SofiaMrsComparison

MRS Comparison Discussion

This page captures some of the discussion on MRS comaprison at the Sofia Meeting. The discussion was led by FrancisBond using slides by MathieuMorey and had many participants.

MRS comparison using graph matching algorithms

(MathieuMorey)

• Context and motivation:
  • work on the Tanaka Corpus
  • development of Jacy: crosslingual parse ranking
  • development of Jaen: acquisition of transfer rules
• Compare the MRSs produced by the ERG and Jacy+Jaen and
  • find the most similar ones: parse ranking
  • bridge the gap between similar MRSs,
    • i.e. find how transform one into the other:
    • acquisition of transfer rules.
• As MRSs are (directed acyclic) graphs, these problems can be seen as
  • inexact graph matching problems
• Differences between MRSs can be formulated in terms of graph edit operations, with associated costs:
  • insertion/deletion of EPs,
  • insertion/deletion of ARG links,
  • substitution of a relation following the type hierarchy, ...
• Transfer rules then correspond to sequences of graph edit operations.

Pros and Cons of graph matching approach

• Pros:
  • graph matching is more robust and flexible than comparing n-grams of Elementary Dependencies,
  • graph edit operations directly describe transfer rules,
  • MRSs have specific properties (DAGness, relatively small size, finite number of possible labels) that (should) help to limit the algorithmic complexity.
• Cons:
  • finding optimally interesting/useful edit costs is not trivial,
  • automatically partitioning the set of edit operations (between two big MRSs) into linguistically meaningful transfer rules is tricky (guide with patterns e.g. N+ADJ $\rightarrow$ N+N)
  • a lot of options to explore

Current State

• Implementation in python (still in preliminary stages):
  • graph matching,
  • graph visualisation (currently graphviz dot),
  • native graph persistency (currently neo4j)
• Wish and will do:
  • make it easy to play with edit costs and to add new graph matching algorithms,
  • push the code to an online repository: python-delphin ?
  • integrate as much as possible with existing libraries: DELPH-IN, nltk, ... \includegraphics[width=\textwidth]{1090410-3-0_0-0.png} (Buggy but colorful)