RmrsDistributional

Applications

Discussion from Cambridge meeting (23/5/2008 -- thanks Ann!).

Interaction between domain knowledge (NE) and syntax/semantics:

• not adequate to pre-NER a string and pass on the output to ERG, etc, due to punctuation, pluralisation, etc.
• logical metonymy: "a pyridine" = a member of the class of compounds which contain a pyridine ring; "the pyridine" = the pyridine ring in a particular compound
• discussion of kind vs. substance readings of chemical compounds -- necessary for downstream processing?
• "three beers" being default portion vs. "three rices" being default kind the sort of thing we want to underspecify rather than represent as distinct readings (c.f. compound nouns)
• RMRS: to index or not to index? Within Sciborg, indexing only chemicals under the assumption that every query will contain a compound (constrain search space to only those documents containing that compound, then play around with the full documents). Similarly for GENIA (index only terms, events)

Distributional similarity and grammar engineering:

• "distributional similarity" = corpus co-occurrence-based vector similarity

• use in interpreting compound nouns (e.g. interpret "cat food" via similarity with "dog meal", etc), verb clustering (learning Levin clusters), etc

• distributional similarity and relation extraction: relation extraction based on training data and pattern "learning" for particular relations, whereas distributional similarity simply compares context vectors and churns out a similarity prediction

• interface between distributional similarity and grammar engineering:

  • QA in context of pattern matching
    • similarity over words or predicates?
    • what about WSD? sentence similarity (union of semantics of words in sentence) can help to disambiguate
  • applications in sentence-level disambiguation of, e.g., "some paper from 1946" (count) vs. "some paper from Kinkos" (mass), where the grammar doesn't tell you anything
  • distributional similarity as way of getting around sense enumeration (via fuzzy clustering)

• compound nouns specifically: applications?

  • MT into Romance languages (also Japanese, Norwegian)
    • need to commit to particular representation for given language pair? Possibly not but don't really know.
    • any systematic study of NN translation divergences in different languages?
  • QA: currently potentially doing worse than BOW approach because of insisting on particular syntax

• interpretation of 3-ary and larger NNs? (semi-)solved task, and parse selection models don't have access to the corpus counts that unsupervised bracketing methods do

• importance of context in interpreting NNs? important, but in limited cases; possibility of distributional similarity helping out (e.g. "helicopter radio" in sense of "radio-controlled")

• cute example of domain example with (out of domain) counterintuitive bracketing: "identical retention time and mass spectrum" -> "(identical ((retention time) and (mass spectrum)))"

• interfacing standalone bracketing, e.g., with parse selection? chicken and egg problem in terms of identifying where to plug in the standalone

• "can we stop now?" (Alex)