RmrsParaphrasing

This page discusses some of the known issues with paraphrasing, with especial reference to (R)MRS. It also tells you how to paraphrase using MRS in the [wiki:LkbTop LKB].

Paraphrasing in the LKB

Paraphrasing using rules

Put your paraphrasing rules in a file paraphraser.mtr, and load it in the lkb/script file.

;;; transfer rules for paraphrasing
#+:mt
(mt:read-transfer-rules 
 (list
  (lkb-pathname (parent-directory) "paraphraser.mtr"))
 "Paraphrasing Rules"
 :filter nil :task :paraphrase)

Then when you parse something, a menu item Rephrase should appear, and this will apply the paraphrase rules to the selected parse, and then generate from the paraphrase.

Paraphrasing rules are done with the same [wiki:TransferRules transfer rules] used in MT and generation. An example is given below. This will paraphrase "We gave a demo of the system" -> "We demoed the system".

give+demo_mtr := light_verb_mtr &
[ INPUT.RELS <! [ PRED "_give_v_1_rel" ],
                [ PRED "_demo_n_rel", ARG1 #arg1 ],
                [ PRED _of_p_sel_rel ], 
                [ PRED _a_q_rel ] !>,
  OUTPUT.RELS <! [ PRED "_demo_v_rel", ARG2 #arg1 ] !> ].

Paraphrasing using Underspecification

• Underspecified lexical types (or rules) effectively lead to paraphrasing

  • parsing Which dog barked and then generating gives

    • Which dog barked?

    • What dog barked?

    • Which dog did bark?

    • What dog did bark?

General Discussion

From the meeting in Cambridge 2008-05-24

Dan showed off the paraphraser in the ERG.

• the rules can rewrite features.
• and whole clauses
• they can be made bidirectional if there is no CONTEXT or FILTER
• nice use of relational nouns in the light verb rule
  • "we gave a demo of the system" -> "we demoed the system"

Two current approaches

• underspecification: in/on/at_rel
• explicit mapping via rules

Uses of paraphrasing

• we can map representations to a canonical form
  • e.g. reducing variation before we do a task such as MT
    • (the hourglass model)
  • not always possible
• query expansion
• relaxed input to the generator
  • e.g. allow "possible_rel" which maps to "can_rel" and "be_able_to_rel". Non-English users shouldn't have to care.

How is working on Paraphrase/rule extraction?

• Mischa, Wang, Eric, Stephan (dictionary) all MT
• It would be nice to share tools
• * Laurie: big dictionary building project at Washington
• * FCB: we could use wordnet

What are some phenomena we can handle now? (can we map them to situation types: LH?)

• light verbs,
• compounds,
• serial verbs,
• classifiers,
• alternations
• relational nouns/prepositions,
• modal verbs and tense/aspect,
• "swim across the river"/"cross the river by swimming"
• by/gerund,
• verbal/nominal, "opening the door he found the revolver/he opened the door and
  • found the revolver",
• done by underspecification --- adjective/adverb

Meta-discussion

• currently we don't have perfect world models so we normally overgenerate and rank

  • want to be able to make things that work now && support deeper processing when people can do it (AAC)

  • the statistical models on overgeneration leads to models useful for analysis

• good to map paraphrase classes to situation types not just syntactic types

• sometimes we need anaphora resolution

  • "Tom ate because he was hungry" -> "Tom's hunger made him eat"
  • no-one is working on this within the lkb
  • we can do some of this with paraphrase rules