Language Generation From Conceptual Structure: Synthesis of German in a Japanese/German MT Project

J. Laubsch, D. Roesner, K. Hanakata, A. LesniewskiACL 1984

TL;DR

This paper introduces SEMSYN, a novel module for generating German text from semantic representations, specifically designed for a Japanese-German machine translation (MT) project. It details a three-stage generation process that leverages distinct knowledge sources, moving from language-independent conceptual structures to syntactically and morphologically correct German output.

What Problem Does This Paper Address?

The core challenge addressed is the generation of natural-sounding German text from a machine-readable semantic representation, a crucial step in machine translation. Specifically, the authors tackle the problem of bridging the gap between abstract semantic meaning and the complex linguistic structures of German, particularly within the context of translating Japanese technical documentation. They aim to demonstrate the feasibility of using a semantic representation as an "interlingua" for practical MT applications, moving beyond simple title translation to potentially handle abstracts and full texts.

Methods

SEMSYN employs a three-stage generation process:

  1. Conceptual Stage (Stage 1): This stage is language-independent and focuses on mapping the input semantic net (produced by Fujitsu's ATLAS/II system) to instantiated knowledge base schemas (IKBS). These schemas represent concepts and actions using roles and transformation rules. The key innovation here is the use of both case schemata for actions and concept schemata for nouns/qualities, allowing for a rich representation of meaning.
  2. Realization Stage (Stage 2): This stage translates the IKBS into instantiated realization schemata (IRS). This involves selecting appropriate German lexical entries and mapping semantic roles to functional grammatical roles. The paper highlights the use of "realization schemata" (RS) which specify syntactic categories (like CLAUSE or NP) and their functional roles, drawing inspiration from Winograd's functional grammar. A significant aspect is the handling of ambiguity by checking fillers of semantic roles to select the most adequate German word sense (e.g., choosing "implementieren" over "realisieren" based on the "Instrument" role).
  3. Surface Generation (Stage 3): While not detailed in this paper, this stage constructs the final German surface string using knowledge about syntax, morphology, and style.

The paper also introduces a mechanism for handling references and focusing using "descriptor" fillers, allowing references to parts of other schemata, which can lead to relative clauses or focused phrases.

Datasets

The paper uses Japanese technical documentation titles and abstracts from the field of "Information Technology" as input for the MT system. These are processed by Fujitsu's ATLAS/II system to produce the semantic nets that serve as input to SEMSYN. No new benchmarks are introduced; rather, existing technical documentation is used for evaluation.

Key Insights and Ideas

  • Semantic Representation as an Interlingua: The paper strongly advocates for the use of a rich semantic representation as a language-independent interlingua for MT, demonstrating its practical application.
  • Hierarchical Knowledge Representation: The distinction between case schemata (for actions) and concept schemata (for nouns/qualities) provides a structured way to represent meaning, allowing for more nuanced generation.
  • Context-Sensitive Lexical Choice: The system's ability to select the most appropriate German word sense based on the fillers of semantic roles (e.g., "realisieren" vs. "implementieren") is a crucial insight for generating natural language.
  • Flexible Realization Schemata: The use of realization schemata that can be inherited and adapted allows for efficient generation and handling of various linguistic constructions.
  • Handling References and Focus: The descriptor mechanism for referencing sub-structures within the semantic representation is a clever way to manage complex linguistic phenomena like relative clauses and focus.

Key Results

The paper focuses on describing the process and design of the first two stages of SEMSYN, rather than presenting quantitative results. The key qualitative finding is the successful demonstration of a structured, multi-stage approach to generating German from semantic representations, capable of handling complex semantic relationships and leading to potentially natural output. The example walk-through illustrates how the system can generate sentences like "Es wird gewünscht, dass ein Sprecher mehrere Zwecke mit einer einzelnen Äußerung erreicht" from a Japanese input.

Limitations and Future Work

The authors explicitly state that this paper describes the first two stages of generation, with the third stage (surface generation) not detailed. They acknowledge that the extensibility to abstracts and full texts is envisioned but not fully demonstrated in this work. Future work would involve completing the surface generation stage and further testing the system's capabilities on more complex texts. The paper also mentions experiments with other generation modules (SUTRA, IPG, MUMBLE) to explore potential front-ends for SEMSYN, suggesting an ongoing effort to integrate and refine the generation process.

Summary updated

BibTeX
@inproceedings{laubsch1984language,
  author = {J. Laubsch and D. Roesner and K. Hanakata and A. Lesniewski},
  title = {{Language Generation From Conceptual Structure: Synthesis of German in a Japanese/German MT Project}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1984},
}
Figure