<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1112">
  <Title>GENERALIZED MEMORY MANIPULATING ACTIONS FOR</Title>
  <Section position="3" start_page="0" end_page="474" type="metho">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> Literature on parsing natural language has recently been concerned with topics which are unfamiliar to traditional Computational Linguistics such as functional similarities of different syntactic constructions (I), reanalysis of 'indelible' structures (9), manipulation of 'indelible' trees (7) Must of the linguistic problems involved are connected with unbounded dependencies and antecedent-referent binding. Grammatical formalisms have been developed, which represent certain linguistic phenomena in terms of discontinuous constituents (10), or generate unboundedly dependent symbols on the basis of specific conventions (4,5l However, the corresponding parsing algorithms are simple modifications of traditional parsers, where the extentions of h)rmalisms are not adequately accounted for on the procedural side Our approach to the problem of parsing Natural Language is to identify a set of processing strategies which may not only process, but also represent the alluded phenomena within the frame of a psychologically motivated CF parsing algorithm, ~_. Generalized memory manipulating actions EXAM-LC (EXtondend Access and Manipulation of the Left Context) is a model of linguistic structures storing and accessing designed according to such a psychological procedural approach. It relies upon the following assumptions -. the left context, ie. the structure corresponding to the currently analyzed part of the input string, is often affected in some way by newly incoming information and may contain information which affect further analysis Therefore, it must be accessible beyond possible limits imposed by the structure of the parser. In fact, in many cases the data within the scope of the current constituent or rule are more freely accessible than the others. On the contrary, the scope of accessibility to the left context should be specified according to how far it is affected or affects the current analysis; -a set of general actions can be defined corresponding to mental operations actually accomplished during the process of comprehension. Any surface construction is described both by its (deep) representation and by the operations which perform the mapping These actions can be assumed as linguistic procedural universals: -. it is possible that syntactic phenomena, that have different structural explications, are handled by a common process or sequence of operations A common space of memory is assumed to contain the current hypothesis about the analysis of the parsed segment of the input from the beginning We will refer to such a structured space as Current Global \]lypethesis (CGH) The following set of abstract operations on the space of memory has been defined till now a) an opening and a closing action wsll respectively start and end the storing of the information related to a phrase/clause in a current subspace subsequently merged with the global space The way of storlrtg depends upot't the representation of the output and the corresponding actions are desgined ill accordance Io it b) a t'et)'Jevth,C/ action revolving two partlcipants a symbol that triggers the action {trig,goD and the information to be retrieved (the target ef tile actk)nl will retrieve ent~;e constituents which appear to be possible antecedents fragments of structure, or even simple lexical features The trigger may be a gap a pronoun an ellipsis and any other phenomenon whicil requires the search for an antecedent (,garRet}. or the need for an agreement After the identification of such a t1:4C/Rcr the actton is decomposed in three steps i) extraction of constraints which must guide the search for the target iilscanning of the CG\[I under the specified conditions and iii) retrieving of the requtretl information f)n this functional ground this description fits to more or less all the cases mentioned in ( 1 ) The action of searching back may be constrained by several types of restrictions, including i) morphologicalfeatures ie the gender and number of a pronoun or those required for agreement by the syntactic environment (e g the verb), ii) syntactic idiosyncrasies of a lexical item: this is the case of ~,TRAN.q verbs tie these verbs requiring a complement clause) that determine which of their arguments is to be the subject of the complement, as shown in ( 1 )a A credette di udire tm rumere A thought to hear a noise b A persuase B a fuggire A persuaded B to run away c A ordino' a B di sedate h imposed to B to sit down iii) semantic features or cognitive descriptions that may be introduced in the process, and iv) syntactic determination of the scope of the search. such as, for instance, Suh/aceao&amp;quot; Retrieving of an antecedent may actually correspond to two different operations depending upon whether the antecedent te be bound linearly proceeds or follows the  symbol it is to be bound to. In fact, in many common sentences the antecedent linearly follows its dependent, as in (2)a Ouando_, si arrabbia, Giovanni diventa fosse When (he) gets angry, John becomes red b Se 1o vedi. saluta Giovanni da parte mia If (you) see him, say hallo to John on my behalf In this case, the binding should take place in two steps, the flagging of the need for a forward binding and the moving of the pointer from the antecedent, once detected, to the flag c) Many of the retrieving actions are to end up with a binding or, more generally, with a movlhdC/ action. This can be realized at least in two ways. by actually copying the retrieved constituent into the trigger's subspace or simply moving a pointer from one to the other. Also in this case the choice is a matter of representation.</Paragraph>
    <Paragraph position="1"> d ) A riC/oafih, tlt'ation action is necessary in order to modify an ah'eady (partially) built structure, as new incoming information indicate the need for such a change The types of modifications to be performed again depend upon the representation of the output, but the one required by the relative pronoun is likely to be a general one.</Paragraph>
    <Paragraph position="2"> Relative pronouns need tO be bound to an antecedent and, besides, are the surface signal of an embedding No '.~pecial processing difficulty is proposed by the sentence  (3) il ragazzo che corre the boy whoruns  where the relative pronoun occurs exactly where the embedding begins In this case a scope restriction can limit the search for an antecedent to the immediately proceeding NP But in the case of p/ed-plpiag as in (4) il cane della fedelta' del quale nessuno dubita the dog about the fidelity of which nobody doubts the relative clause bounderyis set three (in english four) words before the relative pronoun. In the framework we have been discussing an action which structurally modifies the left context can be proposed, which should embed the component(s) being processed in a relative clause as the relative pronoun is met e ) A final type of access to the left context is the relabe/ling of a processed component, ah'eady used for the passive transformation.</Paragraph>
    <Paragraph position="3"> 3. An experimental implemnntation An experimental realization of the above discussed ideas has been implemented in the frame of the ATN parsing algorithm (13) running a functional grammar a'la M. Kay (6), ATN has been preferred because it meets the requirements of being a psychologically motivated CF parser. Moreover ATN can be considered a very well stabilized parsing algorithm. The data structure is a list which is mainly accessed with a typical LIFO stack policy.. It represents a unique memory space non splitted into registers. it contains at any point of the process the CGH, i.e. the entire left context literally represented in terms of attribute-value pairs, as required by functional grammar.</Paragraph>
    <Paragraph position="4"> We give hereafter a list of the functions which access the  The basic storing action is ADD which is used to store any incoming piece of structure.</Paragraph>
    <Paragraph position="5"> Extraction of information is done by the forms FIND, which returns a pair, and FINDVAL, which returns only the vMue of a pair. LOCATE works exactly in the same way,,but returns a pointer to a given radix. All the three functions can work in different modes They can search either only the current level (CI.) or through the entire list (T) In this latter case the current level is excluded and, if no further options are specified, the lower (the nearest to the top) occurrence is returned Another option (dtype) returns all the occurrences either appended in a list (L) or one by one.</Paragraph>
    <Paragraph position="6"> non-deterministically (ND). A third option evaluates conditions in order to select the component identified by the specified path.</Paragraph>
    <Paragraph position="7"> The three last actions. PUSH, POP, and INSERT, manipulate the items in the list. PUSlt adds anew (empty) item in front of the list. The elements of the component being analysed (phrases or sentences) are ADDed in this top item, which has been therefore referred to as current level. POP removes the current top-item and embeds it into the new top-item, possibly assigning a label to the corresponding component. Finally INSERT inserts an item, corresponding to a new level, somewhere back between 'item' and the front part of the list, and fills it with 'data'. It is the only instance of reconfiguration action designed till now while others can be introduced according to the studied phenomena and the adopted representation.</Paragraph>
    <Paragraph position="8"> List manipulation takes place independently from the starting or the ending of the process expressed in a subnet. Thus a component can be POPed after the end of its recognition procedure, when also its function is clarified. This list of actions is open-ended and is supposed to be updated as new general operations are identified.</Paragraph>
    <Paragraph position="9"> The retrieving action scans the entire left context according to specified searching constraints and is able to return any required fragment, regardless of its structure. In particular, information associated to the loxical items already inserted in it structure can be extracted for further processing at any point, without having previously been saved in a register or raised to higher nodes, as often lexical features are. This turns out to be particularly usofull for the treatment of lexical idiosyncrasies. We have already mentioned that an STRANS verb determines which of its arguments is to be the subject of the complement clause. Many other words have special linguistic behaviours. The possibility of treating such syntactic peculiarities of lexical items by associating particular fragments of grammar to lexical entries and processing them when needed has already been experimented (3) A wider and easier use of such a device contributes to our parser certain advantages of word expert parsing still sticking to a syntactic model of analysis.</Paragraph>
    <Paragraph position="10"> The details of this implementation and some examples of processing have been discussed in ( I 1 ).</Paragraph>
  </Section>
  <Section position="4" start_page="474" end_page="474" type="metho">
    <SectionTitle>
4. An intuitive psychological argument
</SectionTitle>
    <Paragraph position="0"> It has been shown (7, 11) that a recoafiguration action.</Paragraph>
    <Paragraph position="1"> which modifies the structure of the already analysed part of the input string limits backtrack, since a certain class of parsing errors can be corrected without actually restarting the processing of the input string. Under a psychological viewpoint the hypothesis is that during the comprehension of a sentence guesses (CGH's) are progressively enriched and stored in a space of memory. In this process errors may occur For some of them, such as relative clauses (11) and adjective attachment (7), it is enough to modify the previous guess while for others a real backtrack and reanalysis is necessary A possible explanation is that in the activity of sentence comprehension a phase of structuring is distinguished from a phase of perception. Errors occuring in the former are remedied by modifying a guess, while those occuring in tile latter need backtrack and the choice of another strategy</Paragraph>
  </Section>
class="xml-element"></Paper>