📄 Motivations and Methods for Text Simplification

Summary

The paper argues that long, syntactically dense sentences, which are common in newswire, cause trouble for NLP systems such as parsers, MT, IR and summarizers. Splitting them into several short, self-contained sentences could help. The authors treat simplification as two steps. First, build a structural representation of the sentence. Second, apply rewrite rules at "articulation points", such as phrase boundaries, punctuation, conjunctions and relative pronouns. Full parsing is avoided because it is slow and brittle on exactly the complex sentences that need simplifying.

They compare two lightweight front ends:

• a Finite State Grammar chunker that yields flat noun and verb groups, from the Vaakya translation prototype;
• a Dependency-based model (DSM) built on LTAG supertagging, which adds constituent and dependency information.

On newswire text, the DSM handled relative clauses and appositives far better than the FSG approach. Its rules were also more general. The work matters as an early statement of rule-based syntactic simplification. It also spells out the discourse problems that simplified output creates: sentence order, referring expressions and tense.

Key points

• Articulation points include phrase starts and ends, punctuation, subordinating and coordinating conjunctions, and relative pronouns. Domain-specific points can be added.
• The FSG system (Vaakya) works in three steps:
  • It POS-tags words, then chunks them into NP and VP groups, marking number, tense, voice and aspect.
  • It applies manually ordered rules repeatedly until none fire.
  • On news stories it turned 224 sentences into 369 simplified ones.
• The FSG approach fails on several sentence types: long-distance or crossed dependencies, stacked appositives, and unclear attachments.
• The DSM uses a trigram supertagger based on LTAG, which runs in linear time. Slot polarity in the supertags is used to derive dependency links.
• On newswire, the two models compared as follows (relative clause and appositive simplification only):

Interesting ideas

• I1. Shallow but structured input instead of full parsing. Supertagging gives near-parse information (argument slots, clause types) at tagger speed. This is a useful middle ground wherever full parsers are fragile, for example on long Ukrainian sentences.
• I2. Rules over constituent types, not surface sequences. One DSM rule for relative clauses works in any position (subject, object, embedded). The FSG approach needed separate variants for each case. Abstracting rules this way cuts rule-set size and coverage gaps.
• I3. Rule ordering as a design control. The FSG rules are ordered by hand so that frequent transformations fire first and harmful interactions are avoided. They are then applied recursively until nothing changes. This is a simple, transparent recipe for any cascade of rewrite rules.
• I4. Choose the output type by downstream use. If the simplifier only feeds a parser, the split pieces can stay incomplete and be recombined later. Gaps such as the relativized noun need filling only when the output must be standalone text. This saves effort when readable output is not required.
• I5. Evaluate per construction, counting false positives. Scoring recall separately for each phenomenon (relative clauses, appositives) and counting spurious extractions gives a sharper diagnostic than a single overall score.