📄 Stephen Pollington 1997: First Steps in Old English. Norfolk: Anglo-Saxon Books

Summary

This is a short review in SELIM 6 by Alicia Rodríguez Álvarez of Stephen Pollington's self-study textbook First Steps in Old English (1997). It is not an NLP paper. The book is for complete beginners with little background in linguistics who want to read Old English prose on their own. It focuses on morphology, keeps syntax and spelling light, and pushes sound changes, dialects and palaeography into notes.

The course runs through eighteen lessons. Each one mixes several grammatical categories, so learners can build whole sentences almost immediately, and it ends with comprehension questions, drills and graded translation exercises with a key. Later parts add grammar reference tables, then annotated prose readings whose annotation is gradually withdrawn.

The reviewer praises the book as accessible and better than comparable beginner books. She criticises two things: vocabulary is spread across three places, and irregular forms caused by i-mutation are not explained.

Key points

• The book has four parts:
  • an 18-lesson course, each lesson with topics, questions, "Practice", "Vocabulary" and "Exercises";
  • a grammar reference with pronunciation, a glossary of grammar terms and "Accidence" paradigm tables;
  • selected prose texts with parsing aids;
  • an answer key for the exercises.
• The exercises are translations in both directions, into and out of Old English, ordered by difficulty.
• In the reading section, the author labels each element's form and syntactic function at first, then drops the labels as the reader gains confidence.
• The book was inspired by Macrae-Gibson's Learning Old English. Its readings are also recorded on the audiotape Ærgeweorc.
• The reviewer compares it with three other works:
  • Blakeley's Teach Yourself Old English (1964), which is more traditional and heavier on phonology;
  • Magennis & Herbison's Discovering Old English (1990), which teaches one category at a time and is more text-centred;
  • Jebson's 1997 web guide, which has a similar exercise placement.

Interesting ideas

• I1. Teaching nouns, verbs and adjectives together rather than one category at a time lets a learner handle real sentences early, which keeps practice close to actual use.
• I2. Readings first come with explicit form-and-function labels, which are then removed step by step as the reader improves. This scaffolding-then-fading pattern suits any learner that starts out dependent on hints.
• I3. The reviewer argues that irregular forms feel arbitrary when shown only as exceptions, and become clear once the underlying sound change is explained. A single rule can replace a long list of oddities.
• I4. She also points out that one consolidated glossary beats several scattered word lists, because lookup becomes faster and duplicates do not drift apart.
• I5. Graded exercises in both directions, checked against a key, test comprehension and production separately.

For Mova

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

• I1 (learn): Training and self-play curricula could mix construction types from the first stage instead of finishing one phenomenon before starting the next. This would expose the parser to realistic clause interactions sooner.
• I2 (learn): Silver training data could begin with heavy teacher hints, such as LLM-provided heads or roles. The hints would be withdrawn in later rounds to test whether the model has really absorbed the structure.
• I3 (tokens): The lemmatiser could encode stem-vowel alternations in irregular plurals and verbs as named, explainable rules rather than opaque lookup entries. Each lemma decision would then trace back to a stated cause.
• I4 (world): Keep a single lexical source of truth from which the WordNet, VerbNet and Wikidata-derived layers are referenced. This avoids per-module vocabularies that silently diverge.
• I5 (back): The text → graph → text compressor could be evaluated in both directions on difficulty-ranked sets with reference outputs, so analysis and generation failures can be told apart.