📄 Tense and Aspect Error Correction for ESL Learners Using Global Context

Summary

This short paper addresses automatic correction of tense and aspect errors in essays by learners of English. These errors are common but have received less attention than article or preposition mistakes, largely because the right verb form often depends on context beyond the sentence. Earlier correction systems classified each word on its own. The authors instead treat the problem as sequence labeling over all verb phrases in a document, so that each verb's predicted tense/aspect can depend on neighbouring verbs. Because existing annotated learner corpora were too small for machine learning, they built a large corpus from native-speaker corrections on Lang-8. A CRF uses local syntactic and lexical features plus features taken from the preceding verb phrase, especially temporal expressions. A re-scoring step adjusts how cautious the system is about changing the learner's original form. On precision-recall curves, the CRF beat SVM and MAXENT baselines for detection and correction of tense, aspect and their combination. This suggests that cross-sentence context gives a modest but real benefit.

Key points

• Corpus: A tense/aspect corpus was derived from Lang-8 corrections. After cleaning it has about 120,000 English entries and roughly 2,000,000 verb phrases, of which about 750,000 carry corrections. Labels cover 12 tense × aspect combinations: past/present/future crossed with none, perfect, progressive and perfect progressive.
• Data split: 100,000 entries were used for training and 1,000 for testing. The test set contains 16,308 verb phrases, 1,072 of them (6.6%) erroneous. Features come from Stanford Parser 1.6.9.
• Models: A CRF was built with CRF++ 0.54. The baselines were MAXENT (also CRF++) and a one-versus-rest SVM (LIBLINEAR 1.8). All used default parameters, identical features and the same T-correction output step. The CRF performed best across all error types.
• Features: Local features include the learner's surface tense/aspect, lemma, neighbouring words, dependency relations (nsubj, dobj, aux, pobj, advmod) and conjunction/clause-type information. Global features take the temporal adverb of the previous verb phrase (p-tmod′, norm-p-tmod′) and combine it with the surface tenses of both verbs.
• Error analysis (α = 0.1): The most frequent error is present used where past is required (211 cases); the system detected 61 and fixed 52. The reverse error (94 cases) was detected only 9 times.

Interesting ideas

• I1. Corpus from a learner platform: Corrections that native speakers post on a language-learning social network can be harvested to create large annotated error data. This approach could be reused for low-resource languages or other error types where hand-annotated corpora are tiny.
• I2. Document-level sequence labeling: Linking verb phrases across sentences in a chain lets one decision inform the next. The same framing could suit other discourse-dependent choices, such as pronoun or article selection.
• I3. Normalized temporal cues: Temporal expressions are mapped to a few coarse values (past, present, future, "this") via keyword lists. This reduces sparsity, and the coarse value is then passed on to the following verb as context. It is a cheap trick applicable to any time-sensitive tagging task.
• I4. T-correction for tunable recall: The model's confidence in the label matching the learner's original form is multiplied by a weight α below 1. This simple knob counteracts a correction system's natural bias toward leaving text unchanged, and it can be applied post hoc to any classifier that outputs confidences.
• I5. Fair comparison with shared components: All classifiers receive the same features and the same output re-scoring, and results are shown as full precision-recall curves rather than single points. This isolates the effect of modelling context and is a good evaluation pattern when operating points can be tuned.