JP2011198075A - Natural language analysis device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a natural language analysis device, method, and program, for determining modification relations for each character, without morphologically analyzing a sentence to be analyzed to the end.SOLUTION: A natural language analysis device 10 acquires characters constituting the sentence to be analyzed, in units of characters, and determines dependency relations for each acquired character. Then the natural language analysis device 10 stacks dependency-undefined characters in a dependency-undefined character stack 107 in a process of determining dependency relations for each character in order from the first character in the sentence and, after the modification of the characters is determined by dependency determination, determines dependency relations of the characters stacked in the dependency-undefined character stack 107 to determine modifications of characters.

Description

  The present invention relates to a natural language analysis apparatus, method, and program.

  2. Description of the Related Art Conventionally, in Japanese sentence analysis, it has been the mainstream to summarize word (morpheme) strings that have been subjected to morphological analysis as phrases and to express the analysis results as dependency relationships (dependencies) between the phrases. In this case, the dependency analysis method needs to perform morphological analysis by scanning from the beginning of the sentence to the end of the sentence in advance, and after returning to the beginning of the sentence after the morpheme analysis, it scans from the beginning of the sentence to the end of the sentence, collects the sentences, After summarizing, we return to the beginning of the sentence and scan from the beginning to the end of the sentence for dependency analysis.

  As a technique for analyzing the dependency relationship between Japanese phrases as described above, Japanese Patent Application Laid-Open No. H10-228867 is known in which phrase grouping and dependency analysis are performed in one scan.

  The technology disclosed in Patent Document 1 inputs a sentence to be analyzed that is decomposed into morphemes, and determines the dependency between each word (morpheme) in the morpheme string (determination between the relation source and its relation destination, relation relation) The process of deciding the two types) is performed using the stack. Here, the determined relationship type also represents a segment break. Therefore, the technique disclosed in Patent Document 1 can perform the phrase grouping and dependency analysis at the same time by performing such a process of determining the dependency, and the phrase grouping and dependency analysis are independent. The module is not required, and the processing speed can be increased.

JP 2009-176062 A

  However, the technique disclosed in Patent Document 1 is a technique for analyzing a sentence whose sentence end is clear, and after analyzing the sentence to be analyzed into morphemes up to the end of the sentence and performing morphological analysis, It determines the dependency relationship between them. Therefore, in the technique disclosed in Patent Document 1, when the sentence end is unclear, the morphological analysis is not performed, and the dependency relationship between phrases cannot be determined.

  Therefore, there is a need for an apparatus that can determine the dependency relationship between clauses without decomposing sentences into morphemes until the end of the sentence and performing morphological analysis.

  An object of the present invention is to provide a natural language analysis apparatus, method, and program capable of determining a dependency relationship for each character without performing a morphological analysis of a sentence to be analyzed to the end of the sentence.

  The present invention provides the following solutions.

  (1) A natural language analysis apparatus for analyzing a natural language sentence, a character acquisition means for acquiring characters constituting the sentence to be analyzed in character units, and a character dependency analysis for determining a dependency relationship for each acquired character And the character dependency analyzing means stacks the characters whose dependency destinations are undetermined in the process of determining the dependency relationship for each character in order from the first character of the sentence to be analyzed. A natural language analyzing apparatus that determines a dependency of a character by determining a dependency relationship of characters accumulated in a stack after determining a character dependency destination by determining the relationship.

  According to the configuration of (1), the natural language analyzing apparatus according to the present invention acquires characters constituting the sentence to be analyzed in character units, and determines the dependency for each acquired character. In the process of determining the dependency relationship for each character in order from the first character of the sentence to be analyzed, the natural language analyzer stacks the characters whose dependency is not yet determined, and determines the character's dependency by determining the dependency. Is determined, the dependency of characters stored in the stack is determined to determine the dependency of the characters.

  Therefore, since the natural language analysis apparatus according to the present invention performs dependency analysis for each character while accumulating characters whose dependency is not yet determined in the stack, even if the sentence to be analyzed is not analyzed to the end of the sentence, the character is analyzed. The dependency relationship can be determined for each.

  (2) Dependency determination in the character dependency analysis unit is based on a grammar definition table in which types of determination results corresponding to character types of dependency sources and dependency destination candidates and character position relationships are associated with each other. The natural language analysis device according to (1).

  According to the configuration of (2), the natural language analyzing apparatus determines the dependency relationship for each character, and determines the type of determination result according to the character type of the dependency source and the dependency destination candidate and the relationship between the character positions. Is performed based on the grammar definition table associated with.

  Therefore, since the natural language analyzing apparatus determines the dependency relationship for each character based on the grammar definition table, it is possible to determine the dependency relationship for each character without performing morphological analysis of the sentence to be analyzed up to the end of the sentence. it can.

  (3) The natural language analysis apparatus according to (1) or (2), wherein the dependency relationship in the character dependency analysis unit is determined based on a grammatical rule machine-learned by SVM.

  According to the configuration of (3), the natural language analyzing apparatus performs the determination of the dependency for each character based on the grammatical rules machine-learned by SVM (Support Vector Machine). Therefore, the natural language analyzing apparatus can determine the dependency for each character by SVM.

  (4) A natural language analyzing apparatus is a natural language analyzing method for analyzing a natural language sentence, and a step of acquiring characters constituting a sentence to be analyzed in character units and a dependency relationship for each acquired character are determined. A character dependency analysis step, wherein the character dependency analysis step stacks a character whose dependency destination is undetermined in the process of determining the dependency for each character in order from the first character of the sentence to be analyzed. A natural language analysis method for determining a character dependency by determining a dependency relationship of a character accumulated in a stack after determining a character dependency destination by determining a dependency relationship.

  Therefore, in the natural language analysis method according to the present invention, the natural language analysis apparatus performs dependency analysis for each character while accumulating the characters whose dependency is undetermined in the stack. Even without analysis, the dependency relationship for each character can be determined.

  (5) A natural language analysis program for analyzing a natural language sentence, in which a computer obtains characters constituting the sentence to be analyzed in character units, and character dependency for determining a dependency for each acquired character. And the character dependency analysis step stacks characters whose dependency destinations are undetermined in the process of determining the dependency for each character in order from the first character of the sentence to be analyzed. A program for determining a character dependency by determining a dependency relationship of characters stored in a stack after determining a character dependency destination by determining a dependency relationship.

  Therefore, by introducing the natural language analysis program according to the present invention into the natural language analysis device and executing it, the natural language analysis device can perform dependency analysis for each character while accumulating the characters whose dependency relationship is undetermined in the stack. Therefore, the dependency relationship for each character can be determined without performing morphological analysis of the sentence to be analyzed until the end of the sentence.

  According to the present invention, in the natural language analysis process, the dependency relationship for each character can be determined without performing a morphological analysis of the sentence to be analyzed up to the end of the sentence.

It is a figure which shows the structural example of the natural language analyzer which concerns on one Embodiment of this invention. It is a figure which shows the structural example of the dependence relationship determination part at the time of using SVM in the natural language analyzer which concerns on one Embodiment of this invention. It is a figure which shows the example of the analysis character storage part in the natural language analyzer which concerns on one Embodiment of this invention. It is a figure which shows the example of the grammar definition table in the natural language analyzer which concerns on one Embodiment of this invention. It is a figure which shows the specific process example of the dependency analysis process part of the natural language analyzer which concerns on one Embodiment of this invention. FIG. 6 is a diagram illustrating a specific processing example of the dependency analysis processing unit following FIG. 5. FIG. 7 is a diagram illustrating a specific processing example of the dependency analysis processing unit following FIG. 6. FIG. 8 is a diagram illustrating a specific processing example of the dependency analysis processing unit following FIG. 7. FIG. 9 is a diagram illustrating a specific processing example of the dependency analysis processing unit following FIG. 8. FIG. 10 is a diagram illustrating a specific processing example of the dependency analysis processing unit following FIG. 9. FIG. 11 is a diagram illustrating a specific processing example of the dependency analysis processing unit following FIG. 10.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  This embodiment is applied to a computer and its peripheral devices. Each unit in the present embodiment is configured by hardware included in a computer and its peripheral devices, and software that controls the hardware.

  The hardware includes a storage unit, a communication device, a display device, and an input device in addition to a CPU (Central Processing Unit) as a control unit. Examples of the storage unit include a memory (RAM: Random Access Memory, ROM: Read Only Memory, etc.), a hard disk drive (HDD: Hard Disk Drive), and an optical disk (CD: Compact Disc, DVD: Digital Versatile Drive, etc.). It is done. Examples of the communication device include various wired and wireless interface devices. Examples of the display device include various displays such as a liquid crystal display and a plasma display. Examples of the input device include a keyboard and a pointing device (mouse, tracking ball, etc.).

  The software includes a computer program and data for controlling the hardware. The computer program and data are stored in the storage unit, and are appropriately executed and referenced by the control unit. The computer program and data can be distributed via a communication line, or can be recorded on a computer-readable medium such as a CD-ROM and distributed.

  FIG. 1 is a diagram illustrating a configuration example of a natural language analysis apparatus 10 according to an embodiment of the present invention. In the following, an example for Japanese will be described. However, the present invention can be similarly applied to any language that can be divided into phrases and has a dependency relationship between phrases.

  In FIG. 1, the natural language analysis apparatus 10 includes an analysis target character input unit 101 that acquires characters constituting an analysis target sentence in units of characters, and characters acquired from the characters input by the analysis target character input unit 101. And a dependency analysis processing unit 102 for determining the dependency of each. The dependency analysis processing unit 102 determines the dependency for each character in order from the first character of the sentence to be analyzed, and stores the analysis result in the analysis character storage unit 108. The data structure of the analysis character storage unit 108 will be described with reference to FIG.

  The natural language analyzing apparatus 10 also includes a dependency source character ID storage area 105, a dependency destination candidate character ID storage area 106, a dependency destination pending stack 107, and a dependency stack used as a work data area by the dependency analysis processing unit 102. A dependency determination unit 103 used for determining dependency between characters in the processing of the receiving analysis processing unit 102, and an analysis result output unit 104 that outputs an analysis result are provided.

  The dependence source character ID storage area 105 holds a character ID that becomes a dependence source in the processing process, and the dependence destination candidate character ID storage area 106 holds a character ID that becomes a dependence destination candidate in the processing process. In 107, the dependence source character ID for which the dependence destination has not been determined is held in last-in first-out (LIFO).

  In addition, the dependency relationship determination unit 103 has a grammar definition table 110. Then, the dependency analysis processing unit 102 determines the dependency relationship of the characters accumulated in the dependency destination pending stack 107 after the dependency determination unit 103 determines the character dependency destination. Here, the grammar definition table 110 holds the relationship between the dependence source and the dependence destination as a data structure in the IF-THEN format or the like, and also includes the case where it is held as a model generated by machine learning by SVM. An example of the grammar definition table 110 will be described later with reference to FIG.

  FIG. 2 is a diagram illustrating a configuration example of the dependency relationship determination unit 103 when SVM is used in the natural language analysis apparatus 10 according to an embodiment of the present invention.

  In FIG. 2, a dependency relationship determination unit 103 is a teacher data input unit 1031 that inputs teacher data for machine learning, and a machine that performs machine learning by SVM based on the teacher data input by the teacher data input unit 1031. A learning unit 1032 and a grammar rule storage unit (for example, corresponding to a grammar definition table 110 described later in FIG. 4) 1033 holding a model generated by machine learning are provided.

  In addition, the dependency relationship determination unit 103 receives a dependency relationship determination request receiving unit 1034 that receives a dependency source character ID and a dependency destination candidate character ID from the dependency analysis processing unit 102 (FIG. 1) as arguments. Based on the dependency determination request received by the dependency determination request receiving unit 1034, the dependency determination is performed using the grammar rule storage unit 1033, and the determination result is returned to the dependency analysis processing unit 102. It has.

  FIG. 3 is a diagram illustrating an example of the analysis character storage unit 108 in the natural language analysis apparatus 10 according to an embodiment of the present invention. The analysis character storage unit 108 includes, for each character constituting one sentence, a “character ID” field for identifying the character, a “character code” field for storing the character, and a character attribute (for example, character type ( For example, a “character type” field indicating a character type that is hiragana, katakana, alphanumeric characters, kanji, symbols, etc., a “dependence destination” field indicating a dependency destination character ID, and a dependency type between the dependency destination characters And a “type” field indicating In the initial state of character analysis, the “character ID” field is filled, and the “character code”, “character type”, “dependence destination”, and “type” fields are blank. The example shown in FIG. 3 indicates that the dependency analysis by the natural language analysis apparatus 10 is completed, the character ID of the character dependency destination is stored in the dependency destination, the word break is set, and the phrase break is set. It is an example. Further, the example shown in FIG. 3 is an example in which an arrow 221 indicates a character destination by a character ID, an arrow 211 indicates a word break, and an arrow 212 indicates a phrase break.

  FIG. 4 is a diagram showing an example of the grammar definition table 110 in the natural language analysis apparatus 10 according to an embodiment of the present invention. The grammar definition table 110 associates conditions such as the character types and positional relationships of the dependence source and dependence destination candidates (relationship between the characters in the characters constituting the sentence) with the type of the determination result.

  Here, the determination result type “W” is a type indicating that the dependence source character and the dependence destination candidate character constitute a continuous character in the same word. The determination result type “B” is a type indicating that the dependence source character and the dependence destination candidate character constitute a word continuation within the same phrase. The determination result type “D” is a type indicating that “the clause whose dependent source character ends and the clause whose dependent destination candidate character ends are in a dependency relationship between clauses”. The determination result type “O” is a type indicating “no dependency”. The determination result type “E” is a type indicating “end of sentence”.

  Furthermore, as an example of the above-mentioned type of determination condition, for example, in the case where “the character type of the dependency source character and the character type of the dependency destination candidate character are the same AND the dependency destination candidate character immediately after the dependency source character”, the dependency source character is W is determined. In addition, in the case of “dependence source character and dependency destination candidate character are different in character type AND dependency destination candidate character immediately after dependency source character”, the dependency source character is determined to be B. In addition, when the “dependent source character is“ NO ”AND the dependent destination candidate character is a character used as a case particle”, the dependent source character is determined to be D. Otherwise, it is determined as O. If the sentence ends, it is determined as E. Here, such a condition is an example, and the present invention is not limited to this. Such conditions are created by machine learning and stored in the grammar definition table 110.

  Here, the analysis processing of the natural language analysis apparatus 10 according to an embodiment of the present invention is shown in a high-level language.

procedure analysis (m, h, t)
var s: stack
begin
Push (-1, s)
m [0] = get_token ()
Push (0, s)
m [1] = get_token ()
i = 1
while (m [i]! = EOS) do begin
j = Pop (s)
m [i + 1] = get_token ()
while (j! =-1 && (Dep (j, i, m, t) || (m [i + 1] == EOS)) do begin
h [j] = i
j = Pop (s)
end
Push (j, s)
Push (i, s)
++ i;
end
j = Pop (s)
h [j] = i
t [j] = “E”
end

  In the above-described analysis processing, m is an array of morphemes (characters), h is an array that stores the relations, and t is an array that stores the type of the determination result. Push (value, s) is a function that pushes a value onto the stack, get_token () is a function that acquires one character, Pop (s) is a function that pops a value from the stack, and Dep (j , I, m, t) is a function for determining whether the j-th character depends on the i-th character.

  That is, the Dep function corresponds to the dependency determination unit 103, accepts a dependency determination request using the jth character (dependence source character ID) and the i th character (dependence destination candidate character ID) as arguments, and The grammar definition table 110 that stores the determination conditions for the original character and the dependence destination candidate character is used to determine and store the character type and various attributes, and the dependency relationship is determined. The Dep function returns “True” when it is determined that the j-th character is dependent (related) on the i-th character, and “False” when it is determined that the j-th character is not dependent (not related).

  In addition, although the said example showed the character type which a Dep function judges for simplicity, it is not restricted to this. Specifically, other examples of character types include characters that can be Chinese numerals (symbols ○, one or two of the Chinese characters, 壱, 弐, one hundred, one thousand, ten thousand, etc.) Is considered a symbol, but it should be treated in the same way as kana (a dakuten, semi-dakuten, long-sound symbol, etc.), or a character that has a high probability of being used as part of a name (child, Yu, Hiroshi, Akira, ...) etc. are also included. Further, as an example of the determination of the Dep function, when the type of character to be analyzed is a character that can be a Chinese numeral, the Dep function may determine that the character constitutes a continuous character in the Chinese numeral. Further, when the type of character to be analyzed is a character that has a high probability of being used as a part of the name, it may be determined that the character constitutes a continuous character in the name. Moreover, it may be determined that one character has a plurality of character types. Specifically, this is a case where it is determined that “◯” has two character types that are a symbol and a part of a Chinese numeral. In this way, the Dep function determines the dependency by referring to various attributes of the character.

  Hereinafter, the above-described analysis process will be described with reference to FIGS. 5 to 11 for a specific example of the morphological analysis result “Meg gave his pen to him” (FIG. 3). FIG. 5 is a diagram illustrating a specific processing example of the dependency analysis processing unit 102 of the natural language analysis apparatus 10 according to the embodiment of the present invention. 6 to 11 are diagrams illustrating specific processing examples of the dependency analysis processing unit 102, which are subsequent to the previous diagrams.

  In FIG. 5, when the dependency analysis processing unit 102 starts the processing, it pushes “−1” to the dependency-destination pending stack 107, and the character “me” with the character ID “0” is stored in the character of the analysis character storage unit 108. Store in code. Subsequently, the dependency analysis processing unit 102 pushes “0” to the dependency destination pending stack 107 and stores the character “G” of the character ID “1” in the character code of the analysis character storage unit 108. Next, the dependency analysis processing unit 102 sets the dependence destination candidate character ID in the dependence destination candidate character ID storage area 106 to an initial value “1”.

  In FIG. 6A, the dependency analysis processing unit 102 determines that the value “0” popped from the dependency destination pending stack 107 because the character code indicated by the dependency destination candidate character ID in the dependency destination candidate character ID storage area 106 is not EOS. Is set as the dependency source character ID in the dependency source character ID storage area 105. Next, the dependency analysis processing unit 102 stores the character “GA” of the dependency destination candidate character ID + 1 (ie, 1 + 1 = “2”) in the character code of the analysis character storage unit 108.

  Next, in FIG. 6B, the dependency analysis processing unit 102 determines that the dependency source character ID of the dependency source character ID storage area 105 is not “−1”, and the determination result of the dependency relationship determination unit 103 is “True”. Therefore, the following processing is performed. More specifically, the dependency relationship determination unit 103 stores “katakana” as the character type of the character “me” with the character ID “0” and sets “katakana” as the character type of the character “gu” with the character ID “1”. Since “Meg” is in the word dictionary, it is determined that “Me” is “related” to “G” (determination by machine learning), and “True” is returned. Since the determination result of the dependency relationship determination unit 103 is “W”, the dependency analysis processing unit 102 sets “W” as the determination result as the dependency source character type of the analysis character storage unit 108 and depends on the dependency destination. The dependence destination candidate character ID “1” in the destination candidate character ID storage area 106 is set. Next, the dependency analysis processing unit 102 sets the value “−1” popped from the dependency destination pending stack 107 as the dependency source character ID in the dependency source character ID storage area 105.

  In FIG. 6 (3), since the dependency source character ID of the dependency source character ID storage area 105 is “−1”, the dependency analysis processing unit 102 stores the dependency source character ID storage area 105 in the dependency destination pending stack 107. The dependence source character ID “−1” is pushed. Further, the dependency analysis processing unit 102 pushes the dependence destination candidate character ID “1” in the dependence destination candidate character ID storage area 106. Next, the dependency analysis processing unit 102 increments the dependence destination candidate character ID in the dependence destination candidate character ID storage area 106 to (1 + 1 =) “2”.

  In FIG. 7A, the dependency analysis processing unit 102 determines that the value “1” popped from the dependency destination pending stack 107 since the character code indicated by the dependency destination candidate character ID in the dependency destination candidate character ID storage area 106 is not EOS. Is set as the dependency source character ID in the dependency source character ID storage area 105. Next, the dependency analysis processing unit 102 stores the character “hi” of the dependence destination candidate character ID + 1 (ie, 2 + 1 = “3”) in the character code of the analysis character storage unit 108.

  In FIG. 7B, the dependency analysis processing unit 102 determines that the dependency source character ID of the dependency source character ID storage area 105 is not “−1” and the determination result of the dependency relationship determination unit 103 is “True”. Perform the process. More specifically, the dependency relationship determination unit 103 stores “Hiragana” as the character type of the character “GA” of the character ID “2”, “MEG” is in the word dictionary, and “GUGA” is in the word dictionary. No, “Guga” is different in character type, etc., “G” is a word break, “G” is information of particle, etc. Is determined (determination based on machine learning), and “True” is returned. Since the determination result of the dependency determination unit 103 is “B”, the dependency analysis processing unit 102 sets “B” of the determination result as the dependency source character type of the analysis character storage unit 108 and depends on the dependency destination. The dependence destination candidate character ID “2” in the destination candidate character ID storage area 106 is set. Next, the dependency analysis processing unit 102 sets the value “−1” popped from the dependency destination pending stack 107 as the dependency source character ID in the dependency source character ID storage area 105.

  In FIG. 7 (3), since the dependency source character ID of the dependency source character ID storage area 105 is “−1”, the dependency analysis processing unit 102 stores the dependency source character ID storage area 105 in the dependency destination pending stack 107. The dependence source character ID “−1” is pushed. Further, the dependency analysis processing unit 102 pushes the dependence destination candidate character ID “2” in the dependence destination candidate character ID storage area 106. Next, the dependency analysis processing unit 102 increments the dependency destination candidate character ID in the dependency destination candidate character ID storage area 106 to (2 + 1 =) “3”.

  In FIG. 8A, the dependency analysis processing unit 102 determines that the value “2” popped from the dependency destination pending stack 107 since the character code indicated by the dependency destination candidate character ID in the dependency destination candidate character ID storage area 106 is not EOS. Is set as the dependency source character ID in the dependency source character ID storage area 105. Next, the dependency analysis processing unit 102 stores the character “ni” of the dependence destination candidate character ID + 1 (ie, 3 + 1 = “4”) in the character code of the analysis character storage unit 108.

  In FIG. 8 (2), the dependency analysis processing unit 102 determines that the dependency source character ID of the dependency source character ID storage area 105 is not “−1” and the determination result of the dependency relationship determination unit 103 is “False”. Perform the process. More specifically, the dependency relationship determination unit 103 stores “Kanji” as the character type of the character “he” of the character ID “3”, and determines that “ga” is “not related” to “he” (machine) Judgment by learning) and returns “False”. The dependency analysis processing unit 102 sets the determination result “O” as the dependency source character type in the analysis character storage unit 108 and does not set the dependency destination because the determination result of the dependency relationship determination unit 103 is “O”. (To be decided)

  In FIG. 8 (3), the dependency analysis processing unit 102 pushes the dependency source character ID “2” of the dependency source character ID storage area 105 to the dependency destination pending stack 107. Further, the dependency analysis processing unit 102 pushes the dependence destination candidate character ID “3” in the dependence destination candidate character ID storage area 106. Next, the dependency analysis processing unit 102 increments the dependency destination candidate character ID in the dependency destination candidate character ID storage area 106 to (3 + 1 =) “4”.

  In FIG. 9A, the dependency analysis processing unit 102 determines that the value “3” popped from the dependency destination pending stack 107 since the character code indicated by the dependency destination candidate character ID in the dependency destination candidate character ID storage area 106 is not EOS. Is set as the dependency source character ID in the dependency source character ID storage area 105. Next, the dependency analysis processing unit 102 stores the character “A” of the dependency destination candidate character ID + 1 (ie, 4 + 1 = “5”) in the character code of the analysis character storage unit 108.

  In FIG. 9B, the dependency analysis processing unit 102 determines that the dependency source character ID of the dependency source character ID storage area 105 is not “−1”, and the dependency determination result of the dependency relationship determination unit 103 is “True”. Therefore, the following processing is performed. More specifically, the dependency relationship determination unit 103 stores “Hiragana” in the character type of the character “ni” of the character ID “4” and determines that “he” is “related” to “ni” (machine learning). And return “True”. Since the determination result of the dependency determination unit 103 is “B”, the dependency analysis processing unit 102 sets “B” of the determination result as the dependency source character type of the analysis character storage unit 108 and depends on the dependency destination. The dependence destination candidate character ID “4” in the destination candidate character ID storage area 106 is set. Next, the dependency analysis processing unit 102 sets the value “2” popped from the dependency destination pending stack 107 as the dependency source character ID of the dependency source character ID storage area 105. Next, since the dependency source character ID of the dependency source character ID storage area 105 is not “−1” and the determination result of the dependency relationship determination unit 103 is “False”, the dependency analysis processing unit 102 determines that the dependency source character ID is “False”. Process. More specifically, the dependency relationship determination unit 103 performs determination (determination by machine learning) that “ga” is “not related” to “ni”, and returns “False”. The dependency analysis processing unit 102 sets the determination result “O” as the dependency source character type in the analysis character storage unit 108 and does not set the dependency destination because the determination result of the dependency relationship determination unit 103 is “O”. (To be decided)

  In FIG. 9 (3), the dependency analysis processing unit 102 pushes the dependency source character ID “2” of the dependency source character ID storage area 105 to the dependency destination pending stack 107. Further, the dependency analysis processing unit 102 pushes the dependence destination candidate character ID “4” in the dependence destination candidate character ID storage area 106. Next, the dependency analysis processing unit 102 increments the dependence destination candidate character ID in the dependence destination candidate character ID storage area 106 to (4 + 1 =) “5”.

  Similarly, the dependency analysis processing unit 102 stacks character IDs whose dependency destinations are not yet determined, and after determining the character dependency destinations by determining the dependency relationship between characters, the dependency analysis processing unit 102 determines the character IDs stored in the stack. Dependency is determined and character dependency is determined.

  Processing near the end of the sentence will be described with reference to FIG. In the example shown in FIG. 10, three character IDs whose dependency destinations are yet to be determined and the next processing target character ID are stacked on the dependency destination undecided stack 107. This is a state where the destination candidate character ID is “13”.

  In FIG. 10A, the dependency analysis processing unit 102 determines that the value “12” popped from the dependency destination pending stack 107 since the character code indicated by the dependency destination candidate character ID in the dependency destination candidate character ID storage area 106 is not EOS. Is set as the dependency source character ID in the dependency source character ID storage area 105. Next, the dependency analysis processing unit 102 stores the character “EOS” of the dependence destination candidate character ID + 1 (ie, 13 + 1 = “14”) in the character code of the analysis character storage unit 108.

  In FIG. 10B, the dependency analysis processing unit 102 determines that the dependency source character ID in the dependency source character ID storage area 105 is not “−1” and the dependency destination candidate character ID in the dependency destination candidate character ID storage area 106. Since the character code indicated by ID + 1 is EOS, the following processing is performed. More specifically, the dependency relationship determination unit 103 stores “symbol” in the character type of the character “.” Of the character ID “13” and determines that “ta” is “related” to “.” (Machine learning) And return “True”. Since the determination result of the dependency determination unit 103 is “B”, the dependency analysis processing unit 102 sets “B” of the determination result as the dependency source character type of the analysis character storage unit 108 and depends on the dependency destination. The dependence destination candidate character ID “13” of the destination candidate character ID storage area 106 is set. Next, the dependency analysis processing unit 102 sets the value “9” popped from the dependency destination pending stack 107 as the dependency source character ID in the dependency source character ID storage area 105.

  In FIG. 10C, the dependency analysis processing unit 102 determines that the dependency source character ID in the dependency source character ID storage area 105 is not “−1” and the dependency destination candidate character ID in the dependency destination candidate character ID storage area 106. Since the character code indicated by ID + 1 is EOS, “D” is set as the type of the dependency source character ID “9” in the analysis character storage unit 108, and the dependency destination candidate character in the dependency destination candidate character ID storage area 106 is set as the dependency destination. ID “13” is set. Next, the dependency analysis processing unit 102 sets the value “4” popped from the dependency destination pending stack 107 as the dependency source character ID of the dependency source character ID storage area 105.

  In FIG. 11A, the dependency analysis processing unit 102 determines that the dependency source character ID of the dependency source character ID storage area 105 is not “−1” and the dependency destination candidate character ID of the dependency destination candidate character ID storage area 106. Since the character code indicated by ID + 1 is EOS, “D” is set as the type of the dependency source character ID “4” in the analysis character storage unit 108, and the dependency destination candidate character in the dependency destination candidate character ID storage area 106 is set as the dependency destination. ID “13” is set. Next, the dependency analysis processing unit 102 sets the value “2” popped from the dependency destination pending stack 107 as the dependency source character ID of the dependency source character ID storage area 105.

  In FIG. 11B, the dependency analysis processing unit 102 determines that the dependency source character ID of the dependency source character ID storage area 105 is not “−1” and the dependency destination candidate character ID of the dependency destination candidate character ID storage area 106. Since the character code indicated by ID + 1 is EOS, “D” is set as the type of the dependency source character ID “2” in the analysis character storage unit 108, and the dependency destination candidate character in the dependency destination candidate character ID storage area 106 is set as the dependency destination. ID “13” is set. Next, the dependency analysis processing unit 102 sets the value “−1” popped from the dependency destination pending stack 107 as the dependency source character ID in the dependency source character ID storage area 105.

  In FIG. 11 (3), the dependency analysis processing unit 102 determines that the dependency source character ID in the dependency source character ID storage area 105 is “−1”. The dependence source character ID “−1” is pushed. Further, the dependency analysis processing unit 102 pushes the dependence destination candidate character ID “13” in the dependence destination candidate character ID storage area 106. Next, the dependency analysis processing unit 102 increments the dependence destination candidate character ID in the dependence destination candidate character ID storage area 106 to (13 + 1 =) “14”. Then, since the character code indicated by the dependence destination candidate character ID in the dependence destination candidate character ID storage area 106 is EOS, the dependency analysis processing unit 102 sets the dependence destination of the value “13” popped from the dependence destination pending stack 107. The dependence destination candidate character ID “14” in the dependence destination candidate character ID storage area 106 is set, the type of the dependence source character ID “13” in the analysis character storage unit 108 is set to “E”, and the process is terminated. In this way, the natural language analyzing apparatus 10 acquires, for example, characters that make up a sentence “Meg gave him that pen.” For each character, and shows the dependency for each acquired character as shown in FIG. ).

  According to the present embodiment, the natural language analyzing apparatus 10 acquires characters constituting the sentence to be analyzed in character units, and determines the dependency for each acquired character. Then, the natural language analyzing apparatus 10 stacks characters whose dependency destinations are undecided in the dependency destination undecided stack 107 in the process of determining the dependency relationship for each character in order from the first character of the sentence to be analyzed, After the character dependency destination is determined by the dependency relationship determination, the character dependency relationship stored in the dependency destination pending stack 107 is determined to determine the character dependency relationship. Furthermore, the determination of the dependency relationship of the natural language analyzing apparatus 10 is performed based on the grammar definition table 110 in which the types of the dependency source and dependency destination candidates and the determination result types corresponding to the character context relationship are associated with each other. . Furthermore, the dependency relationship of the natural language analysis apparatus 10 is determined based on the grammatical rules machine-learned by the SVM. Therefore, since the natural language analysis apparatus 10 performs dependency analysis for each character while accumulating characters whose dependency relationship is undetermined, the dependency relationship for each character can be obtained without performing morphological analysis of the sentence to be analyzed to the end of the sentence. Can be determined.

  As mentioned above, although embodiment of this invention was described, this invention is not restricted to embodiment mentioned above. The effects described in the embodiments of the present invention are only the most preferable effects resulting from the present invention, and the effects of the present invention are limited to those described in the embodiments of the present invention. is not.

DESCRIPTION OF SYMBOLS 10 Natural language analyzer 101 Analysis object character input part 102 Dependency analysis process part 103 Dependency determination part 104 Analysis result output part 105 Dependent source character ID storage area 106 Dependent destination candidate character ID storage area 107 Dependent destination undecided stack 108 Analyzed character Storage unit 110 Grammar definition table 1031 Teacher data input unit 1032 Machine learning unit 1033 Grammar rule storage unit 1034 Dependency determination request reception unit 1035 Dependency determination execution unit

Claims (5)

A natural language analyzer for analyzing natural language sentences,
Character acquisition means for acquiring characters constituting the sentence to be analyzed in character units;
Character dependency analysis means for determining the dependency for each acquired character,
The character dependency analysis means includes:
In the process of determining the dependency relationship for each character in order from the first character of the sentence to be analyzed, the character is determined as the dependency destination is stacked, and the character dependency destination is determined by determining the dependency, Determine the dependency of the characters by determining the dependency of the characters stored in
Natural language analyzer.   The determination of the dependency relationship in the character dependency analysis unit is performed based on a grammar definition table in which a type of a determination result according to a character type of a dependency source and a dependency destination candidate and a character position relationship is associated. The natural language analysis apparatus according to claim 1.   The natural language analysis apparatus according to claim 1, wherein the dependency determination in the character dependency analysis unit is performed based on a grammar rule machine-learned by SVM. A natural language analyzer is a natural language analysis method for analyzing a natural language sentence,
Obtaining characters constituting the sentence to be analyzed in character units;
A character dependency analysis step for determining a dependency relationship for each acquired character,
The character dependency analysis step includes:
In the process of determining the dependency relationship for each character in order from the first character of the sentence to be analyzed, the character is determined as the dependency destination is stacked, and the character dependency destination is determined by determining the dependency, A natural language analysis method for determining the dependency of characters by determining the dependency relationship of characters stored in. A natural language analysis program that analyzes natural language sentences.
Obtaining characters constituting the sentence to be analyzed in character units;
A character dependency analysis step for determining a dependency relationship for each acquired character; and
The character dependency analysis step includes:
In the process of determining the dependency relationship for each character in order from the first character of the sentence to be analyzed, the character is determined as the dependency destination is stacked, and the character dependency destination is determined by determining the dependency, A program that determines the dependency of characters by determining the dependency relationship of characters stored in.

Images