Abstract
Collocations are notoriously difficult for non-native speakers to translate, primarily because they are opaque and cannot be translated on a word-by-word basis. We describe a program named Champollion which, given a pair of parallel corpora in two different languages, automatically produces translations of an input list of collocations. Our goal is to provide a tool to compile bilingual lexical information above the word level in multiple languages and domains. The algorithm we use is based on statistical methods and produces p-word translations of n-word collocations in which n and p need not be the same; the collocations can be either flexible or fixed compounds. Testing and evaluation of Champollion on one year's worth of the Hansards corpus yielded 300 collocations and their translations, evaluated at 77% accuracy. In this paper, we describe the statistical measures used, the algorithm, and the implementation of Champollion, presenting our results and evaluation.
Introduction
In linguistics, a collocation is a sequence of words that co-occur more often than would be expected by chance. For non-native speakers and machine translation (MT) systems, these phrases are often "opaque": one cannot simply translate the individual components and expect a natural result. For example, in English, we "make" a decision, but in French, the equivalent is "take" a decision (prendre une décision).
While single-word translation was a maturing field in the early 1990s, translating flexible, multi-word units remained a significant hurdle. Existing tools often struggled with "flexible" collocations where words might be separated by other text (e.g., "to take several unnecessary steps"). Champollion was developed to automate the creation of bilingual lexicons that include these complex expressions by leveraging parallel corpora.
The Champollion System
Named after the linguist who deciphered the Rosetta Stone, Champollion follows a robust statistical pipeline to identify equivalents across languages:
- Preprocessing: The system utilizes a parallel corpus aligned at the sentence level.
- Source Extraction: It accepts a list of source-language collocations, typically identified by tools like Xtract.
- Statistical Correlation (The Dice Coefficient): To identify the best translation, Champollion calculates the correlation between the source phrase and potential target words. The researchers specifically chose the Dice Coefficient over Mutual Information (MI). They argued that MI over-weights low-frequency events and is too symmetric, whereas the Dice Coefficient better highlights "one-to-one" matches in translation contexts.
- Iterative Algorithm: To avoid an exponential search of every possible word combination in the target language, Champollion uses an iterative approach. It identifies highly correlated single words, then pairs, then triplets, pruning any combinations that fall below a correlation threshold.
- Word Ordering: Finally, it examines the corpus to determine if the translated words appear in a fixed or flexible order, labeling the resulting entry accordingly.
Experimental Results and Evaluation
The system was tested using the Hansard Corpus, containing bilingual transcripts of the Canadian Parliament. The researchers used two primary datasets: DB1 (8 months of data from 1986) and collocation lists (C1, C2) extracted from different years to test the system's robustness across time periods.
As shown in the table above, the system achieved 77% accuracy (Overall) when the source collocations and the translation database were from the same time period (C1/DB1). However, accuracy dropped to 61% when testing on a different year's data (C2/DB1), highlighting that collocations are highly sensitive to the specific domain and time period of the corpus. Evaluation was conducted by three bilingual human judges, who showed high consistency in their ratings.
Translation Examples
Champollion is capable of producing translations where the number of words in the source and target do not match, and it handles both fixed compounds and flexible phrases.
The examples in the table demonstrate the system's ability to handle complex mappings, such as "employment equity" becoming "équité ... matière ... emploi" or "to take steps" becoming "prendre ... mesures."
Discussion and Limitations
While Champollion represents a significant step forward in automated lexicography, it faces certain challenges:
- The "Preposition Problem": "Closed-class" words like prepositions (of, to, in) appear so frequently in every context that their statistical correlation is often "noisy," leading to translations that sometimes omit necessary functional words.
- Data Sparsity: Statistical measures fail if a collocation does not appear frequently enough in the training corpus. Increasing the corpus size is the primary method to alleviate this.
- Dependency on Input: Champollion requires a pre-existing list of collocations in the source language; it finds equivalents rather than discovering new collocations autonomously.
Despite these limitations, the system provides a powerful tool for compiling bilingual lexical information above the word level, bridging a critical gap in machine translation and language learning resources.
Summary updated