{ "paper_id": "1993", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:36:07.218727Z" }, "title": "Reducing Complexity in A Systemic Parser", "authors": [ { "first": "Michael", "middle": [], "last": "O'donnell", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Sydney", "location": {} }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Parsing with a large systemic grammar brings one face-to-face with the problem of unifica tion with disjunctive descriptions. This paper outlines some techniques which we employed in a systemic parser to reduce the average-case complexity of such unification.", "pdf_parse": { "paper_id": "1993", "_pdf_hash": "", "abstract": [ { "text": "Parsing with a large systemic grammar brings one face-to-face with the problem of unifica tion with disjunctive descriptions. This paper outlines some techniques which we employed in a systemic parser to reduce the average-case complexity of such unification.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": ", PROTEUS (Davey, 1978) , SLANG (Patten, 1986) , GENESYS (Fawcett -Tucker, 1990 ) and HORACE (Cross, 1991) (e.g., Halliday, 1985; Halliday -Hasan, 1976; Martin, 1992) .", "cite_spans": [ { "start": 2, "end": 23, "text": "PROTEUS (Davey, 1978)", "ref_id": null }, { "start": 32, "end": 46, "text": "(Patten, 1986)", "ref_id": "BIBREF8" }, { "start": 49, "end": 79, "text": "GENESYS (Fawcett -Tucker, 1990", "ref_id": null }, { "start": 93, "end": 106, "text": "(Cross, 1991)", "ref_id": "BIBREF2" }, { "start": 107, "end": 129, "text": "(e.g., Halliday, 1985;", "ref_id": null }, { "start": 130, "end": 152, "text": "Halliday -Hasan, 1976;", "ref_id": null }, { "start": 153, "end": 166, "text": "Martin, 1992)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Systemic grammar has been used in several text generation systems, such as PENMAN (Mann", "sec_num": null }, { "text": "Parsing with systemic grammar has not, how ever, been as successful. To date, there have been six parsing systems using systemic gram mar: Winograd (1972 ), McCord (1977 ), Cum mings -Regina (1985 , Kasper (1988a , 1988b , 1989 ), O'Donoghue (1991a , 1991b and Bateman et al. \u2022 (1992) . However, each of these systems has been limited in some way, either resorting to a simplified formalism (Winograd -Cummings -Mc Cord) (Matthiessen -Mann, 1985 , Matthiessen -Bateman, 1992 . As this resource is ava ilable, it is desirable to use it for parsing. However, complexity problems have so far made this impossible, except by pre-parsing with another formalism.", "cite_spans": [ { "start": 139, "end": 153, "text": "Winograd (1972", "ref_id": "BIBREF9" }, { "start": 154, "end": 169, "text": "), McCord (1977", "ref_id": "BIBREF3" }, { "start": 170, "end": 196, "text": "), Cum mings -Regina (1985", "ref_id": null }, { "start": 199, "end": 212, "text": "Kasper (1988a", "ref_id": null }, { "start": 213, "end": 220, "text": ", 1988b", "ref_id": null }, { "start": 221, "end": 227, "text": ", 1989", "ref_id": null }, { "start": 228, "end": 248, "text": "), O'Donoghue (1991a", "ref_id": "BIBREF5" }, { "start": 249, "end": 256, "text": ", 1991b", "ref_id": "BIBREF6" }, { "start": 261, "end": 284, "text": "Bateman et al. \u2022 (1992)", "ref_id": "BIBREF0" }, { "start": 391, "end": 420, "text": "(Winograd -Cummings -Mc Cord)", "ref_id": null }, { "start": 421, "end": 445, "text": "(Matthiessen -Mann, 1985", "ref_id": null }, { "start": 446, "end": 474, "text": ", Matthiessen -Bateman, 1992", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": ". Sys temics has proved useful in generation for sev eral reasons: the orientation of Systemics towards representing language as a system of choices, the strongly semantic nature of the grammar, and the extensive body of systemic work linking dis course patterns and grammatical realisation", "sec_num": null }, { "text": "more efficient at a general level. Before becom ing more specific, the Systemic formalism is intro duced (section 3). Section 4 explores one method of avoiding complexity -reducing the size of the disjunctive description by working with sub descriptions rather than the whole description. Section 5 presents three ways of making expan sion, when necessary, more efficient. We conclude the paper with a brief summarisation of our work.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "In the last fe w years, we have developed a parser for Systemic grammar, particularly for use with the Nigel grammar. The parser han dles the full Systemic formalism, and does not depend on another formalism for segmentation. The parser uses a bottom-up, breadth-first algo rithm. A chart is used to handle some of the non-determinism. This paper focuses on some methods we have used in the parser to reduce the complexity prob lems associated with using the Nigel grammar. In particular, we focus on the means used to make disjunctive unification more efficient. Section 2 discusses the problem of disjunc tive expansion, and some means of making it", "sec_num": null }, { "text": "Parsing with a systemic grammar involves much unification of disjunctive descriptions. The usual way to unify such is as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Unification with Disjunc tive Descriptions", "sec_num": "2" }, { "text": "1. Expand out the disjunctive descriptions to Disjunctive Normal Form (DNF) -a form with all disjunction at the top level of the description -a disjunction of non disjunctive forms.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Unification with Disjunc tive Descriptions", "sec_num": "2" }, { "text": "DNF expansion of a description is however an expensive task -the process takes exponen tial time in the worst case (Kasper -Rounds, 1986) . Space is also a problem -DNF expansion is a transformation whereby a disjunctive descrip tion is replaced with a set of descriptions each of which contains no disjunction. For a description containing a high level of disjunction, the size of the DNF form can be excessive.", "cite_spans": [ { "start": 115, "end": 137, "text": "(Kasper -Rounds, 1986)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Unify each term of the first DNF form with each term of the other.", "sec_num": "2." }, { "text": "Space has not however been a problem in our processing, but time has. Systemic parsing is very slow. We thus focus on means for speeding up, or avoiding, the unification process.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Unify each term of the first DNF form with each term of the other.", "sec_num": "2." }, { "text": "There have been proposals for unification with out DNF expansion. Karttunen, for instance, has proposed an algorithm which \"uses constraints on disjuncts which must be checked whenever the disjunct is modified\" (Kasper, 1987, p81). How ever, as noted by Kasper ( 1987, p61), Karttunen's unification algorithm works only for a limited type of disjunctive description, and not for gen eral disjunction as is needed in the present work.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": ".1 A voiding Expansion", "sec_num": "2" }, { "text": "Kasper has proposed a method of re representing disjunctive descriptions which in some cases avoids the need for expansion. His approach separates a disjunctive description into two parts -a definite component (which con tains no disjunction), and an indefinite compo nent ( containing the disjunctive information of the description). A unification process can first check whether the definite components of two de scriptions unify, and only proceeds to unify the indefinite components if the definite components unify successfully. The unification of the indefi nites is avoided if the unification of the definites fails .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "O'DONNELL", "sec_num": null }, { "text": "The Kasper-Rounds form also allows us to de lay expansion until a later time. When two de scriptions are unified, only the definite compo nents need to be checked for compatibility. The result of a Kasper-Rounds unification contains the indefinite descriptions from both descriptions without expansion. At some point in the pro cessing it may be necessary to resolve the indefi niteness, and the disj unctive components are then expanded. However, in many cases, the definite component of the description may become incon sistent before this is necessary, expansion is thus avoided.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Delaying disjunctive expansion until necessary", "sec_num": "2.2" }, { "text": "If DNF-expansion is required, then it should be performed as efficiently as possible. We here dis cuss some methods to achieve this goal:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "When expansion is necessary, expand efficiently", "sec_num": "2.3" }, { "text": "1. Reducing the disjunctiveness of the description: By reducing the extent of the description, we reduce the amount of dis junction to be expanded, and thus speed up the expansion process. We use two methods to reduce the size of descriptions:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "When expansion is necessary, expand efficiently", "sec_num": "2.3" }, { "text": "(a) Extracting descriptions for special purpose: we segment the grammar de scription into sub-descriptions for par ticular purposes. We found that dif ferent parsing processes drew upon ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "When expansion is necessary, expand efficiently", "sec_num": "2.3" }, { "text": "The parser makes extensive use of caching -when any expansion is calculated which is likely to be used again, the result is stored away for later re use. Precompilation has also been a useful tech nique to improve parsing efficiency. Precompila tion is basically a pre-caching of all the values which might be used in the parsing process. By performing most of the DNF expansion of the Systemic grammar (e.g., Halliday, 1985 , Hudson, 1971 , Matthiessen -Mann, 1985 uses an in heritance network to organise grammatical types (or 'feature' in Systemics 1 ), and their structural consequences. A Systemic inheritance network is called a system network.", "cite_spans": [ { "start": 403, "end": 424, "text": "(e.g., Halliday, 1985", "ref_id": null }, { "start": 425, "end": 439, "text": ", Hudson, 1971", "ref_id": null }, { "start": 440, "end": 465, "text": ", Matthiessen -Mann, 1985", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Caching and precompilation: avoiding repeating the same ex pansion.", "sec_num": "2.4" }, { "text": "A system network is used to organise the co occurrence potential of grammatical types, show ing which types are mutually compatible, and which are incompatible. It consists of a set of sys tems, which are sets of mutually exclusive types. There is also a covering relation between the types of a system, meaning that if the entry con dition of the system is satisfied, then one of the types in the cover must be selected. Figure 1 ", "cite_spans": [], "ref_spans": [ { "start": 422, "end": 430, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Caching and precompilation: avoiding repeating the same ex pansion.", "sec_num": "2.4" }, { "text": "Each type inherits the properties of types to its left in the network. Note that the system net work may be logically complex, since entry con ditions {the logical condition on a system) may consist of conjunctions and disjunctions of types. [singular] or [plural] , and [human] (e.g., \"I\" , \"you\" , \"he\" ) or [nonhuman] (e.g., \"it\" , \"that\" ). Only [human] pronouns can fill the Ac tor role of a clause.", "cite_spans": [ { "start": 242, "end": 252, "text": "[singular]", "ref_id": null }, { "start": 256, "end": 264, "text": "[plural]", "ref_id": null }, { "start": 271, "end": 278, "text": "[human]", "ref_id": null }, { "start": 310, "end": 320, "text": "[nonhuman]", "ref_id": null }, { "start": 350, "end": 357, "text": "[human]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "shows a system network for a sim ple grammar of English. It includes 11 systems, representing various grammatical distinctions, for instance, between clause and word, between tran sitive and intransitive clauses, or between nomi native and accusative pronouns.", "sec_num": null }, { "text": "The realisation operators used in the formal ism are as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "shows a system network for a sim ple grammar of English. It includes 11 systems, representing various grammatical distinctions, for instance, between clause and word, between tran sitive and intransitive clauses, or between nomi native and accusative pronouns.", "sec_num": null }, { "text": "Insert e.g., Finite = [ J: indicates that the function Finite must be present in the structure.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "shows a system network for a sim ple grammar of English. It includes 11 systems, representing various grammatical distinctions, for instance, between clause and word, between tran sitive and intransitive clauses, or between nomi native and accusative pronouns.", "sec_num": null }, { "text": "Conflate e.g., Modal/Finite: indicates that the two functions Modal and Finite are filled by the same grammatical unit.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "shows a system network for a sim ple grammar of English. It includes 11 systems, representing various grammatical distinctions, for instance, between clause and word, between tran sitive and intransitive clauses, or between nomi native and accusative pronouns.", "sec_num": null }, { "text": "Order e.g., Subject \" Finite: indicates the se quencing of functions in the surface structure. In this example, the Subject is sequenced directly before the Finite. Any number of elements can be sequenced in a single rule.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "shows a system network for a sim ple grammar of English. It includes 11 systems, representing various grammatical distinctions, for instance, between clause and word, between tran sitive and intransitive clauses, or between nomi native and accusative pronouns.", "sec_num": null }, { "text": "Partition e.g., Thing ... Event ... End: An other sequence operator, specifies that the appear in this order, but not necessarily immediately ad jacent (linear precedence).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "shows a system network for a sim ple grammar of English. It includes 11 systems, representing various grammatical distinctions, for instance, between clause and word, between tran sitive and intransitive clauses, or between nomi native and accusative pronouns.", "sec_num": null }, { "text": "Preselect e.g., Subject: nominal-group : indi cates that the Subject element must be filled by a unit of type nominal-group.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "shows a system network for a sim ple grammar of English. It includes 11 systems, representing various grammatical distinctions, for instance, between clause and word, between tran sitive and intransitive clauses, or between nomi native and accusative pronouns.", "sec_num": null }, { "text": "Lexify e.g., Deict = \"the \": used to assign lex ical items directly to elements of structure. Note that lexify overrides any preselect which may ap ply to the same element of structure.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "shows a system network for a sim ple grammar of English. It includes 11 systems, representing various grammatical distinctions, for instance, between clause and word, between tran sitive and intransitive clauses, or between nomi native and accusative pronouns.", "sec_num": null }, { "text": "For the purposes of the expansion of this gram mar, we re-express it in a logical formalism. Each of these uses makes only partial use of the grammar description. Thus, rather than ex panding out the entire grammar, we can simplify the process by extracting out sub-grammars, one for each of these applications. Since the size of each sub-grammar is smalle_ r, the complexity problem is reduced. This section looks at these three sub-descriptions in more detail.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Grammar of the", "sec_num": null }, { "text": "It has proved useful to separate the type logic component of the grammar from the role logic. The two logic components have different patterns of use -type logic is used to test whether two par tial type-paths can unify. We never try to unify a partial type description with the type grammar as a whole. The type-logic component of the gram mar thus does not need to be DNF-expanded. The role logic, on the other hand, does need to be expanded. We expand the role-logic compo nent to produce a set of non-disjunctive structure rules which can be applied during parsing (some times termed 'chunking'). ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Separating Type Logic fr om Role Information", "sec_num": "4.1" }, { "text": "Thse two components of the description have different properties: type logic is acyclic, while role logic is potentially cyclic. Type logic is con strained such that types are always in disjoint coverings (which allows efficient negation), while role logic doesn't have this constraint.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "O'DONNELL", "sec_num": null }, { "text": "Because of these differences in properties and uses, it has proved efficient to treat these two logics separately. Logical Form I of the sys temic grammar provided in Figure 3 ", "cite_spans": [], "ref_spans": [ { "start": 167, "end": 175, "text": "Figure 3", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Unification of Type Descriptions", "sec_num": "4.1.1" }, { "text": "The parser uses the type-logic component of this grammar without fully expanding it. Partial ex pansion, however, is performed, whereby the type path (the logical-entailment of a system, i.e., the logical expression of types leading back to the root of the network) 2 is pre-compiled for each system. 3 The negation of each type in the system is also pre-compiled, which speeds up unification involving negation of types.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "can be re represented in the equivalent Logical form II \u2022 shown in Figure 4, separating out the type and", "sec_num": null }, { "text": "Type-paths are represented in the form pro posed by Kasper (1987) , and his unification al gorithm is used when two type-paths are uni fied. The main use of the type-logic component is checking the compatibility of two types or type paths.", "cite_spans": [ { "start": 52, "end": 65, "text": "Kasper (1987)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "can be re represented in the equivalent Logical form II \u2022 shown in Figure 4, separating out the type and", "sec_num": null }, { "text": "Type logic has thus been simplified using three strategies: ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "can be re represented in the equivalent Logical form II \u2022 shown in Figure 4, separating out the type and", "sec_num": null }, { "text": "Another use made of the grammatical description in parsing is to assign a set of structural roles to a unit. \u2022 Conflation rules: used to discover which functions a unit can serve simultaneously, and thus, which of the preselection and lex ify rules can combine.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Function Assignment", "sec_num": "4.2" }, { "text": "\u2022 The Type Logic: to show which of these preselection, lexify and conflation rules are systemically compatible.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Function Assignment", "sec_num": "4.2" }, { "text": "Since we have already set up the type-logic for path unification, we can draw upon that re source as needed. We do not need to include the type-logic in the sub-description for the function assignment process.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Function Assignment", "sec_num": "4.2" }, { "text": "For the function-assignment process, we do not need all of the role logic description. We can se lect out only those rules involving preselection, lexify, and conflation. See Logical Form III in Figure 5 . ", "cite_spans": [], "ref_spans": [ { "start": 195, "end": 203, "text": "Figure 5", "ref_id": "FIGREF3" } ], "eq_spans": [], "section": "Extracting the relevant description", "sec_num": "4.2.1" }, { "text": "We next put this description into a form more suitable for DNF-expansion. Note that implica tion can be re-expressed using disjunction, con junction and negation: The order of worst-case complexity of the ex pansion to DNF is easily calculated -it is simply two to the power of the number of disjunctions, which is equal to the number of types which have realisation rules of type conflation, insertion, or preselection.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Implications Out", "sec_num": "4.2.2" }, { "text": "By opting to expand only subsets of the whole grammar, we have reduced the complexity of the description, since the size of n for this sub description is smaller than for the whole descrip tion. However, for a real-sized grammar such as NIGEL, the size of n is still large.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Implications Out", "sec_num": "4.2.2" }, { "text": "From the DNF-form of this description, we can extract out partial-descriptions for each function bundle. We now re-express this logical form in terms of the type constraints on each function bundle, including both the constraint on the type of unit the function-bundle can be part of (the 'parent-constraint'), and the constraint on the filler of the function-bundle ( the 'filler constraint '). We show this as a set of triplets, of the form:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Re-expression in terms of Function Bundles", "sec_num": "4.2.4" }, { "text": "( ) This representation can now be used to assign function-bundles A unit can take on a function bundle if it can unify with the filler-constraint on the function-bundle.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Re-expression in terms of Function Bundles", "sec_num": "4.2.4" }, { "text": "For the instance we started with, \"he\" , with types: ( :and pronoun nominative human singular) , only one triplet would unify. We could thus posit structure for our unit: Note that we have also gained information about the types of the parent-unit of which the unit is a constituent.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Re-expression in terms of Function Bundles", "sec_num": "4.2.4" }, { "text": "Note that there is another simplification we can make to the triplet list. We can take all triplets with identical function bundle and child-type specification, and join them. The parent-types slot is replaced with the disjunction of the two parent-type slots. Thus, elements 2 and 4 above become a single item. This process reduces the number of rules to apply: 2,4. ( (:and clause transitive active)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Reducing the number of Rules", "sec_num": "4.2.5" }, { "text": "Obj ect/Actee accusat ive) ))", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Reducing the number of Rules", "sec_num": "4.2.5" }, { "text": "Another process we use in parsing involves the prediction of what function-bundles can come next in a partially completed structure. With a systemic grammar, this process requires:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting What Comes Next", "sec_num": "4.3" }, { "text": "\u2022 Ordering and Partition rules: to see which function can come next.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting What Comes Next", "sec_num": "4.3" }, { "text": "\u2022 Conflation rules: to see which functions can conflate with the function predicted to come next.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting What Comes Next", "sec_num": "4.3" }, { "text": "\u2022 The type logic: to show which of these or dering, partition and conflation rules are systemically compatible.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting What Comes Next", "sec_num": "4.3" }, { "text": "The processing of this sub-description, and any others, is exactly the same as for function assignment.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting What Comes Next", "sec_num": "4.3" }, { "text": "1. Extract from the role logic description the relevant realisation rules;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting What Comes Next", "sec_num": "4.3" }, { "text": "2. Replace implications with disjunction and negation;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting What Comes Next", "sec_num": "4.3" }, { "text": "3. Expand out the grammar;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting What Comes Next", "sec_num": "4.3" }, { "text": "4. Index the rules in a form useful for the pro cessing.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting What Comes Next", "sec_num": "4.3" }, { "text": "Another means of reducing the overall complexity of the descriptions involves eliminating from the grammar parts which are unlikely to be utilised in the target texts. In systemic terms, we apply register restrictions to the grammar. For example, in a domain of computer manu als, the description of interrogative structures is not likely to be drawn upon. 4 By eliminating this sub-description, we reduce the degree of disjunc tion in the whole description, and thus speed up the parsing of the forms which do appear in the text.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Register Restriction", "sec_num": "4.4" }, { "text": "The method of deriving the register restrictions was as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Register Restriction", "sec_num": "4.4" }, { "text": "1. We parsed by hand 5 a chapter of the com puter manuals we were attempting to parse, building up a register-profile of our target texts.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Register Restriction", "sec_num": "4.4" }, { "text": "2. An automatic procedure then extracted out all the grammatical types which occur in these sentences.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Register Restriction", "sec_num": "4.4" }, { "text": "O'D ONNELL", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The process used this information to dis-", "sec_num": "3." }, { "text": "Section 4 has proposed techniques which reduce the size of the description which needs to be ex panded. However, for large-sized descriptions, the expansion is still complex. This section briefly explores two methods which increase the effi ciency of the expansion process. If we can't avoid full expansion, then at least we can make the ex pansion process more efficient.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Improving the Efficiency of Expansion", "sec_num": "5" }, { "text": "This section assumes a disjunctive description of the following form:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "\"Structure Sharing\" in Expan sion", "sec_num": "5.1" }, { "text": "( 4. The process then eliminated these types and their realisations from the description.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "\"Structure Sharing\" in Expan sion", "sec_num": "5.1" }, { "text": "We were thus left with a restricted grammar which was capable of parsing the sentences in the sample, and also parsing many which were not in the sample (under the assumption that the gram matical forms in the sample were representative of the forms found in the manual as a whole). We reduced the size of the grammar by approxi mately 60% using this method.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "\"Structure Sharing\" in Expan sion", "sec_num": "5.1" }, { "text": "By extracting out sub-descriptions from the full description, we reduce the complexity of the description-to-be-expanded.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Summary", "sec_num": "4.5" }, { "text": "The brute force method for expanding this form involves:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Summary", "sec_num": "4.5" }, { "text": "1. Find all combinations of terms, taking one term from each disjunction.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Summary", "sec_num": "4.5" }, { "text": "2. Test compatibility of each combination, eliminating combinations which are inter nally inconsistent.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Summary", "sec_num": "4.5" }, { "text": "Step 1 of this process produces the following DNF form:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Summary", "sec_num": "4.5" }, { "text": "(:xor (:and ACE ) (:and ACF ) (:and ADE ) (:and AD F) (:and BCE ) (:and BC F) (:and B DE ) (:and B D F) )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Summary", "sec_num": "4.5" }, { "text": "\u2022 the failure of a combination of terms early in the unification process eliminates a large number of expansions by the end of the pro cess.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Summary", "sec_num": "4.5" }, { "text": "A third approach aims at maximising the degree of 'sharing' unifications in the expansion. The disjunctions in the description are split into pairs, and unified. The results of these unifications are then unified in the same pair-wise manner. This expansion for a conjunction of four disjunctions is shown in Figure 9 . The advantage of this approach is that we are maximising the amount of structure-sharing in the unification.", "cite_spans": [], "ref_spans": [ { "start": 309, "end": 317, "text": "Figure 9", "ref_id": null } ], "eq_spans": [], "section": "Binary Organisation of Expansion", "sec_num": "5.1.3" }, { "text": "We compared the number of unifications which take place using each of these methods for vari ous numbers of disjunctions ( all disjunctions hav ing two disjuncts). One can see from Table 1 that the worst-case score for the full expansion method is far worse than the other methods. It is not a practical method.", "cite_spans": [], "ref_spans": [ { "start": 181, "end": 188, "text": "Table 1", "ref_id": null } ], "eq_spans": [], "section": "Comparison of Expansion Approaches", "sec_num": "5.1.4" }, { "text": "Comparing the worst-case for the 'tree' and the 'binary' expansion method, we see that the bi nary method clearly comes out better, by around 50%.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Comparison of Expansion Approaches", "sec_num": "5.1.4" }, { "text": "We also did a simulation to check an aver age case score, since the worst-case score doesn't take into account that many later unifications are avoided when early unification proves inconsis tent. We found that while the binary method still seems superior, in some instances the tree method requires fewer unifications. More work is needed here.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Comparison of Expansion Approaches", "sec_num": "5.1.4" }, { "text": "Since these types won't unify with each other, we can also know that types which inherit from one of these basic types will not unify with the the sub-types of another basic type. We thus do not need to try to unify descriptions which differ in their basic type. If we split any disjunctive de scription into sub-components for each basic type, we know a priori that there is no unification be tween these sub-components.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "These categories are all types in the system net work, just like any other types.", "sec_num": null }, { "text": "Before trying any of the expansion techniques outlined in this paper, the whole grammar is segmented into sub-descriptions, one for each of these basic types. The complexity of the expan sion of each of these sub-grammars is less than for the grammar as a whole.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "These categories are all types in the system net work, just like any other types.", "sec_num": null }, { "text": "Other principles can be used to locate sets of rules which will not unify. These can be applied also.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "These categories are all types in the system net work, just like any other types.", "sec_num": null }, { "text": "While the techniques outlined here have been ap plied in ways particular to a systemic grammar, and for a particular implementation, there are principles behind the re-representations which are general to all implementations:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "6" }, { "text": "Kasper's unification algorithm.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Avoid DNF-expansion where possible, as in", "sec_num": "1." }, { "text": "to be inconsistent in the definite compo nent.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Delay expansion to a later time -informa tion gained later may show the description", "sec_num": "2." }, { "text": "(a) Try to extract out sub-descriptions which can be used , rather than ex panding the entire grammar.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "When expansion is necessary,", "sec_num": "3." }, { "text": "(b) Expand out first disjunctions which are most likely to conflict, since this will reduce the total number of terms which will n\ufffded to be multiplied.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "When expansion is necessary,", "sec_num": "3." }, { "text": "As a result of the application of these techniques (and others not here mentioned), we have been able to implement a parsing system which parses using a large systemic grammar.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "( c) A void expanding terms that can be known to be incompatible.", "sec_num": null }, { "text": "1. We start with the Nigel grammar, as used in the Penman Generation System, slightly modified for parsing purposes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "( c) A void expanding terms that can be known to be incompatible.", "sec_num": null }, { "text": "2. This grammar is then reduced by applying register-restrictions, leaving a less complex grammar, but a grammar which still han dles the bulk of the phenomena in the target texts.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "( c) A void expanding terms that can be known to be incompatible.", "sec_num": null }, { "text": "3. Sub-descriptions of the grammar tailored for particular processes are then extracted, and expanded out as a precompile step, pro ducing a set of 'chunks' which can be used in parsing. This expansion takes approxi mately 2 minutes using Sun Common Lisp on a Sun Spare II. Future work will attempt to reduce this parsing time. Three directions are being followed:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "( c) A void expanding terms that can be known to be incompatible.", "sec_num": null }, { "text": "\u2022 Streamlining the parsing process to further reduce the parsing time.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "( c) A void expanding terms that can be known to be incompatible.", "sec_num": null }, { "text": "\u2022 Moving more processing to the pre compilation stage.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "( c) A void expanding terms that can be known to be incompatible.", "sec_num": null }, { "text": "\u2022 Reducing the complexity of the description without reducing its coverage.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "( c) A void expanding terms that can be known to be incompatible.", "sec_num": null }, { "text": "\u2022 Incorporating heuristics to resolve ambigu ities without full expansion.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "( c) A void expanding terms that can be known to be incompatible.", "sec_num": null }, { "text": "Note that the term 'feature' is used distinctly from its use in most unification paradigms. In Systemics, a feature is what Functional Unification Grammar would call a value, e.g., active, transitive and noun are features.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Note that some of the forms we restrict through register restriction may actually appear in any one text, although quite rarely. We are trading off between speed for the majority of sentences, and ability to parse all sentences in a text.5 The hand-parsing is really computer-assisted, -a tool was developed to traverse the system network for each sen tence ( and each constituent of the sentence) asking the human which feature was appropriate for the target string. This process guaranteed that the human-analysis conformed to the computer grammar.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Systemics prefers the term 'nominal-group' over the equivalent term 'noun-phrase'.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Note that when the parser is given a less complex systemic grammar, the parsing time is under two seconds for this sentence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "The parser discussed in this paper was partially developed in the Electronic Discourse Analyser project, funded by Fujitsu (Japan). The devel opment was aided by discussions with the mem bers of that project: Christian Matthiessen, John Bateman, Zeng Licheng, Guenter Plum, Arlene Harvey and Chris Nesbitt. Thanks to Cecile Paris for profuse comment ing on this paper, and teaching me Latex, and to Vibhu Mittal, who solved the trickier Latex bugs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgements", "sec_num": null }, { "text": "The problem with this approach is with the incompatibility checkingthe same checks will be repeated over and over again. For instance, the incompatibility check between A and C is re peated twice: (:and A C E) and (:and A C F). This repetition occurs for every pair of terms in the conjuncts. The-problem gets worse exponen tially as we add more disjuncts.To avoid this redundancy, we need something like a chart in parsing, a method to record the re sults of each unification and thus avoid repeating any unification.Unfortunately, DNF expansion is not quite like parsing. We can test the consistency between any two pairs of terms ( for instance A and C in the above), but we also need to know about the consistency of terms in combination e.g., the pairs: A&C, A&E and C&E may be consistent, but the combination A&C&E may not be.The rest of this section describes two tech niques which allow some redundancy reduction, sometimes known as structure-sharing. 3. Expand out the results of (2) against each other.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "annex", "sec_num": null }, { "text": "Sometimes, it is possible to tell without full uni fication that a set of rules will not unify with an other set. For instance, assume a larger gram mar than the one we have been using, a grammar which includes clauses, nominal-groups 6 , prepo sitional phrases, adverbial phrases and words. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Avoiding Expansion of Incompatible Terms", "sec_num": "5.3" } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "The non directional repre sentation of Systemic Functional Grammars and Semantics as Typed Feature Structures", "authors": [ { "first": "John -Martin Emele -Stefan", "middle": [], "last": "Bateman", "suffix": "" }, { "first": "", "middle": [], "last": "Momma", "suffix": "" } ], "year": 1992, "venue": "Proceedings of COLING-92", "volume": "III", "issue": "", "pages": "916--920", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bateman, John -Martin Emele -Stefan Momma ( 1992) \"The non directional repre sentation of Systemic Functional Grammars and Semantics as Typed Feature Structures\" in Proceedings of COLING-92 , Volume III, Nantes, France, 916-920.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Ch oice in Text: A Systemic-Functional Approach to Computer Modelling of Va riant Te xt Production", "authors": [ { "first": "Marilyn", "middle": [], "last": "Cross", "suffix": "" } ], "year": 1991, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Cross, Marilyn (1991) Ch oice in Text: A Systemic-Functional Approach to Computer Modelling of Va riant Te xt Production, Ph.D. thesis submitted June 1991, Macquarie Uni versity.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Procedural Systemic Grammars in", "authors": [ { "first": "Michael", "middle": [], "last": "Mccord", "suffix": "" } ], "year": 1977, "venue": "Int. J. Man-Machine Studies", "volume": "9", "issue": "", "pages": "255--286", "other_ids": {}, "num": null, "urls": [], "raw_text": "McCord, Michael (1977) Procedural Systemic Grammars in Int. J. Man-Machine Studies, 9, 255-286, London: Academic Press.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Implementing Systemic Classification by Unification", "authors": [ { "first": "Chris", "middle": [], "last": "Mellish", "suffix": "" } ], "year": 1988, "venue": "Computational Linguistics", "volume": "14", "issue": "1", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mellish, Chris (1988) \"Implementing Systemic Classification by Unification\" , Computational Linguistics, Vol. 14, Number 1, Winter 1988.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "The Vertical Strip Parser: A lazy approach to parsing", "authors": [ { "first": "Tim", "middle": [ "F" ], "last": "O'donoghue", "suffix": "" } ], "year": 1991, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "O'Donoghue, Tim F. (1991a) \"The Vertical Strip Parser: A lazy approach to parsing\" Research Report 91.15, School of Computer Studies, University of Leeds, Leeds, UK.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "A Semantic Inter preter for Systemic Grammars", "authors": [ { "first": "Tim", "middle": [ "F" ], "last": "O'donoghue", "suffix": "" } ], "year": 1991, "venue": "Proceedings of the AGL Workshop on Reversible Gram mars", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "O'Donoghue, Tim F. (1991b) \"A Semantic Inter preter for Systemic Grammars\" in Proceedings of the AGL Workshop on Reversible Gram mars, University of California at Berkeley, 217", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "A formal model of Systemic Grammar", "authors": [ { "first": "Terry", "middle": [], "last": "Patten", "suffix": "" }, { "first": "Graeme", "middle": [], "last": "Ritchie", "suffix": "" } ], "year": 1986, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Patten, Terry and Graeme Ritchie (1986) \"A formal model of Systemic Grammar\" , paper presented at 3rd International Workshop on Language Generation, Nijmegen, August 19- 23, 1986.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Interpreting Systemic Grammar as A Computational Representa tion: A Pro blem Solving Approach to Text Generation", "authors": [ { "first": "Terry", "middle": [], "last": "Patten", "suffix": "" } ], "year": 1986, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Patten, Terry (1986) Interpreting Systemic Grammar as A Computational Representa tion: A Pro blem Solving Approach to Text Generation, Ph. D. dissertation, University of Edinburgh.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Un derstanding Natural Language", "authors": [ { "first": "Terry", "middle": [], "last": "Winograd", "suffix": "" } ], "year": 1972, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Winograd, Terry (1972) Un derstanding Natural Language. New York: Academic Press.", "links": null } }, "ref_entries": { "FIGREF0": { "num": null, "type_str": "figure", "uris": null, "text": "Types of the system network are associated with structural realisations -the structural conse quence of the type.Figure 2shows the realisa tions of the types inFigure 1." }, "FIGREF1": { "num": null, "type_str": "figure", "uris": null, "text": "Fig ure 3 shows Logical Form I of this grammar, in cluding the structural constraints embedded in the form. Note that :xor indicates exclusive dis junction. Logical Form I of the Grammar" }, "FIGREF2": { "num": null, "type_str": "figure", "uris": null, "text": "single-subj ect Subj ect : singular) (:implies plural-subj ect Subj ect : plural) )) Figure 4: Logical Form II of the Grammar role logic." }, "FIGREF3": { "num": null, "type_str": "figure", "uris": null, "text": "since entry conditions of systems can be logically complex, the path itself can contain disjunctions and conjunctions. \u2022 3 Paths are stored with systems rather than types, since the path of all types in a system are identical. Logical Form III: The Function Assignment Sub-Description" }, "FIGREF4": { "num": null, "type_str": "figure", "uris": null, "text": "Using this rule, we can re-express the logical form III as Logical Fo rm IV, as shown inFigure 6. to expand Logical Form IV into DNF (see section 5.1). A small part of the result appears in Logical Form V of the grammar, shown inFigure 7. Logical Form Fo rm V: The Function Assignment Sub-Description in DNF" }, "FIGREF8": { "num": null, "type_str": "figure", "uris": null, "text": ":and (:xor A B) (:xor CD ) (:xor E F) ) Logical form V introduced above was of this form. Much of the pre-processing in the parser in volves the DNF-expansion of disjunctions in this form. cover the types not occurring in the sample." }, "FIGREF9": { "num": null, "type_str": "figure", "uris": null, "text": "Figure 9: Binary expansion method" } } } }