REVISING AN ATN PARSER Giacomo Ferrari, Irina Prodanof Istituto dl Linguistica Computazionale, CNR, Pisa, Italy

1. An ATN parser for Italian has been developed and tested in a series of experiments on complex sentences taken from narrative texts (1,2,3,4). The construction of a complex graemmz as well as the results of our experiments showed some limits and inadequacies of Woods" ATN as it stands (11), especially when used to parse Italian, a relatively freeorder language with a well developed morphological system.

1.1. "Parameter-passing" actions such as SENDR and LIFTR have been often Judged "dirty" operations under a formal view-point. Moreover it is almost impossible to keep the control of the cross-level pa~sed information, unless the ~umber of registers is increased. Such an increase is otherwise unmotivated. This difficulty is more evident in Italian relative clauses such as: L" uomo la cui crudelta" tutti conoseiamo The man (the) whose cruelty everyone knows L" uomo la crudelta" del quale tutti conoscismo The man the cruelty of whom everyone knows L ° uomo del desiderio di vendetta del quale ti he parlato The man about the desire for revenge of whom I spoke to you in which a very heavy use of such actions, as well as of flags, is required. These phenomena, together with the richness of morphological inflexion, increase the necessity for morphological and morpho-syntactic short, long-distance and cross-level tests. - I01 -

1.2, Another problem is proposed by the structural ambi8~ity of sequences of PP's, for which both the "embedding" and the "same-level" hypotheses are presented! a sentence with ) PP's has 5 parses accordlng to the schema (PPI (PP2 (PP3))), (PPI (PF2)(PP3)), (PFI (PI~) PP3),(PP1 P~ (PP3)), (PP1 P~ PP3) An attempt to limit th~s explosion was made by adding oertaln. functional labels depending upon the verb frame. This experiment showed that attention should be focussed on developing an adequate system of functional tests.

1.3. A statistically based, dynsmic heuristic mechanism has been put to work in the traversal of the network. Those paths which have been more frequently traversed in previous parsed frasments of a given text are attempted first, during the analysis of the CUZTent sentence (6).

In this way, the concept of performance based syntactic expectation has been introduced and tested. It appeared that ATN is very suited to the introduction of heuristic techniques and also that these techniques should include a scheduling procedure based on arc types (10).

2. A revision of the linguistic formalism as well as of the parsing alSorithm is now in progress, and should take all these considerations into account. Furthermore, those linsuistio phenomena related to conversation such as anaphora and the analysis of partial and partially correct input must be taken into account in the desis~ing of the formalism and the ATN actions and forms, not only for applicative purposes, but above all in order to enlarge the range of phenomena to be included in a ~ammar.

2.1. The parser relies both on structural and functional Information (7), i,e. a limited collection of surface structures and functional labels. A set of fixed (control) pattsz~s might also be included in this static syntactic knowledge - 102 (B, 9)• The output representation, however, is in terms of functional labels• The utilization of these three data types will be fully discussed.

2.2. It is not our intention to substantially modify the general structure of the algorithm• Storing and retrieving will be modified, in accordance with our observations in 1.1. A single register external to the levels of computation will be progressively set to the results of the analysis during parsing. Access to precise location of this structure is guaranteed by the functional labels. Traditional ATN registers can be locally used for special purposes. A detailed description of our whole mechanism including examples will be given•

Already assigned functional labels cannot be chanEed! left concatenation of them has been preferred (8). A mechanism has also been provided to move pointers from constituents to gape or anaphorical pronouns (inside the sentence borders) (5). This is an attempt to treat by a uniform action or sequence of actions equ/-NP deletion, gapplng, coordination.

2.3. Some local procedures connected with the lexicon have been added to these mechanisms. These procedures include the transmission of feature's to higher level nodes in the structure, the proceduralisation of idiosyncrasies of single leKical entries and "paraphrase'. The latter is expected to reduce nominalizatlon to the corresponding verb with its frame components and to account for linguistic units which have identical distribution and syntactic behavlour.

2.4• The present list of "linguistically* motivated processes is open-ended, as we are now trying to identify others, which can be considered to be linguistic (procedural) °um/versals "• - 103 (1) (2) (3) (4) (5) (6) (7) (8) (9) A.Oappelll, G. Perrari, L. Moretti, I. Prodanof, O. Stocks ~Parsing an Italian text with an ATN parser', report ILC, Flee@ 1978. A.Oappelli, G. Ferrari, L. Morettl, I. Prodanof, "An ATN parser for Italian: an experiment', Atti COLING 78, Stocoolma (microflches). A.Cappelll, G. Ferrari, L. Morettl, I. Prodanof, "Analisi di un testo Italiano con un analizzatore Riviera dl Informatioa, X 3 (1980), pp. 5-13. A.Cappelll, G.Ferrari, L. Morettl, I. Prodanof, "Analisi automatica dell" italiano: alcu.ne oeservazioni*, in La ricerca di base in peicologia. Attl del XVIII Congresso de@ll Psicologl Italianl, Aoireale 29/X - 2/XI 1979, 1980, pp. 246-267. A. Cappelli,G.Perra~i, L. Morsttl, I.lh~odanof, 0.Stookl "Ii trattamento di alcuni fenomenl anaforici mediante un ATN', Atti del semlnario "L'anafora'tenuto preeso l'Accademia della Orueoa, Pirenze 14-16/12/78. In corso di S~ampa, G.Parrarl, O. Stock: "Strategie selection for an ATN Syntactic Parser', in The 18th Annual Meeting of the Assoolation for Computational Linguistics and parasesslon on Topics in Interactive Discourse. Proceedings of the Conference, June 19-22, 1980, Philadelphia, pp. 113-115. R. Kaplan, J. Breenan: Lexieal-Functlonal Grsmmar: A formal system for grammatical representation, to appear in J.Bresnan (ed.), The mental representation of grammatical relations, MIT Press. M. Kay: "Functional Grammar*, in Proc. of the 5th meeting of the Berkeley Linguistic Society, Berkeley 1979, pp. 142-I 58. S.Kwasny, N. Sondhelmer: "Relaxation techniques for parsing ill-formed input', in AJCL 7, 2 (april-June 1981), - 104 O, Stocks del O.Stook: ATN °, O. Stock:

pp. 99-108. (10) E. Weaner, ~. Maratsos: "An ATN approach to comprehension"

in ~° Halle, J. Bresnan, G. Miller (eds.) Linguistic

Theory and PsyChological Reality, NIT Press 1978. (11) W° Woods: "Transition Network Grammars for Natural lanS-

uase Analysis', in CAOM 13~I0, New York, 1970, - 105 -