📄 Language Muse: Automated Linguistic Activity Generation for English Language Learners

Summary

Language Muse is a free web tool from ETS for subject-area teachers in the U.S. who teach English language learners (ELLs), most of whom speak Spanish. These students struggle with the complex texts that newer standards require, and their teachers often lack the linguistic training to spot the difficult parts. Teachers can pick one of 33 bundled texts or upload their own. An NLP backend then finds linguistic features in three areas: lexical items, syntactic structures, and rhetorical/discourse relations. The tool combines hand-built resources with statistical methods, and the hand-built resources often act as filters on the noisier statistical output, because classroom use demands high precision. From these features the system builds 24 editable activities, grouped at the word, sentence, and paragraph level. It suggests the activities for which the text offers enough material, and for many of them it also fills in the answer key automatically. Compared with earlier work on automatic question generation, it covers more language constructs and gives teachers much more control over the output.

Key points

• The backend returns a JSON bundle of features. Examples:
  • Lexical: cognates (a hand-made dictionary checked by a native Spanish speaker), academic words with Wordnik definitions, multiword expressions found by rank-ratio collocation over Google Web1T, synonyms, antonyms (WordNet), and homonyms.
  • Synonyms: a thresholded mix of WordNet, Lin's distributional thesaurus, and SMT-derived paraphrases.
  • Syntax: relative clauses, multi-clause sentences, complex NPs, and complex verb groups, all detected with regular expressions over constituency parses.
  • Discourse: cause-effect, compare-contrast, and evidence/details relations, found by an adapted rule-based discourse analyzer.
• Morphological variants are produced in two steps: rules first generate too many candidates, then co-occurrence statistics from Gigaword filter them.
• There are 24 activities in total, and 15 of them come with an automatic answer key. Uploaded texts are limited to 5000 words and can be plain text, .doc, or .docx.
• Activities are marked in three ways:
  • starred as recommended when at least 5 questions can be built,
  • selectable but unmarked when fewer questions are available,
  • greyed out when the text offers none.
• Teachers can change almost everything: instructions, which words or sentences are selected, answers, and distractors. A paragraph can also be hidden or kept only as context.

Interesting ideas

• I1. Generate many candidates with cheap rules, then keep only those that corpus co-occurrence statistics support; this keeps recall high without letting noise through.
• I2. Combine several independent synonym sources (a curated lexicon, a distributional thesaurus, paraphrase tables) and accept a pair only when their combined evidence passes a threshold; agreement between sources stands in for precision.
• I3. Where precision matters, use hand-made resources mainly as a filter over statistical output rather than as the only source; this keeps the coverage of statistics while gaining the reliability of curation.
• I4. Expose one detected feature through several exercise formats at different granularities, such as word matching versus a paragraph cloze; this exercises different facets of the same skill.
• I5. Since the system created the material itself (for example, the synonyms it inserted), it already knows the correct answers, so answer keys and distractors come for free; also offer a task only when the text yields enough instances.

For Mova

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

• I1 (tokens): Proposed inflected or derived forms in crate en could be checked against corpus frequency before entering the lexicon, so impossible forms never reach the tagger.
• I2 (world): Synonym and near-synonym links in the compiled knowledge layer could require agreement between WordNet and at least one other source, which would make substitution-based reasoning safer to explain.
• I3 (learn): Silver data from witness parsers could pass through curated constraint lists, such as verb-class lexicons, before it is used for training.
• I4 (answer): One extracted fact or relation in a fable could yield several probes (a cloze, a multiple-choice question, a paraphrase check) to test whether a reader truly understood it rather than matched a pattern.
• I5 (learn): Self-play could generate tests from its own meaning graph with known gold answers and skip tests whose source text has too few examples to say anything meaningful.