JP2012185790A - Dependency analysis support device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a dependency analysis support device capable of accurately and efficiently determining dependency of a long sentence through person's interaction.SOLUTION: The dependency analysis support device is characterized in performing correct dependency analysis by repeating the steps of dividing the sentence with punctuations, carrying out morphological analysis and syntactic analysis on divided parts, visually confirming dependency analysis results of the respective parts, and selecting another dependency rule from a previously prepared rule set if a wrong part is found; and in performing correct dependency analysis between parts by repeating the steps of performing dependency analysis between divided parts when the dependency analysis on the respective parts is completed, visually confirming dependency analysis results between the parts, and selecting another dependency rule from the previously prepared rule set if a wrong part is found.

Description

  The present invention relates to a dependency analysis support apparatus suitable for dependency analysis of long sentences in natural language.

  In a system related to natural language applications such as machine translation and natural language understanding, it is necessary to correctly analyze the input natural language.

  Usually, Japanese analysis is performed by morphological analysis in which word levels are cut out and part of speech is determined and grouped into phrases, and syntactic analysis is performed to determine dependency relationships between phrases. When these processes are completed, a semantic analysis may be performed to further extract a semantic relationship between clauses. The number of combinations of dependency becomes the factorial of the number of phrases when roughly estimated as the number of phrases increases. Of course, because there are constraints, it actually decreases, but if the number of clauses exceeds 10, it is said that it will be very difficult. In the current Japanese analysis, it is an important issue to raise the accuracy of dependency analysis to a practical level.

  So far, many attempts have been made towards that goal. In [Patent Document 1], focusing on a predicate in a sentence, a pattern that can be grouped around the predicate is extracted using a rule, analyzed by the group, and subsequently a dependency between groups is used using the rule. Is going. [Patent Document 2] is a method of patterning a semantic description unit of a sentence and structuring the sentence based on the pattern. One factor that causes a long sentence is the case where a parallel phrase is included. For example, [Non-Patent Document 1] has developed a method of extracting a parallel phrase by focusing on the similarity of the arrangement of words in a sentence. In [Patent Document 3], a claim description pattern is extracted from a claim having many long sentences, and the claim is analyzed based on the extracted pattern. [Patent Document 4] discloses a method of dividing a long sentence as pre-editing of machine translation.

  JP 07-056919 A   Japanese Patent Laid-Open No. 06-295308   Japanese Patent No. 3908261   JP-A-08-87504

  "Syntactic analysis of long Japanese sentences based on parallel structure detection", Ikuo Kurohashi, Atsushi Nagao, Natural Language Processing, Vol. 1, No. 1 1, pp. 35-58, 1994

  The conventional dependency analysis described above has the following problems. In the method of [Patent Document 1], it is necessary to cut out a predicate from a predicate pattern accumulated in advance, and there is no guarantee that this can be performed correctly, so there is no guarantee that a highly accurate analysis can be obtained. [Patent Document 2] is also difficult to collect context patterns that cover natural language sentences, and does not show that it can be used in general sentences. The improvement of the analysis accuracy of long sentences by parallel phrases in [Non-Patent Document 1] remains a problem because there are long sentences in general Japanese sentences that cannot be solved only by the analysis of parallel phrases. [Patent Document 3] is for a limited sentence such as a claim in a patent, and its application to general Japanese long sentence analysis is not guaranteed. [Patent Document 4] specifies a division pattern in order to find a division point of a long sentence, but since the identification of the division point is processing for a long sentence in the first place, the selection man-hour for dependency is not reduced. I have to say that it is difficult.

  In order to achieve the above object, the present invention divides a long sentence into a plurality of short units called “parts” by punctuation marks, analyzes the dependency in each part, and then analyzes the dependency between parts. A dependency analysis support means consisting of

  In order to increase the analysis accuracy, the analysis results within the part and the analysis result between the parts are presented in a simple manner such as a syntax tree, and the user visually confirms it. If there is an incorrect dependency, By specifying a clause, other candidate clauses that become recipient clauses of the clause are presented, or by selecting another rule set, the recipient clause specified by the rule is presented, and either presentation But if it's wrong, it provides repeated means to select new candidates. It is also possible to provide means for presenting the rules to be selected according to priority.

  The dependency analysis support device of the present invention can improve the analysis accuracy of a part of a long sentence by dividing the sentence into short parts, and guarantees an accurate analysis between the divided part and the part. Overall, the dependency analysis accuracy of long sentences can be raised to a practical level.

  In addition, when visual confirmation and correction are required, the recipient's clause is selected, the dependency rule is selected, or the sentence is edited, but learning the application of the rule reduces the number of dependency rules selected. It can be expected that the load of manual correction work will gradually decrease.

  Overall structure of dependency analysis support device   Corresponding schematic diagram of example of partial analysis result of sentence and dependency analysis   Schematic step diagram for dependency analysis of sentences   List of operation keys for dependency analysis support   Input example diagrams for dependency analysis support   Example sentence division result diagram   Dependency analysis result diagram in the part   Dependency analysis result diagram between parts   Example of dependency rules for part matching the main phrase   Figure after selection of dependency rules for other parts   Dependency analysis result diagram obtained with the selected rule   Interim dependency analysis result screen view   Analysis result screen diagram of dependency analysis support end for one sentence   Partial dependency modification explanation diagram   Non-sentence analysis example diagram   Example of set of dependency analysis rules in subpart   Example of dependency analysis rule set

  An embodiment of claim 1 will be described.

  FIG. 1 shows the hardware of the dependency analysis support apparatus in this embodiment, the data in the storage device, and the program configuration. Reference numeral 1 denotes a computer including a central processing unit and a storage device. 2 is a display device, 3 is a keyboard, and 4 is a pointing device such as a mouse. In the figure, 1, 2, 3, and 4 are drawn independently, but a touch panel type integrated computer may be used. Reference numeral 10 denotes main components of the storage device. The input buffer 11 stores an input sentence, the output buffer 12 stores an analysis result, and a sentence division unit 13 that is a program for dividing a sentence into parts. A morpheme analysis / syntax analysis unit 14 that is a program for executing morphological analysis and syntax analysis, an intra-partial dependency analysis unit 15 that is a program for performing dependency analysis of a specified phrase in the part, In-part dependency analysis rule set 16 used when performing dependency analysis, part-part dependency analysis unit 17 which is a program for performing dependency analysis between parts, and when performing dependency analysis between the parts It includes a dependency analysis rule set 18 for parts to be used and a dependency rule management part 19 which is a program for managing addition, update and priority of rules.

  16 illustrates a part of the partial dependency analysis rule set 16, and FIG. 17 illustrates a part of the partial dependency analysis rule set 18. The format of both configurations is the same, and consists of a dependency side 50, 53, a reception side 51, 54, and a dependency relationship 52, 55. If the dependency-side clause and the receiving-side clause match the rule, a dependency relationship is established between the two clauses. ‘IR # i’ in FIG. 16 and ‘OR # j’ in FIG. 17 are internal indexes. “<>” And “[]” in the rule are markers indicating parts of speech and symbols. The character string is described as it is. The expression format of this rule and the method of describing the rule are not limited to this. It is also possible to add information for managing rules by expanding the columns of the rules table.

  FIG. 2 is a conceptual diagram of a data structure in which the sentence S # n is divided into parts each including a word string divided by punctuation marks. The sentence is surrounded by '{}', and Sn # i (in this example, i takes a value from 1 to 3) indicates the divided i-th portion. In order to be a part, the segmented word string is limited to those satisfying the following conditions. That is, the non-intersection principle is applied to the dependency, and one is that the analysis within the part is closed. In other words, there is only one dependency relationship from the part to the external part. The other is to allow a plurality of dependency relationships from the external part to the words in the part. The case where there is no dependency from outside is also allowed. In addition, the main clause in the part includes what is called a function word or an attached word. The main clause is a clause that is not related to other clauses in the part. In Japanese, the rightmost part of the part is the main clause. Therefore, it can be said that the first condition is a clause in which only the main clause modifies the external part. Further, the partial dependency analysis 20 and the partial dependency analysis 21 shown in FIG. 2 will be described later. Each Sn # i partial dependency analysis 22, 23, 24 is surrounded by '{}' at the end of the arrow, but this depends on morphological analysis and syntax analysis. The analysis result is recorded, but is not limited to a specific expression. In dependency analysis, modification of dependency is often expressed in a tree structure. The dependency information of the main clause of the relevant part is described in the portion enclosed by “[]” at the tip of the arrow from the dependency analysis 25 and 26 for each Sn # i portion.

  FIG. 3 shows the general steps of dependency analysis support. It is divided into a process of analyzing the dependency within the part and a process of analyzing the dependency between the parts. First, it is assumed that a sentence to be processed exists in the input buffer 11. There may be multiple sentences. In the division into parts 31, each sentence is divided into parts with punctuation marks. Then, the nth input sentence S # n is divided into parts Sn # 1, Sn # 2, Sn # 3, and the like. In the morpheme analysis / syntax analysis 32 in the part, each Sn # i is processed by the morpheme analysis / syntax analysis unit 14 to output a dependency structure. In the confirmation of the analysis result in the part and correction 33 if necessary, the analysis result of the part is visually confirmed, and when a grammatical injustice is detected, the mode is shifted to a sentence editing mode. If there is no injustice, it is visually confirmed in order from Sn # 1 whether each part has been correctly analyzed. If an error is found in the dependency, a clause on the dependency side is specified and another phrase candidate on the reception side is presented. When the analysis in the portion is completed, the analysis result between the portions is confirmed and, if necessary, correction 34, the dependency analysis between the portions is performed for each main phrase, and the dependency check is performed in the same manner as in the portion. To redo the dependency and edit the sentence.

FIG. 4 shows a list of keys to be pressed in the dependency analysis support operation. In this figure, the taskbar format and the task menu format are assumed, but they may be assigned to the keyboard. Provides an overview of key functions. [Sentence Division] divides a sentence. [Intra-part analysis] executes morpheme analysis / syntactic analysis in the divided part. [Partial analysis] analyzes the dependency between parts. [Dependency] performs dependency on the specified phrase. [Next candidate] searches for a new candidate for the recipient phrase. [Intra-part advancement] is to forcibly move the receiving clause one by one, and the movement range is limited to the part. [Reverse in part] restores the receiving phrase by one operation in the reverse operation of [Inward in part]. [Progress between parts] moves the receiving phrase to the next main phrase in the dependency analysis between the parts. [Reverse part part] returns the main sentence of the receiving side to the main part of the previous part by the reverse operation of [Progress part part]. [Edit mode] interrupts the dependency analysis and moves to the sentence edit mode. [Confirm] displays the currently analyzed dependency relationship.
The above keys are not limited to these, and a read / write of a file can be added for an interface, or a new composite key can be added by combining key functions.

  Steps 31 to 34 in FIG. 3 will be described in detail with reference to FIG. 5 to FIG. First, in the division into parts 31, [Sentence Division] is pressed to divide the text in the input buffer into parts by punctuation marks. FIG. 6 shows a screen of the division result of the input example sentence of FIG. Next, in the morphological analysis / syntactic analysis 32 in the part, the morpheme analysis / syntax analysis unit 14 is executed on the divided part by pressing [Internal analysis]. If [OK] is pressed here, the syntax tree of the partial analysis result shown in FIG. 7 is displayed. This display method is not limited to this. In confirmation of the analysis result in the portion and correction 33 if necessary, first, the analysis result in the portion is visually confirmed. Since all parts of this example sentence have been correctly analyzed, the process proceeds. In confirmation of the analysis result between the parts and correction 34 if necessary, first, when [Partial analysis] is pressed, the part-part dependency analysis unit 17 is activated, and the part-part dependency analysis rule set 18 is used to determine the dependency between parts. Analysis is performed. When [OK] is pressed here, the syntax tree of the sentence analysis result shown in FIG. 8 is displayed. The display of the analysis result is visually confirmed, and a failure is found by the dependency between the Sn # 2 and Sn # 3 portions. The correction of dependency failure is shown below. FIG. 8 shows a part of a rule set that matches when the cursor is placed on the main phrase “Later,” of Sn # 2 and [Dependency] is pressed. The underlined portion OR # 8 is the currently selected rule, but this correct answer is related to “decline,” so that [Dependency] is pressed to advance to the rule OR # 9 that matches it. As a result, “After, → Ss3 falls, [Time After]” shown in FIG. 10 is obtained. If [OK] is pressed here, the tree structure diagram 11 is displayed.

  When the cursor is placed on the main phrase “falling down” of another failure location Sn # 3 in FIG. 8, the cursor is placed, and [Dependency] is pressed, a part of the matching rule set is displayed. These are OR # 1, OR # 2, and OR # 3 in FIG. OR # 1 is the currently selected rule, but the correct dependency is related to “disrupted”, so [Dependency] is pressed twice to select the rule OR # 3 that matches it. As a result, “fall down, and Ss4 collapse [continuous use stop]” is obtained. If [OK] is pressed here, the tree structure analyzed so far is displayed in FIG. This eliminates the two dependency failures and gives the correct dependency. This result is stored in the output buffer 12.

  A specific example of correction in the case where the dependency analysis in the part that did not appear in the above example fails will be shown. FIG. 14 shows an example in which the dependency analysis in the part is wrong, and there are two ways to correct 41. One is to press [next candidate candidate] and find the next clause that matches the same rule. In this case, “economic” modifies the noun “Iki”, but [next Pressing “Candidates” will be related to the next noun “recovery”. Alternatively, the correct receiving phrase can be reached by [intra-part advance] or [in-part reverse]. The other is to change the selection of the rule. In this case, when [Dependency] is pressed, a part of the rule set IR # 2 and IR # 3 matching the clause on the side of the dependency is shown at 42. Therefore, by selecting IR # 3, “recovery” is related in the same way. In any method, when [OK] is pressed, it becomes correct as shown in 43.

  Another example of failure is when a division occurs that violates the condition of the part, and FIG. 15 shows an example thereof. An example of the part "The safety function described here is also in the evaluation of the reliability of the information system" When analyzing 44, 45 modifies 46, but 45 modifies the external part. Since the analysis of violates the condition that it is closed, it becomes a grammatical error, and by pressing [Edit Mode] to edit this sentence, the dependency is interrupted and the sentence is edited.

  Embodiments of claims 2 and 3 will be described.

  As the dependency analysis processing, rules are applied to the two dependency analyzes within and between the portions using FIG. 8 and FIG. 9, respectively. Both rule sets have rules that match the clauses on the dependency side of the dependency. It is typological in applying. By accumulating dependency rules within and between parts using many example sentences, the number of dependency rules that match one phrase increases, but priority is applied to the order in which dependency rules are presented. To avoid this problem. For example, a history for each user that gives priority to frequency data and statistical data accumulated by past dependency analysis and dependency analysis of texts in a specific field, which presents the most recently used rule, is used. Since the selection of the priority depends on the subject field of dependency analysis and the vocabulary / grammar category, the dependency rule management unit 19 manages various priorities for each rule so that the user can specify the priority. This makes it possible to learn the order in which rules are presented, and reduces the burden of dependency correction work by visual work.

  In the first and second embodiments, Japanese sentences have been described as examples. However, this is not necessarily limited to Japanese, and any natural language that can perform dependency analysis can be applied in common.

  Since the present invention can provide a natural language processing system that guarantees highly accurate analysis results in long sentences, it leads to the practical application of a wide range of natural language processing applications such as machine translation and natural language understanding systems. Also, because sentences are analyzed by dividing them into short units, grammatical errors can be easily found and can be used for supporting proofreading.

DESCRIPTION OF SYMBOLS 1 Computer including central processing unit and storage device 2 Display device 3 Keyboard 4 Pointing device 10 Main components of storage device 11 Input buffer 12 Output buffer 13 Division into sentence parts 14 Morphological analysis / syntax analysis part 15 Partial relation Receiving analysis unit 16 Intra-part dependency analysis rule set 17 In-part dependency analysis unit 18 In-part dependency analysis rule set 19 Dependency analysis rule management unit 20 Headline for explanation of in-part dependency analysis 21 Partial dependency analysis Description headline 22 Sn # 1 partial dependency analysis 25 Sn # partial dependency analysis 31 Step to divide into 31 parts 32 Step to perform morphological analysis / syntactic analysis in part 33 Confirmation of analysis result in part and correction if necessary Step 34 Dependency analysis between parts, confirmation of analysis results and necessary 40 Dependent correction explanation input example sentence 41 Dependency correction explanation input example sentence analysis result 42 Dependency dependency rule selection screen 43 Dependency dependency analysis result screen 44 selection sentence non-sentence sentence example 50 part Dependent clause description column of internal dependency analysis rule 51 Dependent clause description column of partial dependency analysis rule 52 Dependency relationship column of partial dependency analysis rule 53 Interdependent dependency analysis rule dependency column Description column 54 Description column of the receiving clause of the dependency analysis rule for 55 copies 55 Dependency relationship column of the dependency analysis rule for 55 copies

  Steps 31 to 34 in FIG. 3 will be described in detail with reference to FIG. 5 to FIG. First, in the division into parts 31, [Sentence Division] is pressed to divide the text in the input buffer 11 into parts by punctuation marks. FIG. 6 shows a screen of the division result of the input example sentence of FIG. Next, in the morphological analysis / syntactic analysis 32 in the part, the morpheme analysis / syntax analysis unit 14 is executed on the divided part by pressing [Internal analysis]. If [OK] is pressed here, the syntax tree of the partial analysis result shown in FIG. 7 is displayed. This display method is not limited to this. In confirmation of the analysis result in the portion and correction 33 if necessary, first, the analysis result in the portion is visually confirmed. Since all parts of this example sentence have been correctly analyzed, the process proceeds. In confirmation of the analysis result between the parts and correction 34 if necessary, first, when [Partial analysis] is pressed, the part-part dependency analysis unit 17 is activated, and the part-part dependency analysis rule set 18 is used to determine the dependency between parts. Analysis is performed. When [OK] is pressed here, the syntax tree of the sentence analysis result shown in FIG. 8 is displayed. The display of the analysis result is visually confirmed, and a failure is found by the dependency between the Sn # 2 and Sn # 3 portions. The correction of dependency failure is shown below. FIG. 9 shows a part of a rule set that matches when the cursor is placed on the main phrase “Later,” of Sn # 2 and [Dependency] is pressed. The underlined portion OR # 8 is the currently selected rule, but since this correct answer is related to “decline,” press [Dependency] to advance to the rule OR # 9 of FIG. As a result, “After, → Sn # 3 drop, [Time After]” shown in FIG. 11 is obtained. If [OK] is pressed here, a tree structure diagram 12 is displayed.

  When the cursor is placed on the main phrase “falling down” of another failure location Sn # 3 in FIG. 8 and [Dependency] is pressed, a part of the matching rule set is displayed. These are OR # 1, OR # 2, and OR # 3 in FIG. OR # 1 is the currently selected rule, but the correct dependency is related to “disrupt”, so [Dependency] is pressed twice to select the rule OR # 3 that matches it. As a result, “falling down and then Sn # 4 collapse [continuous use stop]” is obtained. If [OK] is pressed here, the tree structure analyzed so far is displayed in FIG. This eliminates the two dependency failures and gives the correct dependency. This result is stored in the output buffer 12.

  As the dependency analysis processing, two dependency analyzes within and between the portions are applied using the rules shown in FIGS. 16 and 17, respectively. Both rule sets match the clauses on the dependency side for the dependency. It is typified by applying. By accumulating dependency rules within and between parts using many example sentences, the number of dependency rules that match one phrase increases, but priority is applied to the order in which dependency rules are presented. To avoid this problem. For example, a history for each user that gives priority to frequency data and statistical data accumulated by past dependency analysis and dependency analysis of texts in a specific field, which presents the most recently used rule, is used. Since the selection of the priority depends on the subject field of dependency analysis and the vocabulary / grammar category, the dependency rule management unit 19 manages various priorities for each rule so that the user can specify the priority. This makes it possible to learn the order in which rules are presented, and reduces the burden of dependency correction work by visual work.

Claims (3)

  Divide the sentence with punctuation marks, perform morphological analysis and dependency analysis on the divided parts, visually check the result of dependency analysis of each part, and if there is a clause with incorrect dependency analysis, The user decides whether to correct the sentence or to edit the sentence. If the sentence is edited, the dependency analysis is interrupted, the editing mode for editing the sentence is entered, and if the dependency is corrected Search for the next dependency candidate that matches the same rule that you applied, or let the user select another dependency rule from a set of prepared rules, and the dependency is still wrong Repeat the selection of another dependency rule, or specify the receiver clause forcibly at any stage during the selection, determine the correct dependency, and correct the dependency analysis of each part. End If the user determines that the dependency analysis has been performed, the dependency analysis between the divided parts is performed using a prepared rule set, and the dependency analysis result between the parts is visually confirmed, and the dependency analysis is incorrect. If there is a clause, the user decides whether to edit the dependency or the sentence, and if editing a sentence, interrupt dependency analysis and enter edit mode to edit the sentence. When correcting a receipt, search for the next dependency candidate that matches the same rule that has been applied recently, or present another dependency rule to the user from a set of pre-created rules. If the dependency is wrong, repeat the selection of another dependency rule, or specify the receiver clause forcibly at any stage during the selection, and determine the correct dependency. of Minutes of dependency analysis dependency analysis support apparatus, characterized in that the support to be correct.   2. The dependency analysis support apparatus according to claim 1, wherein in selecting a dependency rule for the divided portion, a rule is presented in consideration of a priority.   2. The dependency analysis support apparatus according to claim 1, wherein in selecting a dependency rule for the divided parts, a rule is presented in consideration of a priority.

Images