📄 Three English Learner Assistance Systems Using Automatic Paraphrasing Techniques

Summary

Murata and Isahara apply paraphrasing machinery to three tools for English learners, mainly Japanese speakers. The first compares 80 of the first author's papers before and after native-speaker proofreading, using Unix diff, and counts recurring edits to build a personal error profile. The second rewrites text, such as a talk draft, to reduce the "l" and "r" sounds that Japanese speakers find hard. It proposes WordNet 2.0 synonyms, scores them by how many "l"/"r" occur before vowels, and checks that each candidate appears in the BNC within a two-word window on each side. The third marks rare words (BNC frequency under 1000) or words the user has never written, and attaches EDR dictionary definitions inline. The error profile was dominated by article errors, followed by tense and preposition errors. The rewriter often worked but sometimes changed the meaning, so the authors suggest a semi-automatic mode. The paper matters mainly as an early demonstration that simple diff, lexicon and corpus-frequency methods can support personalized language learning.

Key points

• Error-profile tool
  • Over 80 papers, the most frequent corrections were deleting "the" (393), inserting "the" (357), "a"→"the" (146), and deleting "a" (122).
  • Article errors ran in both directions, but tense errors were lopsided: "is"→"was" occurred 68 times while the reverse was rare. The authors attribute this to present tense being the writer's default.
• "l"/"r" rewriter
  • It has two modules: one generates candidates from WordNet 2.0 synonyms, and the other selects among them.
  • Ties are broken by BNC frequency in context. Candidates never seen in context are discarded.
  • Successful swaps included approach→way, structure→composition and length→size.
  • Failures included title→name and corpus→question, where the meaning shifted.
• Definition annotator
  • It uses definitions from the EDR dictionary and WordNet 2.0 for stemming.
  • It lists every sense of a word, marks repeated words with "[Above!]", and flags words missing from the dictionary with "[Caution!]".
• Unexpected typo detection: the "[Caution!]" flag also caught a misspelling ("recordeded").
• Personalization: excluding words from the user's own papers stopped the system from glossing "corpus", a term the user already knows well.

Interesting ideas

• I1. Aligning a text with its corrected version and counting the edit operations turns proofreading into a ranked error profile, and lopsided edit pairs point to a default form the writer overuses.
• I2. A generate-then-filter design keeps a synonym only if the corpus attests it with the surrounding two words on each side, so collocation evidence filters lexical substitution.
• I3. The substitution failures (title→name, corpus→question) show that swapping synonyms without knowing the word's sense can silently change meaning, which is why the authors fall back to having a human choose.
• I4. Word difficulty can be modeled per user from the words they have actually produced, combined with corpus rarity, rather than from a global list.
• I5. A single "not in lexicon" flag served both as a difficulty warning and as a typo detector.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Diffing parser output against LLM-teacher or repair-rule corrections could rank systematic parser mistakes and seed self-induced repair rules, with one-directional patterns showing default-label bias.
• I2 (back): When generation picks a lexical realization for a frame, an attested-context check could reject fluent-looking but unidiomatic word choices.
• I3 (world): Any use of WordNet relations from the compiled ontology should first fix the word's sense in context, and should expose the ranked candidates when confidence is low.
• I4 (answer): Student-facing readers could track the vocabulary a learner has already seen in course texts and gloss or simplify only the new words.
• I5 (tokens): The lemmatizer's out-of-lexicon signal could be surfaced explicitly, so that misspellings are caught before they reach the parser instead of being silently tagged.