Chart-based RRG parsing using an automatically extracted RRG …€¦ · CO NUC V think RP Asterix TNS does CL CL CO NUC V drank RP Obelix PrCS RP What CL CL CO NUC V drank RP Obelix

Design of RRG Grammars Automatic RRG Grammar Extraction Parsing experiments Issues Summary & Outlook

Chart-based RRG parsingusing an automatically extracted

RRG grammar with features

David Arps & Tatiana Bladier & Laura Kallmeyer

International Conference on Role and Reference Grammar19 August 2019

University at Buffalo, The State University of New York

1 / 32


Overview

Design of RRG Grammars

Automatic RRG Grammar Extraction

Parsing experiments

Issues

Summary & Outlook

2 / 32


Introduction

1 RRG annotated treebank:RRGBank [2],

2 extract RRG elementarytrees automatically,

3 use these elementarytrees for RRG parsing.

SENTENCE

CLAUSE

CORE

NP

PRO

it

V

looks

NUC

PP

COREP

NUCP

P

like

DEF-OP

a

N

NUCN

COREN

NP

holiday

3 / 32


Introduction




NP

PRO

it

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

looks

PP↓

PP

COREP

NUCP

P

like

NP↓

NP∗

DEF-OP

a

NP

COREN

NUCN

N

holiday

3 / 32


Introduction




NP

PRO

it

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

looks

PP↓

PP

COREP

NUCP

P

like

NP↓

NP∗

DEF-OP

a

NP

COREN

NUCN

N

holiday

3 / 32


Outline



Parsing experiments

Issues

Summary & Outlook

4 / 32


RRGbank

? Corpus of RRG annotated sentences [2]→ automatically converted from PennTreebank,→ manually checked and validated;

? 395 gold sentences, 1090 silverannotated sentences;

? RRGbank and RRG annotation tool:rrgbank.phil.hhu.de.

SENTENCE

CLAUSE

CORE

NP

COREN

NUCN

N

Japanese

NUC

V

said

CORE

CLM

to

NUC

AUX

be

NP

COREN

AP-PERI

A

heavy

NUCN

N

buyers

TNS-OP

were

V

NUC

CORE

CLAUSETNS

5 / 32

rrgbank.phil.hhu.de


Removing crossing branchesSENTENCE

CLAUSE

CORE

NP

COREN

NUCN

N

Japanese

NUC

V

said

CORE

CLM

to

NUC

AUX

be

NP

COREN

AP-PERI

A

heavy

NUCN

N

buyers

TNS-OP

were

V

NUC

CORE

CLAUSETNS

→

SENTENCE

CLAUSE

CORE

NP

COREN

NUCN

N

Japanese

NUC

V

said

CORE

CLM

to

NUC

AUX

be

NP

COREN

AP-PERI

A

heavy

NUCN

N

buyers

TNS-OP [OP=CL]

were

V

NUC

CORE

CLAUSETNS? we transform the RRG structures to remove crossing branches,

? we mark the original position of the node with [op=cl],? original tree structure is easily recovered.

6 / 32



CLAUSE

CORE

NP

COREN

NUCN

N

Japanese

NUC

V

said

CORE

CLM

to

NUC

AUX

be

NP

COREN

AP-PERI

A

heavy

NUCN

N

buyers

TNS-OP

were

V

NUC

CORE

CLAUSETNS

→

SENTENCE

CLAUSE

CORE

NP

COREN

NUCN

N

Japanese

NUC

V

said

CORE

CLM

to

NUC

AUX

be

NP

COREN

AP-PERI

A

heavy

NUCN

N

buyers

TNS-OP [OP=CL]

were

V

NUC

CORE

CLAUSETNS? we transform the RRG structures to remove crossing branches,? we mark the original position of the node with [op=cl],

? original tree structure is easily recovered.

6 / 32



CLAUSE

CORE

NP

COREN

NUCN

N

Japanese

NUC

V

said

CORE

CLM

to

NUC

AUX

be

NP

COREN

AP-PERI

A

heavy

NUCN

N

buyers

TNS-OP

were

V

NUC

CORE

CLAUSETNS

→

SENTENCE

CLAUSE

CORE

NP

COREN

NUCN

N

Japanese

NUC

V

said

CORE

CLM

to

NUC

AUX

be

NP

COREN

AP-PERI

A

heavy

NUCN

N

buyers

TNS-OP [OP=CL]

were

V

NUC

CORE

CLAUSETNS? we transform the RRG structures to remove crossing branches,? we mark the original position of the node with [op=cl],? original tree structure is easily recovered.

6 / 32


Operator projection and periphery can be recovered

SENTENCE

CLAUSE

CORE

NP

COREN

NUCN

N

Japanese

NUC

V

said

CORE

CLM

to

NUC

AUX

be

NP

COREN

AP-PERI

A

heavy

NUCN

N

buyers

TNS-OP

were

V

NUC

CORE

CLAUSETNS

→

SENTENCE

CLAUSE

CORE

NP

COREN

NUCN

N

Japanese

NUC

V

said

CORECLM

to

NUC

AUX

be

NP

COREN

AP

PERIPHERY

heavy

NUCN

N

buyerswere

V

NUC

CORE

CLAUSETNS

7 / 32


Elementary trees in RRG Grammars

? We follow Kallmeyer et al. (2013) and Osswald & Kallmeyer(2018) [3, 4] in design of the elementary trees in our grammar.

? Three tree composition operations:→ substitution ( argument slot filling)→ wrapping substitution (displaced argument slot filling)→ sister adjunction ( adding operators and periphery elements);

? Such RRG grammars capture long-distance dependencies→ for example, WH-movement.

8 / 32




? Three tree composition operations:

→ substitution ( argument slot filling)→ wrapping substitution (displaced argument slot filling)→ sister adjunction ( adding operators and periphery elements);


8 / 32




? Three tree composition operations:→ substitution ( argument slot filling)

→ wrapping substitution (displaced argument slot filling)→ sister adjunction ( adding operators and periphery elements);


8 / 32




? Three tree composition operations:→ substitution ( argument slot filling)→ wrapping substitution (displaced argument slot filling)

→ sister adjunction ( adding operators and periphery elements);


8 / 32




? Three tree composition operations:→ substitution ( argument slot filling)→ wrapping substitution (displaced argument slot filling)→ sister adjunction ( adding operators and periphery elements);


8 / 32


Combination operations: Substitution and sister adjunction

NP

PRO

I

CORE∗

TNS-OP [OP=CL]

have

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

felt

NP↓

CORE*N

AP-PERI

COREA

NUCA

A

many

NP

COREN

NUCN

N

aftershocks

Sentence: I have felt many aftershocks

9 / 32



NP

PRO

I

CORE∗

TNS-OP [OP=CL]

have

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

felt

NP↓

CORE*N

AP-PERI

COREA

NUCA

A

many

NP

COREN

NUCN

N

aftershocks


9 / 32



NP

PRO

I

CORE∗

TNS-OP [OP=CL]

have

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

felt

NP↓

CORE*N

AP-PERI

COREA

NUCA

A

many

NP

COREN

NUCN

N

aftershocks


9 / 32



NP

PRO

I

CORE∗

TNS-OP [OP=CL]

have

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

felt

NP↓

CORE*N

AP-PERI

COREA

NUCA

A

many

NP

COREN

NUCN

N

aftershocks


9 / 32



NP

PRO

I

CORE∗

TNS-OP [OP=CL]

have

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

felt

NP↓

CORE*N

AP-PERI

COREA

NUCA

A

many

NP

COREN

NUCN

N

aftershocks


9 / 32



NP

PRO

I

CORE∗

TNS-OP [OP=CL]

have

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

felt

NP↓

CORE*N

AP-PERI

COREA

NUCA

A

many

NP

COREN

NUCN

N

aftershocks


9 / 32



NP

PRO

I

CORE∗

TNS-OP [OP=CL]

have

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

felt

NP↓

CORE*N

AP-PERI

COREA

NUCA

A

many

NP

COREN

NUCN

N

aftershocks


9 / 32



NP

PRO

I

CORE∗

TNS-OP [OP=CL]

have

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

felt

NP↓ CORE*N

AP-PERI

COREA

NUCA

A

many

NP

COREN

NUCN

N

aftershocks


9 / 32



NP

PRO

I

CORE∗

TNS-OP [OP=CL]

have

SENTENCE

CLAUSE

CORE

NP↓ NUC

V

felt

NP↓ CORE*N

AP-PERI

COREA

NUCA

A

many

NP

COREN

NUCN

N

aftershocks


9 / 32



SENTENCE

CLAUSE

CORE

NP

PRO

I felt

V

NUC

NP

AP-PERI

COREA

NUCA

A

many

NUCN

N

aftershocks

TNS-OP [OP=CL]

have

COREN


10 / 32


Combination operations: Wrapping substitution

CL

CL↓CO

NUC

V

think

RP

Asterix

TNS

does

CL

CL

CO

NUC

V

drank

RP

Obelix

PrCS

RP

What

CL

CL

CO

NUC

V

drank

RP

Obelix

CO

NUC

V

think

RP

Asterix

TNS

does

PrCS

RP

What

Sentence: What does Asterix think Obelix drank

11 / 32



CL

CL↓CO

NUC

V

think

RP

Asterix

TNS

does

CL

CL

CO

NUC

V

drank

RP

Obelix

PrCS

RP

What

CL

CL

CO

NUC

V

drank

RP

Obelix

CO

NUC

V

think

RP

Asterix

TNS

does

PrCS

RP

What


11 / 32



CL

CL↓CO

NUC

V

think

RP

Asterix

TNS

does

CL

CL

CO

NUC

V

drank

RP

Obelix

PrCS

RP

What

CL

CL

CO

NUC

V

drank

RP

Obelix

CO

NUC

V

think

RP

Asterix

TNS

does

PrCS

RP

What


11 / 32



CL

CL↓CO

NUC

V

think

RP

Asterix

TNS

does

CL

CL

CO

NUC

V

drank

RP

Obelix

PrCS

RP

What

CL

CL

CO

NUC

V

drank

RP

Obelix

CO

NUC

V

think

RP

Asterix

TNS

does

PrCS

RP

What


11 / 32


Elementary trees with features

RRGbank Extracted Elementary Trees

NP

DEF-OP

the

CORE-N

NUC-N

N

average placeholder

NP∗

DEF-OP

the

NP

CORE-N

NUC-N

N

average

[DEF -OPS [NP -]

][DEF +OPS [NP +]

][DEF +]

[OP=NP]

[NUC-O lOPS [NP -]

] [NUC-O rOPS [NP -]

]

? elementary trees are enhanced with features

→ edge features→ node features

12 / 32




NP

DEF-OP

the

CORE-N

NUC-N

N

average placeholder

NP∗

DEF-OP

the

NP

CORE-N

NUC-N

N

average

[DEF -OPS [NP -]

][DEF +OPS [NP +]

][DEF +]

[OP=NP]

[NUC-O lOPS [NP -]


]

? elementary trees are enhanced with features→ edge features

→ node features

12 / 32




NP

DEF-OP

the

CORE-N

NUC-N

N

average placeholder

NP∗

DEF-OP

the

NP

CORE-N

NUC-N

N

average

[DEF -OPS [NP -]

][DEF +OPS [NP +]

][DEF +]

[OP=NP]

[NUC-O lOPS [NP -]


]

? elementary trees are enhanced with features→ edge features→ node features

12 / 32


Edge featuresRRGbank Extracted Elementary Trees

NP

DEF-OP

the

CORE-N

NUC-N

N

average placeholder

NP∗

DEF-OP

the

NP

CORE-N

NUC-N

N

average

[DEF -OPS [NP -]

][DEF +OPS [NP +]

][DEF +]

[OP=NP]

[NUC-O lOPS [NP -]


]

Left and right edge feature structures:

? unify adjacent structures in the derived tree? model ordering constraints? percolate upwards until phrasal nodes

13 / 32



NP

DEF-OP

the

CORE-N

NUC-N

N

average placeholder

NP∗

DEF-OP

the

NP

CORE-N

NUC-N

N

average

[DEF -OPS [NP -]

][DEF +OPS [NP +]

][DEF +]

[OP=NP]

[NUC-O lOPS [NP -]


]

Left and right edge feature structures:? unify adjacent structures in the derived tree

? model ordering constraints? percolate upwards until phrasal nodes

13 / 32



NP

DEF-OP

the

CORE-N

NUC-N

N

average placeholder

NP∗

DEF-OP

the

NP

CORE-N

NUC-N

N

average

[DEF -OPS [NP -]

][DEF +OPS [NP +]

][DEF +]

[OP=NP]

[NUC-O lOPS [NP -]


]

Left and right edge feature structures:? unify adjacent structures in the derived tree? model ordering constraints

? percolate upwards until phrasal nodes

13 / 32



NP

DEF-OP

the

CORE-N

NUC-N

N

average placeholder

NP∗

DEF-OP

the

NP

CORE-N

NUC-N

N

average

[DEF -OPS [NP -]

][DEF +OPS [NP +]

][DEF +]

[OP=NP]

[NUC-O lOPS [NP -]


]

Left and right edge feature structures:? unify adjacent structures in the derived tree? model ordering constraints? percolate upwards until phrasal nodes

13 / 32


Node featuresRRGbank Extracted Elementary Trees

NP

DEF-OP

the

CORE-N

NUC-N

N

average placeholder

NP∗

DEF-OP

the

NP

CORE-N

NUC-N

N

average

[DEF -OPS [NP -]

][DEF +OPS [NP +]

][DEF +]

[OP=NP]

[NUC-O lOPS [NP -]


]

? One feature structure per node:• unify during tree composition• store syntactic or syn-sem interface information.

? Unification successful → accept parse tree

14 / 32


Outline



Parsing experiments

Issues

Summary & Outlook

15 / 32


RRG Grammar extraction algorithm (1)

? Elementary tree extraction inspired by Xia [6] algorithmfor induction of Tree-Adjoining Grammars.

? Top-down extraction of elementary trees.? Heuristics from head-modifier percolation tables.? We use RRG structures from RRGbank for automatic grammarinduction→ rrgbank.phil.hhu.de.

16 / 32

rrgbank.phil.hhu.de


RRG Grammar extraction algorithm (2)

ROOT

CLAUSE

CORE

NP

COREN

NUCN

DEF-OP

The

N

average

V

dropped

NUC NP

COREN

CD

27

NUCN

N

points

TNS-OP [OP=CL]

had

NP∗

DEF-OP[OP = NP,DEF = +]

The

NP

COREN

NUCN

N

average

CORE∗

TNS-OP[OP = CL,TNS = past]

had

ROOT

CLAUSE

CORE

NP↓ NUC

V

dropped

NP↓

COREN∗

CD

27

NP

COREN

NUCN

N

points

17 / 32


Outline



Parsing experiments

Issues

Summary & Outlook

18 / 32


Chart-based RRG Parser TuLiPA

? input: set of elementary trees and sentences to parse;

? output: all derivations that can be derived by combining theelementary trees;

? standard CYK algorithm;? bottom-up, left-to-right traversion of the derived tree;? software: TuLiPA RRG parser [1]

(https://github.com/spetitjean/TuLiPA-frames)TuLiPA = Tübingen Linguistic Parsing Architecture;

? suitable for hand-crafted precision RRG grammars;? suitable for automatically extracted RRG grammars.

19 / 32

https://github.com/spetitjean/TuLiPA-frames



? input: set of elementary trees and sentences to parse;? output: all derivations that can be derived by combining theelementary trees;

? standard CYK algorithm;? bottom-up, left-to-right traversion of the derived tree;? software: TuLiPA RRG parser [1]



19 / 32





? standard CYK algorithm;

? bottom-up, left-to-right traversion of the derived tree;? software: TuLiPA RRG parser [1]



19 / 32





? standard CYK algorithm;? bottom-up, left-to-right traversion of the derived tree;

? software: TuLiPA RRG parser [1](https://github.com/spetitjean/TuLiPA-frames)TuLiPA = Tübingen Linguistic Parsing Architecture;


19 / 32





? standard CYK algorithm;? bottom-up, left-to-right traversion of the derived tree;? software: TuLiPA RRG parser [1](https://github.com/spetitjean/TuLiPA-frames)TuLiPA = Tübingen Linguistic Parsing Architecture;


19 / 32






? suitable for hand-crafted precision RRG grammars;

? suitable for automatically extracted RRG grammars.

19 / 32







19 / 32



Extracted RRG grammar

• removed punctuation• exhaustive parsing (i.e. not probabilistic, overgenerating a lot)

• 2 versions:1 no features2 edge features for operators model adjunction constraints

• do feature structures eliminate parse trees that contradictlinguistic intuitions?

20 / 32


Parsing experiments

Gold Grammar Silver Grammarw feats w/o feats w feats w/o feats

Sentences 395 1480avg. sentence length 6.1 8.0token-supertag pairs 1526 1497 6288 6044avg. number of parses 6.9 12.7 1166 2939

savings 45.1% 39.7%

features decrease number of results by ≈ 45%

21 / 32


Features restrict adjunction of operators and periphery (1)

SENTENCE

PP-PERI

COREP

NUCP

By

NP[NUC-O rOPS [NP -]

][NUC-O lOPS [NP -]

]COREN

CD

9:45

CLAUSE

CORE

NP [NUC-O rOPS [NP -]

][NUC-O lOPS [NP -]

][NUC-O lOPS [NP -]

]AP-PERI

COREA

NUCA

A

industrial

COREN

NUCN

N

average

. . .

NP∗

DEF-OP

the

[NUC-O lOPS [NP -]

][NUC-O lOPS [NP +]

]

Sentence: By 9:45, the industrial average had dropped 27 points.

22 / 32


Features restrict adjunction of operators and periphery (2)

SENTENCE

PP-PERI

COREP

NUCP

By

NP

COREN

CD

9:45

CLAUSE

CORE

NP

AP-PERI

COREA

NUCA

A

industrial

COREN

NUCN

N

average

. . .

NP

DEF-OP

the

Sentence: By 9:45, the industrial average had dropped 27 points.

23 / 32


Outline



Parsing experiments

Issues

Summary & Outlook

24 / 32


Attachment ambiguitiesCORE*

PP-PERI

CORE_P

NUC_P

P

in

CORE↓

NP*

PP-PERI

CORE_P

NUC_P

P

in

NP↓

CORE_N*

PP-PERI

CORE_P

NUC_P

P

in

NP↓

CORE*

PP-PERI

CORE_P

NUC_P

P

in

CORE↓

CORE

NP

CORE-N

NUC-N

N

bank

X∗

PP-PERI

CORE_P

X↓NUC_P

P

in

NP∗

NP

CORE-N

NUC-N

N

Texas

syntactic information needed that might not be in RRGBank25 / 32


Size of the grammar

Gold Grammar Silver Grammarw feats w/o feats w feats w/o feats

Sentences 395 1480avg. sentence length 6.1 8.0token-supertag pairs 1526 1497 6288 6044avg. number of parses 6.9 12.7 1166 2939

the number of parses per sentence increases with the size of the grammar

26 / 32


Extraction of wrapping substitution treesCL

CL↓CO

NUC

V

think

RP

Asterix

TNS

does

CL

CL

CO

NUC

V

drank

RP

Obelix

PrCS

RP

What

CL

CL

CO

NUC

V

drank

RP

Obelix

CO

NUC

V

think

RP

Asterix

TNS

does

PrCS

RP

What

? discontinuous constituents are marked with traces in PTB;

? no special marking in RRGBank;

? transfer traces from PTB to RRG trees in RRGbank?27 / 32


Outline



Parsing experiments

Issues

Summary & Outlook

28 / 32


Summary

? automatically extracted elementary trees from RRGBank;? experiments with exhaustive parsing of 395/1480 sentences;? parsing w/o edge features → too many results;? some edge features already rule out bad results.

29 / 32


Future Work

? adding more edge features = better results;? extract wrapping elementary trees;? 100s of results → not satisfying;? ambiguity and annotation/extraction mistakes have badconsequences;

? use unlexicalized elementary trees (= supertags);? probabilistic grammar and parsing→ A* parsing algorithm ParTAGe by Waszczuk (2017) [5];

? Web GUI.

30 / 32


Thank you!

THANK YOU VERY MUCH FOR YOUR ATTENTION!

31 / 32

References

References I

Arps, D. and Petitjean, S. (2018). A Parser for LTAG and Frame Semantics. In chair), N. C. C., Choukri, K., Cieri, C.,Declerck, T., Goggi, S., Hasida, K., Isahara, H., Maegaard, B., Mariani, J., Mazo, H., Moreno, A., Odijk, J.,Piperidis, S., and Tokunaga, T., editors, Proceedings of the Eleventh International Conference on LanguageResources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).

Bladier, T., van Cranenburgh, A., Evang, K., Kallmeyer, L., Möllemann, R., and Osswald, R. (2018). RRGbank: a Roleand Reference Grammar Corpus of Syntactic Structures Extracted from the Penn Treebank. In Proceedings ofthe 17th International Workshop on Treebanks and Linguistic Theories (TLT 2018), December 13–14, 2018,Oslo University, Norway, number 155, pages 5–16. Linköping University Electronic Press.

Kallmeyer, L., Osswald, R., and Van Valin, Jr., R. D. (2013). Tree Wrapping for Role and Reference Grammar. In Morrill,G. and Nederhof, M.-J., editors, Formal Grammar 2012/2013, volume 8036 of LNCS, pages 175–190. Springer.

Osswald, R. and Kallmeyer, L. (2018). Towards a formalization of role and reference grammar. In Kailuweit, R., Künkel,L., and Staudinger, E., editors, Applying and Expanding Role and Reference Grammar., pages 355–378.Albert-Ludwigs-Universität, Universitätsbibliothek. [NIHIN studies], Freiburg.

Waszczuk, J. (2017). Leveraging MWEs in practical TAG parsing: towards the best of the two worlds. PhD thesis.

Xia, F. (1999). Extracting tree adjoining grammars from bracketed corpora. In Proceedings of the 5th Natural LanguageProcessing Pacific Rim Symposium (NLPRS-99), pages 398–403.

32 / 32

Documents

Chart-based RRG parsing using an automatically extracted RRG …€¦ · CO NUC V think RP Asterix TNS does CL CL CO NUC V drank RP Obelix PrCS RP What CL CL CO NUC V drank RP Obelix