts are not just for understanding; they are vital for generating coherent translations by capturing implicit causal relationships.
  • Causality Dictates Conjunctions: The presence or absence of explicit conjunctions (like "and") in the target language is determined by the causal inferability of the events described. If events are causally linked within a script, connectives can often be omitted.
  • "GIO" as a Violation Marker: Specific words or markers in the target language can signal that an event sequence deviates from the expected script, preventing misunderstandings.
  • Beyond Word-for-Word: Effective translation requires moving beyond literal word or phrase mapping to understanding the underlying conceptual structure and narrative flow.

Key Results

The paper doesn't present quantitative results in terms of accuracy metrics. Instead, it offers qualitative demonstrations through examples:

  • Example 1 (Restaurant Order): "John went to a restaurant and ordered a hamburger." The "and" is omitted in the Chinese translation because ordering a hamburger is a direct, expected event within the restaurant script.
  • Example 2 (Reading and Writing): "John read a book and wrote a letter." The "and" is translated because reading a book and writing a letter are not inherently causally linked events within a common script.
  • Example 3 (Eating and Drinking): "John went to a restaurant where he ate a hamburger and drank a cup of tea." Here, even though both are restaurant events, the "and" (HEH-) is required because eating and drinking are not causally dependent on each other.
  • Example 4 (Violating Script): The translation of "He left without paying the check" uses a marker ("GIO") to indicate that leaving without paying is a violation of the restaurant script, which would not be the case if the marker were omitted.

Limitations and Future Work

The authors acknowledge that their discussion is a "scratch on the surface" and that there are many other types of high-level memory structures beyond scripts. They are currently working on how different memory structures relate to the generation of coherent text. A key limitation implied is the reliance on human intuition to define scripts and causal relationships, which would need to be formalized for a computational system. The paper also doesn't detail how these scripts are computationally represented or accessed for translation.

Summary updated

BibTeX
@inproceedings{1981high,
  title = {{High Level Memory Structures and Text Coherence in Translation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure