📄 Constructing English Reading Courseware

Summary

EFL teachers have far more authentic reading material available than they can screen, and choosing texts that serve a specific course goal takes a lot of work. The authors frame courseware building as an optimisation problem. From a corpus, the goal is to choose a subset of documents that together contain every word of a target vocabulary while keeping the total token count as small as possible. This is a set-cover problem, so they solve it greedily. At each step they add the document with the best score from a BM25-inspired function. The score rewards short texts that contain many vocabulary types not yet covered, and gives a smaller bonus for types already covered.

They applied the method to a 642-word TOEIC vocabulary and short Daily Yomiuri articles, lemmatized so words could be matched. It produced 116 articles covering the full list. Random article sets of the same size reached only about 62% coverage and contained fewer target words per article. Coverage also rises fastest at the start, so students who read only part of the material still benefit. The courseware is now used in university classes. Students found it motivating, though some articles were difficult and some word senses differed from TOEIC usage.

Key points

• Inputs:
  • Chujo's (2003) TOEIC vocabulary of 640 entries, giving 642 matchable words.
  • The Daily Yomiuri, 1989–2001: about 110,000 articles, filtered to about 25,000 of at most 300 words, each with a Japanese counterpart aligned by Utiyama and Isahara (2003).
• Scoring:
  • The gain function g is modelled on Okapi BM25, with k1 = 1.5 and b = 0.75.
  • The mixing weight is α = |V_done| / (1 + |V_done|). This makes one new type worth as much as all the already-covered types in a document.
• Resulting courseware: 116 articles and 20,900 tokens. Per article, the averages were:
  • 180.2 tokens
  • 25.3 tokens shared with the vocabulary
  • 17.4 shared types
  • 5.5 newly covered types
• Baseline: 1,000 random sets of the same token budget (about 111 articles each).
  • These averaged 19.3 shared tokens, 12.8 shared types and 0.616 coverage, against 1.0 for the courseware.
  • The differences in shared tokens, shared types, new types and coverage were all highly significant.
• Coverage curve: 50% of the vocabulary is reached after 18% of the courseware's tokens, and 90% after 67%.
• Known limitation: lemma matching ignores word sense. For example, "agency" in the newspaper often means a government body rather than a service firm.

Interesting ideas

• I1. Selecting a minimal set of texts with greedy set cover plus a length-normalised gain makes it cheap to assemble a short collection that still hits every required item.
• I2. Blending a score for new items with a weaker score for already-seen items keeps each step focused on fresh material while still repeating earlier material for reinforcement.
• I3. A fair baseline is many random subsets drawn with the same token budget and averaged; this gives a mean and spread to test against.
• I4. Plotting coverage against tokens read, and reporting where 50% and 90% are reached, measures how front-loaded a resource is; that matters when only part of it will be used.
• I5. Matching on lemmas alone silently mixes senses when the vocabulary source and the corpus come from different domains.

For Mova

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

• I1 (learn): Use this to pick the smallest batch of sentences covering all VerbNet classes or UD relations before sending them to the LLM teacher tiers, which would cut labelling cost without leaving gaps.
• I2 (answer): Order school-course test items or fable QA so each new item introduces unseen constructions while revisiting earlier ones, giving a curriculum for the student readers.
• I3 (learn): Use equal-budget random samples as a negative control when claiming that a silver-data selection strategy helps, rather than comparing against a single random draw.
• I4 (tree): Measure how quickly a treebank ordering exposes rare dependency relations, so that early epochs or partial training runs of gparse still see most label types.
• I5 (world): Tag compiled lexicon entries with WordNet senses or domain labels so that text selection and frame grounding do not treat a government "agency" and a business "agency" as the same concept.