| { | |
| "paper_id": "W87-0101", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T06:44:38.850851Z" | |
| }, | |
| "title": "", | |
| "authors": [], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "W87-0101", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "Imagine a situation where we want to parse texts and get as output information about the traditional grammatical categories of sent en ces,", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Gunnel", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "i. e. subject, direct and indirect object, finite verb, ad verbials etc. That is an explicit goal of most parsers, and a necessary prerequisite for many applications of computational linguistics. How much information, syntactic and/or other, do we need in order to reach that goal?", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Gunnel", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In this paper I will look at the kind of syntactic information that is mostly given in syntactic parsers (sect. 1) and relate it to different parts of the parsing process (sect.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Gunnel", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "2). In section 3, I present an algorithm for assigning syntactic function to constitu ents without any use of lexical look-up.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Gunnel", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "I discuss its results (sect. 4) and the type of errors that arise (sect. 5), which leads me to the conclusion that parsing can in most cases very well be done without a lexicon containing syntactic information. In section 6 I point to the need for different parsers for different purposes.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Gunnel", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Many parsing systems use information that is centered around the verb.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "W ' a a t kind of information can be given?", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "In the dictionary of the parser, each verb has information about whether it is transitive or intransitive and how-obligatory or optional its complements are. Often some kind of selectional restrictions are also given. They describe properties of the referents of the noun phrases that enter different syntactical roles in relation to the verb.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "W ' a a t kind of information can be given?", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "Such properties are concrete/abstract, animate/inanimate, human/non-human etc. The apparatus demands marking of the verbs with their functional frames, and marking of all words that can enter those frames accordingly, and procedures for checking the different kinds of marking against each other.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "W ' a a t kind of information can be given?", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "to get correct syntactic assignments in la-b and 2a-b one will need verbal frames like 3 and 4.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(la) Mary gave the book to John.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(b) Mary gave John the book. (2a) She drank the wine from a chrystal glass.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(b) The horse drank (the water) (in the bucket).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(3) g i v e(x,y,2 ) where x=subj, y=dir o b j , z=indir obj patterns: x give y PREP z where PREP=to X give z y (4) drink (x,y,z) where x=subj, y=dir obj, z=loc advl pattern; >; drink (y) (PREP z) where PREP=from, in selectioral restrictions: x=animate, y-liquid, z=some kind container of What good is Syntactic Information in the Lexicon of a Syntactic Parser? Gunnel K\u00e4 llgren Proceedings of NODALIDA 1987, pages 5-16", | |
| "cite_spans": [ | |
| { | |
| "start": 121, | |
| "end": 128, | |
| "text": "(x,y,z)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 187, | |
| "end": 195, | |
| "text": "(PREP z)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The classificatory system can then be developed and refined to cover more and more of the ways a word is used, and more and more of the v o cabulary of a language (see K\u00e4llgren 1986) .", | |
| "cite_spans": [ | |
| { | |
| "start": 168, | |
| "end": 182, | |
| "text": "K\u00e4llgren 1986)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Some existing systems use a much more finely graded classification, and some also include e g a classification of the adverbials that can modify a verb or a sentence. Fig. 5 shows two entries from a such system, the Janus system, that seems to be carefully planned and built on a larger scale than most others, see Gumming 1986 for a closer description of its design. 5:name 'LEAD :spelling \"lead\" :", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 167, | |
| "end": 174, | |
| "text": "Fig. 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "features '(VERB INFLECTABLE LEXICAL NOT-CASEPREPOSITIONS O B J E C T P ERMITTED NOT-TOCOMP NOT-QUESTIONCOMP NOT-PARTICIPLECOMP NOT-MAKECOMP N O T -BAREINFINITIVECOMP N O T-COPULA PASSIVE NOT-THATCOMP N O T -A DJECTIVECOMP N O N E -O F -B I T R A N S I T IVE-INDIRECTOBJECT DOVERB DISPOSAL EFFECTIVE OBJECT NOTREQUIRED N O T -O B J ECTNOTPERMITTED SUBJECTCOMP U N I T A R Y S P E L L I N G S-IRR PASTFORM EDPARTICIPLEFORM)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": ":properties '((PASTFORM \"led\") (EDPARTICIPLEFORM \"led\"))", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": ":name 'SIMPLE :spelling \"simple\"", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": ":features '(ADJECTIVE NOT-CASEPREPOSITIONS R-ST DEGREE COMPLEME N TPERMITTED TOCOMP F ORNPPERMITTED SUBJECTHOLD NOT-SUBJECTCOMP N O N E -O F -A P P R O P R I A T E N E S S-POSSIBILITYPROPERTY-OBVIOUSNESS NOT-THATCOMP N O T -P R E D I C A T E O N L Y ) :properties '()", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The lexicon must mai nly be built up b y hand and that is a task that takes good and clear linguistic intuitions to do. Actually, it is quite h ard to imagine all the different kinds of complements that a verb can take.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Sometimes subtle shades of meaning in the verb accompanies differences in the functional frame ('to read a book' but 'to read a child to sleep').", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The verb can then be entered as two homonymous forms, or some other way of entering the information can be designed, but for the parsing process as such this does not n e c e ssarily remove all ambiguity.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "There is always the risk that some p o s sibility is forgotten and the functional frame gets too small, or that some allowed pattern makes the frame too loose and unwanted constructions become allowed.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In short, this way of put t i n g conditions on the parse is hard, tedious and time-consuming without ever guaranteeing total correctness.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "As a small excercise, the reader might want to try to figure out how the frame of the verb 'to read' should be described, given the corpus in 6, which is only a small subpart of the possible readings (!) of that single verb. What are objects and what are adverbials? When is a prepositi on not a preposition but an aspect marker? How can we capture that, in our culture, messages can be written on the side of buses, but not on trains or airplanes? And where precisely goes the borderline between signs that can be read and signs that can be read on? (b) She read on the train. (c) She read on the bus that there was a sale on. (Ambiguous.) (d) She read the street sign to see where she was. (e) *She read on the street sign to see where she was. (f) She read on the sign that the store was closing at 6. (g) *She read on the street sign that the street was 5th Ave. (h) She read on without being disturbed b y the noise. Some systems use an on-line ordinary dictionary instead. This makes large-scale parsing possible to an extent that will p r o bably never be reached with the more handicraft based systems, but it brings with it the problem of extracting the w a nted information from d i c tionary entries that were not w r i tten with this part i c u l a r applica tion in mind. The problem with full coverage of all possibilities also remains, but still this seems to be a more promising way. (Jensen and Binot 1987.)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The basic thoughts behind this technique are clear;", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Parsing really consists of (at least) two parts, corresponding to the important difference between category and function.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "It is one task to identify the constituents of a sentence and assign their category, another task to decide the functional roles that the c o n stituents play and to assign the structure of the whole sentence. Different languages signal categorial and functional information in different ways and to a different extent, a fact that must influ ence the mode of parsing chosen for a language.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "In a language like Finnish, both greimmatical category and syntactic function can often be seen from the morphology of single words (Karlsson 1985) , w h e r e as in English, knowledge about overall sentence structure is often needed in order to assign the correct category to ambigous words. In 7, the two parsing levels clearly presuppose each other: (7a) He judges sentence (7b) to be correct.", | |
| "cite_spans": [ | |
| { | |
| "start": 132, | |
| "end": 147, | |
| "text": "(Karlsson 1985)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "(b) Judges sentence criminals.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "Swedish lies somewhere between Finnish and English in this respect. The morphology gives a good idea of the category of a word, es p e c i ally when its close context is taken into account (K\u00e4llgren 1984), while it is mainly the word order of the full sentence that gives information about the syntactic function of its constituents. part of speech each wor d is and the allowed internal patterns of constituents such as noun phrases, prepositional phrases, and simple and complex verbs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "If for the moment we disregard the p r o blem of homonymy, we can generally say that a sequence of article + adjective + noun constitutes a noun phrase without knowing about the abstractness or animacy of the noun, and a finite verb is a finite verb, no matter what its functional frame may contain.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "For the task of category assignment then, no syntactic information except part of speech and internal structure of certain phrasal categories seems to be needed. All the elaborate apparatus describ ed above must thus be introduced in order to handle the assignment of syntactic function.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "This seems reasonable; the kind of condi tions expressed in verb frames are conditions on sentence struc ture, not on constituent-internal structure. The purpose of adding the verb frames and the syntactico-semantic properties of the nouns is then to aid in the identification of subject, object, main verb etc.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "The question will thus be: Precisely what kind of information and h ow much information is needed in order to reach a good parse?", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "T a g g i n g and p a rsing systems for English that make precisely this separation of tasks have been constructed for the large-scale analysis of the Brown and LOB corpora.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "Their systems for w o r d tagging, i e about the same as what was above called category assignment, are b a s e d on lexicon and morphology. TAGGIT is the system used for the Brown corpus.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "Its degree of correctness in deciding the grammatical class of words in unrestricted text is reported to be 77% (Greene and Rubin 1971).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "For CLAWS, the LOB system w h ich is developed on the basis of TAGGIT, the degree of correctness is 96-97% (Garside 1987).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Category assignment versus functional assignment", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "I have built up a morphologically b a s e d system for category assignment in unrestricted Swedish text. For theoretical reasons, I have tried to limit the lexicon as much as possible, and today the system has a lexicon of less than 300 words that b e l o n g to closed categories or are morphologically highly irregular, which should be compared to the lexicon of 7 200 words in the CLAWS system.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "I ndependently of their work,", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Except for those 300 words, all w o r d tagging in the Swedish system is b a sed on morphological and contextual clues. As stated above, Swedish morphology contains more information and less ambiguous information than is the case with English, so the performance of the lexicon-less Swedish system is around 90% on any arbitrary text. The system is described in K\u00e4llgren 1984a K\u00e4llgren ,b, 1985 . It is written in BETA (see Brodda 1987) and runs on DEC-10/20 computers under TOPS-10/20 and is being implemented on PCs under DOS.", | |
| "cite_spans": [ | |
| { | |
| "start": 362, | |
| "end": 376, | |
| "text": "K\u00e4llgren 1984a", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 377, | |
| "end": 394, | |
| "text": "K\u00e4llgren ,b, 1985", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "I ndependently of their work,", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Its output can give sentences that are analyzed like 8 and 9, or that simply look like the corres p o n d ing structures 10 and 11. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "I ndependently of their work,", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "To reach a full parse, I have constructed an algorithm for func tional assignment in Swedish sentences, which is b a sed on the output from the morphological system.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "The algorithm is now being implemented in BETA and CommonLisp.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "Its task is to decide, on the basis of as little information as in 10 and 11, the major syntactic roles of the constituents.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "I will here present the algorithm and its results and discuss the implications of it.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "The algorithm has been applied to natural Swedish texts that have already been analyzed into sentences and clauses, and with the following major categories identified:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "noun phrases, pronouns in subject and object form, prepositional phrases, finite and nonfinite verbs, auxiliaries, adverbs, prepositions, and conjunctions. As mentioned above, a system that does precisely this category assignment with a high degree of correctness (around 90%) already exists, but to be cible to judge the output from the different parts of the parse separately, we have started from an idealized situation where the category assignment is taken to be 100% correct. Note that the algorithm does not presuppose correct text, it presupposes text that does not need to be corrected.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "(Cf. the discussion in section 6 about the different possible purposes of parsers.) This means that it gives a best-po s s i b l e parse for every sentence, regardless of whether the sentence adheres to grammatical standards or not. A robustness of this kind is n e cessary to be able to deal with unrestricted text.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "We regard each occurrence of a verb and its complements as a simplex sentence. The present version of the algorithm only works with clauses that contain a finite verb. Non-finite clauses pose an extra set of difficulties to any parser.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "Subordinate clauses get their analysis both as to the functional role they play in the superordinate sentence and as to their internal structure. The basic task is then to identify the subject of each simplex sentence and its direct and indirect object, in case any object(s) occur.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "The algorithm is formulated as a set of rules (around 25 at present), of which some describe clear and unequivocal patterns and others give heuristic solutions for situations that can be ambiguous. Some examples: A noun phrase \"inside\" a complex verb, i e b e t ween a finite and an infite verb, must always be the subject of the verb. Examples like 12a-b thus gives the rule 13. is filled by an adverbial (an adverb or a prepositional phrase), the first noun phrase after the finite verb must be the subject. Examples 14a-b lead to rule 15. Where no such rules are applicable, the heuristic simply says that the first noun phrase in a sentence is its subject. Rule 16 gives a correct result in sentences like 17a but not in 17b.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "(16) NP Fin-v X -> NP := subj (17a) Mannen skriver alia brev i a r b e t s r u m m e t . the man writes all letters in the study (b) Alla brev skriver mannen i arbetsrummet. all letters writes the man in the study A set of analogous rules decide the assignment of direct object (mostly simply the noun phrase that is not subject) and the choice betw e e n direct and indirect object where two objects occur.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "An algorithm for functional assignment in Swedish sentences", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "The table below (Fig 18) gives the results for a corpus of 1,451 simplex sentences, taken from different text types. Of the 1,451 simplex sentences, 160 lack a finite verb, so the figures in the table are computed on the basis of the 1,291 sentences that the algorithm applies to.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 16, | |
| "end": 24, | |
| "text": "(Fig 18)", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "Instances where the algorithm (correctly) predicts zero subject, as in imperatives, count as a correct identification of subject. For the identification of objects, two figures are given. The first covers the cases where object(s) occur without a preposition in the sentence, while the second also includes instances with particle verbs ('skriva ner n\u00e5got' write down something) and instances where a prepositional phrase can on semantic grounds be regarded as a direct or indirect object.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "The last category can be difficult to judge, but some clear instances exist ('titta p\u00e5 TV' look at TV, 'ge boken till flickan' give the bo o k to the girl, as compared to 'ge flickan boken'", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "give the girl the book). Generally, a human linguist in ma n y cases has quite a har d time in choosing between particle + object/prepositional object/prepositional adverbial phrase.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "There are often no clear rules-of-thumb and different people come to different decisions, so the resulting figures in themselves are not that important. The important thing is that such instances occur, and that a w e a k point of the algorithm is its inability to decide the syntactic function of prepositional phrases.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "They are at present all regarded as some unspecified kind of adverbials.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "However, this deficiency is not as destructive as one might fear, since there are in fact not that many such o c c u r r e n c e s . The ratios for identification of subject and direct and indirect object without preposition are remarkably high. Indirect objects are surprisingly infrequent.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "In the whole material of almost 1,300 sentences only five indirect objects appear, three without a preposition that are all correctly identified and two instances with a preposition that have been m i ssed b y the algorithm.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "The really disturbing figure is the one for direct objects with preposition.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "The algorithm has missed 106 such objects, in total 15% of all direct objects, plus 8 direct objects without a preposition.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "Most of the latter errors are instances where the algorithm mixes up subject and direct object (cf 17b a b o v e ) . The number of such errors is however low, considering the fact that Swedish often has OSV order for reasons of textual coherence.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some results of the application of the algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "However, the most interesting results of the application of the algorithm are perhaps the errors. What type of structures can this simplistic method not manage? Can it be improved and developed? Would more elaborate systems like those men t i o n e d in section 1 do better? What kind of information -syntactic, semantic, pragmaticwould be needed in a lexically based parser? I will here give one example of every error type occuring in the material and discuss what kind of extra information would 'rescue' them.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some analysis errors occuring in the material", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "The examples below are all chosen because they are typical for the kind of errors they represent.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some analysis errors occuring in the material", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "As can be seen, many sentences are quite weird and some are even ungrammatical, but that is not an argument for excluding them from the corpus. The aim of my algorithm is to handle unrestricted text; the strange sentences have occured in normal texts and should thus be given an analysis. (Cf section 6.) As a matter of fact, there are also ungrammatical sentences that get correct functional assignment by the algorithm.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some analysis errors occuring in the material", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "Below each example I give the correct analysis and an English translation. As a comparison I then show a structurally analogous sentence, where the algorithm's analysis w o uld be the correct one. It is clear that in most cases the latter sentence represents a more common pattern. The algorithm says that if a finite verb comes immediately after a sentence conjunction, the subject of the preceding clause is the subject of that verb. (There is no number agreement between subject and verb in Swedish.) The error in 19 w o uld be solved by the extra information that 'st\u00e5r' (stands) is intransitive, so that 'en b y g g n a d . ..' cannot be a direct object, and, for semantic reasons (however they are to be specified), cannot be a temporal or locative adverbial; thus it must be the subject. At the same time I find it hard to believe that the selectional restrictions of a c a refully d e signed lexicon would allow lawns to be the subject of the verb bend.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some analysis errors occuring in the material", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "Mix i n g of subject and direct object: Sentence 21 has a verb phrase that is as problematic to a human as to a computer. Is the main verb a simply 'har' (has) with a nominal direct object, or is it 'har ... betydelse (\u00e4n)' (is of ... importance (than))? Most analyzers w o uld probably prefer the second solution, w h ich w o u l d have to be given as an entity in a lexicon, with slot for comparative adjective and all. But the situation is then messed up further by the splitting and fronting of the last part of the phrase.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some analysis errors occuring in the material", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "It is highly unlikely than any lexical look-up m e chanism wou l d be able to restore this discontinuous constituent. On rare occasions, a temporal clause can appear as a direct object of some verbs. Those verbs can however also take temporal clauses as adverbials.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some analysis errors occuring in the material", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "I can see no generalizable way of telling when the when-clause is an adverbial and when it is an object in sentences like 23 and 24. The adverbial reading must have an overwhelming probability in its favour. The main verb is the whole phrase 'fick tag i' (got hold of) with 'i' as a verbal particle, but it is interpreted w i t h the noun 'tag' as direct object and 'i' as a preposition starting a prepositional phrase incorporating what should really be the direct object. All those complex verbs -and there are ma n y of them -must be listed as idioms, but there remains the problem that some of them can have literal readings as well.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some analysis errors occuring in the material", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "No instances in the material. In its concrete meaning, a verb like 'avvek' is constructed with a locative adverbial, but it can also have a transferred meaning, where the locative reading of the complement seems less natural and an object interpretation is closer at hand.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Erroneous indirect object:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In these cases, there are mostly clear instances at either end and a grey zone in the (30) Det vet man v\u00e4l vad han g\u00f6r med kniven.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Erroneous indirect object:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "'One sure knows what he does with the knife.'", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Erroneous indirect object:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The particle verb 'g\u00f6r med' with a sense of doing something against s omebody/something has a particle identical to the instrumental preposition.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Erroneous indirect object:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "A lexicon might tell us that knives are typical instruments and persons are not, and thus guide the analyses of 29 and 30; but how w o uld it handle a sentence like 'One sure knows what he does with them'? A full anaphora resolution is necessary to decide whether the antecedent of 'them' is animate or not and, consequently, whether 'them' is to be analyzed as indirect object or instrumental adverbial.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Erroneous indirect object:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This does not always help either: 'books' are certainly not instrument in 'What will you do with all your books when you move?'", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Erroneous indirect object:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Many of the errors (21, 25, 27, 29 ) exemplify a general tendency. Verbs are often used with a transferred meaning that almost always implies an abstraction in comparison to their 'prototypical' meaning, or the meaning of the predicate is given by an 'empty' verb (be, do, have, get) plus a complement. In both cases, the patt e r n of functional roles of the verb/predicate is often affected.", | |
| "cite_spans": [ | |
| { | |
| "start": 19, | |
| "end": 23, | |
| "text": "(21,", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 24, | |
| "end": 27, | |
| "text": "25,", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 28, | |
| "end": 31, | |
| "text": "27,", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 32, | |
| "end": 34, | |
| "text": "29", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Erroneous indirect object:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Thus, from all this we can conclude that most of the functional assignment can be done without any recourse to lexicon at all, and for those cases that remain, the information necessary for a correct assignment is of a complex and often rather dubious nature. From the error analysis and from the senteces with 'read' in ex. 6, we can see that it is an enormous task to think of all the possible constructions, to find minimal properties that are not too ad hoc and that can allow the d e sired sentences and exclude the others, and to b u ild a system that can check all these possibilities and keep track of what it is doing. To get both readings of 6c for instance, we w o u l d have to b u ild into a parser the knowledge of the wor l d that (in this culture) buses, but not trains or airplanes, can be used as m o ving signboards for advertising, and that it is thus possible either to sit on a bus and read, or to stand beside it and read on it. We must ask ourselves if this is the kind of information we w o uld want to have in a syntactic parser and we must be aware of the fact that there will always be examples that we cannot manage. We must also acknowledge the fact that different purposes have different demands on the performance of parsers.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Erroneous indirect object:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Parsing is no longer only a research enterprise in itself, but will more and more become a necessary prerequisite for other kinds of theoretical research or practical applications, and that fact must influence the choice of parsing algorithm.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsers adjusted to different purposes", | |
| "sec_num": "6." | |
| }, | |
| { | |
| "text": "The important question of what to do wit h ungrammatical input should also depend on the purpose of the parser.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsers adjusted to different purposes", | |
| "sec_num": "6." | |
| }, | |
| { | |
| "text": "A parser that is used to test a specific linguistic theory should reject all input that is not in accordance with what the theory predicts. A b ro a d coverage of language is not seen as very ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsers adjusted to different purposes", | |
| "sec_num": "6." | |
| }, | |
| { | |
| "text": "Proceedings of NODALIDA 1987", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "important and there is a clear distinction between sentences that the parser can and can not handle, a distinction which not always corresponds to the intuitions of a human language user. Ungrammatical sentences are rejected, mostly without indication of what caused their ungrammaticality.A parser that is part of a grammar check or critiquing system of some editor should note all deviant sentences and perhaps even suggest how they should be corrected.It can also note stylistic features and compute frequencies for single words as well as c o n s t r u c t i o n s . Such parsers are supposed to cover large parts of the language and to find and diagnose literally all instances of u n g r a m m a t i c a l i t y . This makes their task considerably more difficult than that of the model testers, and truly there are no fullfledged critiquing systems available at present even if some attempts seem promising.A parser that is to be used in connection wit h e g a q u e s t i o n answering system must have the ability to decide not only what is grammatical but also what makes sense.It will ne e d more semantics than the other parsers, often in the form of knowledge about the database to which it is connected.Questions of grammaticality can be of less importance as long as the input is interpretable.The wor k with large-scale corpora that is emerging in today's linguistics demands of a parser that it can analyze sentences, add structural information, and always give a best-p o s s i b l e analysis, with or without at the same time signalling if there is something wrong with the sentence.Its input is more or less unrestricted text and its output can be used for many forms of linguistic research, also of the model-testing kind, as well as for buil di ng and updating databases etc.In this connection there are ma n y purposes for which a fast, simple, and robust parser of the kind suggested here is precisely what is needed. In: Proceedings from the ALLC conference, Goteborg June 1987. Forthcoming, Spr\u00e5kdata, Goteborg university.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "annex", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Roger, 1987. The CLAWS wordtagging system. In: Garside -Leech -Sampson, The Computational Analysis of English.London: L o n g m a n . ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Garside,", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": {}, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "She read the book." | |
| }, | |
| "FIGREF1": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "(i) She read quickly. (j) The book read quickly. (k) She read the book quickly. (l) The thermometer read 20 degrees. (m) She read 20 degrees on the thermometer. (n) Every morning she read the thermometer to see what to wear. (o) She read the child to sleep. (p) She read the book to pass the exam. (q) She (read) (up on the mountain). (r) She (read up) (on the mountain). (s) She read German. (t) She read law." | |
| }, | |
| "FIGREF2": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "(8) S :( N P :(Mannen) F i n -V : (skrev) NP:(ett b r e v ) ) the man wrote a letter (9) S : ( N P : (Kvinnan) F i n -V : (v i s s t e ) S :(C o n j :(a t t ) P r o n -s u b j :(h \u00e4 n ) the wo m a n knew that he F i n -V : (s k r e v ) NP:(ett brev) A d v l :(i d a g ) )) wrote a letter today (10) S:(NP Fin-V NP) (11) S:(NP Fin-V S;(Conj Pron-subj Fin-V NP Advl)) 8 Proceedings of NODALIDA 1987" | |
| }, | |
| "FIGREF3": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "w r i tten today (13) X Fin-V NP Infin-V Y -> NP := subj If the position before the finite verb (the .so-called fundament position)" | |
| }, | |
| "FIGREF5": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "A n a l y z e d in analogy with:(20) P\u00e5 sl\u00e4tten buktar gr\u00e4smattor och pryder sin omgivning.'On the plain b e n d lawns and adorn their surroundings.'" | |
| }, | |
| "FIGREF6": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "than the parliament have during autumn two other processes.' A n a l y z e d in analogy with: (22) St\u00f6rre universitet \u00e4n Link\u00f6ping har under h\u00f6sten tv\u00e5 andra problem. 'Larger universities than L. have during autumn two other p r o b l e m s .'" | |
| }, | |
| "FIGREF7": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "Erroneous direct object with complex verb: (25) ... n\u00e4r den ena fick tag i ett sn\u00f6re. (Subj ) (FV ) (Object ) '... when one got hold of a s t r i n g . ' Analyzed in analogy with: (26) ... n\u00e4r den ena fick buggar i ett program. \"... when one got bugs in a p r o g r a m . '" | |
| }, | |
| "FIGREF8": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "Missed prepositional direct object: (27) ... det avvek fr\u00e4n vad parterna tr\u00e4ffat f\u00f6rlikning om, ... subj FV (Object ) '... it deviated from what the parties had settled, ...' Analyzed in analogy with: (28) Han avvek fr\u00e5n anstalten. 'He deviated from the prison.' = 'He escaped.'" | |
| }, | |
| "FIGREF9": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "m i d d l e . Missed prepositional indirect object: (29) (Det vet man v\u00e4l) vad han g\u00f6r med oss. Obj Subj FV (Indir O b j ) '(One sure knows) what he does to us.' 13 Proceedings of NODALIDA 1987 Ana l y z e d in analogy with;" | |
| } | |
| } | |
| } | |
| } |