Notes on meetings - kornai/4lang GitHub Wiki

Notes on meetings

Universal dependencies

Core dependents of clausal predicates

  • nsubj csubj nsubjpass csubjpass 1, AGT
  • dobj ccomp xcomp 2, PAT
  • iobj DAT. If the clausal predicate does not have a DAT, insert a (binary) TO

Non-core dependents of clausal predicates

  • nmod advcl advmod neg 0

Special clausal dependents

  • vocative discourse expl -- as a baseline, 0
  • aux auxpass cop punct 0
  • mark
    • if the head is contentful, Forces engaged in fighting after insurgents attacked engage <-1- after -2-> attack
    • else He says that you like to swim say -2-> swim (1 comes from the dobj)

Noun dependents

  • relational nouns (occasion =TO, duty =POSS)
  • nummod nmod acl amod det neg 0
  • appos Sam , my brother , arrived sam <-0-> brother

Case-marking, prepositions, possessive

  • case (see section intermediate representations)

Coordination, Compounding and unanalyzed

  • use Enhanced Representation
  • conj cc compound mwe goeswith name foreign -- drop the relation
  • punct -- drop the dependent

Loose joining relations

  • do not handle

Roles and cases (9 Feb)

phenomena we want to give more or less account of

  • reducing ditransitives to binary predicates
  • verb classes (motion, psych; Levin, Dowty, Vendler; unaccusative)
  • arguments and adjuncts
  • selectional restrictions
  • unaccusative_verbs, ergative languages
  • quirky case, language specific component

task

  • compute FrameNet frames from dict_to_4lang graphs, for an other language for with there is a relational parser and a FrameNet, e.g. Chinese

further reading

Event structure (bibliography)

Representation (19 Jan 2016)

In theory, the 4lang representation system has three (interrelated) formalisms:

  • logical formulae ((linear) strings)
  • graph (good for visualizaton)
  • Eilenberg machines with partitions (operators. But it is not clear, how Eilenberg machines frame our grasp of semantics.)

In graph terms, 4-lang is a semantic network with labeled nodes and edges. Node labels range over the whole vocabulary, edge labels are 0, 1, or 2. The label of a node that has a 2-labelled out-edge is a binary predicate. A set of primitive nodes (nodes that don't have to be expanded with their defining graphs, e.g. cause, has, after, lack) have to be listed in a separate resource.

Graph frameworks

  • Hyperedge Replacement Grammar
    • "Graph parsing with s-graph grammars" (J. Groschwitz, A. Koller, C. Teichmann), Proc. ACL, 2015.
    • "A Synchronous Hyperedge Replacement Grammar based approach for AMR parsing" (X. Peng, L. Song, and D. Gildea), Proc. CoNLL, 2015.
    • "Parsing Graphs with Hyperedge Replacement Grammars," (D. Chiang, J. Andreas, D. Bauer, K. M. Hermann, B. Jones, and K. Knight), Proc. ACL, 2013.
  • unification
    • HPSG
    • PATR Pereira, F. and S. Shieber (1984) The Semantics of Grammar Formalisms Seen as Computer Languages
    • in logical formulas: JA Robinson (1965) A Machine-Oriented Logic Based on the Resolution Principle

Intermediate representations

  • a path of 0-edges means that all the nodes along the path can be unified (substituted in the first one)
  • similarly, as soon as we have attack <-1- PAT -2-> dragon, this subgraph can be replaced by attack -2-> dragon. Edge labels 1 and 2 correspond to deep cases AGT and PAT resp.
⚠️ **GitHub.com Fallback** ⚠️