TL;DR
This paper tackles the challenging task of aligning sentences in bilingual literary works, where translations often deviate from strict sentence-by-sentence correspondence. They propose a novel multi-pass approach that leverages a discriminative classifier trained on high-confidence alignments to improve recall and identify problematic translation passages, outperforming existing open-source tools.
What Problem Does This Paper Address?
Aligning bilingual texts (bitexts) is crucial for various applications like language learning aids, translation studies, and machine translation. While sentence alignment is generally considered a solved problem for many text types, literary works present a unique challenge. Translations of literature often involve stylistic choices, paraphrasing, and structural changes that break the one-to-one sentence correspondence assumed by many standard alignment algorithms. This leads to a significant portion of the text being left unaligned, losing valuable information for downstream tasks. The paper aims to address this by developing a more robust alignment method specifically for literary texts, focusing on recovering more alignments, including many-to-one and one-to-many correspondences, and improving the precision of these alignments.
Methods
The authors propose a multi-pass approach that builds upon existing sentence alignment tools. The core of their method involves:
- Initial Alignment: They start with the alignments produced by a high-precision, but low-recall, tool (Moore's system). These are considered "sure" one-to-one alignments.
- Discriminative Classifier Training: They train a Maximum Entropy (MaxEnt) classifier using these high-confidence alignments as positive examples. Negative examples are generated by pairing a positive sentence with sentences that immediately follow it in the target text. The classifier learns to predict the probability of a sentence pair being a valid alignment.
- Feature Engineering: The MaxEnt classifier uses a rich set of features, including:
- Lexical Features: Presence of specific word pairs.
- Length Features: Ratio of sentence lengths.
- Cognate Features: Words with similar spelling and likely etymological origin.
- Copy Features: Identical words (e.g., proper nouns) appearing in both sentences.
- Filling Alignment Gaps: The trained classifier is then used to score candidate alignments in the regions not covered by the initial high-precision alignments. A greedy search strategy is employed to iteratively select the best scoring alignment blocks and fill the gaps, aiming to recover many-to-one and one-to-many correspondences.
Datasets
The paper uses two main types of datasets:
- BAF Corpus: A benchmark corpus for sentence alignment, specifically the literary subpart which includes a hand-checked alignment of Jules Verne's "From the Earth to the Moon."
- Small Bilingual Library: A newly compiled corpus consisting of ten 19th and early 20th-century literary works (five English classics translated into French, and five French classics translated into English). These were sourced from public domain repositories like Project Gutenberg and Wikisource. For evaluation on this corpus, a subset of the first two to three hundred sentences of each book was manually aligned by an annotator.
Key Insights and Ideas
- Literary texts are fundamentally different: The assumption of strict sentence-to-sentence alignment breaks down significantly in literary translations, requiring specialized approaches.
- Precision vs. Recall Trade-off: Standard tools often prioritize precision, discarding many potentially valid but non-one-to-one alignments. The proposed method aims to improve recall without sacrificing too much precision.
- Discriminative models are powerful: Using a MaxEnt classifier trained on high-quality examples allows for learning complex patterns beyond simple length or lexical matching.
- Feature engineering is key: The combination of lexical, length, cognate, and copy features provides a robust signal for alignment.
- Identifying problematic passages: The method not only improves alignment but also helps pinpoint areas where translation is particularly divergent, which is valuable for translation studies.
Key Results
- The proposed multi-pass approach significantly outperforms two open-source baseline tools (Moore and Hunalign) on both link-based and sentence-based F-measures, especially in terms of recall.
- For the "real-world" literary corpus, the method salvages many useful sentence pairs that would otherwise be left unaligned by baseline systems.
- While Moore's system maintains high precision even on difficult texts, the proposed method achieves comparable precision with substantially higher recall.
- The performance varies across books, with some (like "Du côté de chez Swann") being easier to align due to closer adherence to the source style, while others (like "Emma") are more challenging due to older translations or significant stylistic departures.
Limitations and Future Work
- Manual alignment effort: The evaluation on the "small bilingual library" relied on manually aligned subsets, which is labor-intensive.
- Difficulty with certain texts: Some books, particularly those with older translations or significant stylistic deviations, remain challenging to align accurately.
- Future Work:
- More resources: Incorporating larger training corpora, bilingual dictionaries, and more sophisticated features (e.g., part-of-speech tags, lemmas).
- Tighter integration: Improving the integration with existing alignment tools like Moore's.
- Exact search: Replacing the greedy search with an exact search algorithm for potentially better results.
- Larger scale evaluation: Testing the approach on a much larger and more diverse set of books across multiple languages.
- Handling OCR errors: Extending the approach to handle noisy OCR outputs.
Summary updated