CRITTER: a translation system for agricultural market reports

Pierre Isabelle, Marc Dymetman, Elliott MacklovitchCOLING 1988 · OpenReview

Developed in the late 1980s at the Centre Canadien de Recherches sur l'Informatisation du Travail, CRITTER represents a significant milestone in the history of machine translation (MT). It was designed specifically to translate agricultural market reports between English and French, focusing on a highly specialized "sublanguage." Unlike many contemporary systems of its era, CRITTER was built on the principle of reversibility, ensuring that the same linguistic rules govern both the parsing of a source language and the generation of a target language.

The Reversibility Paradigm

In the 1980s, MT systems were typically unidirectional pipelines. A system designed to translate English to French could not easily be reversed to translate French to English. This often led to linguistic inconsistencies, such as parsers that were too permissive (overgeneration) or generators that were too rigid (undergeneration).

CRITTER addressed this by employing a reversible architecture. The authors argued that a truly adequate grammar must be bidirectional: if a grammar can parse a sentence, it should be able to generate it. To achieve this efficiently in Prolog, they utilized a "double-compilation" technique, where a single set of declarative rules is compiled into two optimized versions—one for analysis and one for synthesis.

Syntactic Analysis with Feature Structures

CRITTER utilizes Extraposition Grammars, a type of logic grammar that excels at handling complex syntactic phenomena like long-distance dependencies. Instead of using simple labels for nodes, the system employs complex "feature structures." These are sets of attribute-value pairs that allow for the precise coordination of linguistic properties like gender, number, and tense across a sentence.

Figure 2: A syntactic phrase structure tree for a sample sentence, showing the use of feature structures like number (NBR) and tense.

As seen in the syntactic tree above, the system decomposes a sentence like "last week hog prices in Saskatchewan increased..." into a hierarchical structure where each node carries detailed grammatical metadata. This level of detail ensures that the generated output in the target language maintains the correct agreement and inflection.

Semantic Representation and the "Inv" Trick

Once a sentence is parsed, CRITTER represents its meaning as a Directed Acyclic Graph (DAG). To maintain a structure that is easy for recursive algorithms to process, the researchers introduced "inverse" argument labels (e.g., inv-1). This allows the system to treat a verb as the root of a structure even when it is modified by an adverb, maintaining a consistent flow from the head of a phrase to its dependents.

Figure 3: A semantic DAG representing the meaning of the sentence, annotated with semantic types like MOVEMENT, EVENT, and COMMODITY.

This semantic layer is where the system’s "sublanguage" focus becomes most apparent. By targeting agricultural reports, the researchers could define specific semantic validating schemas. For instance, an "increase" event is defined as requiring a measure-function (the price), an increment (the percentage), and a base measure (the dollar amount). This semantic hierarchy, shown in the annotations in Figure 3, acts as a quality filter to ensure the translation is contextually appropriate.

Key Insights and Results

CRITTER demonstrated that focusing on a restricted domain—such as weekly reports from the Canadian Department of Agriculture—allows for much higher translation accuracy. The system successfully neutralized source-language-specific quirks during analysis, allowing the transfer layer to focus on pure meaning.

While the prototype initially launched with a dictionary of 500 lexemes per language, its architecture proved the viability of bidirectional logic grammars for real-world applications. Although the system faced challenges with overgeneration in synthesis mode and required further refinement of its semantic hierarchy, it remains a pioneering example of how formal linguistic constraints can be used to build robust, reversible translation engines.

Summary updated

BibTeX
@inproceedings{isabelle1988critter,
  author = {Pierre Isabelle and Marc Dymetman and Elliott Macklovitch},
  title = {{CRITTER: a translation system for agricultural market reports}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
  url = {https://openreview.net/forum?id=KuQT24H8VCM},
}
Figure