TL;DR
This paper tackles the challenge of accurately translating discourse markers (DMs) between Chinese and English, a crucial but often overlooked aspect of machine translation (MT). By focusing on improving word alignments for DMs, the authors demonstrate significant gains in translation quality, suggesting that better handling of these cohesive devices is key to more natural and coherent MT outputs.
What Problem Does This Paper Address?
Statistical Machine Translation (SMT) systems, while impressive, often struggle with producing high-quality translations, especially for distant language pairs like Chinese and English. A significant reason for this is their tendency to translate sentences in isolation, neglecting the crucial role of discourse markers (DMs). DMs are words or phrases (like "because," "however," "therefore") that connect ideas, signal relationships between sentences, and contribute to the overall flow and coherence of text.
The problem is twofold:
- Divergence in Usage: DMs are used differently across languages in terms of placement, frequency, and the specific discourse relations they signal. This makes direct, one-to-one translation difficult.
- Implicit DMs and Cohesion: Chinese, in particular, relies heavily on implicit discourse relations and uses DMs in ways that can be subtle or even omitted, making them hard for current SMT systems to identify and translate correctly. This lack of explicit markers leads to poor word alignments, which in turn degrades the quality of translation rules and the final output.
The paper aims to improve MT by explicitly modeling these cohesive devices and discourse relations, moving beyond sentence-level translation.
Methods
The core of the paper's approach involves improving word alignments, particularly for discourse markers. The methodology can be broken down as follows:
- Corpus Analysis: The researchers analyzed three parallel corpora (BTEC, FBIS, and TED Talks) to understand the challenges of translating DMs between Chinese and English. This involved identifying instances of DMs and examining how they were translated.
- Identifying and Annotating DMs: The paper discusses existing work on identifying and annotating DMs, highlighting the complexity of Chinese DMs, including interchangeable markers and implicit relations.
- Improving Word Alignments: The key innovation lies in addressing misalignments related to DMs. When an explicit English DM was not aligned correctly with its Chinese counterpart, or when a DM was entirely unaligned, an artificial marker was inserted into the Chinese sentence in the nearest free space. This artificial marker acts as a placeholder, guiding the alignment process.
- Hierarchical Phrase-Based SMT: The experimental systems are built upon hierarchical phrase-based SMT, a standard approach. The improvements are achieved by refining the word alignments that feed into this system.
- Two Main Research Areas: The thesis is structured around two key areas:
- Modelling cohesive devices within sentences: Focusing on how DMs and other cohesive elements function within a single sentence.
- Modelling discourse relations across sentences and wider discourse segments: Extending the analysis to how DMs connect larger chunks of text.
Datasets
The research utilizes the following parallel corpora:
- Basic Travel Expression Corpus (BTEC): Primarily short phrases from travel conversations.
- Foreign Broadcast Information Service (FBIS) corpus: News stories and radio podcasts.
- Ted Talks corpus (TED): Approved translations of live Ted Talks presentations.
These corpora are used for both analysis and training/evaluation of the experimental SMT systems. No new benchmark datasets were introduced in this paper.
Key Insights and Ideas
- Word Alignment is Key: The paper strongly suggests that improving word alignments, especially for discourse markers, is a critical bottleneck for achieving better MT quality. Poor alignments directly lead to poorer translation rules.
- Artificial Markers as a Solution: The novel idea of inserting artificial markers for unaligned or misaligned DMs is a practical way to "force" the alignment process to recognize these important linguistic elements, even when they are implicit or difficult to map.
- Chinese Implicitation: Chinese's tendency to rely on implicit discourse relations poses a significant challenge, as SMT systems are often trained on explicit cues. Making these implicit relations explicit, even artificially, can aid translation.
- Cohesion Matters Beyond Sentences: While the paper focuses on sentence-level improvements, it clearly sets the stage for modeling discourse relations across wider segments, recognizing that true coherence requires understanding text beyond individual sentences.
Key Results
- Improved Word Alignments: The experiments showed positive improvements in word alignments, with artificial markers helping to capture unaligned DMs like "then." Figures 2 and 3 visually demonstrate how the artificial marker improved the alignment of "then" and other words in the sentence.
- BLEU Score Improvements: The experimental systems, particularly the one incorporating all markers simultaneously, showed significant improvements in BLEU scores. The final system achieved improvements of up to +0.44 over the baseline, which is considered a substantial gain in MT evaluation.
- Corpus Variance: Misalignment percentages for DMs varied significantly across corpora, with the FBIS corpus showing the highest error rates, likely due to its longer average sentence length and more complex text.
Limitations and Future Work
- Limitations:
- The paper acknowledges that making implicit markers explicit can sometimes seem unnatural or unnecessary to human readers, but argues it's beneficial for the MT process.
- The current work is still in its early stages, particularly regarding the translation of implicit discourse relations.
- The artificial insertion algorithm is described as "simple" and inspired by examples, suggesting room for refinement.
- Future Work:
- Expand DM Coverage: Explore a wider range of DMs, including more complex ones like "as soon as."
- Refine Variance List: Improve the list of variant translations for marker words and develop automated filtering for difficult DMs (e.g., "and," "so" when not used as specific markers).
- Enhance Insertion Algorithms: Develop more sophisticated algorithms for inserting artificial markers to produce a better range of options and reduce damage to existing alignments.
- Alternative Evaluation Metrics: Explore evaluation metrics beyond BLEU that are better at capturing individual linguistic components like DMs and pronouns.
- Prediction Model: The ultimate goal is to develop a true prediction model using parallel data to guide word alignment, potentially by adding constraints or providing additional implicit annotation.
- Document-Level Modeling: Extend the work to model discourse relations across wider discourse segments and entire documents.
Summary updated