JP2002041513A - Device and method for natural language processing and device and method for natural language dictionary generation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To actualize a device and a method for natural language processing which can impose restrictions on a natural language pattern and evade a great increase in the capacity of a dictionary even when the restrictions are imposed. SOLUTION: The device and method perform syntax analysis and syntax generation by using a natural language pattern composed of an array of a language name, a left side, and a right side, and they have a pattern inspecting means (stage) for inspecting whether or not all or some of natural language patterns have center element information prescribing restrictions by origin and a center pattern element at origin propagation time in the left or right side and also an extracted natural language pattern is suitable to a tree structure in terms of the restriction by the origin and a pattern applying means (stage) where the natural language pattern is applied to the tree structure when it is suitable, and the restriction by the origin are propagated when the natural language pattern has the center element information.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a natural language processing apparatus, a natural language processing method, a natural language pattern dictionary creating apparatus, and a natural language pattern dictionary creating method. For example, machine translation for performing machine translation using a translation pattern It can be applied to devices and machine translation methods.

[0002]

2. Description of the Related Art In a conventional machine translation apparatus, translation knowledge (a system dictionary and a user dictionary) that can be registered by a system or a user includes a word dictionary, an idiom composed of a plurality of words,
Furthermore, even high-performance devices are limited to expressions predetermined by a machine translation device, such as a co-occurrence expression of a verb and a noun such as "burn hands". This is a conventional machine translation device,
The source language sentence is parsed using the source language grammar and the dictionary, the language is converted using the conversion dictionary, and the target language sentence is generated using the target language grammar and the dictionary. This is because the user cannot register the translation dictionary.

[0003] Furthermore, since the source language and the target language are analyzed and generated separately, "[noun phrase] which is intuitive to the user as translation knowledge is shown below.""Following is [noun phrase]. A dictionary based on a pattern composed of a pair of a source language and a target language, such as "(noun phrase is a variable) (P1), cannot be registered.

As a solution to such a problem, there is a translation method and a translation device based on a translation pattern described in Japanese Patent Application Laid-Open No. 5-290082. The invention described in this publication expresses translation knowledge within the scope of context-free grammar, and grammar rules are paired with patterns in the source language and patterns in the target language. By storing grammar rules in a trie dictionary, the parsing speed can be reduced to a practical level. Furthermore, by synchronizing the grammar rules with the patterns of the source language and the patterns of the target language, synchronous derivation is possible, and language conversion and syntax generation can be performed only by simple processing.

[0005] This makes it possible to register a dictionary based on a translation pattern composed of a pair of a source language and a target language, such as the pattern P1 described above, in conformity with the user's intuition. Arbitrary translation knowledge can be registered and the translation process can be performed.

[0006]

However, in the technique disclosed in the above publication, all translation knowledge needs to be described in a context-free grammar.

[0007] In the context-free grammar, it is not possible to constrain each variable (non-terminal symbol). When it is desired to apply a constraint, it is necessary to assign a non-terminal symbol to each constraint and write down a grammar using the non-terminal symbol. As a result, the number of grammars becomes enormous and maintenance becomes very difficult.

For this reason, it is possible to add a restriction to the natural language pattern according to the grammar rules, and even if added, it is possible to avoid a large dictionary, and to perform syntax analysis and syntax generation using the natural language pattern. A natural language processing apparatus and a natural language processing method to be performed are desired, and a natural language pattern creation apparatus and a natural language pattern creation method suitable for the natural language processing apparatus and the natural language processing method are also desired.

[0009]

According to a first aspect of the present invention, there is provided a method for parsing and / or synthesizing a language using a natural language pattern composed of a sequence of a language name, a left side, and a right side. In the natural language processing device that performs the generation, all or a part of the natural language pattern prepared in advance in the pattern dictionary prescribes the restriction by the feature on the left side and / or the right side, and the central pattern element at the time of the feature propagation. The natural language pattern extracted from the natural language pattern prepared in the pattern dictionary in advance as a candidate in the processing at the time of syntactic analysis and / or syntax generation while having the core element information to A pattern checking means for checking whether or not there is a constraint based on features, and if applicable, applying the natural language pattern to a tree structure and Turn and having a pattern application unit for propagating constrained by feature when having a center element information.

Further, the second invention uses a natural language pattern composed of a sequence of a language name, a left side, and a right side, and
In the natural language processing method of performing syntax analysis and / or syntax generation, all or a part of the natural language pattern prepared in advance in the pattern dictionary is, on the left side and / or the right side,
Constraints due to features, and having central element information that defines the central pattern element at the time of feature propagation, and parsing from the natural language pattern prepared in advance in the pattern dictionary, and / or processing during syntax generation A natural language pattern extracted as a candidate of the pattern conforms to the tree structure, whether the pattern is inspected from the aspect of the restriction by the feature, and if the pattern is matched, the natural language pattern is applied to the tree structure, And a pattern application step of propagating the restriction by the feature when the natural language pattern has the core element information.

[0011] Further, the third invention is a language, a left side,
In a natural language pattern dictionary creation device that creates a pattern dictionary applied to a natural language processing device that performs syntax analysis and / or syntax generation using a natural language pattern composed of a sequence of right sides, a left side and / or a right side A source dictionary that stores natural language patterns that are all described in text data, and may have central element information that defines the central pattern element during feature propagation. And a feature information conversion unit configured to convert the feature constraint information of the natural language pattern read from the source dictionary into feature constraint data in a format in which a logical operation can be easily performed and store the feature constraint data in the pattern dictionary.

Still further, the fourth invention is applied to a natural language processing apparatus that performs syntax analysis and / or syntax generation using a natural language pattern composed of a sequence of language names, left and right sides. In a natural language pattern dictionary creation method for creating a pattern dictionary, there is a case where the left side and / or the right side have a constraint by a feature and central element information that defines a central pattern element at the time of feature propagation. Constraint information for converting feature constraint information of a natural language pattern read from a source dictionary storing a natural language pattern described above into feature constraint data in a format in which a logical operation can be easily performed and storing the feature constraint data in the pattern dictionary It has a format conversion process.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS (A) First Embodiment Hereinafter, a natural language processing apparatus, a natural language processing method, a natural language pattern dictionary creation apparatus and a natural language pattern dictionary creation method according to the present invention will be described with reference to a machine translation apparatus and a machine translation machine. A first embodiment applied to the method will be described in detail with reference to the drawings.

(A-1) Configuration of First Embodiment FIG. 1 is a block diagram showing a functional configuration of a machine translation apparatus according to the first embodiment. In practice, for example, a processing program, fixed data, and the like are loaded on an information processing apparatus such as a personal computer, and the machine translation apparatus according to the first embodiment is constructed.

In FIG. 1, a machine translation apparatus 1 according to a first embodiment includes an input unit 101, an output unit 102, a morphological analysis unit 103, a morphological generation unit 104, a syntax analysis unit 105, a syntax generation unit 106, and a pattern check unit. 107, a pattern application unit 108, a pattern dictionary 109, a feature table 110, a dictionary creation unit 111, a feature table creation unit 112, a dictionary source 113, and a feature definition table 114.

It should be noted that, of the above components, the input unit 10
1. The output unit 102, the morphological analysis unit 103, the morphological generation unit 104, the syntax analysis unit 105, the syntax generation unit 106, the pattern inspection unit 107, the pattern application unit 108, and the pattern dictionary 109 constitute a machine translation apparatus body. .

The input unit 101 receives an input sentence (source language sentence) to be translated, and corresponds to not only a keyboard but also a reading configuration from a sentence file.
The morphological analysis unit 103 separates the input sentence for each morpheme (for example, a word). Note that the input unit 101 and the morphological analysis unit 103 are the same as the conventional one. Also,
Illustration of a morphological dictionary used for morphological analysis is omitted.

The syntax analyzer 105 includes a morpheme analyzer 103
Based on the morphological analysis result, the pattern dictionary 109 is looked up in a dictionary, and the pattern inspection unit 107 and the pattern application unit 10
8 is used to perform syntax analysis. On the other hand, based on the result of the syntax analysis on the source language sentence given from the syntax analysis unit 105, the syntax generation unit 106
09 is looked up in a dictionary, and a pattern in the target language is generated using the pattern inspection unit 107 and the pattern application unit 108 as appropriate. The functions of the syntax analysis unit 105 and the syntax generation unit 106 will be clarified in an operation description described later.

The morpheme generation unit 104 includes a syntax generation unit 106
Then, a translation result is obtained by applying a morpheme corresponding to the syntax of the target language generated by. The output unit 102 outputs a translation result, and includes not only a display output and a print output, but also a storage output. Alternatively, the number of translation results to be output (the number of candidates) may be specified by the user. Note that the morpheme generation unit 104 and the output unit 102 are the same as the conventional one. Illustration of a morphological dictionary used for morpheme generation is omitted.

The pattern dictionary section 109 stores translation patterns (pairs of source language patterns and target language patterns) used for syntactic analysis and syntax generation. In the case of the first embodiment, the source language pattern and the target language pattern may each have a feature restriction as described later.

The pattern checking unit 107 checks whether the syntax tree of the syntax analysis result or the syntax tree to be generated conforms to the feature constraints of the pattern obtained by dictionary lookup. On the other hand, when the pattern conforms to the constraint, the pattern applying unit 108 changes the syntax tree to a pattern conforming to the constraint.

The feature table 110, the dictionary creation unit 111, and the feature table creation unit 1 among the constituent elements described above.
12, a dictionary source 113 and a feature definition table 114 constitute a translation pattern dictionary creating apparatus for creating a translation pattern to be stored in the pattern dictionary unit 109. The function of each of these components will be clarified in the operation section.
Briefly, it is as follows.

The feature definition table 114 stores data (see FIG. 5 described later) defining features that can be used for constraints and the like. Feature table 11
0 stores constraint feature data in a format that can be added as a constraint to the translation pattern. The feature table creation unit 112 creates constraint feature data to be stored in the feature table 110 from feature definition data stored in the feature definition table 114.

The feature table 110, the feature table creating unit 112, and the feature definition table 114 function according to the source language and the target language.

The dictionary source 113 stores original translation patterns. The dictionary creation unit 111 converts the original translation pattern stored in the dictionary source 113 into the pattern inspection unit 107 and the pattern application unit 108 based on the constraint feature data stored in the feature table 110 as appropriate.
Is converted into translation pattern data in a form that is easy to process and stored in the pattern dictionary 109.

(A-2) Operation of First Embodiment (A-2-1) Operation for Creating Translation Pattern Dictionary with Feature Constraint Before describing the operation for creating a translation pattern to which a constraint based on features is added, first, The translation pattern will be described.

As shown in FIG. 2, the translation pattern is a pattern in which patterns in a plurality of languages are paired. In the case of a one-way or two-way machine translation device between two languages, the translation pattern is a pair of patterns in two languages. The pattern in each language follows a context-free grammar, and includes a language name that specifies a language type, a left side that represents a syntax category (pattern name), and a right side that constitutes a sequence of other syntax categories that make up that syntax category ( Pattern component).

For example, at the time of parsing, if the right-hand side (a non-terminal symbol or terminal symbol) matches the immediately preceding parsing result or morphological analysis result, it is reduced to the left-hand side (the non-terminal symbol). For example, at the time of syntax generation, the left side (a non-terminal symbol) is replaced with the right side (a non-terminal symbol or a terminal symbol).

FIG. 3 shows an example of a translation pattern to which no restriction by the feature is added. Note that FIG.
Also shows an example of storage in the dictionary source 113.

In FIG. 3, en and ja represent English and Japanese, respectively, and S following the: is a nonterminal symbol name (syntax category name) on the left side. Is a non-terminal symbol name. Also,
NP (noun phrase) and VP (verb phrase) are non-terminal symbol names (syntax category names) constituting the right side, and 1 and 2 added to these non-terminal symbol names are non-terminal symbols having the same number. This indicates that there is correspondence between patterns. Each syntax category (pattern component) constituting the right side is distinguished by a pair of square brackets.

FIG. 4 shows an example of a translation pattern to which restrictions due to features are added. Note that the example of FIG. 4 is an example in which a restriction is added only to an English pattern, and no restriction is added to a Japanese pattern. FIG. 4 also shows an example of storage in the dictionary source 113.

Here, the feature added as a constraint to the translation pattern refers to feature information in various information used in translation, such as morphological information, syntax information, and semantic information.

In FIG. 4, num and pos are feature names, the former is a feature name representing a number (quantity), and the latter is a feature name representing a part of speech. = Sg or = n following the feature name indicates a feature value, the former indicates that the quantity is singular, and the latter indicates that the part of speech is a noun. In addition, * in VP: * in FIG. 4 indicates that the element (syntax category) on the right side with this symbol is the central element on the right side. At the time of parsing, all features of the central element are copied to the nodes of the parse tree corresponding to the left side. After being copied, the feature (type = normal in this example) added to the left side is also set.

Next, the operation up to the creation of a translation pattern with constraint feature data stored in the pattern dictionary 109 will be described.

The translation pattern stored in the pattern dictionary 109 is stored in a translation pattern body as shown in FIGS. 3 and 4 in the form of constraint feature data (for example, bit data) for facilitating processing in a pattern inspection described later. (Sequence data).

First, the operation of the feature table creation unit 112 will be described with reference to the flowchart of FIG.

The feature table creation unit 112 reads the stored contents (feature definition) of the feature definition table 114 (step 5).
01).

FIG. 6 shows an example of feature definition data in the feature definition table 114. FIG. 6 shows an example for one feature definition target. In FIG. 6, the portion between / ** / is a comment on the data, the second line is a feature in a character string (string) format, and the feature name is in the standard form ba.
represents that it is defined as seForm,
The fourth line defines the feature in the word (word) format, and the feature name defines that the feature value of the part of speech pos is n (noun) and adj (adjective), and the sixth line defines the feature in the word (word) format. A feature name is number (quantity) num and the feature value is s
g (single) or pl (plural).

As described above, there are two types of definition of the feature name, the first one having a character string (character string value), and the second one having a predetermined value.

The feature definition may be an existing feature definition that has already been converted into data as shown in FIG.

The feature table creation unit 112 reads the feature definition from the feature definition table 114, and then, for a feature having a predetermined value in the feature definition, as shown in FIG. And a unique number (identification numbers N1 to N4 defining the set) for each set of
2) Write a correspondence table (hereinafter referred to as a "feature value / number correspondence table" or "bit arrangement data") between the set of the feature name and the feature value and the number assigned to the set in the feature table 110 (step 503).

Such numbers are stored in the pattern inspection unit 107.
Are used when checking the constraint of the pattern, as described later, so that the pattern checking unit 107 can check at high speed.

FIG. 7 shows a case in which the number N1 is assigned to the feature value n (noun) of the part-of-speech noun feature pos, and the feature value a of the part-of-speech noun feature pos.
dj (adjective) is assigned a number N2, and a feature value sg of the numeric feature num is assigned a number N3, and a feature value p of the numeric feature num is assigned.
This is an example in which a number N4 is assigned to l. FIG. 7 is merely an example, and the feature types and the number of feature values of each feature type are not limited thereto.

FIG. 8 shows the restrictions on the pattern due to its features.
As shown in FIG.

First, a constraint is given by a feature value described for a certain feature (feature name). Num = s of the second term on the right side in the above-mentioned English pattern of FIG.
g is an example of this first type of constraint, and is a constraint that the numeric feature is singular. The second is to impose restrictions on features other than those described for a certain feature (feature name). For example, num! The constraint of = sg is a constraint that the numerical feature is not singular. Third, a constraint is given by a plurality of feature values described for a certain feature (feature name). For example, the constraint pos = n | adj is a constraint that the part-of-speech feature is a noun or an adjective. Fourth, a constraint is imposed on a feature (feature name) other than a plurality of feature values described.
For example, pos! The constraint of = n | adj is a constraint that the part of speech feature is neither a noun nor an adjective.

Note that one syntax element (syntax category)
May be given to a plurality of features. Num = sg: pos = n of the first term on the right side in the above-described English pattern of FIG. 4 is an example of such a plurality of feature constraints, and the syntax element (syntax category) related to the first term on the right side On the other hand, the number feature is singular, and the part of speech feature is a noun. As described above, the difference is represented by:.

The dictionary source 113 describes a translation pattern. A translation pattern to which a restriction by a feature is given (see FIG. 4) and a translation pattern to which no restriction by a feature is given (see FIG. 3). Is also included.

Next, the dictionary creation unit 111 sends the dictionary source 11
The operation of creating a translation pattern (translation pattern with constraint feature data) to be registered in the pattern dictionary 109 by referring to the storage content of the feature table 110 from the storage content (translation pattern body) of No. 3 is referred to the flowchart of FIG. I will explain while.

The dictionary creation unit 111 firstly receives the dictionary source 11
3 (translation pattern) is read (step 9).
01). Next, the dictionary creation unit 111 sends the feature table 11
While referring to the feature value / number correspondence table of 0, the constraint by the feature is replaced (or added) with the number data (bit array data) and written into the pattern dictionary 109 (step 90).
2). Here, the bit array data is determined according to the type of the constraint shown in FIG. 7 and whether the constraint is applied to the left side or the right side, and the pattern dictionary 1
09 is written.

FIG. 10 shows an example of bit array data (number data) in which a part restricted by a feature in a pattern (source language pattern or target language pattern) is replaced.

FIG. 10A shows bit array data for the feature constraint num = sg: pos = n existing on the right side. The bit array data in this case is AND
It consists of (logical product) value array data and AND mask array data. Since the feature value / number correspondence table is as shown in FIG. 7 described above, the AND value array data is “1010”, while the AND mask array data indicating a meaningful bit portion on the AND value array data is “1010”. ".

FIG. 10B shows a feature constraint num! Existing on the right side. = Sg: bit array data for pos = n. The bit array data in this case is also AND
It consists of value array data and AND mask array data. Since the feature value / number correspondence table is as shown in FIG. 7 described above, the AND value array data is “1000”.
A indicating a significant bit portion on the AND value array data
The ND mask array data is “1010”. That is, since three bits of the AND mask array data are “1”, the third bit “0” in the AND value array data indicates that the numerical feature is not singular.

FIG. 10C shows bit array data for the feature constraint pos = n | adj existing on the right side. The bit array data in this case is OR (logical sum)
It consists of value array data. OR value array data is "1100"
Becomes Note that the OR mask array data for the OR value array data is not created because the OR value array data itself includes information of a valid part.

FIG. 10D shows bit array data for the feature constraint num = sg existing on the left side.
In this case, the bit array data includes ADD value array data and ADD mask array data. Since the feature constraint on the right side may be propagated to the left side (see the second embodiment), ADD value array data is employed. In this example, the ADD value array data is “0010”, and the ADD value array is “0010”.
The mask array data is "0011" so that propagation of other features can be accepted. The ADD format shown in FIG. 10D is applied by the processing of the pattern application unit 108 (see FIG. 16 described later).

Even if the expressions shown in FIG. 8 are combined, the feature constraints can be expressed in any of the three types shown in FIG. 10, namely, an AND array format, an OR array format, and an ADD array format. .

The bit array data created by the dictionary creating unit 111 by replacing the feature constraints is 1 bit.
And 0 tend to line up continuously (for example, when there are many feature types related to constraints or when there are many feature value types),
It may be compressed to increase memory space efficiency and processing efficiency.

(A-2-2) Translation Operation Next, a translation operation using a translation pattern in the machine translation apparatus of the first embodiment will be described.

The operation during which the input unit 101 takes in the source language sentence and supplies it to the morphological analysis unit 103 and the morphological analysis unit 103 performs morphological analysis on the input source language sentence is the same as the conventional one.

Hereinafter, the operation of the syntax analyzer 105 will be described with reference to the flowchart of FIG. Note that the syntax analysis unit 105 of the first embodiment performs a process related to restrictions due to features added to the pattern, in addition to the process described in the conventional machine translation method based on the pattern.

The syntactic analysis unit 105 firstly receives the morphological analysis unit 1
Then, the morphologically analyzed syntax tree is received from step 03 (step 1101). Then, after confirming that a tree still requiring parsing remains (step 1102), the pattern dictionary 109 is consulted to extract a necessary pattern (here, a pattern for the source language) (step 1103). Here, a plurality of patterns may be extracted.

Then, it is determined whether or not there remains an unprocessed pattern for which the step 1105 described below has not been executed in a state where the pattern to be applied has not been determined (step 1104). If not, the process returns to step 1102 described above. On the other hand, if an unprocessed pattern remains, the syntax tree and the unprocessed pattern are passed to the pattern checker 107 to check whether the pattern conforms to the constraint (step 1105). Is determined (step 1106).

Note that, conventionally, the extracted pattern is not inspected because there is no restriction due to the feature. However, in the first embodiment, since the pattern is restricted, the pattern inspection in step 1105 described above is performed. Be executed.

If the result of the test is that it is not suitable,
Returning to step 1104 described above, if the target pattern meets the constraint,
The syntax tree and the pattern are transferred to the pattern application unit 108, and a syntax tree resulting from applying the pattern to the syntax tree is received (step 1107).
Return to 2.

The processing loop consisting of steps 1102 to 1107 is repeatedly executed. When the syntax analysis is completed, a tree of the syntax analysis result is created and passed to the syntax generation unit 106 (step 1108).

FIG. 12 is a flowchart showing the processing of the pattern inspection unit 107 when an inspection request is made from the syntax analysis unit 105. The processing of the pattern inspection unit 107 shown in FIG. 12 corresponds to the inspection request in step 1105 described above.

First, the pattern checker 107 receives a syntax tree and a pattern (step 1201), and sets a processing right side number parameter i to an initial value 0 (step 120).
2).

Then, it is confirmed that the parameter i has not reached the number of right sides of the pattern (step 1203).
I + 1th child node of the syntax tree and i on the right side of the pattern
The + 1st is taken out to determine whether or not the constraint of the pattern is satisfied, and the result of the determination is recognized (step 1).
204, 1205). If there is no constraint on the (i + 1) th right side of the pattern, it is determined that the constraint is satisfied.

If the (i + 1) -th child node of the syntax tree does not satisfy the constraint on the (i + 1) -th right side of the pattern, the pattern checker 107 sends a message to the effect that it did not match to the syntax analyzer 105 ( Step 1208).

On the other hand, if the (i + 1) th child node of the syntax tree satisfies the constraint on the (i + 1) th right side of the pattern, the parameter i is incremented by 1 and the process returns to step 1203 (step 1206). .
If the result of the determination that the parameter i has reached the number of right sides of the pattern in step 1203 is obtained, all the child nodes of the syntax tree satisfy the constraint of the corresponding right side element of the pattern. The unit 107
A reply to the parsing unit 105 is sent to the parsing unit 105 (step 1).
207).

Here, it usually takes a considerable amount of time, processing amount, and cost to check the feature constraint. However, in the first embodiment, as shown in FIG. 10, the feature and its constraint are represented by bit array data. Is held, it is only necessary to compare the bit array data in order, and the time, processing amount and cost are considerably reduced.

Hereinafter, the details of the method for judging the suitability of the constraint using the bit array data will be described with reference to the examples shown in FIGS.

FIG. 13 shows that the restriction due to the feature
The case of the AND format as shown in (A) or (B) is shown. Reference numeral 1301 denotes bit array data (AND value array data and AND mask array data) for a certain right side element. Reference numeral 1302 denotes bit array data of a child node of the syntax tree corresponding to the right side element. Such bit arrangement data of the child node is formed from the feature information of the child node when determining the suitability. It should be noted that the morphological analysis unit 103 may be formed by referring to the storage contents of the feature table 110.

The bit array data for the right side element and the bit array data of the child node are collated (reference numeral 1303).

In the case of the AND format as shown in FIG. 13, specifically, the value array data of the child node is ANDed with the bit array of the AND mask array data, and the collation is required in the value array data of the child node. Only the bit portion is highlighted. If the result is exactly the same as the AND value array data, it is determined that the constraint has been met. If even one bit is different, it is determined that the constraint has not been met (reference numeral 130).
4).

In the case of FIG. 13, the bit array data (reference numeral 1302) of the child node of the syntax tree conforms to the constraint represented by the bit array data (reference numeral 1301) for the right side element of the pattern.

As is clear from FIG. 13, even if the value array data of the child node of the syntax tree relates to a plurality of features, the value array data for the constraint represented by the bit array data for the right-hand side element of the pattern does not match. Only conformance matters.

FIG. 14 shows a case where the feature constraint is of the OR type as shown in FIG. 10 (C). Reference numeral 1401
Indicates bit array data (OR value array data) for the right side element having a pattern. Reference numeral 1402 is
The bit array data of the child node of the syntax tree corresponding to the right side element is shown.

Also in the OR array, the bit array data for the right side element and the bit array data of the child node are collated (1403).

In the case of the OR format as shown in FIG.
Specifically, the OR of the value array data of the child node and the pattern
An AND for each bit of the value array data is taken, and 1 is added to the result.
If there is even one bit, it is determined that the constraint has been met. If all bits are 0, it is determined that the constraint has not been met (1404).

In the case of FIG. 14, the bit array data (reference numeral 1402) of the child node of the syntax tree conforms to the constraint represented by the bit array data (reference numeral 1401) for the right side element of the pattern.

As is apparent from FIG. 14, in the OR format, there are many feature values to which constraints are applied, and even if the value array data of the child node of the syntax tree is related to one of them, the constraint is imposed. It is necessary to

FIG. 15 is a flowchart showing the processing of the pattern application unit 108 during syntax analysis. The processing of the pattern application unit 108 is the same as that of the above-described step 1107 (FIG. 1).
1). FIG. 15 is described from the propagating surface of the feature.

The pattern application unit 108 first receives a syntax tree and a pattern (step 1501). The feature of the child node of the syntax tree corresponding to the central node of the pattern is set to the top node of the syntax tree (step 1502). This is called a feature propagation. Next, the feature information on the left side of the pattern is added to the top node of the syntax tree (step 150).
3).

The processing of the pattern application unit 108 is executed by the inspection by the pattern inspection unit 107.
For patterns where the parse tree meets the constraints,
By the processing of the pattern applying unit 108, the feature information that conforms to the constraint is propagated to the upper node.

FIG. 16 shows a method in which the above-described processing of the pattern applying unit 108 is executed by arithmetic processing on bit array data.

In FIG. 16, reference numeral 1601 denotes bit array data (A
DD value array data and ADD mask array data), and reference numeral 1602 denotes bit array data (value array data and mask array data) of a child node of the syntax tree corresponding to the right-hand side center element of the pattern.

The pattern application unit 108 first receives the reference numeral 160
The bit array data (feature) of the child node indicated by 2 is copied to the top node of the syntax tree (reference numeral 1603). next,
The features in the top node (value array data and mask array data) reflect the restrictions in the pattern. This process
It can be expressed by the following equations (1) and (2) (reference numeral 1).
604).

Value array = ADD value array | (〜 (ADD mask array) & value array) (1) Mask array = ADD mask array | Mask array (2) In the equations (1) and (2), , | Represent OR (logical sum) for each bit, ~ represents NOT (logical NOT) for each bit, and & represents AND (logical product) for each bit. Further, the term of data in the array data is omitted.

Expression (1) is obtained by calculating the feature value (ADD value array) of the feature restricted by the pattern and the feature value (（(ADD mask array) & (Value array) is used as a feature at the top node. In order to cope with such an increase in feature types, a mask array that defines valid bit positions in the value array also has a value of 2 as shown in Expression (2).
This is a combination of different types of mask arrangements.

Reference numeral 1605 denotes a case where the bit array data for the left side element of the constraint conforming pattern is indicated by reference numeral 1601, and the bit array data of the child node of the syntax tree corresponding to the right side central element of the pattern is indicated by reference numeral 1602. This shows the bit array data (value array data and mask array data; features) at the top node after processing when this is shown.

The syntax analysis unit 105 has the pattern inspection unit 107
The syntax analysis result obtained by performing the syntax analysis while appropriately requesting the processing to the pattern application unit 108 is provided to the syntax generation unit 106, and the syntax generation unit 106 generates syntax tree information in the target language.

Hereinafter, details of the operation of the syntax generation unit 106 will be described with reference to the flowchart of FIG. In addition,
The syntax generation unit 106 according to the first embodiment performs, in addition to the processing described in the conventional machine translation method based on a pattern, processing relating to a constraint due to a feature added to the pattern.

Although FIGS. 3 and 4 show an example in which the target language (Japanese) is not restricted by the feature, the target language (Japanese) may be naturally restricted by the feature. Further, there may be a plurality of target language patterns for the same source language pattern.

First, the syntax generation unit 106
5 receives the parsed parse tree (step 17)
01). Then, after confirming that a tree (tree portion) requiring syntax generation still remains (step 1702),
A necessary pattern (here, a target language pattern) is extracted by looking up the pattern dictionary 109 (step 1703).
Here, a plurality of patterns may be extracted.

Then, it is determined whether or not an unprocessed pattern for which the step 1705 described below has not been executed remains in a state where the pattern to be applied has not been determined (step 1704). If not, the process returns to step 1702 described above. On the other hand, if an unprocessed pattern remains, the syntax tree and the unprocessed pattern are passed to the pattern checker 107 to check whether the pattern conforms to the constraint (step 1705). Is determined (step 1706).

In the meantime, conventionally, the extracted pattern is not inspected because there is no restriction due to the feature. However, in the first embodiment, since the pattern is restricted, the pattern inspection in the above-described step 1705 is not performed. Be executed.
At the time of the pattern check at the time of the syntax generation, the syntax generation is performed from the top down, so the pattern inspection unit 107 omits the illustration of the flowchart, but the feature matching between the top node of the syntax tree and the right side of the pattern is performed. To inspect.

If the test result is not conforming,
Returning to step 1704 described above, if the target pattern conforms to the constraint,
The syntax tree and the pattern are passed to the pattern application unit 108, and a syntax tree resulting from applying the pattern to the syntax tree is received (step 1707).
Return to 2.

At this time, the pattern application unit 108 applies each child node of the syntax tree and each element on the right side of the pattern, although illustration of the flowchart is omitted (FIG. 1).
3, see FIG. 14). Also, the feature information of the top node of the syntax tree is set to a child node of the syntax tree corresponding to the central element on the right side of the pattern. This is the propagation of features from the top node to the central node.

The processing loop consisting of steps 1702 to 1707 is repeatedly executed, and when the syntax generation has been completed, a syntax tree of the syntax generation result is created, and the morpheme generation unit 10
4 (step 1708).

The morpheme generation unit 104 applies the morpheme in the target language to the syntax tree generated by the syntax generation unit 106 while reflecting the features and the like, and converts the translation result (translated sentence) in the same manner as in the related art. Then, the translated sentence obtained by the output unit 102 is output.

(A-3) Effects of the First Embodiment According to the machine translation device and the machine translation method of the first embodiment, the following effects can be obtained.

A. In conventional machine translation based on patterns, there are no restrictions due to features and no feature propagation, so if you try to incorporate various constraints (such as constraints by meaning, constraints by number or gender, or constraints by syntax information) into a pattern, For each difference, a non-terminal symbol or terminal symbol must be prepared and written in the pattern. Therefore, the number of patterns becomes enormous.

According to the first embodiment, the constraint can be described in a pattern in a simple form as additional information for a non-terminal symbol or a terminal symbol, so that the number of patterns can be reduced. In addition, since restrictions are introduced in the pattern as additional information for non-terminal symbols and terminal symbols, it is not necessary to increase the types of non-terminal symbols and terminal symbols, and general information of features can be applied to additional information.
As a result, the pattern becomes easy to understand, and the cost of pattern maintenance (processing time, storage capacity, etc.) can be drastically reduced.

B. Previously, there was no constraint processing,
Other translation methods had to be combined, but by using constraint processing, everything including the fine translation processing can now be translated by a pattern-based translation method.
Therefore, a translation engine that can be essentially learned can be created.

C. Since pattern constraint information is registered in the dictionary as bit array data, processing such as pattern inspection to check whether the pattern conforms to the constraint and modification of the syntax tree according to the applied pattern can be performed at high speed. Can be performed.

(A-4) Modified Embodiment of First Embodiment In the above description, a format as shown in FIG. 2 is shown as a pattern expressing method, and a format as shown in FIG. 5 is shown as a feature definition table. However, if there is the same amount of information, another expression form may be used.

In the above description, the bit array data is used for expressing the features in consideration of easiness of collation and synthesis.
As long as the same result is obtained as a result of processing such as collation and synthesis, another expression form may be used.

Further, in the above description, the components including the components 110 to 114 forming the contents to be registered in the pattern dictionary 109 are shown. However, these components are provided in separate devices,
The registered pattern dictionary 109 may be ported to the machine translation device.

Further, in the above description, the feature table 11
Although 0 is shown as existing separately from the pattern dictionary 109, the feature table 110 may be stored in the pattern dictionary 109.

(B) Second Embodiment Next, a natural language processing apparatus, a natural language processing method, a natural language pattern dictionary creation apparatus and a natural language pattern dictionary creation method according to the present invention are applied to a machine translation apparatus and a machine translation method. The applied second embodiment will be described in detail with reference to the drawings.

(B-1) Configuration of the Second Embodiment FIG. 18 is a block diagram showing the functional configuration of a machine translation device according to the second embodiment.

In FIG. 18, a machine translation apparatus 2 according to the second embodiment includes an input unit 1801, an output unit 1802, a morphological analysis unit 1803, a morphological generation unit 1804, a syntax analysis unit 1
805, syntax generation unit 1806, pattern inspection unit 180
7, pattern application unit 1808, pattern dictionary 1809,
It comprises a feature table 1810, a dictionary creation unit 1811, a feature table creation unit 1812, a dictionary source 1813, and a feature definition table 1814.

Here, the input unit 1801 and the output unit 180
2. Morphological analysis unit 1803, morphological generation unit 1804, syntax analysis unit 1805, syntax generation unit 1806, pattern inspection unit 1807, pattern dictionary 1809, dictionary creation unit 181
1. The dictionary source 1813 is the same as the corresponding element of the first embodiment.

In the case of the second embodiment, the feature definition table 181
4. Feature table creation unit 1812, feature table 181
0 and the pattern application unit 1808 are different from those of the first embodiment.

As will be clarified in the operation section described later, the feature definition in the feature definition table 1814 is different from that in the first embodiment.
Are also different from those of the first embodiment. The second embodiment is different in that the pattern application unit 1808 also refers to the feature table 1810.

(B-2) Operation of the Second Embodiment Next, the operation for forming the contents to be stored in the feature table 1810 will be described.

In the feature definition table 1814, a feature definition is obtained by adding the definition portion shown in FIG. 19 to the definition portion shown in FIG. In FIG. 19, comments are omitted.

The parts shown in FIG. 19 are non-terminal symbols and terminal symbols (syntax category; syntax tree element;
The definition of y) and the definition of the feature name that it can take.

The line marked * is a definition line for feature names that can be taken by all non-terminal symbols and terminal symbols, and the feature names define the standard form, the appearance form, and the part of speech.

The line "" is a feature name that can be taken by a terminal symbol, and indicates a utilization form.

The line of NP is a nonterminal symbol name NP (noun phrase)
Defines the possible feature names, which are numeric and semantic features. The VP line is the non-terminal symbol name VP
(Verb phrase) defines possible feature names, which are numeric features and semantic features. The row of S defines a feature name that can be taken by the non-terminal symbol name S (sentence), and the feature names are a numeric feature and a sentence pattern feature.

As described above, in the second embodiment, feature names that can be restricted are defined according to the types of non-terminal symbols and terminal symbols.

The NP row and the like are definitions intended to propagate only the features that can be taken by the non-terminal symbol or terminal symbol at the propagation destination when the feature is propagated. If only the specified features are propagated in this way, unnecessary features are not propagated,
Only the necessary information is included in the syntax tree, which is easy to understand and lean. Also, when a large number of parsing candidates are present during parsing, a parse tree having top nodes with the same feature can be merged to greatly reduce the cost of parsing, but unnecessary features are eliminated. Often, merging is not possible due to propagation. Restriction of feature propagation can also solve the problem that merging cannot be performed due to this unnecessary feature, and leads to a significant improvement in processing efficiency.

The feature table creation unit 1812 creates a feature mask (bit array data) for each non-terminal symbol or terminal symbol as shown in FIG. 20 for the defined portion in FIG. 19, and stores it in the feature table 1810. . Note that the feature table creation unit 1812 performs the processing shown in FIG. 7 described in the first embodiment.
20 (in this second embodiment, a box into which bit values can be inserted as shown in FIG. 20) and store it in the feature table 1810.

FIG. 20 shows a feature mask that propagates for each non-terminal symbol and for each terminal symbol. If the feature data is masked with this when propagating, only the necessary ones are propagated.

In FIG. 20, the feature mask (bit array data) of NP is based on the feature definition of FIG. 19, and the features taking values in NP are the part of speech features (according to the first line of FIG. 19).
Numerical features (according to the third line in FIG. 19) and semantic features (FIG. 19)
In the third line), so that these can be propagated. The number of semantic feature elements is six.
The feature mask of S (bit array data) can be propagated because the features taking values in S are part of speech features, numeric features and sentence features from the feature definitions in FIG. It should be noted that there are two elements of the sentence pattern feature. VP
Is not shown, but in the case shown, it is the same as NP.

FIG. 21 is a flowchart showing the processing of the pattern application unit 1808 during syntax analysis according to the second embodiment. The processing of the pattern applying unit 1808 corresponds to the processing of step 1107 (see FIG. 11) described above.

The pattern application unit 1808 receives the syntax tree and the pattern (step 2101), and sets the features of the child nodes of the syntax tree corresponding to the central node of the pattern to the top node of the syntax tree (step 2102). However, at this time, the feature mask (FIG. 20) of the non-terminal symbol or terminal symbol of the set destination is extracted from the feature table 1810, and the feature of the child node is masked by this feature mask before propagation. Similarly, when adding the feature information on the left side of the pattern to the top node of the syntax tree, it is added after masking with the feature mask (step 2103).

The mask processing in steps 2102 and 2103 may be performed at once. The feature mask may be built in the pattern application unit 1808.

Although illustration of the flowchart is omitted, the pattern applying unit 108 applies each child node of the syntax tree and each element on the right side of the pattern at the time of generating the syntax. Also, the feature information of the top node of the syntax tree is set to a child node of the syntax tree corresponding to the central element on the right side of the pattern. Also at this time, the applicable range is defined for each non-terminal symbol or terminal symbol by a feature mask as shown in FIG.

(B-3) Effects of the Second Embodiment The same effects a to c as those of the first embodiment can be obtained by the machine translation apparatus and the machine translation method of the second embodiment. According to the second embodiment, the following effects can be further obtained.

D. If only the specified features are transmitted, unnecessary features are not propagated, and only necessary information is included in the syntax tree, which is easy to understand and wasteful.

E. When a large number of parsing candidates exist during parsing, parse trees with top nodes with the same feature can be merged to greatly reduce the cost of parsing, but unnecessary features are propagated Often cannot be merged. Restriction of feature propagation can also solve the problem that merging cannot be performed due to this unnecessary feature, and leads to a significant improvement in processing efficiency.

(B-4) Modified Embodiment of Second Embodiment The dictionary creating unit 1811 uses the feature mask information in the feature table 1810 to convert an inappropriate feature into a non-terminal symbol or terminal symbol in a pattern. It is also possible to confirm whether or not it has been added, and when an inappropriate feature has been added, display it as an error and prompt the dictionary creator to correct it.

Further, the dictionary creating unit 1811 may detect the use of an undefined non-terminal symbol by using the information of the feature mask in the feature table 1810 and prompt the dictionary creator to make correction. . This can prevent misspellings of non-terminal symbols.

Further, the format as shown in FIG. 19 is shown as the feature definition table. However, if there is the same information amount, another expression format may be used. The present invention is not limited to the array data, and another expression format may be used as long as the same processing result is obtained.

Furthermore, a method of controlling propagation of only non-terminal symbols and restrictable feature names defined for each terminal symbol is as follows.
This may be performed by a method other than the method using the feature mask.

(C) Third Embodiment Next, a natural language processing apparatus, a natural language processing method, a natural language pattern dictionary creation apparatus and a natural language pattern dictionary creation method according to the present invention are applied to a machine translation apparatus and a machine translation method. A third embodiment applied will be described in detail with reference to the drawings.

(C-1) Configuration of Third Embodiment FIG. 22 is a block diagram showing a functional configuration of a machine translation apparatus according to the second embodiment.

In FIG. 22, a machine translation apparatus 3 according to the third embodiment includes an input unit 2201, an output unit 2202, a morphological analysis unit 2203, a morphological generation unit 2204, and a syntax analysis unit 2.
205, syntax generation unit 2206, pattern inspection unit 220
7, pattern application unit 2208, pattern dictionary 2209,
It comprises a feature table 2210, a dictionary creation unit 2211, a feature table creation unit 2212, a dictionary source 2213, a feature definition table 2214, and a context data storage unit 2215.

That is, a context data storage unit 2215 used by the pattern inspection unit 2207 and the pattern application unit 2208 is added to the second embodiment as a component. Further, when compared with the second embodiment, those whose functions are slightly changed are the feature definition table 2214,
Feature table creation unit 2212, feature table 2210,
A pattern inspection unit 2207 and a pattern application unit 2208 are the same as the corresponding components of the second embodiment.

(C-2) Operation of Third Embodiment In the case of the third embodiment, the translation pattern may include a variable in the feature value (hereinafter, referred to as a feature variable).
FIG. 23 shows an example of a translation pattern including a feature variable. In FIG. 23, num = {NUM} is a feature name n
This indicates that the feature variable NUM is applied as the feature value of um. By applying such a feature variable NUM, in the example of FIG. 23, it is meant that a constraint that the feature value of the feature name num is a common value is given to the right side elements NP and VP on the pattern. .
Note that, for example, at the time of syntax analysis, a specific feature value is inserted into the feature variable portion by bottom-up feature propagation processing by the pattern application unit 2208.

The feature definition table 2214 may include a definition as shown in FIG. This is an addition of the description of “number” shown in FIG. 6 according to the first embodiment, and is a definition of a feature name that can use a feature variable. It is presumed that the introduction of the feature variable makes it easy to distinguish even the same feature, and therefore, a plurality of feature names are assigned to the same feature. That is, FIG. 24 means that num, sNum, and oNum have the same value and can be checked with the same feature variable. As described above, features that are not connected by = in the feature definition table cannot use the same feature variable. Of course, the same feature name can use the same feature variable.

When reading the definition for the feature variable, the feature table creation unit 2212 writes the bit array data as described on the lower side of FIG. 25 to the feature table 2210. Here, any value (sg, p
Since 1) is not known to be inserted, both are set to 1.
The offset value indicates the position of the feature variable in the original bit array data (for example, FIG. 7 according to the first embodiment). The dictionary creation unit 2211 stores the dictionary source 2
If the translation pattern from 213 has a feature variable,
Replacement as shown in

The pattern inspection unit 2207 stores the feature value in the context data storage unit 2215 when the feature variable is used when checking the feature constraint. Also,
At the time of storing, if there is already context data with the same variable name, the variable value in the context data storage unit 2215 is ANDed (bit ANDed), and if all the results become 0, it is determined that the constraint is not satisfied. . If even one bit remains, the result is stored in the context data storage unit 2215.
Store again.

If the feature variable is used, the pattern application unit 2208 fetches the data of the variable name from the context data storage unit 2215, and puts the data at the offset position shown in FIG. 25 of the value array of the node to be set. The context data is set by the size of the array in FIG.

(C-3) Effects of the Third Embodiment The same effects a to e as those of the above-described embodiment can be obtained by the machine translation apparatus and the machine translation method of the third embodiment. According to the third embodiment, the following effects can be further obtained.

F. When it is desired to make the value of a certain feature of a certain node on the right side the same as that of a certain node, in the absence of the third embodiment, patterns are required by the number of possible values of the feature. However, if the feature variables are used, writing can be performed in one pattern, the speed of syntax analysis and the like can be increased, and the maintenance cost can be significantly reduced.

(C-4) Modified Embodiment of Third Embodiment Although a format as shown in FIG. 23 is shown as a translation pattern, another expression format may be used as long as the same information amount is present. Although the format as shown in FIG. 24 is shown as the feature definition table, another expression format may be used if the same information amount is present. Although the bit array data shown in FIG. 25 is used for expressing the feature variables, other expression forms may be used as long as the same result is obtained.

(D) Other Embodiments In each of the above embodiments, the present invention has been applied to the machine translation apparatus and method using a translation pattern. Each process has a feature, and therefore, the technical idea of each of the above embodiments can be applied to other natural language processing devices and methods. For example, the technical idea of the present invention can be applied to a part until the answer sentence is parsed using a pattern (natural language pattern) by the question and answer apparatus. Also, the technical idea of the present invention can be applied to a case where syntax generation processing is performed from a syntax tree of a further question sentence according to an answer sentence.

Further, the present invention can be applied to a natural language processing apparatus and method including only syntax analysis processing, and a natural language processing apparatus and method including only syntax generation processing.

[0152]

According to the present invention, it is possible to add restrictions to natural language patterns related to grammar rules, and even if added, it is possible to avoid an increase in the size of the dictionary. A natural language processing device and a natural language processing method using a language pattern can be realized.

Further, according to the present invention, a natural language pattern creating device and a natural language pattern creating method suitable for the natural language processing device and the natural language processing method can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a functional configuration of a machine translation device according to a first embodiment.

FIG. 2 is an explanatory diagram showing a basic configuration example of a translation pattern.

FIG. 3 is an explanatory diagram showing an example of a translation pattern to which no restriction by a feature is added.

FIG. 4 is an explanatory diagram illustrating an example of a translation pattern to which a restriction based on a feature according to the first embodiment is added;

FIG. 5 is a flowchart illustrating an operation of a feature table creation unit 112 according to the first embodiment.

FIG. 6 is an explanatory diagram illustrating an example of a feature definition according to the first embodiment;

FIG. 7 is an explanatory diagram showing an example of assigning a number to a set of a feature name and a feature value according to the first embodiment;

FIG. 8 is an explanatory diagram illustrating an example of a constraint type based on features according to the first embodiment.

FIG. 9 is a flowchart illustrating an operation of the dictionary creation unit 111 according to the first embodiment.

FIG. 10 is an explanatory diagram illustrating an example of bit array data of restrictions due to features according to the first embodiment.

FIG. 11 is a flowchart illustrating processing of the syntax analysis unit 105 according to the first embodiment.

FIG. 12 is a flowchart illustrating a process of a pattern inspection unit 107 according to the first embodiment.

FIG. 13 is an explanatory diagram (1) of a method for determining constraint suitability using bit array data according to the first embodiment;

FIG. 14 is an explanatory diagram (2) of a method of determining constraint suitability using bit array data according to the first embodiment.

FIG. 15 is a flowchart illustrating processing of a pattern application unit according to the first embodiment.

FIG. 16 is an explanatory diagram of a specific processing method of the pattern application unit according to the first embodiment.

FIG. 17 is a flowchart illustrating processing of a syntax generation unit 106 according to the first embodiment.

FIG. 18 is a block diagram illustrating a functional configuration of a machine translation device according to a second embodiment.

FIG. 19 is an explanatory diagram illustrating an example of a feature definition according to the second embodiment.

FIG. 20 is an explanatory diagram illustrating an example of a feature mask according to the second embodiment.

FIG. 21 is a flowchart illustrating processing of a pattern application unit 1808 according to the second embodiment.

FIG. 22 is a block diagram illustrating a functional configuration of a machine translation device according to a third embodiment.

FIG. 23 is an explanatory diagram illustrating an example of a translation pattern including feature variables according to the third embodiment.

FIG. 24 is an explanatory diagram of a feature definition related to a feature variable according to the third embodiment.

FIG. 25 is an explanatory diagram of bit array data relating to a feature variable according to the third embodiment.

[Explanation of symbols]

 1, 2, 3 ... machine translation device, 105, 1805, 2205 ... syntax analysis unit, 106, 1806, 2206 ... syntax generation unit, 107, 1807, 2207 ... pattern inspection unit, 108, 1808, 2208 ... pattern application unit, 109, 1809, 2209 ... pattern dictionary, 110, 1810, 2210 ... feature table, 111, 1811, 2211 ... dictionary creation unit, 112, 1812, 2212 ... feature table creation unit, 113, 1813, 2213 ... dictionary source, 114, 1814, 2214 ... feature definition table, 2215 ... context data storage unit.

 ────────────────────────────────────────────────── ─── Continuing from the front page (72) Inventor Sayori Shibatata 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (72) Miki Sasaki 1-7-112 Toranomon, Minato-ku, Tokyo Inside Electric Industry Co., Ltd. (72) Inventor Takeshi Fukui 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. F term in Industrial Co., Ltd. (reference) 5B091 AA06 AA15 BA03 CA02 CA05 CA24 CC01 CC04 CC15

Claims (12)

[Claims] 1. A syntax analysis and / or a natural language pattern composed of a sequence of a language name, a left side, and a right side.
In a natural language processing apparatus that performs syntax generation, all or a part of the natural language pattern prepared in advance in the pattern dictionary includes, on the left side and / or the right side, a constraint by a feature and a pattern element at the time of feature propagation. The natural language pattern having the specified core element information and the natural language pattern extracted from the natural language pattern prepared in the pattern dictionary in advance as a candidate in the syntax analysis and / or the processing at the time of syntax generation conforms to the tree structure. A natural language pattern is applied to the tree structure if it matches, and if the natural language pattern has central element information, A natural language processing apparatus comprising: a pattern application unit that propagates a constraint. 2. The feature according to claim 1, wherein the pattern application unit limits the feature to be propagated according to a definition predetermined for each of the non-terminal symbol and the terminal symbol when the feature information is propagated. Natural language processor. 3. The method according to claim 1, wherein a feature variable is also applied as a feature constraint in the natural language pattern, and the pattern inspection means and the pattern application means perform pattern matching inspection and pattern application on the feature variable. The natural language processing device according to claim 1. 4. The natural language pattern registered in the pattern dictionary holds feature restriction information in a format in which a logical operation can be easily executed. The natural language processing device according to 1. 5. A syntax analysis and / or a natural language pattern composed of a sequence of a language name, a left side, and a right side.
In the natural language processing method for generating a syntax, all or a part of the natural language pattern prepared in advance in the pattern dictionary includes, on the left side and / or the right side, a constraint by a feature and a pattern element at the center when the feature is propagated. A natural language pattern having specified central element information and extracted from the natural language pattern prepared in advance in the pattern dictionary as a candidate in a syntax analysis and / or a process at the time of syntax generation conforms to a tree structure. The natural language pattern is applied to the tree structure if it matches, and if the natural language pattern has the core element information, A pattern application step of propagating constraints. 6. The pattern applying step according to claim 1, wherein, when the feature information is propagated, the feature to be propagated is limited according to a definition predetermined for each of the non-terminal symbol and the terminal symbol. Natural language processing method. 7. A method according to claim 7, wherein a feature variable is also applied as a feature constraint in the natural language pattern, and the pattern inspection step and the pattern application step perform pattern matching inspection and pattern application on the feature variable. The natural language processing method according to claim 5. 8. The natural language pattern registered in a pattern dictionary holds feature restriction information in a format in which a logical operation can be easily executed. Natural language processing method described in 1. 9. A syntax analysis and / or a natural language pattern composed of a sequence of a language name, a left side, and a right side.
In a natural language pattern dictionary creating apparatus that creates a pattern dictionary applied to a natural language processing apparatus that performs syntax generation, a central element that defines a constraint by a feature and a central pattern element at the time of feature propagation on a left side and / or a right side. A source dictionary storing natural language patterns, which are all described in text data, which may have information, and feature constraint information of the natural language patterns read from the source dictionary, in a format in which logical operations can be easily executed. A natural language pattern dictionary creation device, comprising: a constraint information format conversion unit that converts the data into feature constraint data and stores it in the pattern dictionary. 10. The constraint information format conversion means includes: a feature definition storage unit that stores feature information definition information including feature names and feature values used for constraints; and executes a logical operation based on the definition information. A feature constraint data format determination unit that determines a data format in a format that is easy to perform, and a conversion unit that converts feature constraint information of a natural language pattern into feature constraint data in a format that is easy to perform a logical operation according to the determined data format. 10. The natural language pattern dictionary creating apparatus according to claim 9, comprising: 11. A natural language for creating a pattern dictionary applied to a natural language processing device that performs syntax analysis and / or syntax generation using a natural language pattern composed of a sequence of language names, left and right sides. In the pattern dictionary creation method, on the left side and / or the right side, there is a constraint by a feature, and there may be central element information that defines a central pattern element at the time of feature propagation, and all natural language patterns already described in text data A feature information conversion process for converting feature constraint information of a natural language pattern read from a source dictionary storing the feature dictionary into feature constraint data in a format in which a logical operation can be easily performed, and storing the feature constraint data in the pattern dictionary. How to create a natural language pattern dictionary. 12. The constraint information format conversion processing determines a data format of a format in which a logical operation is easy to execute, based on feature information defined in advance as feature names and feature values used for constraints. A feature constraint data format determining step, and a conversion step of converting feature constraint information of a natural language pattern into feature constraint data in a format in which a logical operation can be easily executed in accordance with the determined data format. 12. The method for creating a natural language pattern dictionary according to item 11.