File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/91/e91-1024_metho.xml
Size: 14,747 bytes
Last Modified: 2025-10-06 14:12:38
<?xml version="1.0" standalone="yes"?> <Paper uid="E91-1024"> <Title>cation of default inheritance mechanisms. In</Title> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 OVERVIEW OF DATR </SectionTitle> <Paragraph position="0"> DATR (described in detail by Evans/ Gazdar, 1989a; 1989b; 1990)is a declarative language for the definition of semantic networks which allows for defaults as well as multiple inheritance. Its general properties are non-monotonicity, functionality, and deterministic search.</Paragraph> <Paragraph position="1"> A DATR theory (or network description) is a set of axthms (or expressions) which are related to each other by references. Together they define a hierarchical structure, a net. Both regularities and exceptions can be expressed, regularities using default inheritance, and exceptions, overriding.</Paragraph> <Paragraph position="2"> DATR axioms consist of node-path pairs associated with a right-hand side. This can be a value (atomic or lis0, or an evaluable DATR expression if the value is to be inherited from another node, path, or node-path pair. The following DATR theory comprising three node definitions I encodes familiar linguistic information to illustrate some relevant DATR features:</Paragraph> <Paragraph position="4"> LEXIC.AL: <syn major bar> ~ zero.</Paragraph> <Paragraph position="5"> NOUN: <> == LEXICAL <syn major nv n> == yes <syn major nv v> == no.</Paragraph> <Paragraph position="6"> ADJ: o == LEXICAL <syn major nv n> == NOUN <syn major nv v> == <syn major nv n>.</Paragraph> <Paragraph position="7"> The represented information can be retrieved with special DATR queries. These also consist of a node-path pair, whose evaluation returns the value sought. With the above DATR description the following examples show sensible DATR queries and their corresponding yes (inherited from node NOUN) ADJ:<syn major nv v> ? yes (inherited from node NOUN via path <syn major nv n> in node ADJ) ADJ:<syn major bar> ? zero (inherited from node LEXICAL) Seven inference rules and a default mechanism are given for the evaluation of DATR queries. Their precise semantics and properties are described in (Evans/Gazdar, 1989b; 1990).</Paragraph> <Paragraph position="8"> A major feature of DATR is its distinction between global and local inheritance. In the above example only local inheritance is involved, but global inheritance plays a crucial role in one of the later examples. Variables constitute an additional device available in DATR but are assumed to have the status of abbreviations.</Paragraph> <Paragraph position="9"> Despite their syntactic similarities, DATR and PATR differ completely in their semantics, so that there is no obvious way of relating the two formalisms to each other. Some approaches are discussed in the next section.</Paragraph> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 RELATING DATR AND PATR </SectionTitle> <Paragraph position="0"> A PATR system needs to have the lexical information it uses encoded in feature structures consisting of attribute-value pairs. The lexical information represented in the DATR theory above (1) would appear as follows when stated in feature structures:</Paragraph> <Paragraph position="2"> information specific to NO: syn.' or. ~nv r; \[&quot; '1/11 tv: nOllll information specific to ADJO: ~ n: najor.</Paragraph> <Paragraph position="3"> r In The question that arises is how to relate DATR and PATR so that the hierarchically structured lexical information in DATR can be made available in PATR-usable feature structures. null</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.1 A DATR-PATR INTERFACE </SectionTitle> <Paragraph position="0"> The first idea that one might have is to exploit the syntactic similarities between the two formalisms and encode the lexical information in a DATR description like (1). In this way a DATR axiom like NOUN: <~yn major nv n> == yes would be directly equivalent to the path equation <NOUN syn major nv n> = yes in PATR, where the node name in DATR corresponds to the variable name for a feature structure in PATR. Although this looks reasonable, one major problem arises: You must know exactly the path you want to query, i.e. all its attributes and their precise order. If such a query is posed, the answer will be the atomic value yielded by the DATR evaluation.</Paragraph> <Paragraph position="1"> Such an approach requires an interface with the following functions: Queries that the grammar writer has stated explicitly have to be passed on to DATR. Every query together with the resulthag value has to be transformed into a PATR path equation (that partially describes a feature structure) and passed on to the PATR system. What is most disturbing about this strategy is the fact that for every distinct PATR path you have to know the corresponding DATR query. It is tempting to think one could simply check which paths are defined for a given node, but this doesn't work because of inheritance: the entire network is potentially relevant. So in effect all the PATR structures except the atomic values have to be defined twice: once in the DATR statements and once in the queries. This redundancy cannot be eliminated unless types for the feature structure are declared which are consulted in formulating the queries.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.2 USING DATR OUTPUT DIRECTLY </SectionTitle> <Paragraph position="0"> A completely different approach is to formulate a DATR theory which gives the lexical information in a PATR-usable format (i.e. a feature structure) as the result of the evaluation of a DATR query. Thus, the DATR description reflects the hierarchical structure of the lexical information and consequently meet.~ one of the main requirements for an adequate representation that cannot be met by a simple PATR formalism. The resulting feature structures include all the information necessary for PATR but neglect the inheritance structure, although the latter is involved in their construction (i.e. the evaluation of queries). There are various DATR-programming techniques that realize these ideas. Three examples will b:: presented here which cover the lexical information encoded in (1).</Paragraph> <Paragraph position="1"> The first technique, which is illustrated in (4) 2 , uses global inheritance (represented with double quotation marks) to store the node at which the query originates. This also allows other information in the global node to be.</Paragraph> <Paragraph position="2"> accessed.</Paragraph> <Paragraph position="4"> BAR: <> == MAJOR <majpaths> == ( bar ':' &quot;<bar>&quot; ). BAR0: o == BAR <bar> ~ zero.</Paragraph> <Paragraph position="5"> This DATR theory makes it possible to get the feature structure associated with the node NOUN, i.e. the evaluation of the DATR query NOUN:<>.</Paragraph> <Paragraph position="6"> To evaluate this DATR query the nodes NV, MAJOR, and SYNTAX are visited. In the node SYNTAX part of the corresponding feature specification is constructed and the evaluable path <synpaths> refers back to the original node NOUN. Then the query NOUN: <synpaths> is evaluated in the same way up to the node MAJOR, where the next part of the feature structure is built and the evaluable path <majpaths> refers again to the global node NOUN. At the end of the evaluation the feature structure \[syn:\[maj:\[nv: ln:yes,v:no1111 emerges. null Lexical entries defined with the DATR network above have the form FROG: <> == (&quot;NOUN .... BARO&quot;), which means intuitively that the lexeme frog is an nO. Given the network in (4), the value of the query FROG:<> will inherit the information of the global nodes NOUN and BARO. Thus, the global environment is changed in the course of the evaluation. As a declarative language, DATR is independent of the procedural evaluation strategies embodied in particular DATR-implementations. Nevertheless, DATR theories like (4) may themselves reflect different evaluation strategies (just as different search strategies may be implemented in pure PROLOG, independently of the particular PROLOG implementation). null The evaluation strategy in (4) can be described as top-down depth-first and is rather costly because of the cyclic returns to the global nodes. A more efficient strategy is illustrated in (5). This DATR description embodies a breadth-first search and uses variables (designated by the prefix $) instead of changing the global environment.</Paragraph> <Paragraph position="7"> Here an appropriate query would be SYNTAX: <noun barO>. At the origin of the query the outer layer of the feature structure is already constructed. The rest of the feature structure results from evaluating MAJOR:<$NV $BAR>, where SNV is instantiated with noun and $BAR with barO as in the original query. We then obtain the feature structure \[syn:\[maj:\[nv:\[n:yes,v:no\],bar:zero\]\]\] as the result of the evaluation. Unlike the network in (4), it is not possible to ask for just a part of this feature structure: Neither the information about the N/V-scheme nor the information about the bar level can be queried separately. An entry for the lexeme frog given the network (5) would have the form FROG:<> == SYNTAX::<noun barO>, which of .course also means that the lexeme frog is an nO. But this time the information is inherited from the - 140 node SYNTAX, where the value provides the frame for the resulting PATR feature structure. Apart from the differing DATR techniques employed, the resulting feature structures for a lexical entry also differ slightly. While the first is nearer to a set of PATR paths which has to be collapsed into a single feature structure, the second has exactly the form required by the PATR system we use.</Paragraph> <Paragraph position="8"> The third technique is illustrated in (6).</Paragraph> <Paragraph position="9"> An appropriate query for this DATR theory would be NOUN:<>, the value of which is \[syn:lmaj:\[nv:\[n:yes,v:no1,bar:zero111. The evaluation of this query is similar to the one in (5) in that the value of SYNTAX:<> constitutes the frame of the resulting PATR-usable feature structure. Unlike (5), no variables are used; instead, information from the global node is used via global path inheritance to specify the values. Notice that whereas with (4) the global node is changed, it remains unchanged during the evaluations with (6).</Paragraph> <Paragraph position="10"> The advantages of (6) are obvious.</Paragraph> <Paragraph position="11"> Since neither variables nor global nodes are used, fewer DATR facilities are involved. Nevertheless, the required PATR feature structures Can be defined. For example, the lexical entry for frog would be FROG:<>==NOUN, where the noun-specific information is inherited from NOUN.</Paragraph> <Paragraph position="12"> This third approach forms the base for our current lexicon. Some of the related issues are raised in the next section.</Paragraph> </Section> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> 5 THE DATR LEXICON </SectionTitle> <Paragraph position="0"> It has been shown above that DATR theories can serve as a lexicon for a PAT R system where the lexemes are represented as DATR nodes and the returned values of queries are the corresponding feature structures. In a lexicon which is formulated as in (6), aparl; from the lexical nodes (i.e. nodes like FROG which define lexemes) two other kinds of nodes can be distinguished: nodes like SYNTAX or NV, which correspond to PATR attributes, and nodes like NOUN or LEX/CAL, which represent a kind of type information (see Pollard/Sag, 1987). The lexemes inherit this information through reference to the type nodes, while the lexeme-specific information is as~ciated direct. ly with the lexical nodes.</Paragraph> <Paragraph position="1"> There are several differences between these three kinds of nodes. Whereas it is appro. priate to pose a query like FROG:<> or NOUN:<>, such queries make no sense for nodes like SYNTAX. In this respect lexemes and types are related.</Paragraph> <Paragraph position="2"> Another property distinguishes lexical nodes from type nodes. The latter are hierarchically structured, while the former are unstructured in the sense that they refer to types but not to other lexemes. The structuring of the type nodes reflects the above mentioned regularities as well as irregularities.</Paragraph> <Paragraph position="3"> The following DATR theory is a lexicon fragment for a possible classification of intransitive verbs in German. Regular verbs (e.g. schlafen ',sleep') take a nominative subject and inherit all type-specific information from the node INTRANS_VERB. One exception are verbs with expletive subject (e.g. regnen 'rain'), another those with nonnominative (accusative or dative) subject (e.g. dilrsten 'suffer from thirst' with accusative). These verbs refer to the types nodes INTRANS_VERB_EXPL and IN-TRANS_VERB_ACC, respectively. The latter types inherit from the node INTRANS_VERB but override some of its information.</Paragraph> </Section> <Section position="7" start_page="0" end_page="0" type="metho"> <SectionTitle> 6 CONCLUDING REMARKS </SectionTitle> <Paragraph position="0"> We have seen that it is possible to formulate the lexicon of a PATR system as a DATR theory. That is, given a lexical entry in DATR, a corresponding feature structure can be derived. A system postulating new entries for unknown words on the basis of contextual information during parsing (Kilbury, 1990) must be able to convert a given feature structure into a corresponding lexical entry in DATR so that the new lexeme is located and integrated in the lexical network. To solve this problem the concept of type nodes can be exploited.</Paragraph> <Paragraph position="1"> A final difficulty involves certain PATR-specific devices like disjunctions and reentrancies for which no obvious DATR facilities are available. At present we still have only ad hoc solutions to these problems.</Paragraph> </Section> <Section position="8" start_page="0" end_page="0" type="metho"> <SectionTitle> 2. The colons in single quotes, the commas, and the square </SectionTitle> <Paragraph position="0"> brackets are DATR atoms, not part of the language itself.In contrast, the parentheses of DATR enclose a list value.</Paragraph> </Section> class="xml-element"></Paper>