📄 An English Dictionary for Computerized Syntactic and Semantic Processing Systems

Summary

The authors argue that natural-language processing systems of the early 1970s fail partly because their lexicons are tiny, usually only a few hundred items. They plan a large computerized English dictionary that can serve two roles: a component inside language-processing programs, and a database for syntactic and semantic research. The paper identifies the main obstacles. Conventional dictionaries are circular, a realistic lexicon needs very large storage, and pointer-chasing lookups are slow. It then surveys the linguistic information such a resource should hold. Its main sources are the Soviet "explanatory-combinatory" dictionary of Apresjan, Mel'čuk and Žolkovsky, Fillmore's case frames and lexical requirements, McCawley's semantic atoms, factivity, genericity and specificity, and graded (fuzzy) features from Ross and Lakoff. The authors propose a model meant to reduce storage and search costs. Their appendices show machine-sorted adjective definitions drawn from a standard dictionary. The work matters as an early argument that a lexicon should encode relations, argument structure and alternations, not only glosses.

Key points

• The authors estimate that a usable English dictionary needs at least 200,000 entries, including inflected forms.
• They give storage arithmetic to show why naive designs fail:
  • About 20 definition words per entry would mean 4,000,000 stored words.
  • Chomsky-style entries would need roughly 20 features each.
• Definitions in standard dictionaries are circular. Roughly 50% of the vocabulary shows up inside definitions (Olney, personal communication).
• The paper reviews the Apresjan/Mel'čuk/Žolkovsky entry format, which has ten fields, including syntactic potential, lexical functions, lexical universe and synonym discrimination. It lists lexical functions such as:
  • Syn, Conv, Anti and Gen
  • S0, A0, Sinstr and Mult
  • Figur, Oper, Fin and Son
• The paper draws on Fillmore's set of case roles, which covers extent, time, place, goal, source, object, instrument, experiencer and agent. It also lists his eight kinds of information a lexicon must expose. The planned uses are AI systems, computer-assisted instruction (PLATO), lexicon testing and information retrieval.

Interesting ideas

• I1. The Conv lexical function pairs predicates that describe one situation with swapped roles, such as sell/buy, precede/follow and beat/lose to. Storing these pairs explicitly lets different surface sentences converge on the same meaning.
• I2. Oper links an event noun to its typical support verb, as in lend support, suffer defeat and experience a recession. The verb then adds little meaning, and the noun carries the event.
• I3. Syntactic potential records which arguments may surface as subject. For example, "break" lets the object become subject ("the window broke"), but "hit" does not.
• I4. Predicates can constrain the role of their subject. "Be happy" demands an experiencer, while "be warm" accepts an object, place, time or instrument, and adverbs like "personally" act as diagnostics.
• I5. Non-circular definitions built from a small set of semantic atoms, as in redden = cause to come to be red, give an exact paraphrase instead of a gloss.

For Mova

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

• I1 (world): A Conv table over verb classes would let "A sold B to C" and "C bought B from A" produce one transfer event with identical holdings.
• I2 (frames): When a UD object of a known Oper verb is an event noun, build the frame from the noun and map the subject to the noun's participant role.
• I3 (frames): Encoding causative and instrument-subject alternations per verb would let role assignment correctly label a patient or instrument that appears in nsubj position.
• I4 (frames): Role constraints on adjectival predicates could disambiguate the subject's role and flag odd readings, keeping explanations explicit.
• I5 (back): Decomposing verbs into atoms like CAUSE/BECOME/STATE would give the graph-to-text step a principled way to paraphrase or lexicalize events.