Mapping between English Strings and Reentrant Semantic Graphs knight 3/12/15
Strings, Trees, and Graphs Automata greatly simplify design of NLP systems Finite-state speech recognition, tagging, string string Transducer transliteration, etc. (FST) Tree machine translation, tree tree Transducer summarization, etc. string (LNT, etc.) Not yet semantic interpretation, graph graph thoroughly meaning-to-text tree studied for string NLP apps
Strings, Trees, and Graphs String Automata Tree Automata Graph Algorithms Algorithms Algorithms N-best … … paths through an WFSA … trees in a weighted forest (Jiménez & Marzal, 2000; (Viterbi, 1967; Eppstein, 1998) Huang & Chiang, 2005) EM training Forward-backward EM Tree transducer EM training (Baum/Welch, 1971; Eisner (Graehl & Knight, 2004) 2003) Determinization… … of weighted string acceptors … of weighted tree acceptors (Mohri, 1997) (Borchardt & Vogler, 2003; May & Knight, 2005) Intersection WFSA intersection Tree acceptor intersection string WFST WFSA tree TT weighted tree Applying transducers acceptor Transducer WFST composition Many tree transducers not composition closed under composition (Pereira & Riley, 1996) (Maletti et al 09) General tools Carmel, OpenFST Tiburon (May & Knight 10)
[Banarescu et al 13] Graph/String Data “Pascale was charged with public intoxication and resisting arrest.” 14,000 sentences have been annotated with Abstract Meaning Representation (AMR). Freely available portion at http://amr.isi.edu/download/amr-bank-v1.4.txt
[Braune, Bauer & Knight 14] Graph/String Data • 10,000 smallest semantic graphs composed of: – Predicates BELIEVE and WANT – Entities BOY and GIRL • Plus 10 English string realizations of each graph He wants her to believe he wants her. He wants her belief that he wants her. For her to believe he wants her is wanted by him. etc. Freely available at http://amr.isi.edu/download/boygirl.tgz
Graph/String Data • Graphs created synthetically by enumeration. • English string realizations then created synthetically by arbitrary program. • Microworld, but: – graphs are highly re-entrant, enabling focus on handling entities playing multiple roles – strings involve pronouns, case, reflexives, zero pronouns, nominalizations, passives , etc. – not easy to map graphs to strings & vice-versa!
Goal • Concisely capture all graph/string pairs in the corpus • Using a formalism with nice theoretical and computational properties
Candidate Formalisms • Unification grammar [Kay 84; Shieber 86, Moore 89] • Synchronous Hyperedge Replacement Grammar (SHRG) [Drewes et al 97; Chiang et al 14] • DAG-to-Tree transducer (D2T) [Kamimura & Slutzki 82; Quernheim & Knight 12ab] • Tree transducer (xLNT, xLNTs) [Rounds 70; Thatcher 70; Maletti et al 08] ... and cascades of these devices.
Unification-Based Semantics syn sem pat cat io want io agt S agt go RULE B: io x0 x1 x2 ;; VP VB VP-INF RULE C: boy (x0 sem) = (x1 sem) x0 x1 x2 ;; S NP VP (x1 syn inf) = x2 (x0 sem) = (x2 sem) (x1 syn subj) = (x2 syn subj) = (x0 syn subj) RULE C (x2 syn subj) = x1 syn sem pat subj cat io RULE A: want io agt VP x0 x1 x2 ;; VP-INF TO VB sem agt go (x0 sem) = (x2 sem) (x0 syn subj) = (x2 syn subj) RULE B syn sem cat subj io go VP-INF agt sem syn sem pat subj cat io syn sem inf want syn sem RULE A VB cat agt io sem syn cat subj io NP boy sem cat go VB agt sem TO the boy wants to go
Desired Properties • Concisely captures linguistic phenomena – huge sections of transformation grammar not replicated & tweaked over and over! • Polynomial membership checking • Reversible, bidirectional application (forward and backward) • Weighted or probabilistic version • Efficient N-best generation • Efficient EM training
end
backup slides
Hyperedge Replacement Grammar [Drewes et al 97] • Represents a possibly infinite set of graphs • Start with graph containing – terminal edges: part of final output – non-terminal edges: to be expanded • Recursively replace non-terminal edges by rule, until none remain
HRG Derivation LET’S DERIVE THIS: instance ARG1 ARG0 ARG1 instance ARG1 instance WANT BELIEVE ARG0 WANT G B = boy wants girl to believe that he is wanted
HRG Derivation LET’S DERIVE THIS: “the boy wants ARG1 something instance instance ARG1 involving himself” ARG0 ARG1 instance ARG0 X ARG1 instance WANT WANT BELIEVE ARG0 WANT G B B
HRG Derivation LET’S DERIVE THIS: “the boy wants ARG1 something instance instance ARG1 involving himself ” ARG0 ARG1 instance ARG0 X ARG1 instance WANT WANT BELIEVE ARG0 WANT G B B
HRG Derivation “the boy wants the girl to believe LET’S DERIVE THIS: something involving him ” ARG1 instance instance ARG1 ARG0 ARG0 ARG1 instance X instance ARG1 instance WANT BELIEVE WANT BELIEVE ARG0 ARG0 WANT G G B B
HRG Derivation LET’S DERIVE THIS: ARG1 instance instance ARG1 ARG0 ARG0 ARG1 instance X instance ARG1 instance WANT BELIEVE WANT BELIEVE ARG0 ARG0 WANT G G B B “something involving B”
HRG Derivation LET’S DERIVE THIS: ARG1 instance instance ARG1 ARG0 ARG0 ARG1 ARG1 instance instance ARG1 ARG1 instance instance WANT BELIEVE WANT BELIEVE ARG0 WANT ARG0 WANT G G B B FINISHED!
HRG Derivation 2 LET’S DERIVE THIS: ARG1 instance ARG1 WANT ARG1 ARG0 instance WANT ARG0 BELIEVE ARG0 instance ARG1 WANT X ARG0 G B “the boy wants B something involving himself ” (= boy wants girl to believe that he wants her)
HRG Derivation 2 this new hyperedge LET’S DERIVE THIS: labeled X has two tails ARG1 instance ARG1 WANT ARG1 ARG0 instance WANT ARG0 instance BELIEVE ARG0 BELIEVE instance ARG1 ARG0 X WANT ARG0 G G B B “the boy wants the girl to believe something involving them both ”
HRG Derivation 2 LET’S DERIVE THIS: ARG1 instance ARG1 WANT ARG1 ARG0 instance WANT ARG1 ARG0 instance BELIEVE ARG0 BELIEVE instance ARG1 ARG0 instance WANT ARG0 WANT G ARG1 G B ARG0 B FINISHED!
Synchronous Hyperedge Replacement Grammar [Chiang et al 13] • Represents a possibly infinite set of graph/tree (or graph/string, or graph/graph) pairs • Each SHRG rule outputs a graph fragment and a tree fragment simultaneously
SHRG Derivation S instance X ARG0 B wants INF WANT “the boy wants something involving himself” B
SHRG Derivation S ARG1 instance ARG0 B wants INF instance X G to believe S WANT BELIEVE ARG0 G B “something involving B”
SHRG Derivation ARG1 instance S ARG0 instance B wants INF ARG1 instance WANT BELIEVE G to believe S ARG0 WANT he is wanted G B FINISHED!
DAG-to-Tree Transducer [Kamimura & Slutzki 82] • Bottom-up transformation of graph to tree „Girl believes girl wants girl to want boy“
DAG-to-Tree Transducer [Kamimura & Slutzki 82] Graph node GIRL transformed into three tree nodes (NP-she, NP-she, NP-0), and labeled with different states.
Top-Down Tree Transducers (Rounds 70; Thatcher 70, Maletti et al 08) xLNT (tree-to-tree) S q S r NP q VBZ s NP NP VP S ga wa SBAR enjoys PRO PRO VBZ NP he VBG VP ga he enjoys SBAR wa listening P NP VBG VP to music listening P NP to music
Top-Down Tree Transducers (Rounds 70; Thatcher 70, Maletti et al 08) xLNTs (tree-to-string) q S r NP q VBZ s NP NP VP , , wa , , ga SBAR enjoys PRO kare wa PRO VBZ NP ongaku o kiku no he VBG VP ga daisuki desu he enjoys SBAR listening P NP VBG VP to music listening P NP to music
Solutions We Designed and Tested graph SHRG string xLNTs graph DAG2Tree tree string (take yield) xLNT xLNT xLNTs DAG2Tree (introduce tree graph tree (introduce tree string (take yield) (tree-ify) verbs) pronouns) All transducer cascades are bidirectional : we run forwards for NL generation task, and backwards for NL understanding task.
Issues with SHRG Long distance interactions need to be encoded in • nonterminal set. S S inf_arg0masc instance ARG1 S inf_arg0masc he wants instance her to believe NP arg0masc ARG0 WANT BELIEVE NP arg0masc ARG0 instance S inf_arg0masc girl indicates that the realization of the arg0 of the next verb will be masculine boy
Issues with DAG2Tree All surface forms for reentrant entity nodes must • be generated at the same time The node GIRL is realized by The node GIRL is realized by she she she she she 0
Cascading with xLNT xLNT xLNT xLNTs DAG2Tree (introduce tree graph tree (introduce tree string (take yield) (tree-ify) verbs) pronouns) Breaks down complex mapping into simpler tasks • Most concise solution • Still very low coverage for the understanding task!
Recommend
More recommend