JPH0415750A - Analyzing device for natural language sentence - Google Patents

Abstract

PURPOSE:To facilitate the operation of an analyzing device for natural language sentence by extracting the information to discriminate the analysis structure of a natural language sentence from other analysis structures. CONSTITUTION:An input sentence received from an input part 2 which inputs the natural language sentences is divided into the morphemes with use of a morpheme division dictionary of a dictionary part 4. Then plural component element structures are produced based on the contents of a syntax rule part 5 and the function structures are produced from the component element structure. In this case, a processing part 1 applies the limit by means of the limit of a limit part 6 and a limit application dictionary of the part 4 and then performs multivocal dissolution processing jobs after a conversation. Then the information is extracted for discrimination of the analysis structure of a natural language sentence from other analysis structures in an interactive form. Thus this analysis method can be applied a Japanese-English machine, translation system which is operated by a Japanese operator. Then the operator' s load is reduced.

Description

[Detailed Description of the Invention] The present invention relates to a natural language sentence analysis device, and particularly relates to a natural language sentence parsing device for a system that receives natural language sentences as input, such as a machine translation device. This article relates to an interactive syntax parsing device.

There are many sentences in natural language whose meanings are ambiguous. For example, (a) I read the book I bought yesterday.

It is unclear whether it means "I bought it yesterday" or "I read it yesterday." This ambiguity is ``required''
It is called ambiguity. In other words, the meaning of ``yesterday'' is ambiguous. Also, (b) I eat fish.

Regarding this, it is unclear whether ``the fish eats something'' or ``what eats the fish''. This ambiguity is called ``role'' ambiguity. In other words, "eat" in "r fish".
This means that it is ambiguous whether the role of the word is the subject or the object.

For humans, these ambiguities in natural language are
In speech, it is thought that the situation of the utterance, intonation, pauses, etc. are resolved as information.
In a syntax analysis device used for machine translation, etc., only character string data input through an input unit is considered as input data, so it is difficult to eliminate the above-mentioned ambiguity. Therefore, many proposals have been made for techniques in which the operator intervenes to obtain the results intended by the operator. Regarding machine translation processing, which is an application field of natural language sentence analysis devices 1 For example, Japanese Patent Application Laid-Open No. 63-300360
No. 59-140582, Japanese Patent Application Laid-open No. 1982-140582
-18073, etc. Hereinafter, the technology for narrowing down the analysis results of such ambiguous sentences will be referred to as the technology for "resolving ambiguity."

First, the prior art described in Japanese Patent Application Laid-Open No. 63-300360 and similar technologies are called post-edit methods, in which a natural language sentence analysis device forcibly selects one of the possible analysis results. The conversion generation unit of the machine translation device converts the solution into the target language. Errors in translation results due to incorrect selection of an analysis device can be corrected by editing the target language itself.

The prior art described in Japanese Patent Application Laid-Open No. 59-140582 and similar technologies are called pre-editing methods, which remove ambiguity by adding information to resolve ambiguity to an input sentence input to a natural language sentence analysis device. Eliminate.

The prior art described in JP-A No. 61-18073 and similar technologies are called interedit methods, in which a natural language sentence analysis device shows the analysis result of the input sentence to the operator, and the operator can check if the analysis result is incorrect. If so, it will be corrected by itself, and the corrected result will be processed for conversion generation.

On the other hand, when parsing a natural language sentence, it is difficult to perform precise analysis just by constructing a structural data structure such as a parse tree. Therefore, in addition to the structural data structure, a semantic data structure is constructed, and as a grammar theory that provides a framework for describing grammar complementary to these two data structures, the Mental Representat
ion of GrammaticalRelation
ns, J (Bresnan, J, MIT Pres.
There is a word-to-function grammar disclosed in J. S., 1982).

The problem with the post-editing method is that knowledge of the target language is required in order to edit the target language. For example, if a Japanese person wants to use a Japanese-English machine translation device, knowledge of English is also required.

Furthermore, although the pre-editing method does not require knowledge of the target language, it is necessary to anticipate the ambiguity of the input sentence. In order to add enough information to resolve ambiguity, it is necessary to add a lot of information to various parts of the input sentence, and for this to be necessary and sufficient, sufficient knowledge of the language and system processing are required. requires sufficient knowledge of In other words, it is difficult to add the minimum amount of information necessary to obtain the analysis results intended by the operator.
This inevitably adds redundant information, which is wasteful. Further, the work itself of adding information to each input sentence is complicated.

According to the above distinction, the present invention can be said to be a syntactic analysis device used in the interedit system. However, the above-mentioned Japanese Patent Application Publication No. 6
According to Japanese Patent No. 1-18073, a tree structure or a similar structure is displayed as an analysis structure of an input sentence, and the operator corrects and confirms the display. However, there is a problem that the operator cannot understand even if the complicated structure of the input sentence is shown to him.

The present invention was made in view of the above-mentioned circumstances.
When the natural language sentence analysis device detects ambiguity in the input sentence when analyzing the input sentence, it asks for enough information to resolve the ambiguity at that point in the most efficient order, and this question Since the questions are "receive" or "role" questions, knowledge of the target language is not required, and the questions are asked when ambiguity is detected, and the questions are asked in the most efficient order, so the questions are ambiguous. The content of the question is necessary and sufficient to solve the problem, and the question is related to dependencies related to the input sentence.

It concerns the role of syntactic elements, and was developed with the purpose of providing a natural language sentence analysis device that can eliminate interfaces that are difficult for operators to understand, such as showing analysis structures.

In order to achieve the above object, the present invention includes an input section for inputting a natural language sentence, and a morpheme segmentation dictionary of a dictionary section that divides the input sentence from the input section into morphemes. For the morpheme-divided morphemes, the content of the syntactic rule part is divided to create multiple constituent structures, and a functional structure is created from each of the constituent structures. When creating the functional structure, the constraint part is It consists of a processing section that applies constraints using the constraints from and the constraint application dictionary of the dictionary section, and then performs ambiguity resolution processing through dialogue, and an output section that outputs the analysis results of the processing section. This method is characterized by extracting one or more pieces of information that distinguishes the parsing structure from other parsing structures from the parsing structure of a linguistic sentence, and performing one or more of the pieces of information in an interactive manner.

In the natural language sentence analysis device according to the present invention, grammatical properties are expressed as constraints, and information regarding the grammatical strength (few exceptions) of each constraint is added to each constraint. On the other hand, a structure corresponding to the functional structure in the known overhead functional grammar is constructed for the input sentence, and several constraints for realizing the functional structure are checked. Since each constraint has information on the number of exceptions, the information on the number of exceptions mentioned above is comprehensively evaluated for constraints that are not satisfied, the evaluation is expressed numerically, and it is regarded as an incompatibility of the functional structure. This incompatibility is called a penalty.

When an ambiguous sentence is input, a plurality of functional structures are obtained by calculating penalties for each. The penalty expresses the low probability that the functional structure is the correct analysis result. In the present invention, a structure representation vector is constructed by extracting information sufficient to distinguish each of these plurality of functional structures from other analysis results. Here, the structure expression vectors extracted from each functional structure sufficiently express each functional structure. The most efficient question for identifying one vector from the multiple penalized structure representation vectors obtained in this way is that the structure representation vector sufficiently represents the functional structure. This is the most efficient question for identifying a single functional structure. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining an embodiment of a natural language sentence analysis device according to the present invention, in which 1 is a processing section, 2 is an input section, 3 is an output section, 4 is a dictionary section, and 5 is a syntax rule section, 6 is a constraint section, and 7 is a function relationship section.

The input unit 2 includes input from an input device such as a keyboard or a storage device. The input unit 2 also serves as a means for inputting information from the operator during ambiguity resolution processing through dialogue.

The output unit 3 is a display or the like for conveying the analysis result or question content to the operator, and includes cases where an application using a syntax analysis device attempts to use the analysis result, such as in the case of machine translation.

FIGS. 2(a) to 2(c) show the dictionary and the syntax rule R1 constraint C, respectively.

Each element of the dictionary shown in FIG. 2(a) is expressed as a headword, word category (part of speech), and feature list. The feature list is a list of features, and each feature is a pair of the form (feature name, feature value). Here, the feature is the extracted property of the word.

The syntactic rule R shown in FIG. 2(b) is written in a labeled phrase structure grammar that is an extension of the notation of a known phrase structure grammar. Each element on the right-hand side is either a labeled nonterminal or an unlabeled nonterminal.

In the case of (N P ; case) in R1, NP is a non-terminal symbol and the label case is added. Also, the non-terminal symbol on the right side, the lowercase alphabetic character, is the front terminal symbol (word category),
Uppercase letters are nonterminal symbols other than preterminals. Also, the labels in the syntax rules represent function names. Here, the term "function name" is used in the same manner as in the known language function grammar. That is, the description of R1 is the same as the following Rho in the language function grammar.

Each element of constraint C shown in FIG. 2(c) is a production rule with a function name and a penalty.

The notation is in the form of a (Function name: Penalty) constraint rule. Here, the penalty is a numerical value representing the grammatical strength of each constraint, and the larger the value, the stronger the grammatical constraint is, that is, the fewer exceptions there are. Each constraint rule references information in the functional structure, specifically features from the dictionary.

Here, the functional structure is used in the same manner as in the known overhead function grammar, and is a recursive matrix in which one function name is an attribute name and the functional structure is an attribute value. FIG. 6 shows an example of the functional structure.

FIG. 3 is a flowchart of the processing section of the natural language sentence analysis device according to the present invention.

Kei Tanfu: The processing unit receives the Japanese sentence “solid writing” from the input unit and divides the input sentence into a list of morphemes. For example, if the input sentence is (aO) He plays in the park, then the result of morpheme segmentation is.

(al) (he: n) (ga: p) (Ko II: n) (
So: p) (play: v). Here, each form is expressed in the form (headword 2 miscellaneous expenses range ml). In morpheme segmentation, a dictionary for morpheme segmentation is extracted from the dictionary section and used.

Next, a configuration structure is created using the contents of the syntax rule section for the sequence of morphemes after the morphological analysis. Here, the compositional structure is used in the same way as in the well-known expense function grammar, and is a labeled tree structure as shown in FIG. Labels attached to the tree structure represent function names, and can be obtained by referring to the labels described in the syntax rules when applying the syntax rules.

Regarding the process of creating a compositional structure from a sequence of morphemes,
Although many methods have been proposed and are known, here, the CKY method is used from the bottom up. When creating a compositional structure from a morpheme sequence, when there are no more morphemes to process, all compositional structures that have reached the final state are processed in 5 steps.
2, and there are generally multiple solutions.

Here, whether the final state has been reached is determined by whether the root node of the compositional structure created for the entire sentence is a non-terminal symbol S.

One of the construction structures created from the morpheme sequence of (al) (
a2) is shown in FIG.

%; A functional structure is created from each of the plurality of constituent structures obtained in step 2 above. This process may follow the process of creating a functional structure from a constituent structure in the well-known S-search functional grammar, or it may be possible to follow the process of creating a functional structure from a constituent structure in a well-known S-search functional grammar, or to write a child node without a label in a constituent structure as a main word in the entry section, and to write it to a labeled child node. For a node, a functional structure may be created recursively by using its label as an attribute name and the functional structure of a child node as an attribute value. The functional structure (a3) created from the component structure (a2) shown in FIG.
As shown in the figure. In this step 3, the functional structure action dictionary is extracted from the dictionary section and used. The contents of this dictionary are part of the feature list in the dictionary. Since there is a one-to-one correspondence between the configuration structure and the functional structure, there are multiple solutions for the functional structure in the output of 5tep 3.

Katsura Tori: Constraints are applied to the functional structure using the constraints from the constraint section and the constraint application dictionary from the dictionary section.

The procedure for applying constraints is as follows.

■For all functions in the functional structure, all constraints indexed by each function are applied, and for constraints that are not satisfied, the penalty of that constraint is added. At this time again,
Functional structures for which the penalty is sufficiently large are treated as analysis failures and no further analysis is performed.

■For a function to which constraints have been applied, refer to the function-related section and change it to a child function of that function. If there are multiple child functions at this time, the functional structure is copied to multiple functions.

■Repeat steps (2) and (2) above for the changed function name. Continue these ■, ■, and ■ until the function name cannot be changed.

The above steps (1) to (2) are performed recursively for the recursive structure of the 'M&N structure. Here, each constraint has a function name and a penalty as shown in constraint C, so it is easy to refer to. Furthermore, since the function names are managed in a tree structure as shown in FIG. 4 in the function-related section, it is easy to change each function to a child function.

An example of the processing in step 4 is executed for the functional structure (C3) shown in FIG. The function name corresponding to the phrase "in the park" is initially a case given by the syntax rule R1. Regarding this Ca5e function, constraint CO is applied according to the above-mentioned item (2), but it is satisfied because the main word in the parent functional structure is the term "play". Next, the case function is changed to its child functions 5ubj, obj, obj2°obl according to the above item (2). In this case, since there are multiple child functions, the functional structure is copied to correspond to each child function. For example, when the function is changed to 5ubj, constraints 01 to c4 are applied to the 5ubj function according to the above-mentioned section. Among these, C3 is not satisfied because the value of the case marker feature in the 5ubj case is "de". Therefore, a penalty of 200 is added to this functional structure, but this penalty is sufficiently large that "in the park" is 5u.
This functional structure, which has a bj case, results in an analysis failure and no further analysis is performed. On the other hand, constraints c5 and C6 are applied to the functional structure where the case function is changed to the ob1 function, but since both are satisfied, the penalty is 0.
The function name is further changed in step (■) above. The above process is also performed for the phrase "he", and an example of the finally obtained functional structure is shown in FIG. The penalty for this functional structure is O
It is. The process in step 4 creates a plurality of functional structures for one functional structure. Said 5tep 4
Since there are a plurality of functional structures in the distance, the output of 5tep 4 is a plurality of functional structures as shown in FIG. 7, and a penalty has been calculated for each functional structure.

Dan: The functional structure is selected based on the penalty. Specifically, one or more functional structures with sufficiently low penalties are selected.

5te for unambiguous sentences like (aO)
There is only one functional structure selected in step 5 (penalty o) as shown in FIG. 7, so there is no need to perform ambiguity resolution processing through dialogue, and nothing is done in the next step 5.

Finally, the functional structure after the process of eliminating ambiguity through dialogue is presented in the output section as an analysis result.

For the example sentence (al), the functional structure shown in FIG. 7 is output.

The above was about the unambiguous input sentence ``He plays in the park.'' but.

(bO) As for "I read the book I bought yesterday," as mentioned above, it is unclear whether "yesterday" relates to "bought" or "read." Figure 8 shows the constituent structure of ``yesterday'' related to ``bought''.

In Figure 8, the function rentai#obj is the function renta.
It is a child function of i, and the function rentai is an adnominal modification clause (
(embedded sentence), the function rentai#obj means that it matches the objective case in the modified cylinder. In this example, ``1 book'' means that it is the objective case of ``bought.'' A functional structure created from the assembled element structure of FIG. 8 is shown in FIG. 9 (bl). Now, in (bO), the dependency of "yesterday" was ambiguous, so the functional structure after processing shown in Fig. 7 is obtained as shown in Fig. 9 and that shown in Fig. 10 (b2). (Each penalty is O). Here the first
In Figure 0, the features are omitted for simplicity. It can be seen that the difference between the two functional structures is the position of ``yesterday'' in the functional structure.

FIG. 11 is a flowchart of ambiguity resolution processing through dialogue. Below, each step will be explained in order.

In case there are multiple functional structures, ambiguity is resolved through dialogue in Step 5 of Figure 3. In that case, a structure representation vector is obtained for the multifunctional structure. This means that for (bl), (b3) penalty = O (2adv 4 rentai#obj 4 rent
ai#obj6 obj 6 obj Ohead 0
head ) For (b2), (b4) penalty = 0 (6adv 4 rentai#obj 4 rent
ai#obj6 obj 6 obj Ohead 0
head). Here, the odd-numbered elements of each vector represent the morpheme number of the modifier light of each morpheme in order from the front. For example, the first “2” in (b3) above is the morpheme “
This indicates that ``yesterday'' modifies the second morpheme ``buy''. Furthermore, rQJ indicates that there is no modifying morpheme. The even-numbered elements of each vector represent the role of each morpheme in the sentence in order from the front. Here the "role" in the sentence
is expressed as a function name. The second element in (b3) above, radvJ, has the functional name of the morpheme "yesterday" as ra
dν". Here, each structure representation vector is sufficient to restore the functional structure using dictionary information, that is, it sufficiently represents the functional structure. In other words, differences in functional structure can be seen by comparing the structure expression vectors. In (b3) and (b4) of the above example, "yesterday" is related to "
You will soon realize that there is a difference between "buying" and "reading".

Since there are multiple functional structures, the structure representation vector is based on multiple sides. From now on, we will narrow down this vector to one based on the information obtained through dialogue.

If the number is narrowed down to one, a functional structure corresponding to the structure expression vector is output in step 515.

If there are multiple structure expression vectors, find the sequence with the largest variation from the set of structure expression vectors. In the above example, the elements other than the first element are the same, so it can be seen that it is the first column.

7: Ask the operator about the column found in step 12 above. In the example above.

(b5) Who is responsible for "r yesterday"? or.

(b6) Is "r yesterday" related to "buying" or "reading"? Just ask.

Only the structure expression vectors that match the answer to the question obtained in step 5 above are retained, and the other vectors are treated as failures in analysis. If the answer to question (b5) or (b6) is "buy," only the structure expression vector (b4) is left.

扛肚上I: Since the structure expression vector has been narrowed down to one at the time of step 14, the functional structure (bl) corresponding to the structure expression vector is output.

Shyness According to one aspect of the present invention, as is clear from the above description.

It has the following effects.

(1) Since knowledge of the target language is not required, it can be applied to, for example, a Japanese-English machine translation system operated by a Japanese person.

(2) The information that the operator should add is necessary and sufficient to eliminate ambiguity, so there is no need to add unnecessary information, and the order of questions is the most efficient.
The burden on the operator can be reduced.

(3) The content of the information that the operator should add is easy to understand, so
The content of the questions is related to dependencies regarding input sentences and the roles of syntactic elements, and it is possible to eliminate interfaces that are difficult for the operator to understand, such as showing analysis structures.

[Brief explanation of drawings]

FIG. 1 is a block diagram for explaining an embodiment of a natural language sentence analysis device according to the present invention, and FIG. 2 shows a dictionary, syntax rules,
3 is a flowchart of the processing section of the natural language sentence analysis device according to the present invention; FIG. 4 is a diagram showing the contents of the functional relationship section; FIG. 5 is a diagram showing the constituent structure; 6th
7 is a diagram showing the functional structure, FIG. 8 is a diagram showing other component structures, FIGS. 9 and 10 are diagrams showing other functional structures, and FIG. 11 is a diagram showing dialogue. It is a figure which shows the flowchart of the ambiguity resolution process by. 1... Processing section, 2... Input section, 3... Output section, 4
... Dictionary section, syntax rule section, 6. Constraint section, 7. Functional relation section.

Claims (1)

[Claims] 1. An input section for inputting natural language sentences, and a morpheme segmentation of the input sentence from the input section using a morpheme segmentation dictionary in the dictionary section, and the content of the syntax rule section for the morpheme segmentation. A plurality of constituent structures are created using the above-mentioned constituent structures, and a functional structure is created from each of the constituent structures, and when creating the functional structure, constraints from the constraint section and a dictionary for constraint application of the dictionary section are used. It consists of a processing section that applies constraints and then performs ambiguity resolution processing through dialogue, and an output section that outputs the analysis results of the processing section. A natural language sentence analysis device characterized by extracting one or more pieces of distinguishing information and performing one or more of the pieces of information in an interactive format.