Introduction
In the early 1990s, the development of robust machine translation systems and comprehensive bilingual lexicons was a primary goal in computational linguistics. While methods existed for aligning sentences or individual words, they often overlooked the underlying linguistic structure. This paper addresses the challenge of automatically aligning noun phrases (NPs) between English and French versions of the same text.
The core difficulty lies in two areas: the "Long Tail" of language, where many words appear only once (hapax legomena), and structural mismatches, where a single English word might correspond to a multi-word French collocation. Julian Kupiec proposed an iterative statistical algorithm that uses noun phrases as the primary unit of alignment, providing a linguistically grounded approach to processing bilingual data.
Methodology
The proposed system utilizes a pipeline that combines symbolic natural language processing with statistical estimation.
The process begins with tagging and recognition. English and French Hidden Markov Model (HMM) based taggers assign part-of-speech categories to the text. Subsequently, finite-state recognizers extract "simple" noun phrases, such as "the Prime Minister" or "le Premier Ministre."
The heart of the methodology is an iterative re-estimation algorithm, a specific application of the Expectation-Maximization (EM) framework (similar to the Baum-Welch algorithm). Starting from a state of "total ignorance"—where every English NP in a sentence is assumed equally likely to correspond to every French NP in the aligned sentence—the algorithm iteratively boosts the probability of pairs that co-occur frequently across the entire corpus. To maintain computational efficiency in long sentences, a "windowing" technique is employed, restricting the search for correspondences to a specific range based on the relative position of the source phrase.
Dataset and Implementation
The algorithm was tested using the Canadian Hansards, a standard bilingual corpus containing official records of the Canadian Parliament in English and French. For the study, approximately 2,600 sentence alignments were used, from which roughly 4,900 distinct English NPs and 5,100 distinct French NPs were extracted.
Key Insights
Several critical insights emerged from this approach:
- Linguistic Anchoring: By focusing on NPs, the search space for correspondences is significantly reduced, ensuring syntactically meaningful results.
- Robustness to Low Frequency: The algorithm can successfully map phrases appearing only once in the corpus, provided they are the sole NPs in their respective sentence alignments.
- Rapid Convergence: Mappings typically stabilize within just two to four iterations.
- Reversibility: The mathematical model is inherently reversible, allowing for seamless French-to-English or English-to-French mapping.
Results and Evaluation
Manual inspection of the 100 highest-ranking correspondences revealed a high precision rate of 90%. The system demonstrated a remarkable ability to map complex, multi-word phrases accurately.
As shown in the results, the algorithm correctly identified correspondences such as "Atlantic Canada Opportunities Agency" and its French counterpart, "Agence de promotion économique du Canada atlantique." It also handled acronyms like "DREE" to "MEER" effectively.
Limitations and Future Directions
Despite its successes, the algorithm faces challenges with non-literal or "creative" translations. For instance, it struggled with alliterative phrases like "roaring rabbit" when the translation shifted literal meanings to maintain a rhyme or stylistic effect.
Structural complexities also posed issues; the current NP recognizer sometimes fails on conjunctive modifiers (e.g., "off-the-job training"), leading to fragmented mappings. Furthermore, prepositional phrase attachment (e.g., "of State") remains ambiguous. Future work may involve using the algorithm itself to resolve these structural ambiguities by comparing likelihoods across the corpus or applying the method nestedly to align individual words within the identified phrases.
Summary updated