Statistical Machine Translation (SMT) has long been a foundational approach for Grammatical Error Correction (GEC), treating the correction process as a translation task from "erroneous English" to "fluent English." However, standard phrase-based SMT often suffers from a "local vision" problem, where it excels at fixing short-range errors but fails to maintain global grammatical consistency, such as long-distance subject-verb agreement.
This paper introduces a novel two-stage approach to address these limitations by moving beyond the standard practice of selecting a single "best" translation. Instead, the authors propose an "edit selection" method that synthesizes a superior correction by combining the most reliable individual edits from across multiple candidate hypotheses.
Methodology: Beyond the 1-Best Hypothesis
The core of the proposed system is a two-stage pipeline designed to inject global context into the SMT framework.
1. The Edit Classifier
The system first extracts every "edit" (the specific change from the source sentence to a candidate correction) from the SMT system's $n$-best hypotheses. A binary classifier is then trained to distinguish between valid corrections and erroneous changes. Unlike the base SMT system, this classifier leverages features that capture broader linguistic context:
- Syntactic Context: Utilizing a parser to identify the nearest noun phrase (NP) or verb phrase (VP) heads to ensure grammatical coherence (e.g., checking if a verb change aligns with its distant subject).
- Language Model Scores: Evaluating the fluency of the sentence before and after an edit using a massive 5-gram model trained on Wikipedia.
- Lexical and POS Features: Analyzing specific word choices and their parts of speech to identify common error patterns.
2. Exploiting the N-Best List
With a trained edit classifier, the authors propose two strategies for improving the final output:
- Reranking: The classifier's scores are used as a new feature to re-sort the $n$-best list, allowing the system to pick a better overall sentence than the original SMT decoder.
- Edit Selection: This novel approach identifies all "valid" edits across the top $n$ hypotheses. It then greedily selects the highest-scoring, non-overlapping edits and applies them to the original source sentence. This allows the system to "Frankenstein" a new sentence that may be more accurate than any single hypothesis produced by the decoder.
Experimental Results
The system was evaluated on the CoNLL-2014 shared task benchmark, a standard for GEC performance. The results demonstrate that both reranking and edit selection provide significant gains over the baseline SMT system.
As shown in the results, the Edit Selection (5-best) method achieved the highest $F_{0.5}$ score of 41.19, a substantial improvement over the 40.58 baseline. For comparison, the top-performing system in the original CoNLL-2014 shared task (CAMB) reached an $F_{0.5}$ of 37.33. The primary driver of this improvement was increased precision—the system became significantly better at avoiding "bad" corrections that would introduce new errors.
Insights and Future Directions
A key takeaway from the ablation studies is that "Context Features" and "Language Model Features" are the most critical contributors to the classifier's success. This confirms that the primary weakness of traditional SMT in GEC is indeed its lack of global structural awareness.
While the system is currently limited by the SMT decoder's ability to generate good edits in its $n$-best list, the "edit selection" framework is inherently general. Future work could involve combining hypotheses from diverse GEC architectures—such as rule-based, neural, and statistical systems—to leverage the unique strengths of each approach within a single, unified correction framework.
Summary updated