JPH05298355A - Natural language sentence analyzer - Google Patents

Abstract

PURPOSE:To attain machine translation without requiring the knowledge of an objective language. CONSTITUTION:A natural language senstence is inputted from an input part 1 and converted into an aggregate of analytical structure by a syntax analyzing part 3. The analyzing part 3 has an analytical structure feature extracting part for extracting features sufficiently for distinguishing other analytical structure from a certain analytical structure. An interactive processing part 4 outputs one or plural features out of the aggregate of features extracted by the extracting part to an operator as questions, solves the various meanings of the analytical structure by using answers from the operator and executes interaction based upon the entropy of features.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a natural language sentence analysis device, and a natural language sentence syntax analysis device used in the realization of a system such as a machine translation device for inputting a natural language sentence. The present invention relates to an interactive syntax parsing device.

[0002]

2. Description of the Related Art Many natural language sentences have ambiguous meanings. For example: (a) I read the book I bought yesterday. Regarding, it is ambiguous whether "I bought yesterday" or "I read yesterday". This ambiguity is called "dependency" ambiguity. In other words, it means that the party to which "Kiyo" belongs is ambiguous. Also, (b) eat fish. Is ambiguous as to whether "the fish eats something" or "something eats the fish". This ambiguity is called "role" ambiguity. In other words, it is unclear whether the role of "fish" for "eat" is the subject or the object.

In the case of human beings, it is considered that such ambiguity in natural language eliminates the utterance situation, intonation, and how to make a space as information. However, in a natural language parsing device used for machine translation or the like, since the character string data from the input unit is the input information, it is difficult to eliminate the ambiguity as described above. Therefore, many proposals have been made for the technique in which the operator intervenes in the natural language sentence analysis device to obtain the result intended by the operator. Also in the machine translation technology, which is an application field of the natural language sentence analysis device, for example, JP-A-63-300360, JP-A-59-140582, and JP-A-61-18073.
There is a bulletin, etc. Hereinafter, a technique for narrowing down the analysis result of such ambiguous natural language sentence is referred to as a technique for “elimination of ambiguous”.

First, in order to improve the efficiency of the translation / editing work, the one disclosed in Japanese Patent Laid-Open No. 63-300360 is
The editing control unit responds to edit command information issued during translation continuation in response to edit function information issued during translation continuation, while the translation unit continues translation and responds to edit command information issued during translation continuation. Has a function of editing a translated sentence that has already been completed, and such a conventional technique and a similar technique are called a post-edit method, and the natural language sentence analysis device forces one of the possible analysis results. , And the conversion generator of the machine translation device converts the unique solution into the target language. The correction of the error in the translation result due to the incorrect selection of the analysis device is performed by editing the target language itself.

In Japanese Patent Laid-Open No. 59-140582, there is a possibility that ambiguity may occur in natural language analysis by classifying cases of polysemy in advance. Translation processing is performed by inserting instruction data in advance in the relevant part of the input language according to an instruction method that eliminates ambiguity before the parsing processing. Is called the pre-edit method, and ambiguous is resolved by adding information for disambiguation to the input sentence input to the natural language sentence analysis device.

Further, the one disclosed in Japanese Patent Laid-Open No. 61-18073 performs proofreading at the end of the analysis of the first language in the translation process, and displays the analysis result as an analysis tree, and the node number thereof, for example. Proofreading is performed by using a macro command for a series of frequently used proofreadings. Such a conventional technique and a similar technique are called an interedit method. Indicates to the operator the analysis result or intermediate state of the input sentence,
If the output is incorrect, the operator corrects it by himself, and the corrected result becomes the conversion generation processing target.

Further, as an interactive natural language sentence analysis device which resolves ambiguousness by a dialog with an operator, for example,
"On a method of interactive translation" (Aoyama, 4 members, IEICE Technical Report, NLC90-14, P.17-
24, 1990). This is called an interactive method. In the above-mentioned conventional technology, the operator tried to resolve the ambiguity by the editing method, but in the interactive method, when the natural language sentence analysis device detects the ambiguity, the operator is asked for information for the resolution and the operation is performed. It is solved by the person who answers it.

On the other hand, when parsing a natural language sentence,
Precise analysis is difficult only 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 there is a vocabulary functional grammar in the framework that describes the grammar complementarily by these two data structures. This is `` The Mental Representation of
Grammatical Relations, '' (Bresnan, J. The MIT Pres
s, 1982). The post-edit method has a problem that knowledge of the target language is required to edit the target language. For example, if a Japanese person wants to use the Japanese-English machine translation system, he or she will also need to have sufficient knowledge of English (No. 1
Problem).

Further, the preedit method does not require knowledge of the target language, but the operator has to predict the ambiguity of the input sentence. In order to add enough information to resolve the ambiguity, a lot of information must be added to each part of the input sentence. Moreover, it is difficult to add the minimum necessary information to obtain the analysis result intended by the operator. That is, in order to obtain the intended analysis result, redundant information is inevitably added, which is wasteful. Further, the work of pre-editing itself is complicated (second problem).

In the inter-editor method, a tree structure or a similar structure is displayed as an analysis structure or an intermediate structure of the input sentence, and the operator corrects and confirms the display. However, in this method, even if a complicated structure relating to an input sentence is shown, the operator does not know, and the editing work of the tree structure and the like is complicated (third problem). As described above, the conventional interactive method is complicated because all the ambiguous meanings generated in each analysis phase of the natural language sentence analysis apparatus are eliminated on the spot. In addition, the number of dialogues was increased in order to resolve the ambiguity in the order in which they occurred, which placed a heavy burden on the operator. Furthermore, since it is a grammatical question, the burden on the operator who has insufficient grammatical knowledge is heavy. For example, if the question "re" is used in the sentence "Bought book is ..." in Japanese is "passive" or "respect", the operator's answer will be lenient. .. In the conventional dialogue system, such a problem has not been taken into consideration (fourth problem).

[0011]

[Object] The present invention has been made in view of the above-mentioned circumstances, and when a natural language sentence analysis device detects an ambiguous word when analyzing an input sentence, it is most efficient to provide information sufficient to eliminate the ambiguous word. Ask questions in order. The dialogue by "asking questions in the most efficient order" is called the maximum efficiency dialogue. Questions are "dependency" questions about input sentences and "role" questions of syntax elements. Therefore, the question content solves the first problem regardless of the target language, and when the polysemy is detected, the dialogue is conducted in the most efficient order, so that the dialogue eliminates the polysemy of the input sentence. It is necessary to solve the second problem, and the question is related to the role of dependency and syntax element, and eliminates the interface that is difficult for the operator to understand such as showing the parse structure. Therefore, it is possible to solve the third problem, and to carry out the dialogue in the most efficient order in consideration of all the meanings regardless of the phase of analysis of the input sentence. Solve the complexity of. Further, for each question, information about the ease of the question, in other words, the certainty of the operator's answer to the question is held, and the fourth problem is solved by reflecting the information in the maximum efficiency dialogue. It was made for the purpose of providing the natural language sentence analysis device.

[0012]

In order to achieve the above object, the present invention provides (1)
An input unit that takes a natural language sentence as an input, a parsing unit that converts the input sentence by the input unit into a set of analytic structures, and the analytic structure from each analytic structure by the analytic unit to another analytic structure. An analysis structure feature extraction unit that extracts enough features to distinguish them, and one or more questions are asked from the set of extracted features for the set of analysis structures, and the answer is used to resolve the ambiguity of the analysis structure. And an output unit for outputting an analysis result by the dialogue processing unit, wherein the dialogue processing unit conducts a dialogue based on the entropy of the feature, and (2) the syntax analysis unit Calculating a value relating to the incompatibility of the analytic structure for each analytic structure, and (3) the dialogue processing unit has information regarding the certainty of the obtained answer, and the entropy is the certainty of the answer. Based on information about It is attributed calculation, or,
(4) An input unit for inputting a natural language sentence, a parsing unit for converting the input sentence by the input unit into a set of parsing structures, and another parsing structure from each parsing structure by the parsing unit An analysis structure feature extraction unit that extracts a feature sufficient to distinguish it from the analysis structure, and one or more questions are asked from the set of extracted features for the set of analysis structures, and the answer is used to determine the analysis structure. The dialogue processing unit for eliminating the ambiguity, and the output unit for outputting the analysis result by the dialogue processing unit, the dialogue processing unit performs the dialogue so as to reduce the number of elements of the set of analysis structure to the minimum, further, (5) In (4) above, the syntactic analysis unit calculates a value regarding incompatibility of the analytic structure for each analytic structure, and further,
(6) In (4) above, the dialogue processing unit has information on the certainty of the obtained answer, and when the dialogue is made so as to reduce the number of elements in the analysis structure in the morning, the certainty of the answer is obtained. It is characterized by being used for information about sex.

In the natural language sentence analysis apparatus according to the present invention, grammatical properties are expressed as constraints, and information about the grammatical strength of the constraints (the number of exceptions is small) is added to each constraint so that the grammatical phenomenon is recognized. Express. On the other hand, the analysis device constructs a structure corresponding to the functional structure in the known vocabulary functional grammar for the input sentence, and further checks some constraints in order to examine the suitability of the functional structure. Since each constraint has information on the number of exceptions, all the information on the exceptions of unfulfilled constraints are comprehensively evaluated and numerically expressed. This numerical value is the incompatibility of each functional structure and is called the penalty.

When an ambiguous (ambiguous) natural language sentence is input, a plurality of functional structures are obtained by calculating the respective penalties. Penalty is incompatibility of its functional structure,
In other words, it indicates that the possibility that the functional structure is a correct analysis result is low. In the natural language sentence analysis device according to the present invention, a structure expression vector is constructed by extracting sufficient information from each of these functional structures with penalties so as to distinguish them from other functional structures. From the set of structural expression vectors with penalties obtained in this way, questions are asked in the most efficient order to identify one vector.

In this maximum efficiency dialogue, entropy is calculated for each column of the structure expression vector, and the question having the larger entropy value is queried. According to another implementation, the question is asked from the one having a larger expected value of the number of structural expression vectors which is reduced by determining one column of the structural expression vector. Such a method of realizing maximum efficiency dialogue is called dialogue planning. Of the two dialogue plans, the former is called "maximum efficiency dialogue using entropy" and the latter is called "maximum efficiency dialogue using the number of functional structures". Furthermore, in the dialogue planning, information regarding the ease of question and the certainty of answer is added to each question for calculation. As described above, according to the dialogue plan that is most efficiently performed to identify one vector from a plurality of structural expression vectors, the structural expression vector sufficiently expresses the functional structure, so In order to specify one functional structure, the most efficient dialogue will be carried out in consideration of the grammatical knowledge of the operator.

FIG. 1 is a block diagram for explaining an embodiment of a natural language sentence analysis apparatus according to the present invention. In the figure, 1 is an input unit, 2 is an output unit, 3 is a syntax analysis unit, and 4 is a dialogue. Processing unit, 5
Is a dictionary part, 6 is a syntax rule part, 7 is a restriction part, and 8 is a functional relation part. A natural language sentence is input to the input unit 1 from an input device such as a keyboard or a storage device. Also input unit 1
Is also a means for inputting information from the operator in the interactive processing. The output unit 2 is a display device or the like for showing the analysis result or the question content in the interactive processing to the operator. This also includes the case where another module is added as an output device, as in the case of a machine translation system.

2A, 2B and 3 show a dictionary D which is a part of the contents of the dictionary part, a syntax rule R which is a part of the contents of the syntax rule part, and a part of the contents of the constraint part. 4 shows a constraint C, and FIG. 4 shows a part of the contents of the functional relation part. Each element of the dictionary D shown in FIG. 2A is a headword,
It is expressed by a vocabulary category (part of speech) and a feature list. The feature list is a list of features, and each feature is a pair of the form (feature name, feature value). Here, a feature is a feature extracted from a vocabulary. The syntax rule R shown in FIG. 2B is described by a labeled phrase structure grammar, which is an extension of the well-known phrase structure grammar notation. Each element on the right side is either a labeled non-terminal symbol or an unlabeled non-terminal symbol. For example, in FIG.
In the case of (NP; case) in rule R1 in (b), NP is a non-terminal symbol and a label case is added. In each rule, lower-case letters on the right-hand side represent lower-case letters, which represent front-terminal symbols (vocabulary categories), and upper-case letters represent non-terminal symbols other than front-terminal symbols. The label in the syntax rule represents the function name. Here, the function name is similar to the usage in the known vocabulary functional grammar, and for example, the description of the rule R1 is the same as the following R10 in the vocabulary functional grammar.

Each element of the constraint C shown in FIG. 3 is a production rule with a function name and a penalty. The notation is in the form of (feature name: penalty) constraint rules. Here, the penalty is a numerical value representing the grammatical strength of each constraint, and the larger the value, the stronger the constraint is. That is, there are few exceptions. Each constraint rule refers to information in the functional structure, especially features. Here, the functional structure is the same as the definition in a known vocabulary functional grammar, and is a recursive matrix in which a functional name is an attribute name and a functional structure is an attribute value. FIG. 7 shows an example of the functional structure.

FIG. 5 is a flow chart of the syntactic analysis unit and the dialogue processing unit of the natural language sentence analysis apparatus according to the present invention. step1 ; The syntactic analysis unit, to which the Japanese sentence "solid writing" is input from the input unit, divides the input sentence into a list of morphemes according to the morphological analysis process. For example, if (a0) He plays in a park is an input sentence, the result of morpheme division is (a1) (he: n) (ga: p) (park: n) (de: p) (play: v). Here, each morpheme is expressed in the form of (entry word: vocabulary category). In the morpheme division, a morpheme division dictionary is extracted from the dictionary unit and used.

[0020] step2; now creates a constituent structure using the contents of the syntax rules portion relative morpheme string after morphological analysis. Here, the constituent structure is similar to the usage in the known vocabulary functional grammar, and is a tree structure with labels as shown in FIG. Each label represents a function name, and the label described in the syntax rule is referred to when the syntax rule is applied to obtain the constituent structure. Although many methods for creating a constituent structure from a sequence of morphemes have been proposed, here, the CKY method is used from the bottom up. When a constituent structure is created from a morpheme string, when there are no more morphemes to be processed, the constituent structures that have reached the final state are all output from step 2, and generally there are multiple constituent structures. Whether or not the final state has been reached is determined by whether or not the root node of the constituent structure created for the entire sentence is the non-terminal symbol S (see syntax rule R0). FIG. 6 shows one of the constituent structures (a2) created from the morpheme string (a1). step3 : A functional structure is created from each of the plurality of constituent structures obtained in step2. This process follows the process of creating a functional structure from a constituent structure in a known vocabulary functional grammar. FIG. 7 shows a functional structure (a3) created from the constituent structure (a2) shown in FIG. In step 3, the functional structure action dictionary is extracted from the dictionary section and used. Since there is a one-to-one correspondence between the constituent structure and the functional structure, there is generally a plurality of functional structures in the output of step 3.

[0021] step4; performing application of constraints using the constraint application dictionary from constraints and the dictionary portion of the restriction portion for functional structures of the respective. The procedure for applying constraints is as follows. For all functions in the functional structure, all the constraints indexed by each function (while looking at the function name in constraint C) are applied, and for the constraints that are not satisfied, the penalty of the constraint is added. At this time, the functional structure with a sufficiently large penalty is not analyzed and the subsequent analysis is not performed. For the function for which constraint application has been completed, refer to the function-related part and change it to a child function of that function. If there are multiple child functions, the functional structure is copied to multiple functions. The above is further performed for the changed function name. Continue the above steps until the function name cannot be changed.

The above steps 1 to 3 are recursively performed on the recursive structure of the functional structure. Since each constraint has a function name as shown in constraint C, the above reference is easy. Moreover, since the function name is managed in a tree structure as shown in FIG. 4 in the function-related part, the above-described processing for changing each function into a child function is easy.

An example of the processing in step 4 is executed for the functional structure (a3) shown in FIG. The function name corresponding to the phrase "in the park" is initially the case given by the syntax rule R1. For this case function, the constraints indexed by case are applied as described above. That is, the constraint C0 of FIG. 3 is applied, but the main word in the functional structure of the parent is a play word, which is satisfied. Then, according to the above, the case function is changed to its child functions, subj, obj, obj2, and ob1.
In this case, since there are a plurality of child functions, the functional structure is copied so as to correspond to each of them. As an example, consider the case where the function is changed to the subj function. Regarding the subj function, the constraints C1 to C4 are applied according to the above. Within these four constraints, C3 is not satisfied because the value of the case marker feature in the subj case is "at". Therefore, this functional structure adds a penalty of 200. Since this penalty value is large enough, the functional structure in which "at the park" is the subj function fails in analysis and no further analysis is performed. On the other hand, case
The constraints C5 and C6 are applied to the functional structure in which the function is changed to the ob1 function, but the penalty value is 0 because both are satisfied, and the function name is changed.

FIG. 8 shows an example of the functional structure finally obtained by performing the above processing for the phrase "hega". Since this functional structure satisfies all the constraints, the penalty value is 0. The processing in step 4 creates one or more functional structures for one functional structure. The step
Since the input to 4 is a plurality of functional structures, the output of step 4 is a plurality of functional structures shown in FIG. 8, and the penalty value is calculated for each functional structure.

[0025] step5; to select the original functionality structure the penalty value. Specifically, one or more functional structures having sufficiently low penalty values are selected. For an unambiguous sentence such as (a0), the functional structure selected in step 5 is only one shown in FIG. 8 (penalty value =
0), therefore, there is no need to perform ambiguous resolution processing by dialogue, and the next step 6 does nothing. step6 ； The ambiguousness is not resolved by dialogue. Finally, the analysis result is output to the output unit. For the example sentence (a0), the functional structure shown in FIG. 8 is output.

The above is a processing example for an unambiguous input sentence "He plays in the park". However, (b0) Regarding reading a book that was bought yesterday, it is unclear whether "yes" relates to "buy" or "read" as described above. "Yesterday"
FIG. 9 shows a constituent structure related to “buying”. In Figure 9, the function rentai # obj is a child function of the function rentai, and the function rentai has the adnominal modifier clause (embedded sentence).
rentai # obj means that the modified noun matches the objective in the modifier clause. In this example, "book" means the purpose of "buying". FIG. 10 shows (b1) a functional structure created from the constituent structure of FIG.
Now, in (b0), since the reference of "Kiyo" was ambiguous, the functional structure after the processing of step 5 of FIG. 5 becomes two as shown in FIG. 10 and (b2) shown in FIG. Has a penalty value of 0). Here, in FIG. 11, the coincidence of the features is omitted for simplification, but it can be seen that the difference between the two functional structures is the position in the functional structure of “yes”.

FIG. 12 is a flowchart of the ambiguous resolution processing by the dialogue performed by the dialogue processing unit in step 6 of FIG. Hereinafter, each step will be described in order. step10 ; It is when there is a plurality of functional structures that the ambiguousness is resolved by dialogue. In that case, a structure expression vector with a penalty is obtained for each functional structure. here
For (b1), (b3) penalty = 0 (2, adverb, 3, rentai # obj, 4, obj, 0, head), and for (b2) (b4) penalty = 0 (4, adverb, 3) , Rentai # obj, 4, obj, 0, head). Here, for each structural expression vector, the odd-numbered elements sequentially represent the clause number of the modification destination of each clause.

For example, the first "2" in the above (b3) indicates that the phrase "Kinoyu" modifies the second phrase "Bought". Further, "0" represents that there is no clause to be modified (it is the main clause of the whole sentence). The even-numbered elements represent the role of each clause in the sentence in order from the front. Here, the "role" in the sentence is expressed by a function name. The second element “adverb” in (b3) above indicates that the role (function name) of the clause “Kinoyu” is “adverb”. Each structure expression vector has enough information to restore the functional structure using the dictionary information. Therefore, the difference in functional structure can be seen by comparing the structural expression vectors. (B3) and (b in the above example
Comparing 4), it is immediately clear that the difference between the functional structures of the two is the difference between "buy" and "read" in the "yes".

Step 11 ; Since there are a plurality of functional structures that are the targets of the interactive processing, a plurality of penalty- added structure expression vectors are obtained. After that, this vector is narrowed down to one based on the information obtained by the dialogue. At this time, the dialogue is performed by efficiently querying a sequence of vectors based on the penalty structure expression vector (maximum efficiency dialogue). With regard to the maximum efficiency dialogue, when it is narrowed down to one, the functional structure corresponding to the structure expression vector is output in step 15. step12 ; When there are a plurality of structural expression vectors, the column with the largest variation is obtained from the set of structural expression vectors. In the above example, since the elements other than the first element of the structural expression vector are the same, the first column is queried.

Step 13 ; The operator is queried for the row obtained in Step 12 above. In this case, (b5) What is the person in charge of "Yesterday"? Alternatively, (b6) the person in charge of “Yes” read “I bought” and “Read”
Which one? And so on. step 14 ; Only the structural expression vector that matches the answer obtained from the operator in step 13 is left, and the other vectors are failed in analysis. When the answer to the question of (b5) or (b6) is "buy", only the structural expression vector (b4) is left. step15 : Since the structure representation vector is narrowed down to one at the time of step 14, the functional structure (b1) is output.

Although there is the ambiguity that "I read the book I bought yesterday", the ambiguity was about the input sentence, which is one of the references of "Kino". However, (e0) I read the book he bought yesterday. about,
It is ambiguous whether "yes" is "buried" or "read", and ambiguous whether "he" is "buried" or "read". There is ambiguity. That is, in step 5 of FIG. 5, for example, three functional structures are output. The following three structure expression vectors are calculated by the dialogue processing unit from each of them. (E1) Penalty value = 0 (3, subj, 3, adverb, 4, rentai # obj, 5, obj, 0, head) (e2) Penalty value = 10 (5, subj, 3, adverb, 4, rentai # obj, 5, obj, 0, head) (e3) Penalty value = 15 (5, subj, 5, adverb, 4, rentai # obj, 5, obj, 0, head).

The meaning of this result is that, as the interpretation of (e0), it is most likely that (e4) "someone (probably the speaker) read the book he bought yesterday." Recognize. (E2) corresponds to the interpretation that "he read the book that (someone) bought yesterday.", And (e3) that he read the book that (someone else) bought yesterday. It can be seen from the penalty value that these two interpretations are less likely than the interpretation of (e4). In this case, as the dialogue processing unit, based on the penalty value of each structure expression vector, (e5) What is the contact of "he is"? (E6) What is the person in charge of "Kinou"? From among the possible questions, select (e5) and ask the operator.
The reason is as follows.

Based on the penalty value, the dialogue processing unit expects that the functional structure having the structure expression vector (e1) is correct. Therefore, the operator's reply is considered to be an answer that matches (e1). If you ask (e6) first, you will probably get the answer that you have bought. In that case, if you do not continue to ask question (e5) (e5
1) It is unknown whether the functional structure corresponding to (e2) is intended by the operator. On the other hand, if the question (e5) is asked first and the answer is "buy", it is understood that the functional structure corresponding to (e1) is intended by the operator regardless of (e6). That is, in maximum efficiency dialogue,
It is better to ask (e5) first.

The above will be described in more detail using mathematical expressions. The set of functional structures is Γ = {S _i } (i = 1, 2, ... N) (1). Each functional structure (eg, S ₃ ) can be uniquely vector-expressed based on its syntactic structure and semantic properties.

[0035]

[Equation 1]

M is the dimension of the vector, indicating that each functional structure can be expressed in M dimensions. Furthermore, each functional structure has a penalty. If the negativeness of the functional structure S _i is e _i (0 < e <∞), the relative probability p of the functional structure is

[0037]

[Equation 2]

[0038] Here base is the "bottom" when entropy is returned to probability, and base> 1. This relative probability p is normalized so that it is 1 in the entire set of functional structures.

[0039]

[Equation 3]

Therefore, the ambiguity of the solution set is completely expressed by the N × M matrix x _ij which uniquely expresses the heat of the functional structure and the Nth-order vector [p ′ _i ] which represents the penalty. 1
The information that can be obtained by one interaction is one column of the matrix (eg [x _i5 ]). If the dialog reveals that x _i5 = v, this is expressed as a constraint on the functional structure. For example, a functional structure S _i such that x _i5 ≠ v
Is an analysis failure. By the way, in the maximum efficiency dialogue using entropy, entropy for each column is calculated, and the value of the column with the highest entropy is obtained by the dialogue. For example, the entropy E ₅ of the Nth order vector [x _i5 ] representing the fifth column can be calculated as follows. The variation of the value of [x _i5 ] is
Let {v ₁ , v ₂ , ..., V _L }. Here, 1 < L < N. At this time, the probability that the matrix component x _i5 takes the value v _k is

[0041]

[Equation 4]

It is Using this P,

[0043]

[Equation 5]

It becomes Further, in the maximum efficiency dialogue using the number of functional structures, the dialogue value is obtained such that the expected value of the number of functional structures which can be reduced by knowing the value of the column is maximized. Specifically, the equation (6) in the previous section is

[0045]

[Equation 6]

It is assumed that Σ _{xi5 = vk} 1 means that the value of x is v _k
It just counts the number of functional structures. The above is the mathematical definition of the maximum efficiency dialogue. It is mathematically assured that the maximum efficiency dialogue is used for the most efficient dialogue when there are a plurality of analysis results of the input sentence.

Next, (f0) I read the book bought yesterday. Regarding, it is ambiguous whether "yes" is related to "bought" or "read". It is also ambiguous whether "bought" means "passive" or "respect". Therefore, the syntactic analysis unit outputs, for example, four functional structures. The following four structure expression vectors are calculated by the dialogue processing unit from each of them. (F1) Penalty value = 0 (2, adverb, 3, rentai # pass # obj, 4, obj, 0, head) (f2) Penalty value = 10 (4, adverb, 3, rentai # pass # obj, 4, obj, 0, head) (f3) Penalty value = 0 (2, adverb, 3, rentai # reg # obj, 4, obj, 0, head) (f4) Penalty value = 10 (4, adverb, 3, rentai # (reg # obj, 4, obj, 0, head)

Here, the function name rentai # pass # obj is the passive usage of the phrase “Bought” and “Book”, and the “Book” is the object of “Buy”. It means that. Also, the function name rentai # reg # obj is the phrase “Bought”
Indicates that the book is adnominally modified, meaning "respect", and that "book" is the object of "buy".

The most efficient question for choosing one of these four structural representation vectors is that the maximum efficiency dialogue considering only the penalty value indicates that the first column of the vector, namely the "yes" The ambiguity and the ambiguity of the meaning of the fourth column, that is, "ru" are considered with the same priority. However, since the dialogue processing unit according to the embodiment of the present invention has a question pattern information amount table as shown in FIG. 14 in advance, referring to this table, as compared with the question "inquiring about the person in charge", It turns out that the question has only 0.2 minutes of information. So, what is the contact point of (f5) “Kiyo”? Ask. It is more efficient to answer this question first. After that, the meaning of “ru” may be asked, or the question may be stopped because the amount of information for the question asking the meaning of “ru” is sufficiently small. When the question is stopped, it means that it is not necessary for the natural language grammar solver to separate the meaning of "ru" in consideration of the grammatical knowledge of the operator. In this case, the output of the analysis device is compressed, as shown in FIG. 13, without determining the meaning of "ru". If necessary, the meaning of “yes” may be determined to be “passive” according to a known default process.

[0050]

[Effect] As is apparent from the above description, the present invention has the following effects. (1) Since it does not require knowledge of the target language, it can be applied to, for example, a Japanese-English machine translation system in which a Japanese operator is used. (2) Since the information added by the operator is the minimum necessary to resolve the ambiguity, it is not necessary to add unnecessary information, and the order of the questions is based on the maximum efficiency dialogue. The burden on the person can be reduced. (3) Since the content of the information added by the operator is easy to understand, the question content is related to the dependency structure of the input sentence,
This is related to the role of the syntax element, and it is possible to eliminate an interface which is difficult for the operator to understand, such as showing a parsing structure or an intermediate structure. Further, it is possible to place less importance on answers to questions that are difficult for the operator to understand. (4) Since the operator does not need more grammatical knowledge than is necessary, the simplicity of the question is calculated for each question, and the maximum efficiency dialogue is performed so that the easier the question is, the easier the question is asked earlier. Questions that are difficult for the operator to understand are asked when performing more detailed analysis. The analysis device attaches more importance to the operator's answer to a simple question.

[Brief description of drawings]

FIG. 1 is a configuration diagram for explaining an embodiment of a natural language sentence analysis device according to the present invention.

FIG. 2 is a diagram showing a dictionary and syntax rules according to the present invention.

FIG. 3 is a diagram showing constraints according to the present invention.

FIG. 4 is a diagram showing the contents of a functional relation part according to the present invention.

FIG. 5 is a flowchart of a processing unit of the natural language sentence analysis device according to the present invention.

FIG. 6 is a diagram showing a constituent structure according to the present invention.

FIG. 7 is a diagram showing a functional structure according to the present invention.

FIG. 8 is a diagram showing a functional structure according to the present invention.

FIG. 9 is a diagram showing another constituent structure of the present invention.

FIG. 10 is a diagram showing another functional structure of the present invention.

FIG. 11 is a diagram showing another functional structure of the present invention.

FIG. 12 is a diagram showing a flowchart of ambiguous resolution processing by dialogue according to the present invention.

FIG. 13 is a diagram showing still another functional structure of the present invention.

FIG. 14 is a diagram showing a question pattern information amount table of the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Input part, 2 ... Output part, 3 ... Syntax analysis part, 4 ... Dialog processing part, 5 ... Dictionary part, 6 ... Syntax rule part, 7 ... Restriction part, 8 ...
Functional section.

Claims (6)

[Claims] 1. An analytic structure is obtained from an input unit for inputting a natural language sentence, a parsing unit for converting an input sentence by the input unit into a set of analytic structures, and an analytic structure for each analytic structure by the parse unit. An analysis structure feature extraction unit that extracts a feature sufficient to distinguish it from other analysis structures, and one or more questions are asked from the set of features extracted for the set of analysis structures, and the analysis is performed using the answer. A dialogue processing unit that eliminates the ambiguity of the structure,
A natural language sentence analysis device comprising: an output unit that outputs an analysis result by the dialogue processing unit, wherein the dialogue processing unit performs a dialogue based on the entropy of the feature. 2. The natural language sentence analysis device according to claim 1, wherein the syntactic analysis unit calculates a value regarding incompatibility of the analysis structure for each analysis structure. 3. The natural language according to claim 1, wherein the dialogue processing unit has information regarding the certainty of the obtained answer, and calculates the entropy based on the information regarding the certainty of the answer. Sentence analyzer. 4. An analytic structure is selected from an input unit for inputting a natural language sentence, a parsing unit for converting the input sentence by the input unit into a set of analytic structures, and an analytic structure for each analytic structure by the parse unit. An analysis structure feature extraction unit that extracts a feature sufficient to distinguish it from other analysis structures, and one or more questions are asked from the set of features extracted for the set of analysis structures, and the analysis is performed using the answer. A dialogue processing unit that eliminates the ambiguity of the structure,
An apparatus for outputting a natural language sentence, comprising: an output section for outputting an analysis result by the dialogue processing section, wherein the dialogue processing section conducts a dialogue so as to reduce the number of elements of a set of analysis structures to a minimum. 5. The natural language sentence analysis apparatus according to claim 4, wherein the syntactic analysis unit calculates a value regarding incompatibility of the analysis structure for each analysis structure. 6. The dialogue processing unit has information on the certainty of the obtained answer, and when the dialogue is made so as to reduce the number of elements of the analysis structure in the morning, the information on the certainty of the answer is provided. 5. The natural language sentence analysis device according to claim 4, wherein the natural language sentence analysis device is used.