📄 Simplifying Text for Language-Impaired Readers

Summary

This short EACL'99 paper describes a prototype pipeline that makes English newspaper articles easier to read for people with aphasia, a language disorder usually caused by stroke or head injury. Earlier work by Devlin showed that such readers have trouble with rare words and with certain syntactic structures. Newspaper prose adds its own obstacles: dense opening paragraphs, long sentences, noun compounds, stacked adjectives and heavy use of the passive. Within the EPSRC-funded PSET project, the authors chain existing tools into a sequence of stages: tagging, morphological analysis, parsing, anaphora resolution (still unimplemented), syntactic simplification, lexical simplification and morphological generation. Syntactic rules convert passives to actives and are being extended to split conjoined and embedded clauses. Lexical simplification swaps content words for more frequent synonyms. The paper reports no simplification results. Instead, it lays out a planned user-centred study with aphasic readers. The authors expect the system to help other limited readers as well, such as non-native speakers.

Key points

• The input is articles downloaded automatically from the online edition of The Sunderland Echo, a local newspaper in north-east England.
• The tagger (Elworthy, 1994) is a first-order HMM with a lexicon tuned to newspaper text. Its unknown-word guesser is about 85% accurate.
• The parser combines a feature-based unification grammar over PoS and punctuation tags (Briscoe and Carroll) with probabilistic LR disambiguation. It produces shallow phrase structure and covers roughly 80% of a substantial general-text corpus.
• The lexical simplifier pulls synonyms from WordNet and looks up their Kucera-Francis frequencies in the Oxford Psycholinguistic Database, keeping the most frequent candidate.
• Anaphora resolution is planned to build on CogNIAC but had not yet been implemented. The evaluation with aphasic participants was also still in the planning stage.

Interesting ideas

• I1. Applying rewrite rules until nothing changes. Syntactic simplification runs repeatedly until no rule can fire. This lets each rule stay small while still handling nested structures, which suits any rule-based normalisation task.
• I2. Fixing pronouns after splitting sentences. When a sentence is split, a pronoun in it is replaced by the noun it refers to, so no reference crosses the new sentence boundary. Any sentence-splitting or summarisation tool could adopt this to keep the output coherent.
• I3. Getting the generator for free from the analyser. The morphological generator is produced by automatically inverting the finite-state analyser. Any improvement to the analyser therefore carries over to the generator with no extra work, which saves effort for morphologically rich languages.
• I4. Letting the user set the simplification level. The user chooses what share of the synonym list is considered before picking the most frequent word. This gives a simple control over how aggressive the substitution is, and could be used to adapt text to different reader levels.
• I5. Evaluating with the target readers themselves. Participants are checked for adequate reading ability and screened out if they have visual reading problems, but they are not filtered by aphasia type. Their understanding is then tested with questions on both explicit and implicit content. This is a model for evaluating accessibility tools beyond automatic metrics.