JPH0713968A - Syntax checking device - Google Patents

Abstract

PURPOSE:To support the generation of a document by indicating the cause or position of a grammatical error included in the document with high precision. CONSTITUTION:The document is segmented by each sentence by a sentence segmenting part 1 and transmitted to a syntax analysis part 2 by a sentence unit. When the analysis of the whole sentence becomes failure in the syntax analysis part 2, prepared partial trees (plural) and information on the failure of rule application is transmitted to a non-sentence analyzing part 5. The non- sentence analyzing part 5 applies a non-sentence analysis rule 6 for investigating the error with the combination of those as an argument. As a result, an error kind is specified and a message to a user is outputted from a message generating part 7 in accordance with the kind. A priority processing part 9 for deciding the priority of the partial trees is added before the non-sentence analyzing part 5 so that a processing can be executed at high speed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a syntax check device in document processing and natural language processing.

[0002]

2. Description of the Related Art Conventionally, techniques for pointing out erroneous sentences in order to support document creation include spell checking that points out spelled character strings that are not found in dictionary headings as unknown words, and erroneous word arrangements. Some match patterns and point out grammatical mistakes.

[0003]

There are spelling errors and grammatical errors as errors, but here we will consider grammatical errors. With the method of matching the pattern of word arrangement and pointing out an error, it is possible to point out when the input matches the pattern completely, but if the pattern is disturbed by the influence of modifiers or conjugation , I can't match because I can't match. In order to point out such an error, you have to prepare another pattern,
Not only does the number of rules grow enormous, but it also increases the possibility of accidental hits. Also, if the matching is hit, an error can be pointed out, but if the matching is not hit, no information regarding the error can be obtained. Furthermore, it is difficult to point out missing words because it is difficult to describe them as patterns. It is theoretically possible to predict all possible substitutions, omissions, additions, etc. for grammatical mistakes, but it is theoretically possible, but since these possibilities have a huge number of combinations,
The data of dictionaries and syntax rules become large, and the processing time becomes too long to be practical.

The present invention solves such a conventional problem, and an object of the present invention is to provide a syntax check device capable of greatly improving various errors in a sentence as compared with simple pattern matching. To do.

[0005]

SUMMARY OF THE INVENTION In order to achieve the above object, the present invention parses an input document sentence by sentence, and parses the clipped input sentence using a parsing rule. , If the input sentence has an error and the parsing fails, the parser outputs the subtree up to that point and the error in the input sentence using the non-sentence analysis rule from the subtree output by the parser. Is provided with a non-sentence analysis unit for detecting a message and a message creation unit for outputting a message corresponding to the detected error.

According to the present invention, a non-sentence analysis unit is given a high priority to a substructure that is not included in another subtree from a set of substructures that form a subtree created by the syntactic analysis unit. It is provided with a priority processing unit for sending to.

[0007]

According to the present invention, with the above configuration, the input document is first cut out one sentence by the sentence cutout unit and sent to the syntax analysis unit. The parsing unit parses the extracted input sentence. Partial parse tree (plural) created in the process of parsing when the input sentence is incorrect and parsing fails
And the failure information to the non-sentence analysis section. The non-sentence analysis unit successively applies non-sentence analysis rules for detecting errors to these syntax trees. If an error is found as a result, the message creating unit creates a corresponding message and points it out to the user. If no error can be found by the non-sentence analysis rule, the user is presented with information regarding a likely error position based on the position of the created subtree. Also, the priority processing unit gives priority to the subtrees output from the syntax analysis unit, and applies the non-sentence analysis rules in descending order of priority.

[0008]

1 is a block diagram of a syntax check device showing an embodiment of the present invention. 1 is a sentence cutout unit that cuts out the input document one sentence at a time, 2 is a syntax analysis of the cut out input sentence, and if there is an error in the input sentence and the syntax analysis fails, the subtree up to that point Parser that outputs
Is a parsing rule used for parsing, 4 is a word dictionary that holds word information, 5 is a non-sentence analysis unit that detects a non-sentence cause from the subtree output by the syntactic analysis unit 2, and 6 is when searching for an error The non-sentence analysis rule describing the condition of No. 7, a message creating unit for creating a message corresponding to the detected error, 8
Is a message dictionary used when composing a message. 9 indicates a substructure that is not included in another subtree in the set of substructures (nodes) that make up the subtree passed to the non-sentence analysis unit 5 because the parsing unit 2 failed in the parsing. This is a priority processing unit that gives a high priority to the non-sentence analysis unit 5 for processing.

Next, the operation of the above embodiment will be described. The document to be pointed out as a erroneous sentence is first the sentence segmentation unit 1
Entered in. The sentence cutout unit 1 cuts out a document into sentences one by one based on a delimiter and the like, and sends it to the syntax analysis unit 2. When the input sentence is input to the syntactic analysis unit 2, the syntactic analysis unit 2 starts the analysis from the beginning of the sentence and grows the syntactic analysis tree while drawing the syntactic analysis rule 3 and the word dictionary 4. When a certain rule in the parsing rule 3 is successful, the resulting subtree is saved. Hereinafter, this subtree will be referred to as a “success goal”. Since the success goal is a subtree, it includes information such as the part of speech and conjugation of each word included in the subtree, the category and role of each branch, and the position in the sentence. If the parsing algorithm can generate the same subtree many times, check that the exact same subtree is not duplicated. In the present embodiment, an example will be described in which a parser that is a syntactic analysis program uses a syntactic analysis algorithm that grows a tree bottom-up while predicting a category. Sometimes a category that is in the middle of analysis is predicted but the tree rooted in that category may fail to be generated. In this case, the category name that failed and the position in the sentence are saved. Hereinafter, the trace of failure that has two pieces of information on the category name and the position will be referred to as a "failure goal". In this embodiment, the syntax analysis unit 2 grows the tree from the beginning of the sentence. Therefore, if there is an error in the sentence, the subtree may not be formed in the subsequent portions. Most of the attributes of a tree that has grown greatly are inherited from the attributes of those that play a major role in the subtree of the tree. In this way, the attribute as the tree (category) can be registered in every subtree according to the structure and the included words. If the parsing is successful, processing ends there. When the syntactic analysis fails, the created success goal and failure goal are sent to the non-sentence analysis unit 5.

The non-sentence analysis section 5 applies the non-sentence analysis rule 6. Each non-sentence analysis rule 6 includes (1) some category names (including part of speech), (2) positional relationship between them, and (3)
It is possible to describe three of the success conditions and the failure goals of each category as the activation conditions. For example, in the case of rules for extracting contradictions concerning English articles and the singular / plural of such nouns, (1) is the article and the noun phrase main part, and (2) is these articles in the order of the article and the noun phrase main part. They are adjacent, and both (3) are "success goals". This rule can be applied to mistakes such as "a boys". Here, the main part of the noun phrase is formed as a subtree even if a modifier such as an adjective is applied, so "a good boys" that cannot be used for pattern matching such as "words with a + s"
An error can be detected even for a form such as "." In addition, attributes such as whether or not they are plural are registered as tree attributes when the parsing unit 2 creates a parse tree, so by paying attention only to “tree attributes” 1 even if the pattern is different on the surface
An error can be determined by one rule, and an increase in the number of rules can be prevented.

The non-sentence analysis unit 5 extracts from the input all combinations of success goals and failure goals that meet the conditions (1), (2), and (3) of each non-sentence analysis rule 6, and then extracts the rules. Apply. The method to extract all combinations is
A well-known algorithm such as sorting subtrees by position or root category may be used. Inside each non-sentence analysis rule 6, an attribute is inspected for each goal of rule (1) if necessary, and a contradiction is extracted. In English, for example, tenses and singular / plural matches are checked. There are also rules that do not check attributes. If a discrepancy between categories is detected, it is checked for other non-conflicting interpretations if necessary.
This is often done by checking for positions in the sentence that have the same category name but different attributes. If no other inconsistent interpretation is found by this inspection, the goal associated with the error type (identifier) is output to the message creating unit 7.

When the identifier is sent from the non-sentence analysis unit 5, the message creation unit 7 extracts a message according to the type of contradiction from the message dictionary 8. The corresponding word or phrase is extracted from the input sentence or subtree, and the final message is composed by performing processing such as embedding in the message. If no contradiction can be detected by any of the non-sentence analysis rules 6 in the above process, the range in which the success goal is created in the sentence is presented as a message to the user. This can be realized by generating a special error identifier when the non-sentence analysis unit 5 ends the process without detecting an error.

Next, the above operation will be described based on an actual example sentence. The target language is English. In the following description, for English part-of-speech, a noun may be expressed as N, a general verb as V, a be verb as BE, a preposition as P, a pronoun as PN, an article as DET, and an adjective as ADJ. In addition, although the parse tree plays an important role, the category name that becomes the node is expressed as S, noun phrase NP, verb phrase VP, infinitive phrase INF, noun phrase main part NH, etc. There is.

Taking the input sentence "He drank a cold water." As an example, this is grammatically erroneous in that the indefinite noun "a" is applied to the non-additive noun "water". There is. (In colloquial terms "a cup of wate
There may be an interpretation such that the r "is omitted, but this is a grammatical error.) This sentence is separated from the preceding and following sentences by the sentence cutout unit 1 and input to the parsing unit 2. Since this input sentence contains an error,
The parsing unit 2 fails in complete parsing. However, as a result of partial analysis, a subtree as shown in FIG. This is shown in FIG. 2B as a success goal and a failure goal. The success goals are shown in the category name and the range of control. It actually has more internal structure and grammatical attributes. The failure goal is indicated by the starting position and category name in the sentence. The description method of success goals and failure goals is the same for other figures.

Here, it is assumed that there is a parsing rule as shown in FIG. When applying each rule, the attribute match is checked, and if the conditions are not met, it fails. This rule will be described by applying it to the process of forming a subtree as shown in FIG. At the beginning of the sentence, rule 1 predicts NP and rule 2 predicts PN, but the attribute of the word "He" is PN.
Therefore, the prediction is satisfied, the successful goal PN can be made, and further the successful goal NP can be made. As a result, VP, which is the next element of rule 1, is predicted, and further V (transitive verb) is predicted by rule 6, and "drank" is a transitive verb.
A success goal V is made, and the next element of rule 6, NP, is predicted. Further, according to Rule 3, the success goal DET and the success goal NH are predicted and created. However, at this stage, when trying to create an NP based on Rule 3, the creation of the success goal NP fails due to the mismatch of attributes (addition / non-addition). At this time, information that the NP starting from the position of “a” has failed is registered as a “failure goal”. By now, it can be seen that success goals and failure goals as shown in Fig. 2 can be made. Although the above-mentioned rules and actions of the parsing unit 2 are only important for performing non-sentence analysis, the number of rules is much larger than this in order to cover many language phenomena. Must be,
Since there are often multiple parts of speech and attributes for a single word, success goals and failure goals can be other than those listed above. In particular, there are so many failed goals that I don't know at this point which goals are the ones that lead to mistakes.

The success goal and the failure goal of FIG. 2B created as described above are sent to the non-sentence analysis section 5. The non-sentence analysis unit 5 applies the non-sentence analysis rules to the categories adjacent to each other in FIG. Figure 4 Example of non-sentence analysis rules
Shown in. The combination of adjacent success goal DET and success goal NH, which are arguments of rule A in FIG. 4, is the success goal DET (= "a") and success goal NH (= "cold water") in the input. Fits. Therefore, the rule A is applied with these as arguments. At this time, in the success goal DET is "a"
The attributes "addition", "single", etc. are registered. The attribute of the noun "water" which is the main part of the success goal NH is inherited as the attribute of the entire success goal.
Attributes such as "no addition" are registered in the success goal.
These are referred to in the rule A, and "addition" and "non-addition" are respectively detected, so that it is determined that "error 1" is established. Next, it is examined whether there is another interpretation, but it is found that "water" has no other meaning such as an additive noun, and there is no other interpretation that is consistent. As a result, the identifier “error 1”, the success goal DET, and the success goal NH are sent to the message creating unit 7.

In the message creating section 7, a message corresponding to the type of error is retrieved from the message dictionary 8 and
Include success goal information in the message for clarity. In the case of this sentence, the identifiers "error 1" to "
Addition / non-addition of * and * is inconsistent. Message is extracted from the message dictionary 8, and the word "a" from the success goal DET and the central word "water" from the success goal NH input into the message dictionary 8 are incorporated, and "" a "
And "addition / non-addition of" water "are inconsistent" is created and output to the user.

Next, another operation will be briefly described by taking the input sentence "This is pen." As an example. In this case, the article "a" is missing just before "pen". Such a lack cannot be detected by word-level pattern matching. The analysis result in the syntax analysis unit 2 is shown in FIG. The failure goals are actually generated in addition to those shown in the figure, but are omitted here. In the parsing unit 2, N is put after "This is"
P is predicted, but the NP cannot be constructed by "pen" alone, so a failure goal NP can be made at this position. In the non-sentence analysis unit 5, the rule B in FIG. 4 is successfully applied with the success goal NH and the failure goal NP at the position of "pen" as arguments,
“Error 3” is detected and sent to the message creating unit 7. In the message creating unit 7, the message “The article is missing before *” corresponding to “Error 3” is retrieved from the message dictionary 8 and the success goal NH is followed by “pe” in the surface layer.
n "is extracted and" the article is missing before "pen""
Message is synthesized and output to the user.

As a final example, "I went have been to the
Describes the action for the erroneous sentence "island."
At first I tried to write "I went to the island."
It is assumed that you tried to fix "nt" to "have been", but forgot to delete "went". An example of analysis in this case is shown in Fig. 6. In this case, "I went" All predictions at this position fail because an unexpected word comes immediately after that, so no success goals can be achieved after heve, and no clear error can be detected by combining success goals and failure goals. When the success goal has not been established in this way, the unit 5 creates a special identifier "detection failure" and sends the range in which the success identifier has been created to the message creation unit 7. In response to this, the message creating unit 7 issues a message "" have be
en to the island. "Unable to analyze" is synthesized and output. In this way, even when a clear error feature cannot be extracted, information about the error position can be shown to the user.

Next, the operation of the priority processing section 9 will be described. In the syntax check device, if the parsing rule or the ambiguity of the part of speech or attribute for one word increases, the number of combinations of positions to which the rule should be applied increases, and the processing time increases.
In order to deal with this, the processing can be speeded up by adding the priority processing unit 9. This speedup is
The property that "a large subtree has many successful predictions in the process of creating the tree, and the structure is likely to be intended by the user", and "the large subtree constitutes it" By using the property of the syntax tree "inherit the important attributes of a small subtree", the subtree is prioritized so that the error position is reached as soon as possible.
This is to reduce the average processing time.

First, the success goals and failure goals output from the syntax analysis section 2 are once input to the priority processing section 9. The priority processing unit performs grouping on the input success goals according to the start position and end position in the sentence. Success goals with the same start position and end position belong to the same group. A "column" is formed by using those whose positions are adjacent to each other in the group. The fact that the group A and the group B are adjacent to each other means that the end position of the group A and the start position of the group B are the same.
A "column" is created by several adjacent groups. The column is created so that there is no group whose start position is the beginning of the sentence and whose start position is the end position. Anything that violates this is not counted as a row. Hereinafter, the number of groups belonging to a row will be referred to as "priority", and the smaller number will be referred to as "higher priority". The priority processing unit 9 first extracts only the success goals belonging to the group forming the column having the highest priority. Only the highest priority column needs to be created here, but for this, only "groups that do not have other groups with ranges that completely include their range" are extracted, It is advisable to combine only the extracted groups to form a row. All of the highest priorities are included in this column. If there are multiple columns with the highest priority, the success goals are extracted from all of them so that they do not overlap. Moreover, only the failure goal whose start position matches any one of the start positions of the groups in the extracted column is extracted. It should be noted that which success goal and failure goal are extracted is stored in the priority processing unit 9. The extracted success goal and failure goal are sent to the non-sentence analysis unit 5. The non-sentence analysis unit 5 receives this as input and performs the same processing as above. If an error is detected, a message is output to the user and the processing relating to this sentence ends. In this case, the priority processing unit 9 also ends the processing relating to the sentence. When the contradiction is not detected and ends, the priority processing unit 9 is notified of the contradiction. In this case, the success goal and the failure goal are stored in the non-sentence analysis unit 5. The priority processing unit 9 receives this notification and continues the processing. First, it is checked whether or not there is a column having, as a priority, the number obtained by adding 1 to the previous priority value. To create such a column, let n be the previous priority, collect all columns that have priority n + 1 among the columns that were created previously, and It is sufficient to generate all the columns obtained by dividing any of the groups forming the column into two. If you can create a column of this priority in this way, you can extract the groups that have not been selected previously from the groups included in that column, and the success goals included in those groups and these groups. The failure goal having the same start position is sent to the non-sentence analysis unit 5. The non-sentence analysis unit 5 applies a rule by combining a newly input goal and a previously existing goal. Here, the rule is applied only when the newly entered goal is included as one of the arguments of the application rule. This is to prevent the rule from being applied with exactly the same arguments as before. Thereafter, similarly, the processing is continued until the contradiction is detected while decreasing the priority one by one. If no contradiction is detected, the processing may be terminated at a specific relative priority (relative to the first priority).

In FIG. 7A, "The stars are saw at nigh
The result of analysis of the erroneous sentence "t." is shown. The operation of the priority processing unit 9 will be described by taking this as an example. This sentence becomes a correct sentence if the past participle form "seen" is used for saw. 8) shows that the success goals of () are divided into groups, and in FIG.8, only the groups a, d and e are "there are no other groups that completely include their own range". a and d,
Since "d" and "e" are adjacent to each other and "a" is the start position of the sentence and there is no group whose start position is the end position (sentence) of "e", "ade" is the highest priority column 1
Become one Since there is no combination of a, d, and e that satisfies the condition of "column" except "ade", this is the only column of "priority 3". As described above, FIG.
Success goals and failure goals as shown in (b) are first selected and sent to the non-sentence analysis unit 5. In the non-sentence analysis unit 5, "ar
The rule is applied between the BE composed of e "and the VP composed of" saw at night ", and an error is determined because the attribute of the VP is not a past participle. Since an error was determined, this sentence is described here. In this way, it is possible to narrow down the goals to be applied in the non-sentence analysis unit.

As described above, according to the above-described embodiment, the target sentence is first parsed so that the partial tree generated as an intermediate result remains even if the sentence as a whole fails. Holds information about partial trees in and determines why those trees could not grow anymore by referring to the attributes of that tree and the attributes of other trees around it. Specifically, various kinds of errors are detected by preparing a rule in which errors that are easy for humans to make are described by the relative positions of trees and the attributes of trees. Since the error search is performed in units of subtrees, the pattern disorder is absorbed by the diversity of the subtree structure, and many errors can be detected with fewer rules compared to word-level matching, greatly improving accuracy. be able to.

Further, even if the registered error rule is not hit, it can be understood that the error has been included somewhere due to the failure of the parsing. It is possible to provide the user with information regarding the position where an error has occurred by checking whether or not it is distributed in the range.

Furthermore, since the rules can be created based on human-prone mistakes, the search space is much smaller than when considering all the substitutions, omissions, and additions, and the practical amount There is an advantage that it can correspond to the dictionary and syntax rules of.

[0026]

As described above, according to the present invention, it is possible to point out an error without being influenced by the collapse of a pattern such as a decoration, and to point out a type of error that cannot be handled conventionally, such as a missing word. Further, even if the error detection fails, information about the position of the error can be output, and the syntax check can be executed quickly for a large number of rules and vocabularies.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a syntax check device according to an embodiment of the present invention.

[Fig. 2] (a) Parsing diagram showing a partial syntax tree of the input sentence "He drank a cold water." (B) Parsing diagram showing success and failure goals of the input sentence "He drank a cold water."

FIG. 3 is a rule structure diagram showing an example of parsing rules.

FIG. 4 is a rule structure diagram showing an example of a non-sentence analysis rule.

[Figure 5] Parsing diagram showing the success and failure goals of the input sentence "This is pen."

[Figure 6] Parsing diagram showing the success and failure goals of the input sentence "I went have been to the island .."

[Fig.7] (a) Input sentence "The stars are saw at night."
Parsing diagram showing the success and failure goals of (b) Parsing diagram showing the selected success and failure goals of the input sentence "The stars are saw at night."

[Fig. 8] A diagram showing groups of success goals of the input sentence "The stars are saw at night." And their inclusion relationships.

[Explanation of symbols]

 1 sentence cut-out part 2 syntactic analysis part 3 syntactic analysis rule 4 word dictionary 5 non-sentence analysis part 6 non-sentence analysis rule 7 message creation part 8 message dictionary 9 priority processing part

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Hideko Mori 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd.

Claims (2)

[Claims] 1. A sentence segmentation unit for segmenting an input document one sentence at a time, and parsing the segmented input sentence using a parsing rule. When the input sentence has an error and the syntactic analysis fails, A parser that outputs the subtree up to that point, a non-sentence analyzer that detects an error in the input sentence using the non-sentence analysis rules from the subtree that the parser outputs, and the detected error A syntax check device having a message creating section for outputting a message. 2. A priority given to a non-sentence analysis unit by giving a high priority to a substructure not included in other subtrees from a set of substructures forming a subtree created by the syntactic analysis unit. The syntax check device according to claim 1, further comprising a degree processing unit.