| { | |
| "paper_id": "1993", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:37:25.479423Z" | |
| }, | |
| "title": "Handling Syntactic Extra-Grammaticality", | |
| "authors": [ | |
| { | |
| "first": "Fuliang", | |
| "middle": [], | |
| "last": "Weng", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Mexico State University", | |
| "location": { | |
| "postCode": "88003", | |
| "settlement": "Las Cruces", | |
| "region": "NM" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "This paper reviews and summarizes six different types of extra-grammatical phenomena and their corresponding recovery principles at the syntactic level, and describes some techniques used to deal with four of them completely within an Extended GLR parser (EGLR). Partial solutions to the remaining two by the EGLR parser are also discussed. The EGLR has been implemented.", | |
| "pdf_parse": { | |
| "paper_id": "1993", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "This paper reviews and summarizes six different types of extra-grammatical phenomena and their corresponding recovery principles at the syntactic level, and describes some techniques used to deal with four of them completely within an Extended GLR parser (EGLR). Partial solutions to the remaining two by the EGLR parser are also discussed. The EGLR has been implemented.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Extragrammatical phenomena in natural lan guages are very common and there has been much effort devoted to dealing with them ( Car bonell -Hayes, 1983; DARPA 1991 DARPA , 1992 . Al though (Generalized)LR parsers have many mer its when applied to NL, most progress with extra grammatical phenomena has been through rule based systems, in contrast to the applications of LR parsers in programming languages. In this pa per some techniques are developed to extend the ability of a ( G )LR parser in dealing with extra grammatical phenomena, though similar tech niques can also be applied in other parsers. The extended GLR (EGLR) parser is implemented.", | |
| "cite_spans": [ | |
| { | |
| "start": 125, | |
| "end": 151, | |
| "text": "( Car bonell -Hayes, 1983;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 152, | |
| "end": 162, | |
| "text": "DARPA 1991", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 163, | |
| "end": 175, | |
| "text": "DARPA , 1992", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In section 2, six types of extra-grammatical phenomena at the syntax level are classified, and then five recovery principles are introduced. Sec tion 3 then describes the techniques used in the EGLR parser. A correctness theorem is given, and a property is also presented, showing that any sentence can be accepted if the relaxation parameter is set large enough. Section 4 gives some examples to show how the EGLR parser works. Section 5 briefly discusses the sixth type of extra-grammatical phenomenon and relevant issues, and other people's work is also compared.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Among discussions of extra-grammaticality in natural language processing, (Carbonell -Hayes, 1983 ) gave a comprehensive and complete overview in this area . Here we try to rephrase these extra-grammatical phenomena from the viewpoint of purely structural possibilities at the syntactic level. Following (Carbonell -Hayes, 1983) , we use extra-grammaticality to refer to phenomena in which sequences of words are not covered by cur rent grammar rules. In doing so, we try, at this moment, to avoid unnecessary debate about the possibility of drawing a clear line between gram matical and ungrammatical phenomena. At the end of the paper, we shall show an interesting re sul t with a deviation degree parameter, produced by our parser that may provide some hint for understanding (un) grammaticality. In what fol lows, extra-grammatical phenomena contain both ungrammatical and uncovered grammatical phe nomena.", | |
| "cite_spans": [ | |
| { | |
| "start": 74, | |
| "end": 97, | |
| "text": "(Carbonell -Hayes, 1983", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 304, | |
| "end": 328, | |
| "text": "(Carbonell -Hayes, 1983)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 779, | |
| "end": 783, | |
| "text": "(un)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Categorizing Extra-gram maticality at the Syntax Level", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "The six types of syntactic extra grammaticalities are: 1 Phenomenon 1: the absence of a word's cate-1 Phenomenon 1 was discussed in (Tomita, 1985) and might be also considered to be at the lexical level. Phenomena 2, 3 and 4 and the similar recovery principles were also discussed in (Aho -Peterson, 1973; Saito -Tomita, 1991) .", | |
| "cite_spans": [ | |
| { | |
| "start": 132, | |
| "end": 146, | |
| "text": "(Tomita, 1985)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 284, | |
| "end": 305, | |
| "text": "(Aho -Peterson, 1973;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 306, | |
| "end": 326, | |
| "text": "Saito -Tomita, 1991)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Categorizing Extra-gram maticality at the Syntax Level", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "xyz is not present in the English dictionary and therefore the sentence is not covered by a grammar.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "gories. For example, xyz leads arbitrariness., w9\ufffdm", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Phenomenon 2: category switching.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "gories. For example, xyz leads arbitrariness., w9\ufffdm", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For example, Th e man on the left was talk ing non-sense., when in the dictionary en try, left only has a category ADJ but not N, and therefore the sentence can't be covered by a grammar with only N as its head in NP rules. Phenomenon 6: non-extragrammaticality fail ure: a constituent is covered by rules but not intended by the speaker (or writer). For example, Th e man sit on the river bank, given the only rule NP +-Det N for NP formation. In this case, the river will form a NP, instead of the river bank.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "gories. For example, xyz leads arbitrariness., w9\ufffdm", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We put aside the last phenomenon for a mo ment since it needs the coordination of multiple levels besides the syntactic one.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "gories. For example, xyz leads arbitrariness., w9\ufffdm", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Corresponding to these phenomena, we pro pose five principles which try to remedy the fail ures caused by them. When a parser fails, at a WENG point, we hypothesize several alternatives accord ing to the following five principles, which corre spond to the first five types of phenomena:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "gories. For example, xyz leads arbitrariness., w9\ufffdm", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Principle 1: hypothesizing all the categories as the categories of the absent of the failed word.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "gories. For example, xyz leads arbitrariness., w9\ufffdm", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Principle 2: hypothesizing the complement of all the categories of the failed word.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "gories. For example, xyz leads arbitrariness., w9\ufffdm", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Principle 3: hypothesizing a pseudo word which has all the categories.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "gories. For example, xyz leads arbitrariness., w9\ufffdm", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Principle 4: hypothesizing that the failed word is improperly added. In the next section, we will present some tech niques to incorporate the first four principles into an extended GLR parser, and a limited version of the fifth principle can be achieved by setting the relaxation parameter to 2, a number which will be explained later.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "gories. For example, xyz leads arbitrariness., w9\ufffdm", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The Parser (s, t) in the action ta ble may either contain a set of actions (i.e., A( s, t) = { ai} ) , where s stands for a particular state, t stands for a particular terminal, and ai stands for actions which could be shifting, reducing or ac cepting, and A( s, t) can also be empty. The last case (i.e., A(s,t) is empty) indicates that, in state s, it is impossible to meet terminal t if the input is grammatical: in other words, the parser will report an error if, for any current state s and any category c associated with the current word, A(s, c) is empty. Like the GLR parsers when an input sentence appears, the EGLR algorithm reads one word after another from left to right, goes from one state to another, and does what the two ta bles specify: reduce, shift , etc. And the actions of the EGLR parser, i.e., reduce, shift and accept, are very similar to the ones in GLR parsers, ex cept in places that will be specified later. The main differences between the GLR parsers and the EGLR parser are caused by the additional requirement of EGLR, i.e., allowing different hy potheses to compete when an error is detected ( or a mismatch occurs) .", | |
| "cite_spans": [ | |
| { | |
| "start": 11, | |
| "end": 17, | |
| "text": "(s, t)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "3", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The realization of different hypotheses is not so direct if we still want to stick to the following idea: allow different principles to compete at a same time, whenever it is possible, since different principles create different new sentence lengths, i.e. principles 1 and 2 do not change the length of the input sentence, principle 3 increases the length by one unit, and principle 4 reduces the 321 length by one unit. The alignment of these length discrepancies and the compensation of its effect are realized as follows ( Among the four steps, the first three are for the parser itself and the fourth is for the input sentence. Since there might be more than one er ror in an input sentence, the fourth step can be repeated. We use a relaxation parameter to char acterize the maximum number of times allowed in performing such relaxation.", | |
| "cite_spans": [ | |
| { | |
| "start": 524, | |
| "end": 525, | |
| "text": "(", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Notice that in the fourth step, the four com binations of the two successive category sets, i.e. { * Spall * } { * Span * }, cati { * Spall * }, { * Spall * } cati and cati cati , are exactly the first four principles 3 plus the original sentence. Since the original sentence is blocked at that point the reconstruc tion realizes the first four principles at the same time.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The first four principles are all in favor of lo cal mismatches. In order to accommodate the fifth principle better, which is non-local, simple alignment is not sufficient and a more compli cated mechanism needs to be used. One method for this is to use the notion of parameterization in a universal grammar. Notice that swapping two linguistic units requires the same underlying mechanism as resetting a parameter in X-bar the ory (Chomsky, 1980; Gibson, 1989; Nyberg III, 1989) .", | |
| "cite_spans": [ | |
| { | |
| "start": 432, | |
| "end": 447, | |
| "text": "(Chomsky, 1980;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 448, | |
| "end": 461, | |
| "text": "Gibson, 1989;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 462, | |
| "end": 479, | |
| "text": "Nyberg III, 1989)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Although the parser only encoded the first four principles directly, it is not difficult to see that a limited version of the fifth principle, i.e. swapping two adjacent words, is implied by the first four principles with the relaxation parame ter being 2. The details will be illustrated by an example in section 4. One may notice that in the phenomena and the principles enumerated, we not only allow words but also constituents to be manipulated, while up till this moment in the EGLR only words are taken as the basic units. One way to deal _ with constituents is as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We introduce a notion called virtual termi nals, abbreviated as vtm, that contain informa tion about the corresponding non-terminals; a normal grammar is appended with a set of rules with vtms: A.__ * VtmA , where A E NTM and NT M is the set of non-terminals 4 ; since * VtmA will never occur as a category of any word, the parser will not take any path containing * VtmA unless we reconstruct the input category sequence when necessary, which is very much like the case when we deal with word as unit. An example with the extended rules and their GPTs is shown Like (G)LR parsers, the EGLR parser has two main components: a parsing table generator and a parsing algorithm. The generator takes a context-free grammar as its input and produces an action table and a goto table; while the parsing algorithm takes a sequence of words and a value of the relaxation parameter as its input and pro duces results, usually with four parts:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "1. a sequence of triples, whose first element is the index of the triple in the sequence, whose second element is the set of the cate gories attached to the word in the input sen tence, and whose third element is the word itself;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "2. the history of the sequences of the triples, whose first element is the index of the triple in the sequence like in part 1, whose second element is either the same set of the cate gories as in part 1 when it is parsed, or the derived one based on the alignment proce dure when a mismatch happens, and whose third element is also the same as in part 1 except in the case when a mismatch occurs and then NIL is supplied;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "3. (partially) successful hypothesized se quences of categories if there are mis matches, and the number of times the re laxation process is actually performed (i.e., the value of the deviation parameter) ;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "4. a shared-packed forest representation for the sequences in item 3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Items 2 and 3 will not be present if the original category sequence is covered by the grammar.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We now present some simple examples in or der to explain better to the reader the underlying ideas, using the grammar and its corresponding GPTs given in Table 1 and Table 2. The grammar in Table 1 does not contain any rules for pronouns. Assuming that a pronoun oc curs in an input sentence and the lexicon does provide a category ( *PRON) for the pronoun (i.e., a case when phenomenon 2 happens), how will the parser behave? Example sentence 1 ( They read the book.) shows this case, where part 1 con tains a sequence of sets of categories attached to the words in the sentence together with their in dices in the sentence. The indices start at 0. 1-st hypothesized sentence line in part 2 gives a guess when the parser found that the grammar does not cover *PRON, according to the principles in section 2. After feeding the new sentence to the parser internally, the parser returns the successful category sequence in part 3 and its parsing forest in part 4. A few more words about the shared packed forest representation are given here: each row in the list is a description of a node in the forest, and it consists of two parts: the first part is the index of a node, and the second part is a tuple; the first argument of the tuple is either a terminal or a non-terminal; if it is a terminal, a T and the word 6 associated with the terminal are followed; if it is a non-terminal the follow-up lists include all its child node sets, e.g., 10 (NP (8 9)) means that node 10 is a NP having nodes 8 and 9 as its children. ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 154, | |
| "end": 174, | |
| "text": "Table 1 and Table 2.", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 190, | |
| "end": 197, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "GLR", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Example sentence 2 shows the phenomenon 1 and the result produced by the parser. It is simi lar to example sentence 1, and hypotheses a set of categories for the unknown word researchers and produces a parsing forest for that guess. N (NIL -1 0) ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 233, | |
| "end": 245, | |
| "text": "N (NIL -1 0)", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "*", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(S (5 10)) the end of the forest .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "11", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Example sentence 3 shows a combination of several possible phenomena. The category set se quence derived from sentence the man lives in in the house can not get parsed without relaxation, and \u2022 the parser detects an error at the second in indexed as 4-th in the input sequence. Then its first hypothesized sequence is proposed and shown in part 2. Because the default value of the relaxation parameter is 1, this relaxation process is permitted and the indexed sequence is inter nally resubmitted to the parser. This time, it gets parsed. The parsing forest is shown in part 4 and a set of successfully hypothesized sequences are given in part 3. The above examples only show that if there is one extra-grammatical place in a single sen tence the parser can deal with it . Actually, the ability of the parser is not limited to this. An op tional relaxation parameter is offered ( default is 1) by the parser. So, if there are multiple extra grammatical places in one sentence, by setting this relaxation parameter properly the parser can still proceed with various guesses. Example 4 shows this situation. Sentence the man home likes is not covered by the grammar given in Table 1 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 1173, | |
| "end": 1180, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "NIL *", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "When the parser tries to parse the category set sequence derived from sentence the man home likes with the relaxation parameter being 1, it detects an error at the word home indexed as 2nd in the input sequence. Then its first hypoth esized sequence is proposed and shown in part 2. Because the value of the relaxation parame ter is 1, this relaxation process is permitted and the indexed sequence is submitted internally to the parser. With the reconstruction of the input sequence, the parser proceeds and goes through word likes, and then it detects another error and can not go further because the relaxation param eter is 1 and no further relaxation is allowed.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "NIL *", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "A next example is about the same sentence be ing presented to the parser with the relaxation parameter being 2. The same process as the one in the previous example happens until the parser detects the second error at position 5. This time, another hypothesized sequence gets proposed and allowed to be internally resubmitted to the parser because the relaxation parameter is 2. It gets parsed. The parsing forest is shown in part 4.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "NIL *", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The two successfully hypothesized sequences in part 3 need a little bit more explanation. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "NIL *", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "HOME), we actually could realize a limited ver sion of the fifth principle, i.e., swapping two adja cent words, as mentioned before. A finer degree of relaxation related to the fifth principle can also be classified as below, although its usefulness may likely be finally decided through multiple knowl edge interaction: 1. When the deleted category set (e.g., cat(HOME) in the current example) has no intersection with the inserted category set (e.g., (*N) in the current example), rank it as bad; the relaxation may not be allowed in this case; 7", | |
| "cite_spans": [ | |
| { | |
| "start": 354, | |
| "end": 370, | |
| "text": "(e.g., cat(HOME)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 446, | |
| "end": 457, | |
| "text": "(e.g., (*N)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "325", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "2. When the deleted category set has intersec tion with the inserted category set, but not inclusion, rank it as acceptable;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "325", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "3. When the deleted category set is contained the inserted category set, rank it as good;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "325", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "4. When the deleted category set contains the inserted category set, rank it as good;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "325", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The second hypothesized sequence is similar to the first one. The difference between the first hypothesized sequence and the second one is that in the first relaxation, instead of deleting princi ple, the inserting principle survives. All the rest of the two hypothesized sequences are more or less the same and we are not going to explain further. the grammar does not accept this sentence ! here is a partial result maticality and the deviation/relaxation param eter. The hints provided by our parser seem to suggest that grammaticality is a graded and individual-related notion, i.e. it depends not only on the individual's knowledge about language but also the ease of absorption of certain phenomena into her or his knowledge. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "325", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The curly parentheses indicate sets, LJ set union, and the overline on a set refers to the set complement operation w.r.t. the set of all the terminals, terms.3 The first two principles share much in common and are dealt with in a same way.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Some non-terminals may not be able to derive a string containing only terminals, we can use a standard algorithm to detect them in (Hopcroft -Ullman, 1979) Ch 4.4, and there is no need to have the corresponding rules.5 The proof of the two properties are given in (Weng, 1993).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "If the deleted category set is empty or contains only unrecognized categories, the ranking may be different.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "The author would like to express his thanks to Dr. Y. Wilks for his helpful comments on the content and presentation of this paper. Thanks also goes to the anonymous referees for valuable comments.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Acknowledgments", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The following hypothesized sentences get parsed:) and the value of the deviation parameter is 2 Part 4:it is accepted the root of its parse forest is (19) , and the forest is :", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "annex", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(NP (7)) , 1975; Fass -Wilks, 1983; Slater -Wilks, 1990 ). ", | |
| "cite_spans": [ | |
| { | |
| "start": 9, | |
| "end": 16, | |
| "text": ", 1975;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 17, | |
| "end": 35, | |
| "text": "Fass -Wilks, 1983;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 36, | |
| "end": 55, | |
| "text": "Slater -Wilks, 1990", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "9", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "A", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "References Aho A. -T. Peterson (1972)", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": {}, | |
| "ref_entries": { | |
| "FIGREF1": { | |
| "text": "the indexed sentence = 6 NIL is filled if the terminal is from a hypothesized category set", | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null | |
| }, | |
| "FIGREF4": { | |
| "text": "The first sequence ( (*DET THE) (*N MAN) (*V LIKES) (*N NIL)) is created as follows: after the parser parsed ((*DET) THE) and ((*N) MAN), it meets ( (* ADJ * ADV *N) HOME) and detects an error as described earlier. The parser then uses the four principles to create hypotheses, and after the parser processes ((*V) LIKES), the deleting principle survives. Since the next symbol in the input sequence is the end of the sequence and the parser detects another error. Further hypotheses are made and the one with *N finally survives. Back linking the *N with the one deleted (i.e.,", | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null | |
| }, | |
| "TABREF4": { | |
| "type_str": "table", | |
| "text": ". for the set of grammar rules and its GPTs given inTables 1 and 2.", | |
| "content": "<table><tr><td>One may not agree with the locality principle, i.e. that the position of real error(s) is not nec essarily the place where there is no entry in the generalized action table (GAT), in other words, there are cases that can be characterized by phe nomenon 6, which we will call extra-grammatical garden path, abbreviated as xgp. Obviously, xgp can also occur in combination with the first 5 types of phenomena. One solution to the xgp problem is to make some changes to cati -I \u2022 Here, we can reconstruct the input sentence at cati -I in step 4 of the alignment procedure, instead of cati .</td><td colspan=\"2\">To conclude this section, a correctness theorem and a property concerning the EGLR parser are presented as follows: 5 Theorem 1. Let inputo = C1 , C2 , ... , C n be a sequence of cate gory sets without * Span * , input 1 = C1 , C2 , ... , Ci -1 , ci+l, ... , C n , and input2 = C1 ,C2 , ... ,Ci LI { * Span * }, ... , C n . Then a GLR parser accepts input 0 or input 1 with grammar G iff the EGLR parser ac cepts input2 with the same grammar and relaxation parameter being 0.</td></tr><tr><td colspan=\"2\">It is not necessary that cati -1 should be mod ified. In general, as a k-step xgp, cati -k is modi fied, where k \ufffd 1 and the choice of k can be based on knowledge sources at other levels, such as se mantics and discourse, when wardi -k seem to be a particularly odd fit in the local context. Note too that step 4 in the alignment proce d ure can be replaced by any reasonable hypothet-ical sequence of categories, including the pseudo-4 terminal * Span * . We consider the one presented here as a good candidate.</td><td>Property: Given any sentence and context free grammar, we can always find a value for the relaxation param eter such that the EGLR parser ac cepts the sentence based on the four recovery principles. Some Results Produced by the EGLR</td></tr></table>", | |
| "html": null, | |
| "num": null | |
| }, | |
| "TABREF5": { | |
| "type_str": "table", | |
| "text": "The Extended GPTs.", | |
| "content": "<table><tr><td colspan=\"3\">the extended generalized action table: I stt-id *DET *N *PREP *V $ *SPAN* *VTMNP *VTMPP *VTMS *VTMVP 0 s3 s4 s0 s5 s6 1 s9 s11 s1 s10 s12 2 s9 a s2 s10 3 s14 s3 4 rl r1 rl s4 rl rl 5 r2 r2 r2 s5 r2 r2 6 r6 r6 s6 r6 7 r7 r7 s7 r7 8 r3 r3 r3 s8 r3 r3 9 s3 s4 s9 s5 10 r5 r5 r5 s10 r5 r5 11 s3 s4 s9 s11 s5 s10 12 r12 r12 s12 r12 13 r8 rs s13 r13 14 rO rO rO s14 rO rO 15 r4 s9 r4 r4 s15 r4 s10 r4 16 rll rll s16 rll 17 rlO s9 rlO s17 rlO s10 the VP 0 1 2 1 8 7 2 13 3 4</td></tr><tr><td>5 6 7 8 9 10 11 12 13 14 15 16 17</td><td>15 17</td><td>16 8 8</td></tr></table>", | |
| "html": null, | |
| "num": null | |
| } | |
| } | |
| } | |
| } |