<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-3084">
  <Title>A PARLOG Implementation of Government-Binding Theory</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
SCOPE AND OBJECTIVES
</SectionTitle>
    <Paragraph position="0"> In oMer to explore ttle relationsllips between GB-subsystems and their realization in a parallel parser, a development goal is one of modularity ill that each GB.-module should be transparently encoded. 2 This perlnits investigation of processes within and across GI:~ components to be made explicit, Since the focus of tile parscr is on GB principles and since GB itself is a theory of core gralnmar, file coverage of the parser is restricted to a subset of English that reflects syntactic processes explained by the theory.</Paragraph>
    <Paragraph position="1"> While it may be premature to speak of psychologically-real parallel parsers, many of the cognitive presuppositions that are the basis of the work of/Berwick and Weinberg 1984/./Marcus 1980/, and/Milne 1983/have been the motivation for several of the design decisions that have been incorporated in the system.</Paragraph>
    <Paragraph position="2"> More specifically, the parser is deterministic and has the capability of delaying certain actions, e.g., projection of nodes and resolution of lexical ambiguity, until more information is available, rather than computing competing parses or backtracking uncontrollably. These features together with the goal of transparent representation of GB principles should provide a foundation for future research into tile cognitive plausibility of parallel parsers.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
AN IMPLEMENTATION NOTE
</SectionTitle>
    <Paragraph position="0"> The parser is being developed in PC-PARLOG: requiring an IBM PC, XT, AT or compatible machine with tit least 512Kb of memory and two floppy disk drivcs or a hard disk.</Paragraph>
    <Paragraph position="1"> Although tile inlplenlentation sinmlates para!lelism by atin~esharing scheduler, tile parser run~ on a single processor nlachinc and, therefore, lacks true parallelism, ttowever, insofar as the purpose of the parser is to specify concurrency of G13-subsystems at a high level of abstraction, tile analyses are in ternls of compulalional processes rather than processors. As parallel haMware becomes tnore readily available, a valuable By-product of this research is that tile system coukl be ported to parallel logic machines with little or no additional elTort.</Paragraph>
    <Paragraph position="3"> mmg language based on logic (/Con lon 1989/, /Gregory 1987/, and/Shapiro 1988/). &amp;quot;File basic fornl of a clause is (I) head &lt;-- guard: body where tile head is a goal r(tl ..... tn) and both guard (optiomfl) and hod)(obligatory) are conjunctions of goals. Each procedure nlust be preceded by a mode declaration specifying input ('?) and  output (A) arguments. Sequential a xl para lel co lit lctio a (AND) and sequential and parallel disjunction (OR) are represented by &amp;quot;'&amp;&amp;quot;, &amp;quot;,&amp;quot;, &amp;quot;';&amp;quot;, &amp;quot;'.&amp;quot;, respectively.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
THE PARSER
</SectionTitle>
    <Paragraph position="0"> The parser combines top-dov, n and bottom-up strategies and recovers a set of licensing relations directly; phrase structure is considered derivative and is not cornputed.+ The basic operations rely on the current governing category as the left bounded context and are deterministic in the sense that once a node is typed or licensed it cannot be altered. The output of the parser is a constructed list of structures  (2) CP = lcp(C(mlplementizer),ext arg(Subject),</Paragraph>
    <Paragraph position="2"> where cp represents the complementizer phrase, Ihe head of the CP or S-bar, inll denotes the inflectional elemenl, tile head of S, and predicate is tile verb. Tile struclures ext_arg, int_arg_l, and int__arg 2 are tile arguments of the verb, the lirst being the external argument (subject) and the last two being tile (direct) internal arguments (objects). The indireet_arg is a prepositional phrase (PP) that a verb may license and there may be several of these structures depending on the lexical specifications of the verb. While the cp, ext_arg, infl, and predicate are obligatory, the internal and indirect arguments Irre contingent upon the 0-grid of the verb and, therefore, are optional.</Paragraph>
    <Paragraph position="3"> The basic objects of the system are nodes where a node is a structure of the form xp(Word,x_bar(Features),lndex). Word is a lexical item, an empty category PRO, trace, or variable, or an empty complementizer or inflectional element, and Featm'es represents the type of the node in terms of x-bar features_+N,_+V.</Paragraph>
    <Paragraph position="4"> Every node receives a unique index unless it is bound (coindexed) to another via Binding or Control Theory. Each of the terms of (2) are nodes except for the arguments which are a list of nodes that represent the specifier, head, and complement structures. Specifiers may be determiner or adjectival phrases, and complemeuts may be PPs or, in the case of retative clauses, a CP, which would be reflected in another list whose structure is analagous to (2).</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
THE LEXICON
</SectionTitle>
    <Paragraph position="0"> In accordance with GB Theory, the lexicon plays :t cent,'al role.</Paragraph>
    <Paragraph position="1"> Each lexical item contains the idiosyncratic features of the lexeme and they direct many actions of the parser. The lexicon is a database of PARLOG assertions which may be searched in parallel. A small subset of the lexicon is  (3) mode lexeme(item?,x_bar ^, feature^).</Paragraph>
    <Paragraph position="2"> a. lexeme(man, \[x_bar(nl,vo) l,\[pl(men)\]).</Paragraph>
    <Paragraph position="3"> b. lexeme(put, \[x_bar(no,v l)\],\[...,s(puts), ing(putting), 0-grid(int_arg_ 1, locative_PP),, i.\] ).</Paragraph>
    <Paragraph position="4"> c. lexeme(plan, \[x._bar(no,v l),x bar(n l,vo)\], \[ .... 0-grid (proposition), tenseless,subject control,... \] ).</Paragraph>
    <Paragraph position="5">  The first te,'m of each lexeme is the lexical entry and the second term is a list of structures of the form x bar(F I ,F2), where F 1 and F2 are x-bar primitives, _+N, +_V, with &amp;quot;l&amp;quot; and &amp;quot;0&amp;quot; denoting &amp;quot;+&amp;quot; and &amp;quot;-&amp;quot;, respectively. Lexically ambiguous items, e.g., (3)c., have more tlmn one x-bar structure. The third term is a list of symbols and structures that depend on the lexical item and its type, i.e., x-bar features. For instance, as (3)a. illustrates, nouns (x_bar(nl,vo)) have a plural form as It member in their list of features while verbs (x bar(no,vl )) have morphological derivatives, e.g., participial forms or ing, in their features list as (3)b. shows. Verbs also have 0-grids which characterize argument structures and other features indicating specifics of complement structure such as tenseless (infinitival) subcatcgorizations or control criteria, e.g., subject_control ((3)c.).&amp;quot; In the currant version of the parser, derivational morphology is minimal.</Paragraph>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
THE PARSING ENGINE
</SectionTitle>
    <Paragraph position="0"> The basic actions of tile parser lrre to identify gaps, to construct maximal projections, and to license these projections, and these are encoded by tile respective procedures, detect gap, project, and license. The parsing loop is called recursively until tile sentemial input, a list, is exhausted. During a parse, governing categories are produced and used as left-bounded context for&amp;quot; certain procedures before being shunted to the list that will ultimately represent the output.</Paragraph>
  </Section>
  <Section position="8" start_page="0" end_page="395" type="metho">
    <SectionTitle>
GB MODULES
</SectionTitle>
    <Paragraph position="0"> The GB-modules of Trace, Binding, Control, and Bounding Theories are incorporated in the parser, especially in goals detect_gap and license. Detect_gap identifies the presence, if any, of an implicit elemeut. This may be an argument, viz., trace, PRO, or variable, or a non-lexical item, viz., an empty complementizer or inflectional element. A fi'agment ~' of the PARLOG code for detect_gap is  (4) mode detect gap(sentence?,gov__cat^,empty_cat^).</Paragraph>
    <Paragraph position="1"> a. detect_gap(\[WordlWords\],Gov cat,Empty cat)&lt;--</Paragraph>
    <Paragraph position="3"> Tile first relation has a guard to determine if the current word has passive morphology. In formally, the guard passivemorphology checks the inflectional element of the CtuTent governing category fro&amp;quot; a form of BE and, in parallel, determines if the current word is a verb ot'the passive form. If the guard succeeds, then the output substitution is complete with Empty_eat being unified with trace. In (4)b., the guards provide checks with the current token (Word) and governing category for the features of the verb and possible presence of an overt subjec~ in order that detect gap may detect a PRO. The rest of the specifications for detect gap have similar strategies for finding variables and empty inflections and complementizers. Each of these clauses are computed in parallel together with one that determines an absence of an empty category in tile current position of the sentence.</Paragraph>
    <Paragraph position="4"> The next goal in parse is project, which constructs a maximal projection, xp(ltem, Type). In the case that Item is lexicalty ambiguous, project has calls to lexical disambiguation routines which are invoked in parallel to attempt a resolution.</Paragraph>
    <Paragraph position="5"> The goal license indexes and assigns 0-roles to arguments of predicates. Binding Theory (/Cllomsky 1981/) has three principles, frequently labeled A, B, and C m the GB literature, that specify- co-indexation procedures for irnaphm's, pronominals.</Paragraph>
    <Paragraph position="6"> and referential (R-) expressions, respectively. Since the principles are independent, they are canclidates for parallel execution. Thus, Binding is specified in the system as</Paragraph>
    <Paragraph position="8"> i s pronom in a l ( Proj ): principle_b(Proj,Gov cat, ludexed_pmj). bi nding(Proj,Gov_cat,lnde xed_pmj) +-is r_cxp(Proj): principle c(Proj,Gov_caI, Indexed proj). The purpose of the guards is to determine tile argument type of a particular argument (Proj) which then invokes a call tn a specific method of indexing in accordance with Binding Theory. (There is a numerical indexing scheme that is embedded in each principle.) The result of binding is a node, Indexed_proj, that has an index associated with it.</Paragraph>
    <Paragraph position="9"> The other major part of license is the assigmnem of 0-roles. The parser assigns 0-roles to the respective arguments of the predicates based on their 0-grids. Tile binding and ()-assignment procedures together comprise the licensing procedure. Their communication channels are being explored and stream AND parallelism appears promising.</Paragraph>
    <Paragraph position="10">  predicate(persuaded,x_bar(no,v t),4), int arg_ 1 (trace,x bar(n I ,vo),2), int_arg 2(cp l,cp,5), \[cp_l (emp,cp,5), ext arg(pro,x bar(n I ,vo),2), in fl(to,infl,6), predicate(leave,x bar(no,v 1 ),7)\]1, which illustrates PP, O and trace detection and binding.</Paragraph>
  </Section>
  <Section position="9" start_page="395" end_page="395" type="metho">
    <SectionTitle>
FUTURE DIRECTIONS
</SectionTitle>
    <Paragraph position="0"> The primary focus of the experimental parser will be to include wider coverage within st GB-framework, including adjuncts, and to examine psychological aspects of concurrent language processing.</Paragraph>
  </Section>
class="xml-element"></Paper>