{ "paper_id": "W98-0114", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T06:05:04.821784Z" }, "title": "Tree-Grammar Linear Typing for unified Super-Tagging/Probabilistic Parsing Models", "authors": [ { "first": "Ariane", "middle": [], "last": "Halber", "suffix": "", "affiliation": { "laboratory": "", "institution": "Thomson-CSFt", "location": {} }, "email": "ber@enst.fr" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "We integrate super-tagging, guided-parsing and probabilistic parsing in the framework of an item-based LTAG chart parser. Items are based on a linear-typing of trees that encodes their expanding path, starting from their anchor.", "pdf_parse": { "paper_id": "W98-0114", "_pdf_hash": "", "abstract": [ { "text": "We integrate super-tagging, guided-parsing and probabilistic parsing in the framework of an item-based LTAG chart parser. Items are based on a linear-typing of trees that encodes their expanding path, starting from their anchor.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Practical implementations of LTAG parsing bave to face heavy lexical ambiguity and parsing combinatorial ambiguity. Main techniques to address these issues are super-tagging (Joshi and Srinivas, 1994) , which consists in disambiguating elementary trees before parsing; guided-parsing, like head.-driven parsing (van Noord, 1994) or anchor driven parsing (Lavelli and Satta, 1991; Lopez, 1998) ; and probabilistic parsing (Schabes, 1992; Caroll and Weir, 1997) .", "cite_spans": [ { "start": 174, "end": 200, "text": "(Joshi and Srinivas, 1994)", "ref_id": "BIBREF2" }, { "start": 311, "end": 328, "text": "(van Noord, 1994)", "ref_id": "BIBREF10" }, { "start": 354, "end": 379, "text": "(Lavelli and Satta, 1991;", "ref_id": "BIBREF4" }, { "start": 380, "end": 392, "text": "Lopez, 1998)", "ref_id": "BIBREF5" }, { "start": 421, "end": 436, "text": "(Schabes, 1992;", "ref_id": null }, { "start": 437, "end": 459, "text": "Caroll and Weir, 1997)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "All of tbese approaches exploit specific properties of LTAG to improve parsing efficiency, but none is totally satisfactory.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "Guided-parsing is a very nsefull means to limit overgeneration of spurious items in the chart, but it does not provide a new ambiguity bound. Besides, lexical ambiguity remains the main factor of computational load and this problem is only undirectly addressed by such techniques.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "Super-tagging strength is to discard elementary trees while avoiding to go through actual tree combinations. lt exploits instead. local models of Well-Formedness (WF), as those used for tagging, where parse depencies remain implicit or underspecified. The problem though is that if a single tree is incorrect the parse will fail. To be robust, parsing \"ENST Paris, 46 rue Barrault, 75634 Paris Cedex 13", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "IThomson-CSF, LCR, Domaine de Corbeville, 91404", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "Orsay Cedex, FRANCE 54 must thus take several hypothesis into account. This leaves one with two regrets: first, parsing has still to find some way to tackle combinatorial ambiguity, and second, there is a lack of synergy between super-tagging and parsing , while they seem to share a kuowledge about tree potential-combinations.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "Probabilistic parsing offers a way to tune the compromize between accuracy and speed, by thresholding partial parsing paths according to their current Inside probability. This incurs a well-known bias (Goodman, 1998) : At a given derivation step, the lnside-probabilities of parse constituents estimate the relevance of the derivation past, but do not teil anything about its future. This can be corrected by A* cost functions, or Outside-probability estimates.", "cite_spans": [ { "start": 201, "end": 216, "text": "(Goodman, 1998)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "To meet the weak points mentionned above, at least partialy, we develop a unified framework for tbe tbree techniques, and push their interactions further.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "Sharing a parsing framework We propose an item-based chart-parser, where the parsing scheme is expressed as a deduction system (Shieber, Schabes, and Pereira, 1994) .", "cite_spans": [ { "start": 127, "end": 164, "text": "(Shieber, Schabes, and Pereira, 1994)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "This framework is also amenable for expressing probabilistic parsing (Goodman, 1998) .", "cite_spans": [ { "start": 69, "end": 84, "text": "(Goodman, 1998)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "Sharing models for super-tagging and itempruning. Super-tagging can be seen as a treesequel)Ce WF-model, and extended to score derived item-sequences in the cbart, wrt their likelihood of completing a parse. This yields a sound thresholding technique (Rayner and Carter, 1996; Goodman, 1998) .", "cite_spans": [ { "start": 251, "end": 276, "text": "(Rayner and Carter, 1996;", "ref_id": null }, { "start": 277, "end": 291, "text": "Goodman, 1998)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "Sharing tree-types for item-pruning and guided-parsing. To support the WF parametric mode!, trees and itcms are abstracted by theit linear type, which consists in a list of connectors that represent combination properties. Guided-parsing relies on a specific ordering of these connectors, so that a single type drives the parsing deduction and ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "< \u2022\"x, 1\",,rt !'~lrX(r001 J> JJ(\u2022. 1 ,-,-11L\\(ryl f'df r>a1 1 ,k,Ji.1rl <-rr;ri1r r>o[1,k,Ji .Jr] ff E {IX*, S} .i1\"J1 .-.-1 0(11,k,/1 ,J,I ri (t sl spine(o) 0(1,k,fi ,fr] <\"l\\-1,,,,, 11~1Xroot> \u00df(\"J.-.-1 a!J,k,/ 1 ,/,J r~ E {~X,S} n[J,k,fi./rl <1.\\'lrX\u2022> \u00dff\u2022,J,Ji ,J o(J,k,/\"/d <\u20221 l\\\u2022(\u2022'I) .\\'r1.(. j .J-.\\rirXt'l)>a()r,)r.-.-J < .j. X17fdf r> ofj,k,/r.!d > wrnp-1 < \"IXl1-X \u2022> ;Jl\u2022.J.J 1 ,Jr] o(Jt,Jr.f!,J:l of\u2022.1.1: .1:1 < IXj'll fi ll r> n(1.1.J, ./rl [ \u2022. J./1 ./rl <1X(rool)r, lf r> .il\u2022./r.-.-l f1 (t {'IX, S}", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "wrnp-2 adjunction an sub-tree to Categorial Grammar. Left and right walks are exprcssed as stacks of connectors, so that the extreme connector is the one to connect the dosest to the anchor 1 An illustration is given in table 2 for the tree in figure 2.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "Typing strategy. In it.s own walk, the foot bears the adjunction, with type l or r inversly to the foot side. In the opposit walk, the foot-node may be reached as weil, provided that there is a direct path from root to foot. In the deduction system, in table 1, the foot-node of a left or right auxiliary tree achicves adjunction, but the foot-node of a wrapping auxiliaiy tree creates a gap and passes its adjunction", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "1 The derivation is represented as a fully connected and oricnted graph of trees whose edge labcls arc connector names (pr,,,ided that a sub-tree is extracted to decompose wrapping adjunction, cf. capacities to the root-node, with an opposit type for the opposit. sides.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "lt. can bc noted that each nodc that can receive adjunction yields two linked connectors, which bound t he su b-list of connectors of their su b-trce.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "lntroduction", "sec_num": "1" }, { "text": "3 Deductive Chart Parser \\Ve wish to get elementary-like lypes on derived structure , so as to use a super-tagging-like model to prune derived paths. \\Vc t.ry thus to keep as close as possible to trees when driving thc parsing. But we are not aiming at top-down parsing. since we wish lo evaluate deri,\u2022ed paths that span the input. This lcads to isolating wrapping adjunction from left-an < r;'1r~' > Irr 1\u2022\u2022'1\u2022\u2022' d P.' V \u2022 t , \u00b5ro \u2022 o . in r\"r\u2022 V", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Probabilistic Thresholding", "sec_num": "4" }, { "text": ", t r (Goodman, 1998) proposes two useful approximation of the outside score of itcm [s] . in ordcr to correct t hc inside probabilily hi type: abstraction on connector stacks, removes specialh:ed substitutions: co-Anchor: w-l.-+ LEX.).sub-trec: .\\\"17.).-+ X --l .", "type_str": "figure", "uris": null }, "TABREF0": { "num": null, "html": null, "content": "
Goal:<\u2022'
Axioms:
Anchor
co-Anchor
Rulcs:
Substitution
Right Adjunction
Lcrt Adjunction
Lcft Adj on spine
Sub-tree extrnction
Wrap Adj on spinc
No Lcft Adjunction
Gap crcation
", "type_str": "table", "text": "Jl 1 .JS> n[O,n,-,-1 Anchor(o\u2022) = U!i r,' r r connectcd walks of a" }, "TABREF1": { "num": null, "html": null, "content": "
The estimates the pruning model. Ty\u00b5es are described
in section 2, their use in t.hc deduction systcm, m
section 3. their use for itcm-pruning in section 4.
2 Linear Typing
Guiding the Tree expansion \\\\'e guide the pars-
ing by independent left and right connected-walks,
inspired from ( Lavelli and Satta. 19!)1) bidirectional
parscr and (Lopez, 1998) connected routes. Left and
right connected walks follow respcctively left-and
right-monolonic expansion, out ward. from thc an-
chor to t hc root, as dis\u00b5layed in flgure 2. Thcy list
node operations considercd as connectors.
Link-Gramnrnr expression To express linear
types.
", "type_str": "table", "text": "Deducti \\'e syslcm for an L TAG bidirectional chart-parser, lexica\\ly guided and based majoritarily on trces. thanks to a prccom\u00b5ilation of thcir nodes into left and right walks. acllve ('QJIJlector IS \u00b5opcd on e\u2022treme ldl (resp. right) of its Stack r, (resp. rr)-Each connector is associated with its node ~.thou11h we do not always mark 11. The spme 1s the path from anchor to root.wr&p-1, wrnp-2, wrap-3 identifie the three steps of a" }, "TABREF2": { "num": null, "html": null, "content": "
In a right walk, ix\u2022 cxpreoses an auxiliary root-node and 'IX, a
node expecting adjunct1on, X.j. expresses a Substitution Site and
.jX, the root of an initial tree. In a left walk they work the other
way around.
", "type_str": "table", "text": "Typing thc tree in figure 2." } } } }