JP2000112936A - Language processor and word meaning deciding device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly precisely supplement an abbreviated part and to decide a word meaning in a translation device even if an error exists in an input character string. SOLUTION: Deciding tree learning parts 12-1 to 12-M inductively and mechanically learn an abbreviated part which is to be supplemented from text data where an abbreviated element is already supplemented in a text data memory 21 and deciding trees are obtained for respective attribute tables. Abbreviated element supplement parts 14-1 to 14-M execute abbreviation element supplement processings by using the corresponding deciding trees and decide the abbreviated element to be supplemented. An abbreviated element supplement selection part 15 selects the abbreviated element having maximum frequency and outputs a character string where the abbreviated element is supplemented. Knowledge obtained from the corpus of text data is expressed by the respective deciding trees for the respective attribute tables and is stored in deciding tree file memories 23-1 to 23-M.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for automatically omitting a case element in a Japanese dialogue sentence using an example stored in a Japanese text database including information on the omitted pronoun and its position. The present invention relates to a language processing device such as a Japanese abbreviation element complementing device for complementing a language and a meaning determining device for a translation device.

[0002]

2. Description of the Related Art In the processing of a Japanese dialogue sentence in which the display of a subject or object is not obligatory, it is important to supplement these omitted elements (non-explicit elements) called "zero pronouns". In particular, when translating from Japanese to English or German, the complementing process is indispensable.

[0003] As one of Japanese abbreviation complementing techniques, there have been proposed several techniques using information on appearing words.
For example, prior art document 1, "Maki Murata et al.," Estimation of referents, pronouns, and zero pronouns in Japanese sentences using examples and surface expressions ", (Natural Language Processing, Vol. 4,
No. 1, pp. 87-109, 1997 "(hereinafter referred to as"
This is referred to as a first conventional example. ), All the information necessary for abbreviated completion is manually ruled and scored.
A technique is proposed in which the candidate with the highest score is designated as an instruction target.

[0004] Prior Art Document 2 "Tatsunori Mori et al.," Zero-pronoun anaphora in Japanese manuals based on properties of linguistic expressions ", Information Processing Society of Japan, NL115.
-1, pp. 1-8, 1996 "(hereinafter, referred to as a second conventional example), two of the pragmatic constraints of the linguistic expression itself and the pragmatic knowledge of the target sentence are manually extracted, and as a result, We propose a method of complementing omissions using the obtained information. In particular, in the case of machine translation processing from Japanese to English or German, supplementary processing is essential.

[0005] Furthermore, prior art document 3 "Ikuo Kudo et al.
"On the Completion Mechanism of Omission Using the Characteristics of Japanese Predicates", IEICE Transactions, Vol. J76-D-
II, No. 3, pp. 624-635, 1993)
(Hereinafter, this is referred to as a third conventional example.) A method of manually classifying predicate expressions and verbs required for omission and manually determining abbreviation complement targets for each predicate expression and verb is used. is suggesting.

[0006] There is a wide variety of information related to the omission of Japanese case elements, and it is generally considered that they are complicatedly affected and omitted. In the first conventional example, these elements are manually scored, but it is unclear whether these assigned scores are really correct. In the second conventional example and the third conventional example, each linguistic expression is classified and each item is abbreviated and complemented, but these classification methods are also arbitrary and appropriate. Is unknown. Further, in the second conventional example and the third conventional example, regarding the influence of a plurality of elements, only the priority is given unifiedly between the verb to be complemented and various end-of-sentence expressions. No consideration has been given to the interaction between the two.

[0007] In order to solve the above problems, the present inventor has proposed a supplementary process for omitting case elements in Japanese sentences.
Japanese abbreviation element completion device that automatically grasps the influence of many pieces of information necessary for completion without human intervention, and can determine and supplement omitted pronouns more accurately and automatically. Is disclosed in Japanese Patent Application No. 9-245201. This conventional Japanese abbreviation element complementing device is composed of an abbreviation element supplemented text data consisting of a character string of a Japanese spontaneous utterance sentence, in which the correct person of the abbreviation element of the sentence is complemented in advance, and a verb normal form. , Based on an attribute list of a plurality of attributes including a verb's semantic attributes, a sentence end expression, and speaker information as out-of-language information, all of each sentence of each text data and each attribute of the attribute list A generating unit that collates them for each combination and generates an attribute table indicating whether each attribute exists for each sentence and a correct personal name indicating the personal name of the omitted element; Based on the attribute table, the decision tree is selected so that the attribute having the largest difference between the entropy after division by all attributes and the entropy before division is selected, and a node divided by the selected attribute value is generated. , Generates a decision tree for determining the correct person of an abbreviated element having a tree structure in a binary tree form that is divided depending on the attribute value of each attribute by learning to update until reaching the node Based on the character string of the input Japanese natural utterance sentence, and by referring to the attribute list, performing matching for each attribute to determine an attribute value indicating whether or not each attribute exists. Correctness of the omitted element in the character string of the input natural utterance sentence using the extraction means for extracting and outputting and the attribute value output from the extraction means using the decision tree generated by the learning means And a complementing means for determining a person and outputting a character string in which the omitted element including the correct person is complemented. "

[0008]

However, in the conventional Japanese abbreviation element supplementing device, since the omitted portion is supplemented by using one decision tree, the input to the processing device such as a speech recognition result is performed. When there is a possibility that an error is included in the character string, the reliability of the output complement result is reduced, and as a result, there is a problem that accuracy may be deteriorated.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems, so that even if there is an error in an input character string, it is possible to complement an omitted part and determine the meaning of a word in a translation device with higher accuracy than in the conventional example. An object of the present invention is to provide a language processing device that can perform the processing.

[0010]

According to a first aspect of the present invention, there is provided a language processing apparatus comprising a character string of a predetermined language, wherein information for eliminating ambiguity of a sentence is complemented in advance. Each sentence of each text data and each attribute included in the attribute list are divided into a plurality of sets based on the completed text data and the attribute list of multiple attributes including the language attribute for resolving language ambiguity. For all combinations with each attribute set at the time of division, collate them and determine whether each attribute in each attribute set exists for each sentence,
And a plurality of attribute means indicating correct information for resolving language ambiguity, respectively, a plurality of generating means for each attribute set, based on a plurality of attribute tables generated by the plurality of generating means, The decision tree is selected so as to select the attribute having the largest difference between the entropy after division by all attributes in each attribute set and the entropy before division, and to generate a node divided by the selected attribute value. By learning to update until reaching the node, determine the correct answer information to resolve the ambiguity of the language, which has a tree structure of a binary tree format that is divided depending on the attribute value of each attribute A plurality of learning means for generating a decision tree for each attribute set, and a matching for each attribute by referring to the attribute list based on the input character string in a predetermined language. Extracting means for extracting and outputting an attribute value indicating whether or not each attribute is present, and using each of the decision trees generated by the plurality of learning means for the attribute value output from the extracting means. A plurality of complementing means for determining correct information for resolving language ambiguity in the input character string and outputting the information together with the frequency, and correct information output from the plurality of complementary means and the frequency thereof. And a complement selecting means for selecting a predetermined number of correct answer information having a maximum frequency or a higher frequency based on the information and outputting a character string supplemented with information including the correct answer information.

Further, in the language processing apparatus according to the second aspect, the language processing apparatus according to the first aspect is a Japanese abbreviation element complementing apparatus, and the information-completed text data is:
Omitted element-completed text data consisting of a character string of a Japanese natural utterance sentence, in which the correct person of the omitted element of the sentence has been completed in advance, and the attribute list includes the normal form of the verb, the semantic attribute of the verb, It is an attribute list of a plurality of attributes including an end-of-sentence expression and speaker information as out-of-language information. The attribute table indicates whether each attribute exists for each sentence, and what abbreviation element is the first person. Is an attribute table indicating correct correct person names, each of the decision trees is a decision tree for determining correct correct person names of omitted elements, and each of the complementing means is a correct correct person name of omitted elements in the input character string. Is determined, and the complement selection means selects a predetermined number of correct personal names having the highest frequency or higher frequency, and outputs a character string in which the omitted element including the correct personal name is complemented.

Further, the language processing device according to claim 3 is
3. The language processing apparatus according to claim 2, wherein the omitted element is a subject of a subject or an object in a Japanese natural utterance sentence.

According to a fourth aspect of the present invention, the language processing apparatus according to the first aspect is a semantic determination apparatus for a translation apparatus for translating from a first language to a second language. The information-completed text data is composed of a character string of a sentence in the first language, and is semantically determined text data in which the meaning of the sentence is complemented in advance. The attribute list includes the sentence type, tense, subject, and purpose. An attribute list of a plurality of attributes including word information. The attribute table is an attribute table indicating whether or not each attribute exists for each sentence, and the correct meaning. Each of the decision trees is a correct word meaning. Is a decision tree for determining the correct word meaning in the input character string, and the complement selecting means selects a predetermined number of correct word meanings having the maximum frequency or the higher frequency. do it,
A character string in which the correct meaning is complemented is output.

Further, the language processing device according to the fifth aspect is characterized in that:
5. The language processing apparatus according to claim 4, wherein the correct meaning is a bilingual element of a second language.

[0015]

Embodiments of the present invention will be described below with reference to the drawings.

<First Embodiment> FIG. 1 is a block diagram showing the structure of a Japanese language elimination element complementing apparatus according to a first embodiment of the present invention. The Japanese abbreviation element complementing device of the present embodiment determines the abbreviations to be complemented from the corpus (text data) for which the abbreviation elements have been complemented, by using the decision tree learning units 12-1 to 12-M (M is plural and collectively 12). ), A decision tree is obtained for each attribute table by inductive machine learning, and the omitted element complementers 14-1 to 14-M (collectively referred to as 14).
By using each corresponding decision tree, the omitted element complementing process is performed to determine the omitted element to be complemented, and the omitted element complement selecting section 15 selects the omitted element having the maximum frequency, and the omitted element is complemented. Character strings are output. Here, the knowledge acquired from the corpus is represented by each decision tree for each attribute table and stored in the decision tree file memories 23-1 to 23-M (collectively referred to as 23).

First, a description will be given of a phenomenon of omitting a Japanese dialogue sentence. As an example, consider supplementing the subject of "forget" with the following utterances. (Example 1) I forgot the camera in my room. In this example, the verb “forget” has semantic attributes and “come”, “get”, “ta”, “you”, “n”
A large number of elements, such as the end-of-sentence expression such as "ga", may be related to the subject to be complemented, and it is difficult to clearly describe which of these elements influences subject complementation and to what extent. On the other hand, in the method using the decision tree according to the present embodiment, it is only necessary to enumerate the elements that may influence, and attributes unnecessary for subject complementation as a result of learning are automatically excluded from the attributes of the decision tree. You. Further, it is also possible to automatically learn the influence of simultaneous appearance of a plurality of elements which has not been noticed conventionally.

Generally, it is considered that the following information is necessary to complement a subject omitted in a dialogue sentence. (1) Information in sentences Verbs, declarative / questioning, active / passive, respect / humility, etc. (2) Information before the preamble (contextual information) The flow of the conversation so far in the dialogue. (3) Non-language information Where the sentence was sent to whom, etc.

Here, it should be noted that the information classified as described above is not independent, and the subject is omitted while affecting each other. For example, in a reception at a hotel reception and a conversation between a customer and a guest, they generally “stay”
Is a guest or "general person", but other possibilities are possible in certain special contexts.

Next, the complementing process using the decision tree will be described. As described above, there is a wide variety of information necessary for supplementing the abbreviations of Japanese dialogue sentences. In this embodiment, a method using a decision tree is used as a method for uniformly and automatically complementing the subject with such information. A decision tree is one of the knowledge expression methods for a concept in which multiple elements are complicatedly related, and is represented by a directed tree. Each branch node branches according to an attribute value corresponding to a certain attribute, and a decision is made at each leaf. The decision tree is largely divided into a binary tree and a multi-ary tree according to the number of branches at the branch node.
In the present embodiment, the former is used.

Next, the correct answer data and decision tree learning will be described. In order to automatically and uniquely create a decision tree, a decision tree is created by inductive learning from a corpus (text data). The corpus used in the example of the present embodiment is a travel conversation corpus owned by the present applicant that includes conversation between two parties in ticket reservation, tourist information, and the like. In this corpus, the first person, the second person, the third person, and the unspecified person are added to the verb in which the subject is omitted (given when it is considered that a general person is spoken in mind). 4 types of correct personal information of the subject to be complemented. At this time, these correct answers were given considering only Japanese. In other words, it is not provided from the viewpoint of which subject is used when translating into English or the like. Generally, decision tree learning is NP (Nondeterministic Pol)
Abbreviation for ynomial. In this embodiment, since it is complete, the greedy method (greedy a
lgorithm). No pruning was performed.

Next, an algorithm of a decision tree learning process for constructing a decision tree for determining “correct personal information of the subject to be complemented” will be described. In the decision tree learning process, the validity of each attribute is calculated independently of the other attributes, and the classification order based on the efficient attributes for class determination is determined as a tree structure having a structure divided in the form of a binary tree. To construct. The validity of the attribute is evaluated based on the entropy H after division and classification according to the attribute. The entropy here indicates the priority of the validity of the attribute. That is, the node N
When divided into ₁ and the node N _2, and the entropy H ₀ before the division, and the entropy H of the divided, the entropy H ₁ for node N _1, and the entropy H ₂ to node N ₂ is given by: You.

[0023]

H ₀ = Σ p (tagall) · log ｛p
(Tagall) @tagall

H = p ₁ H ₁ + (1−p ₁ ) H ₂ where:

H ₁ = Ｈp (tagN ₁ ) · log ｛p (t
agN ₁ )｝ tagN ₁

H ₂ = Σp (tagN ₂ ) · log ｛p (t
agN ₂ )｝ tagN ₂

Here, p (tagall) is the frequency probability of the number of cases for all “correct correct person information of the subject to be complemented (hereinafter referred to as correct person information)” before division (this is also referred to as a tag). Or the probability of appearance, tagall
Indicates the sum of all correct personal information before the division. Further, p ₁ is the sum of the frequency probabilities of the number of cases of the correct personal information included when the node is divided into the nodes N ₁ . In addition, p (tagN ₁₎ is the number of cases of frequency probability for all of the correct person information of the node N _1, tag
Σ for N ₁ represents the sum of all the correct person information of the node N _1. p (tagN ₂₎ is the number of instances of frequency probabilities for all correct answers personal information of the node N _2, the Σ of tagn _2, indicates the sum of all the correct person information of the node N _2.

In order to calculate the validity, for each sentence, case information consisting of a set of "attribute, its attribute value, and correct personal information" is preliminarily extracted from the text data for learning. Specifically, for all the case sets, an attribute that minimizes the entropy H after classification is obtained and assigned to the first node. The case set is divided according to the attribute value of this attribute, and corresponding child nodes are created. A tree structure is constructed by repeating the same process at each child node. The condition for stopping the division is that the updated node has a single attribute. Here, a node that is not divided is called a leaf. At the leaf of the learned decision tree, the frequency probability of each correct personal information is calculated from the given case set.

Next, conditions of a robust decision tree model will be described. As described in the section of the prior art, the present inventor has disclosed an abbreviated complementing method using a decision tree as a decision making means in the conventional example. In this conventional method, a single decision tree is introduced, and decision tree learning is performed using an attribute set that may be useful information for complementation.
In the attribute matching performed at the time of learning, a method was used to determine whether a morpheme (or semantic attribute) exists in a certain search space. Considering the robustness of this method, it is expected that it is robust in the following points. (A) A plurality of information sources (attributes) are used during decision tree learning. (B) The matching method is concerned only with the presence or absence of an element in a certain space.

As for the former, for example, the auxiliary verb “tai”
If the decision is made only from a single element, such as "the subject is the first person," the complementation will fail if the element contains any error. In decision tree learning, since probabilistic calculations are performed for all combinations of elements that have appeared, even if "tai" does not appear due to lack or if "tai" is replaced by another word due to replacement, Some robustness can be expected. In the latter case, since only the presence or absence of the element in the search space is used as a clue, robustness can be expected especially for insertion errors.

Next, information sources and decision making means will be described. As described above, the decision tree and its attribute determination are expected to have some robustness, but these are not sufficient. This is probably due to the single source of decision. When an error or an unreliable part in an input sequence cannot be determined, it is dangerous to treat all of these elements as correct elements in a problem in which input reliability cannot be ensured.

One way to improve the reliability of the process is to use multiple decision sources. That is,
Decision maker (Omission element complementing unit 1 in the first embodiment)
4. Corresponds to the semantic determination unit 14a in the second embodiment. ) Are prepared to use the same information source, or separate information sources to be used for decision making, and make a plurality of decisions from each information source using the same decision maker. In the latter case, various options are possible, such as making the information sources to be used exclusively or information sources excluding different information from all elements.

When a plurality of answer decision devices are prepared and a plurality of answer candidates are obtained in this manner, it is important to determine which of them is the final answer. When the type of decision is finite, especially when the number is small, a majority model may be considered. However, it is necessary to adjust the relationship between the decision sources, and the problem remains. Therefore, the decision tree model disclosed in the present embodiment recognizes the reliability or the likelihood corresponding to the reliability from the answers from the respective decision sources, and determines the likelihood when determining the final answer. Consider doing this by comparing. At this time, in the case where the data is output from different decision-making devices, it is necessary to measure or predict the reliability as an objective measure and to compare them. Is not easy.

From the above considerations, in the present embodiment, the following measures have been adopted for constructing a robust model. (A) Prepare a plurality of answers and select an answer from them. (B) A plurality of answers are created by changing information sources from the same decision maker. (C) The answer is selected by comparing the likelihoods from the same decision maker.

Based on the above considerations, this embodiment proposes a model having robustness against input inaccuracy.
This model is an extension of the conventional case element elimination complement model invented by the present invention, and is an MDT (Multiple Decision Tree) having a plurality of decision trees.
e) In the specification of the patent application of Japanese Patent Application No. 9-245201, a case element other than the subject is considered, but in the present embodiment, the discussion is limited to the subject.
However, the discussion of the present embodiment can be similarly applied to other case elements. ). For comparison, the conventional decision tree model is hereinafter referred to as SDT (Single Decision).
(Ion Tree) model. A comparison of these models is shown in FIG.

In the SDT model, subject complementing knowledge is constructed using a knowledge expression technique called a decision tree. In learning a decision tree, a question is asked based on the presence or absence of an attribute prepared in advance from a case having an error-free input sentence and the correct subject information, and the case is classified based on the entropy criterion. The MDT model is a robust model based on the above SDT model. In this MDT model, only the attribute set is changed from the same learning set, and a plurality of decision trees are created by the same decision tree learning method.

Next, a method for selecting one appropriate decision tree from a plurality of decision trees will be described. From a plurality of answer candidates obtained by the MDT model, in the present embodiment, the MCL preference (Maximum-Case Leaf) is selected.
A selection method called “Preference” is newly introduced. First, an assumption regarding the input uncertainty necessary for using this MCL preference is made. <Assumption> When the input of the process is different from the correct answer, the input is unlikely to be somewhere in comparison with the correct answer element sequence. That is, an input character string containing an error is a partially rare element string.

This assumption is a natural assumption when processing the output result of speech recognition is assumed. Because
If the input element string is a natural element string by any standard than the correct element string, it is no longer possible to recognize that the input character string contains an error. This is because, in all the processings of the above, the processing must be performed by regarding those element strings as correct element strings. Therefore, such an input is impossible, or even if it is present, it is an input that needs to be dealt with in speech recognition, and it is natural to remove it from the target input in the present embodiment.

Next, the attribute set will be described. In order to supplement the omitted case element, various information must be considered. In particular, in learning using a decision tree model, what kind of attribute is prepared greatly affects the accuracy. In the present embodiment, a decision tree is created in consideration of the attributes described below.

(A) Semantic attribute of content word: Information on what content word is included in a sentence to be omitted. The content word is broadly divided into information on declinable information and information on case elements (nominal). The middle category (100 attributes) in the Kadokawa New Thesaurus was used as the semantic attribute of the content word. An example is shown in the following table.

[0038]

[Table 1] Examples of verb semantic attributes (from Kadokawa New Thesaurus) ――――――――――――――――――――――――――――――――― 3 actions: 30 actions, 31 traffic, 32 facial expressions, 33 observations, 34 statements,… ――――――――――――――――――――――――――――――― ―― 34 statements: 340 statements, 341 silence, 342 statements, 343 discourse, ―――――――――――――――――――――――――――――――― ― 343 Discourse: 1 speak, 2 speak, 3 voices, 4 speak, 5 speak, 6 speak,… ―――――――――――――――――――――――― ―――――――――― (Note) The numbers are the classification numbers of the verbs.

(B) Appearance of a function word: Information on the appearance of a function word such as a group of attached words following a verb and a particle. The appendix group includes auxiliary verbs such as passive / respect / possible / spontaneous "re" causative "sell" aspect "being" and quasi-form auxiliary verbs "should" indicating the purpose. In addition, a set of verbs indicating respect (to be consumed), humility (to ask), possible (to drink), etc., is one attribute,
Treated as a function word indicating "respect". In addition, received expressions such as "do" and verbs "do" and "become" were also considered as special function words.

Other functional words include case particles, connecting particles,
Final particle, there is a prefix that represents the respect of the verb just before seen in the example of "do you think,""you can offer." In addition, functional words also include the adjective noun "dame" indicating prohibition, the formal noun "gonna" indicating intention, and a set of question words ("where""why"). (C) Out-of-language information: As the out-of-language information, an attribute indicating whether the speaker of the uttered sentence is an information provider or an information receiver is used.

The attribute set used in this embodiment is shown in the following table.

[0042]

[Table 2] Set of prepared attributes ―――――――――――――――――――――――――――――――――― Set Attribute Meaning Attribute Function word Speaker Role total ―――――――――――――――――――――――――――――――― Set A Semantic attribute of predicate 100 Semantic attribute of case element 100 Function Word (end-of-sentence expression) 166 Speaker role 1 (total) 367 ―――――――――――――――――――――――――――――――――― Set C Semantic attribute of predicate 100 Semantic attribute of case element 100 Speaker role 1 (total) 201 ――――――――――――――――――――――――――――――― --- Set F Function words (end-of-sentence expression) 166 Speaker role 1 (total) 167 ―――――――――――――――――――――――――――――― ――――

Next, attribute matching will be described. For the information other than the out-of-language information, attribute matching was performed by matching with the morpheme sequence. That is, information about which position of the following five types appears in the center of the word to be complemented is given to all attributes in advance. (A): before = the morpheme before (including immediately before) the morpheme is included. (B): contains the morpheme “latest = immediately before the declinable word”. (C): here = the word is ... (D): includes a morpheme of next = immediately after the declinable. (E): After = the morpheme after (including immediately after) the morpheme is included.

For example, the attributes relating to a word are: here,
For case particles: before, for prefixes:
Gives the location information of the latest. Regarding semantic attributes, matching was performed based on whether or not a word having a semantic attribute at a certain position was included. When a sentence such as a compound sentence or a compound sentence is composed of a plurality of simple sentences, it is approximately divided into simple sentences. In the division method, a connecting particle (here, when a plurality of connecting particles are continuous, such as "... tar / ba", the connecting particle is the last connecting particle) is used as a dividing position and divided before and after.

In FIG. 1, the omitted element complemented text data memory 21 previously stores the omitted element supplemented text data as shown in the following table.

[0046]

[Table 3] Example of text data with completed omitted elements ――――――――――――――――――――――――――――――――― sentence subject morpheme ――――――――――――――――――――――――――――――――――― Statement 1. (1st person) Decided / Depends on your / * Contact * / Give / Masu / U / K (1st person) Decided / Depends on your / * Contact * / Give / Masu / U / ka Sentence 3. (1st person) Decided / Depending on your / * Contact * / Give / Masu / U / K (1st person) Decided / Depends on your / * Contact * / Give / Musu / U / K (1st person) Decided / Depends on / Guide / * Contact * / Give / Masu / U / ka Sentence 6. (1st person) Decided / Depends on your / * Contact * / Give / Masu / U / K (1st person) Today / no / reservation / available (1st person) Immediately / * Contact * / Give / Ta / N / I / Yo Sentence 9. (1st person) Then / confirmation / take / * take * / make / get / sentence Sentence 10. (1st person) Tomorrow / Your / * Contact * / Give me / Your / Issue Sentence 11. (2nd person) Your / * Contact * / Yes / Tell me / Mus / or Sentence 12. (2nd person) So / to / to / to / * contact * / to / you can / you / masu / ka Sentence 13. (2nd person) Yes / Jay / Memo / But / * Take * / Get / Masu / ka Sentence 14. (2nd person) Yes / As much as possible / Night view / Gaze / Enjoy / Seats // Take * / Tell me / Want / No / Is / But Sentence 15. (Second person) Slowly / Dinner // * Take * / Take / No / Ha / Is / Would // ――――――――――――――――――――――― ―――――――――――――― (Note) The part between () is where the zero pronoun (abbreviated element) is complemented, and the part between * is the normal form part of the verb. is there. Here, / is inserted in order to clarify that the processing is performed in word (normal form) units.

The attribute list memory 20 is a memory storing the above-mentioned attributes, and an example of the attribute set is shown in the following table.

[0048]

[Table 4] Attribute set AS1 in attribute list ――――――――――――――――――――――――――――――――― Attribute A1. Is the verb "contact"? Attribute A2. Is the verb "take"? Attribute A3. Do you include "ga" before the verb? Attribute A4. Do you include "given" after the verb? Attribute A5. Do you include "terechi" after the verb? Attribute A6. Do you include "is" after the verb? Attribute A7. Do you include "mas" after the verb? Attribute A8. Do you include "u" after the verb? Attribute A9. Do you include "ka" after the verb? Attribute A10. Do you include "yo" after the verb? Attribute A11. Do you include the word with the meaning code 41 before the verb? ――――――――――――――――――――――――――――――――――― (Note) “41” is the semantic attribute of the verb in Table 1. Indicates the meaning code "confirmation".

[0049]

[Table 5] Attribute set AS2 in the attribute list ――――――――――――――――――――――――――――――――― Attribute A3. Do you include "ga" before the verb? Attribute A6. Do you include "is" after the verb? Attribute A7. Do you include "mas" after the verb? Attribute A8. Do you include "u" after the verb? Attribute A9. Do you include "ka" after the verb? Attribute A11. Do you include the word with the meaning code 41 before the verb? ―――――――――――――――――――――――――――――――――――

[0050]

[Table 6] Attribute set AS3 in attribute list ――――――――――――――――――――――――――――――――― Attribute A4. Do you include "given" after the verb? Attribute A5. Do you include "terechi" after the verb? Attribute A6. Do you include "is" after the verb? Attribute A7. Do you include "mas" after the verb? Attribute A10. Do you include "yo" after the verb? Attribute A11. Do you include the word with the meaning code 41 before the verb? ―――――――――――――――――――――――――――――――――――

[0051]

[Table 7] Attribute set AS4 in the attribute list ――――――――――――――――――――――――――――――――― Attribute A1. Is the verb "contact"? Attribute A2. Is the verb "take"? Attribute A3. Do you include "ga" before the verb? Attribute A4. Do you include "given" after the verb? Attribute A5. Do you include "terechi" after the verb? Attribute A9. Do you include "ka" after the verb? ―――――――――――――――――――――――――――――――――――

Then, the attribute table generating unit 11 generates each sentence of each text data based on the omitted element complemented text data in the omitted element complemented text data memory 21 and the attribute list in the attribute list memory 20. , For all combinations with each attribute in the attribute list, generate an attribute table indicating whether each attribute exists for each sentence and the correct personal name of the omitted element. Stored in the attribute table memory 22. Here, the attribute set AS1
, An attribute table AST2 is generated based on the attribute set AS2, an attribute table AST3 is generated based on the attribute set AS3, and an attribute table AST4 is generated based on the attribute set AS4. The text data is composed of a character string of a Japanese natural utterance sentence, and the correct personal name of an abbreviated element of the sentence is complemented in advance. The attribute list includes, for example, a verb normal form and a verb semantic attribute. , A sentence end expression, and speaker information as out-of-language information. FIG. 2 shows a flowchart of the attribute table generation processing. An example of the attribute table generated based on Tables 4 to 7 is shown in the following table.

[0053]

[Table 8] Attribute table AST1 ――――――――――――――――――――――――――――――――― Statement 1-6 7 8 9 10 11 12 13 14 15 ――――――――――――――――――――――――――――――――――― Attribute A1 Y N Y N Y Y Y Y N NN attribute A2 N Y N N N N N Y Y Y Attribute A3 N N N N N N N N N Y N N Attribute A4 Y N Y N N N N N N N N Attribute A5 N N N Y N Y N N N Y Y Attribute A6 N N Y N N N N N Y N N Attribute A7 Y Y N Y Y Y Y Y Y N N Attribute A8 Y N N N N N N N N N N N Attribute A9 Y N N N N Y Y Y N Y Attribute A10 N Y Y N N N N N N N Attribute A11 Y N N Y N N N N N N N ―――――――――――――――――――――――――――― ―――― ――― Correct answer 1 1 1 1 1 1 1 2 2 2 2 2 ―――――――――――――――――――――――――――――――――――

[0054]

[Table 9] Attribute table AST2 ――――――――――――――――――――――――――――――――――― Statement 1-6 7 8 9 10 11 12 13 14 15 ――――――――――――――――――――――――――――――――――― Attribute A3 N N N N N N N N N YN attribute A6 N N Y N N N N N N Y N attribute A7 Y Y N Y Y Y Y Y Y N N attribute A8 Y N N N N N N N N N N attribute A9 Y N N N N N Y Y Y Y N Y Attribute A11 Y N N N N N N N N N ――――――――――――――――――――――――――――――――――― Correct person 1 1 1 1 1 2 2 2 2 2 ―――――――――――――――――――――――――――――――――――

[0055]

[Table 10] Attribute table AST3 ――――――――――――――――――――――――――――――――― Statement 1-6 7 8 9 10 11 12 13 14 15 ――――――――――――――――――――――――――――――――――― Attribute A4 Y N Y N Y N N N N NN attribute A5 N N N Y N Y N N N Y Y attribute A6 N N Y N N N N N Y N N attribute A7 Y Y N Y Y Y Y Y Y N N N attribute A10 N Y Y N N N N N N N N Attribute A11 Y N N N N N N N N N ――――――――――――――――――――――――――――――――――― Correct person 1 1 1 1 1 2 2 2 2 2 ―――――――――――――――――――――――――――――――――――

[0056]

[Table 11] Attribute table AST4 ――――――――――――――――――――――――――――――――― Statement 1-6 7 8 9 10 11 12 13 14 15 ――――――――――――――――――――――――――――――――――― Attribute A1 Y N Y N Y Y Y Y N NN attribute A2 N Y N N N N N Y Y Y Attribute A3 N N N N N N N N N Y N N Attribute A4 Y N Y N N N N N N N N Attribute A5 N N N Y N Y N N N Y Y Attribute A9 Y N N N N Y Y Y N Y ---------------------------------------------------------------- 1 1 1 1 1 2 2 2 2 2 ―――――――――――――――――――――――――――――――――――

Further, the decision tree learning units 12-1 to 12-
Each of M is based on the attribute table in the attribute table memory 22 based on “correct correct person information of the subject to be complemented” (in the present embodiment, correct subject personal information of the subject, but the present invention is not limited to the subject. , May be omitted elements such as direct objects, indirect objects, etc.).
4 and generates the corresponding decision tree file memory 23-1.
To 23-M. The decision trees DT1 to DT4
8 to 11 are shown in FIGS. Here, the decision tree DT1 in FIG. 8 is an attribute table A generated based on the attribute set AS1.
The decision tree DT2 of FIG. 9 is a decision tree generated based on the attribute table AST2 generated based on the attribute set AS2, and the decision tree DT3 of FIG. FIG. 1 is a decision tree generated based on an attribute table AST3 generated based on a set AS3.
The first decision tree DT4 is a decision tree generated based on the attribute table AST4 generated based on the attribute set AS4.
The decision tree is a decision tree for determining a correct person of an omitted element having a tree structure of a binary tree format that is divided depending on the attribute value of each attribute.

FIG. 3 shows a flowchart of the decision tree learning process. Further, a specific process of the decision tree generation process of FIG. 3 is shown in a flowchart of FIG. 4, first, all the entropy H ₀ before division entropy H after division by each attribute calculated in step S21, the difference between the entropy (H ₀ -H) selects the greatest attribute in step S22, A node divided by the attribute value selected in step S23 is generated to update the decision tree. Then, it is determined whether the attribute value under the node updated in step S24 is single, that is, whether or not it is a leaf node indicating the determination result. If NO, the node generated in step S25 is set as a processing target and the process proceeds from step S21. Continue processing. On the other hand, if YES is determined in the step S24, the decision tree generating process ends and the process returns to the main routine. That is, the decision tree learning unit 12 selects, based on the attribute table in the attribute table memory 22, the attribute having the largest difference between the entropy after the division by all the attributes and the entropy before the division, and selects the selected attribute. The decision tree is generated by learning to update the decision tree until a leaf node is reached so as to generate a node divided by a value.

Here, an example of the decision tree learning process will be described. The input correct personal name data is $ 11111111112
The initial state entropy at 2222 ° is calculated as follows.

[0060]

(5) − (10/15) × log ₂ (10/15) − (5/1
5) × log ₂ (5/15) = 0.918

Then, when the correct answer is classified according to the attribute AS1, that is, {Yes} No} = {11111
The total entropy when 11122 {11222} is calculated as follows:

[0062]

(10/15) × {− (7/10) × log ₂ (7/1
0) − (3/10) × log ₂ (3/10)} + (5/15) × {−
(2/5) × log ₂ (2/5) − (3/5) × log ₂ (3/5)} =
0.911

Further, the amount of decrease (difference) in entropy when the attribute AS1 is adopted is calculated as follows.

[0064]

0.98−0.911 = 0.007

Similarly, the entropy reduction (difference) is calculated for each attribute to obtain the following table.

[0066]

[Table 12] ――――――――――――――――――――――――――――――――――― Attribution Correct person data Entropy Reduction of entropy ― ―――――――――――――――――――――――――――――――― Attribute AS1 {1111111122} {11222} 0.911 0.007 Attribute AS2 {11222} {1111111122} 0.805 0.113 Attribute AS3 {2} {11111111112222} 0.806 0.112 Attribute AS4 {11111111} {1122222} 0.403 0.515 Attribute AS5 {1222} {11111111122} 718 0.200 Attribute AS6 {12} {1111111112222} 0.905 0.013 Attribute AS7 {111111111222} {122} 0.833 0.085 Attribute AS8 {111111} {111122222} 0.595 0.323 Attribute AS9 {1111112222 } {11112} 0.888 0.030 Attribute AS10 {11} {1111111122222} 0.833 0.085 Attribute S11 {1111111} {11122222} 0.509 0.409 -----------------------------------

In this processing example, the attribute AS4 whose entropy reduction amount is the largest is selected, a node divided by the attribute value is generated, and the decision tree is updated.

On the other hand, the attribute value extracting unit 13 outputs a character string of a natural utterance sentence input using an input means such as a keyboard or recognized by a voice recognition device, or a natural utterance stored in a hard disk memory. Based on the character string of the sentence, by referring to the attribute list in the attribute list memory 20 and comparing each attribute, an attribute value indicating whether or not each attribute exists is extracted, and the buffer memory 24 Through the omission element complementing units 14-1 to 14-M. The attribute value extraction processing of the attribute value extraction unit 13 is shown in the flowchart of FIG.

Further, the omitted element complementing sections 14-1 to 14-1
Each of the decision trees DT1 in the decision tree file memories 23-1 to 23-M corresponds to the input attribute value.
Using DT4 to DT4, the candidate correct answer person of the subject is determined and output to the omitted element complementing selection unit 15, and in response to this, the omitted element complementing selecting unit 15
The answer candidate having the highest frequency among the answer candidates from -1 to 14-M (or a plurality of n-number of answer candidates having the higher frequency (n-best) may be the final answer). To generate a character string in which the omitted element including the correct person is complemented and output to an external device such as a memory, a printer, or a display device.

FIG. 6 is a flowchart showing the omitted element complementing process executed by each of the omitted element complementing units 14-1 to 14-M. In FIG. 6, first, a decision tree is read in step S41, and the attribute values extracted in step S42 are read. Next, in step S43, the roof node of the decision tree is set as the current node to be processed. In step S44, it is determined whether the current node is a leaf node. If NO, the process proceeds to step S45 based on the attribute value of the current node. The corresponding child node (the node located below the decision tree) is set as the current node to be processed, and the process returns to step S44. YES in step S44
In the case of, since the correct person has been determined, the answer candidate of the omitted element including the correct person is output in step S46, and the omitted element complementing process ends.

Here, a case is considered in which, of the words of sentences 1 to 15, the word "present" is erroneously recognized and input as the word "call". That is, an example of processing when a character string shown in the following table is input will be described.

[0072]

[Table 13] ――――――――――――――――――――――――――――――――― Statement 1 '. Decided / Dependent / Your / * Contact * / Eat / Masu / U / ka Sentence 2 '. Determined / Dependent / Your / * Contact * / Eat / Masu / U / ka Sentence 3 '. Decided / Dependent / Your / * Contact * / Eat / Masu / U / ka Sentence 4 '. Decided / Dependent / Your / * Contact * / Eat / Masu / U / ka Sentence 5 '. Decided / Dependent / Your / * Contact * / Eat / Masu / U / ka Sentence 6 '. Decided / Dependent / Your / * Contact * / Eat / Masu / U / ka Sentence 8 '. Immediately / * Contact * / Eat / Ta / N / I / Yo Sentence 10 '. Tomorrow / your / * contact * / eat / have / /----------------

In the conventional method, the subject is complemented using only the attribute set AS1 as a criterion, so that sentence 10 is correctly complemented with the first person, but sentences 1 through 6 and sentence 8 are judged as the second person. Would. That is, seven out of eight sentences output incorrect complement results. On the other hand, in the method of the present embodiment, the answer candidate with the highest frequency is selected as the final answer among the answer candidates output by the omitted element complementing units 14 using the four decision trees of the attribute sets AS1 to AS4. I do. That is,

## EQU8 ## Sentence 1 '. When determined / depending / go / * contact * / enjoy / masu / u / ka is input, in the decision tree DT1 obtained from the attribute set AS1, the second person (frequency 4) is set as the answer candidate. Obtain the results in the table.

[0074]

[Table 14] ―――――――――――――――――――――― Decision Tree Answer Candidate Frequency ――――――――――――――――――― ――― Decision tree DT1 2nd person frequency 4 Decision tree DT2 1st person frequency 7 Decision tree DT3 1st person frequency 1 Decision tree DT4 1st person frequency 4 ――――――――――――――――――― ―――

Of these, the answer of the decision tree DT2, which is the answer candidate with the highest frequency at the time of learning, that is, the first person is selected and complemented correctly. Similarly, sentences 2 'to 6' and sentence 8 'are correctly complemented (sentence 10' outputs an incorrect solution). That is, these sentences, in which seven sentences out of eight sentences were incorrect by the conventional method, can be correctly answered by seven sentences,
Accuracy degradation due to erroneous input can be minimized.

Next, similarly, consider a case where "ka" is mistaken for "ne" for some reason. That is, consider a case where a sentence shown in the following table is input.

[0077]

[Table 15] ――――――――――――――――――――――――――――――――― Statement 1 ”. Contact * / Give / Masu / U / Ne Sentence 2 ". Determined / Dependent / go / * Contact * / Give / Music / U / Ne Sentence 3 ". Decided / Desired / Go / * Contact * / Give / Mus / U / Ne Sentence 4". Determined / Dependent / go / * Contact * / Give / Music / U / Ne Sentence 5 ". Determined / Desired / Go / * Contact * / Give / Music / U / Ne Sentence 6". Determined / Dependent / Your / * Contact * / Give / Mus / U / Ne Sentence 11 ". So / that / until / to / * contact * / do / you can / can / ne sentence 13 ". Yes / then / memo / even / * take * / get / get / ne sentence 15". Slowly / Dinner /// * Take * / Take / No / Ha / How // ---―――――――――――――――――――――――――― ――――――――――

Of these, the decision tree DT of the prior art method
In the method using only 1, sentence 11 ″, sentence 12 ″, sentence 1
While four sentences of 3 ″ and sentence 15 ″ give incorrect answers, the method of the present embodiment outputs correct solutions for all of the above sentences.

The reason why the method of the embodiment according to the present invention works effectively is considered as follows. In the conventional method, there was no need to doubt the existence of attributes included in the decision tree,
If these attributes in the input, for example, "given"
Does not work effectively if there is an error in However, in the embodiment according to the present invention, a decision tree that does not include “given” as an attribute is also created (decision tree DT2), which corresponds to trying complementation from another viewpoint. For this reason, it is possible to avoid being extremely dependent on the existence of "give" as information necessary for complementation.
Even if is deleted or replaced with another word, it can be correctly complemented with other valid attributes such as "including the word with the meaning code 41 before", and as a result, a coherent decision tree is created I have. As a result, in the case where the “given” is not included, the attribute set AS1 or the like can reach a low-frequency leaf node, whereas in this case, the decision tree AS2 that is the only correct decision tree may reach a high-frequency leaf node. Expected to be relatively high.
In the embodiment according to the present invention, by utilizing this property, the complement candidates obtained from a plurality of decision trees are ranked according to the frequency of the leaf nodes at the time of learning.

The attribute table generating unit 11, the decision tree learning unit 12, the attribute value extracting unit 13, the omitted element complementing unit 14,
The omitted element complement selection unit 15 is constituted by, for example, a digital computer having a CPU, and includes an attribute list memory 20.
The abbreviation element-completed text data memory 21, the attribute table memory 22, the decision tree file memory 23, and the buffer memory 24 are constituted by, for example, a hard disk memory.

In the above embodiment, if an item that does not affect the complement of a pronoun is listed as an attribute item,
Since these attributes are automatically rejected by the decision tree generation unit 12 and do not adversely affect the accuracy of the complementation as a result, there is no need to make any arbitrariness and to perform sophisticated linguistic introspection. . Here, the tree is a binary tree, but a multiple tree can also be applied using exactly the same technology. In addition, the present invention can be applied with exactly the same technology other than subject complementation such as omitting an object. The present invention can be applied not only to Japanese but also to languages having a similar omission phenomenon, for example, Korean.

As described above, according to the present embodiment, a technique called a decision tree is introduced for the Japanese case element elimination and complementation in which multiple elements affect in a complicated manner without manual intervention. Automatically grasps the influence relationship of many pieces of information necessary for, and can automatically determine omitted case elements. In addition, a plurality of decision trees corresponding to a plurality of attribute tables are generated, an answer candidate is obtained for each decision tree, and the answer candidate having the highest frequency is used as an omission element of a correct answer. Elements can be complemented with high precision.
In addition, since a plurality of decision trees corresponding to a plurality of attribute tables are generated, an answer candidate is obtained for each decision tree, and the answer candidate having the highest frequency is regarded as a correct meaning, so that an error is included in the input. In addition, the meaning can be determined with higher accuracy than in the conventional example.

<Second Embodiment> FIG. 12 is a block diagram showing a configuration of a meaning determining device for a translation device according to a second embodiment of the present invention. This second embodiment is used for determining the meaning of an English to Japanese translation device, and differs from the first embodiment of FIG. 1 in the following points. (1) Instead of the text data memory 21 with completed omitted elements, a semantically determined text data memory 21a for storing semantically determined text data in which a Japanese translation is added as a meaning to an English sentence is provided. (2) Instead of the attribute list memory 20, an attribute list memory 20a for storing an attribute list of attributes including English sentence type, tense, subject type, and object type is provided. (3) An attribute table list memory 22a is provided in place of the attribute table list memory 22, where the attribute list memory 20a is provided.
The attribute table generated based on the attribute list in is stored. (4) Decision tree file memories 23a-1 to 23a-M instead of decision tree file memories 23-1 to 23-M
(Collectively referred to as 23a), in which a decision tree generated based on each attribute table in the attribute table memory 22a is stored. (5) Instead of the omitted element complementers 14-1 to 14-M,
Meaning determination units 14a-1 to 14a-M (collectively, 14a
Attached. ), And the meaning determining unit 14a respectively determines the input attribute values by using the decision tree file memory 23a-
By using each of the decision trees in 1 to 23a-M, a meaningful answer candidate representing the meaning of the input English sentence is determined and output. (6) A semantic decision selecting section 15a is provided in place of the omitted element complement selecting section 15, and the semantic decision selecting section 15a has the highest frequency among the answer candidates from each of the semantic determining sections 14a-1 to 14a-M. An answer candidate (or a plurality of n answer candidates having a higher frequency (n-best) may be selected) as a final answer, and a Japanese character string whose meaning is determined is selected. Generate and output to an external device such as a memory, printer or display device.

Next, as an example of processing in the second embodiment, the processing of determining the meaning of the English verb “take” (or determining the Japanese translation) will be described. An example of the text data for which the meaning has been determined in the text data memory 21a for which the meaning has been determined is shown in the following table.

[0085]

[Table 16] ――――――――――――――――――――――――――――――――――――――――――――――――― Number Japanese Translation English ―――――― ――――――――――――――――――――――――――――― Statement 101. Such It takes five minutes. Such It takes three days. Such It takes five minutes. Such It takes about an hour. Such It took ten minutes. Such It takes thirty minutes. Sentence 108. It will take us in twenty minutes. Ride I take a bus to the airport. Get on Are you taking a train? Ride I am taking a shuttle bus today. Are you going to take me to downtown? Take it It takes you to the station. ―――――――――――――――――――――――――――――――――――

Next, an example of an attribute set in the attribute list stored in the attribute list memory 20a is shown in the following table.

[0087]

[Table 17] Attribute set AS11 in the attribute list ――――――――――――――――――――――――――――――――― Attribute A101. Is the sentence declarative? Attribute A102. Is the sentence questionable? Attribute A103. Is the statement imperative? Attribute A104. Is the tense present tense? Attribute A105. Is the tense a past tense? Attribute A106. Is the tense a future tense? Attribute A107. Is the subject I? Attribute A108. Is your subject you? Attribute A109. Is the subject it? Attribute A110. Is the object a vehicle? Attribute A111. Is the object time? Attribute A112. Is the object a person? ―――――――――――――――――――――――――――――――――――

[0088]

[Table 18] Attribute set AS12 in the attribute list ――――――――――――――――――――――――――――――――― Attribute A104. Is the tense present tense? Attribute A106. Is the tense a future tense? Attribute A108. Is your subject you? Attribute A109. Is the subject it? Attribute A111. Is the object time? Attribute A112. Is the object a person? ―――――――――――――――――――――――――――――――――――

[0089]

[Table 19] Attribute set AS13 in the attribute list ――――――――――――――――――――――――――――――――― Attribute A101. Is the sentence declarative? Attribute A102. Is the sentence questionable? Attribute A104. Is the tense present tense? Attribute A106. Is the tense a future tense? Attribute A110. Is the object a vehicle? Attribute A112. Is the object a person? ―――――――――――――――――――――――――――――――――――

[0090]

[Table 20] Attribute set AS14 in attribute list ――――――――――――――――――――――――――――――――― Attribute A101. Is the sentence declarative? Attribute A102. Is the sentence questionable? Attribute A105. Is the tense a past tense? Attribute A106. Is the tense a future tense? Attribute A110. Is the object a vehicle? Attribute A111. Is the object time? ―――――――――――――――――――――――――――――――――――

FIGS. 13 to 16 show examples of decision trees DT11 to DT14 generated by the decision tree learning units 12-1 to 12-M based on the attribute sets AS11 to AS14 used in the second embodiment, respectively. It is a tree structure diagram. Attribute table generation unit 11 and decision tree learning units 12-1 to 12
The processing of -M is the same as that of the first embodiment, and is omitted here. The attribute table includes whether each attribute exists in each sentence, and the correct meaning of each sentence.

Next, as an example of processing in the apparatus of the second embodiment, consider the case where all the times of the object part are erroneously recognized. That is, the object of sentence 101 is misrecognized,

Sentence 101 '. Suppose that It takes five minuets. Has been entered.

In this case, in the conventional method, since the meaning is determined based on only the attribute set AS11 as a criterion, the sentence 101 'is determined to be "taken".
Similarly, sentence 102 ′, sentence 103 ′, sentence 104 ′, sentence 10
Also in the four sentences 6 ′, if the object is incorrect, it is determined that “take”. Also, the sentence 105 'is determined to be "taken", and all of them output incorrect meanings.

In the method of the embodiment of the present invention, the attribute set A
Among the answer candidates output by the four decision trees S11, AS12, AS13, and AS14, the answer candidate with the highest frequency is selected as the final answer candidate. That is,

Sentence 101 '. When "It takes five minuets." Is input, in the decision tree DT11 obtained from the attribute set AS11, "take" (frequency 1) is set as the answer candidate, and similarly, the results shown in the following table are obtained.

[0095]

[Table 21] ――――――――――――――――――――――――――――――――― Decision tree Answer candidate frequency ―――――― ――――――――――――――――――――――――――――― Decision Tree DT11 Take 1 Decision Tree DT12 Take 1 Decision Tree DT13 Take 6 Decision Tree DT14 Take 1 ――――――――――――――――――――――――――――――――――――

Among these, the answer of the decision tree DT13, which is the answer candidate with the highest frequency at the time of learning, that is, "take" is selected, and the meaning is correctly determined. Similarly, sentence 1
02 ', sentence 103', sentence 104 ', sentence 105', sentence 10
With respect to the five sentences 6 ′, “Such” in the decision tree DT13 is selected, and the meaning is correctly determined.

In the above embodiment, the semantic determination device for the English-to-Japanese translation device has been described. However, the present invention is not limited to this, and the present invention is not limited to this. Can be applied to a meaning determination device or a translation determination device for a translation device.

As described above, according to the present embodiment, by introducing a technique called a decision tree to the semantic determination processing for a translation apparatus in which multiple elements affect in a complicated manner, human intervention is required. In addition, it is possible to automatically grasp the influence relationship of much information necessary for complementation and automatically determine the meaning of the corresponding translated sentence. In addition, since a plurality of decision trees corresponding to a plurality of attribute tables are generated, an answer candidate is obtained for each decision tree, and the answer candidate having the highest frequency is defined as a correct meaning, so that it is compared with the conventional example. The meaning can be determined with high accuracy. In addition, since a plurality of decision trees corresponding to a plurality of attribute tables are generated, an answer candidate is obtained for each decision tree, and the answer candidate having the highest frequency is regarded as a correct meaning, so that an error is included in the input. In addition, the meaning can be determined with higher accuracy than in the conventional example.

[0099]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The inventor conducted an evaluation experiment in order to evaluate the performance of the Japanese language elimination element complementing apparatus shown in FIG. In the experiments, we evaluated the accuracy against real errors using actual speech recognition results as input and the accuracy against artificial errors that artificially created errors.

First, an experimental result of robustness with respect to a speech recognition result will be described. Table 22 shows the characteristics of the speech recognition result used for the input.

[0101]

[Table 22] Error characteristics of input sentence and performance of speech recognition device ―――――――――――――――――――――――――――――――― Parameter P1 P2 P3 ―――――――――――――――――――――――――――――――― Number of utterances 968 968 968 Average utterance morpheme 14.9 14. 9 14.9 Character recognition rate (%) 78.48 78.89 72.09 ――――――――――――――――――――――――――――――― ――― Average utterance error 3.44 3.49 4.09 (insertion error) 0.56 0.51 0.64 (missing error) 0.76 0.84 0.88 (replacement error) 2.12 2. 14 2.57 ――――――――――――――――――――――――――――――――――

Next, Table 23 shows the performance against real errors. The experiment was performed for two types of input: a correct answer input without a speech recognition error and a speech recognition result of the same sentence set.
The number of experimental sentences is 448. The experiment was performed for both an experiment using an SDT model that performs complementation using a single decision tree and an MDT model proposed in the present embodiment.
Attribute sets in the SDT model were performed for three types: set A, set F, and set C. This is called SD
They are expressed as T / A, SDT / F, and SDT / C. Also, M
Omission element complementer 1 which is a decision-maker in DT model
No. 4 performed processing using three of the above sets A, C, and F.

[0103]

[Table 23] Performance against speech error ―――――――――――――――――――――――――――――――― Correct answer P1 P2 P3 ――― ――――――――――――――――――――――――――――――― SDT / A 0.674 0.625 0.651 0.663 SDT / C 0 0.642 0.602 0.621 0.605 SDT / F 0.691 0.623 0.640 0.638 ―――――――――――――――――――――― ―――――――――― MDT 0.715 0.676 0.688 0.705 ――――――――――――――――――――――――― ―――――――

Next, in order to discuss the relationship between the robustness of the model and the tendency of errors, an experiment was conducted to determine what characteristics the model exhibits with respect to the following artificial errors. The experiment is
The following three types of errors were performed. (A) Insertion error. (B) Missing error. (C) Replacement error. (D) mixing errors.

The insertion error was created as follows. First,
One position where an error is to be inserted is determined at random in a morpheme sequence having no error. At this position, an arbitrary word is randomly selected from the morpheme set of the training conversation in which the decision tree learning has been performed, and this word is inserted. Since the inserted word is determined by the same expected value as the appearance ratio of each morpheme in the training conversation, the possibility that a frequently appearing word such as a case particle is inserted increases. The above is the process of inserting one word. When inserting a plurality of N words, the above process is repeated N times. FIG. 17 shows the relationship between the number of insertion errors and the performance. According to FIG.
It turned out that the T model has almost no performance degradation with respect to insertion errors. Also, with respect to the three types of SDT models, although the accuracy is slightly reduced, the degree of the degree of reduction with the increase in the number of erroneous words is gradual. The fact that the performance of the SDT model does not decrease so much with respect to insertion errors indicates that the element matching method discussed above has robustness, and the SDT model also has some robustness with respect to insertion errors. It was confirmed that. Further, it is considered that the reason why the MDT model hardly deteriorates in performance is that, in addition to the robustness of the above-described SDT, the present method of selecting after performing a plurality of decision making functions effectively.

Next, the missing error was created as follows. A morpheme to be deleted is selected at random by using a morpheme having no error as an input. However, verbs or sa-variant nouns for the omitted subject are excluded from selection. This is because, if the morpheme from which the verb or sa-variant noun is missing is the result of speech recognition, it is impossible to detect omission and is not a target for complementation. The above process is repeated a plurality of times, and a morpheme to be deleted is randomly selected each time.

FIG. 18 shows the relationship between the number of missing errors and performance. As can be seen from FIG. 18, the performance of the MDT model was slightly lower than that of the SDT / C model in the absence of seven or more morphemes, but it is considered that the accuracy is practically the same. In addition, the average of one utterance is 15 morphemes, and the average missing error of the actual speech recognition result is one utterance.
Considering that it is less than a morpheme, the lack of information of seven or more morphemes does not pose a problem in practical use. In FIG. 18, S
The DT / C model has almost no performance degradation, because the target of the missing error does not include a predicate. SD
In the T / C model, the subject is determined by using this information as main information, and therefore, there is not much performance degradation when a morpheme other than a predicate is missing.

Next, a replacement error was created as follows. A morpheme to be deleted is selected at random by using a morpheme having no error as an input. However, the verb or the sub-variant noun for the omitted subject is excluded from the object of the omission for the same reason as the omission error. Thereafter, as in the case of the insertion error, an arbitrary word is randomly selected from the set of morphemes of the training conversation in which the decision tree learning has been performed, and this word is inserted into this missing position.
The above is the process of inserting one word. When inserting a plurality of N words, the above process is repeated N times. FIG. 19 shows the relationship between the number of replacement errors and the performance.

Further, the mixing error was created as follows. First, the type of error is determined for the correct answer input. Errors are randomly determined so that the three types of insertion, deletion, and substitution appear with the same probability. After the type of error is determined,
The above-described processing of insertion, deletion, and substitution error is performed. In the case of an error of a plurality of morphemes, the above processing is repeated a plurality of times, and the error type is randomly selected each time. FIG. 20 shows the relationship between the number of mixing errors and the performance.

FIGS. 17 and 18 show that the MDT model proposed in this embodiment is more robust than the comparison method. In particular, it was confirmed that the MDT model was very robust against insertion errors, and that there was almost no deterioration in complementation performance for insertion of less than 10 morphemes. Also, from Table 23, it was confirmed that, as a result of actual speech recognition, an input including an error was superior to SDT as well. We expect that the MDT model is superior in terms of robustness for the following two reasons. (A) Because attribute matching is performed based on the presence or absence of a morpheme or semantic attribute. (B) A decision tree matched with a relatively rare element is not finally selected by MCL preference.

The former only increases the number of elements to be collated due to erroneous insertion of morphemes, and in the collation method proposed in the present embodiment, guarantees that the attributes that should be collated are always collated by morpheme insertion. . In the latter case, even if they are over-matched, the answer candidates obtained by using them as information sources are ultimately determined by the MCL preference because the attribute over-matched by the above assumption is a rare element in some sense. It means that there is a high possibility that it will not be selected.

As described above, it has been confirmed that the MDT model is robust against speech recognition errors. However, even if there is no error, the MDT model may function effectively. This is because the framework for obtaining answers by MCL preference for answer candidates obtained from different information sources through the decision maker means supporting the answer of the attribute group whose similar case was the largest at the time of learning. is there. This hypothesis was verified in the left column of “correct answer input” in Table 23. According to this, MDT showed a clearly higher complementing performance than any SDT even for correct input.

As described above, speech dialogue processing requires robustness not found in text processing. In the present embodiment, however, the decision tree proposed by us is used for complementing subject abbreviations that frequently appear in dialogues. It was shown that the proposed method was robust against speech recognition errors. In an experiment by the inventor, the performance of the speech recognition result was about several percent lower than that of the correct text input, and was particularly robust against input errors. The findings obtained in the present embodiment are summarized below. (A) The decision tree itself has some robustness. (B) A more robust decision can be made by using a multiple decision tree (MDT) model and MCL preference. (C) The multi-decision tree model has higher performance than the single decision tree (SDT) model even for error-free deterministic element inputs. Therefore, it can be concluded that the model proposed in the present embodiment has sufficient robustness particularly against input string insertion errors.

As described above, the present embodiment has the following unique effects. (A) A decision tree enabling omission complementation or semantic determination can be automatically created without manually selecting an optimal feature pattern, and an appropriate omitted element is complemented by using the created decision tree. Alternatively, the meaning (or bilingual term) for the translation device is determined. (B) A decision tree enabling omission complementation or semantic determination can be created more accurately and automatically without manually adjusting parameters and determining priorities, and is appropriately performed by using the created decision tree. Such omitted elements are complemented, or the meaning (or translated word) for the translation device is determined. (C) A decision tree that enables omission complementation or semantic determination for automatically adding a new feature is automatically created, and appropriate omitted elements are supplemented by using the created decision tree, or the translation device Meaning (or translated word) is determined. (D) Compared to the single decision tree learning in the conventional example, even if there is an error in the input of the process, the deterioration of reliability can be suppressed. Therefore, it is possible to complement appropriate omitted elements with higher precision than in the conventional example, or to determine the meaning (or bilingual term) for the translation device. (E) Even if an input symbol string that may contain noise, such as a speech recognition result or an OCR character input result, is used as an input sentence, deterioration in accuracy can be suppressed. Therefore, it is possible to complement appropriate omitted elements with higher precision than in the conventional example, or to determine the meaning (or bilingual term) for the translation device.

<Modifications> In the above embodiment, the Japanese language omission element complementing device and the semantic determination device for the translation device have been described. However, the present invention is not limited to this, and the following language processing device is applicable. Can be applied. (A) Text data to which information for resolving language ambiguity has been added instead of the text data for which the omitted element has been complemented. (B) The attribute list is a language attribute for resolving language ambiguity, and specifically includes the attribute shown in the first embodiment or the second embodiment. (C) As in the first and second embodiments, the attribute is divided into a plurality of sections, and an attribute table is created for each of them.
In addition, a decision tree is similarly created. Here, the attribute table includes information indicating whether each attribute exists and correct information of information for resolving ambiguity. (D) Instead of the omitted element complementing unit 14, the attribute value extracting unit 13
, Using the decision tree generated by the decision tree learning unit 12 for the attribute value output from the buffer memory 24 through the buffer memory 24, the language ambiguity in the character string of the input natural utterance sentence is resolved. And a plurality of information complementing means for determining information to be output and outputting a character string in which the information is complemented. (E) Instead of the omitted element complementing selection unit 15, the answer candidate having the highest frequency among the answer candidates from the information element means (or a plurality of n answer candidates having the higher frequency (n-best)) is used. ) May be selected as the final answer, and a character string in which the information of the answer is complemented is generated and output.

[0116]

As described above in detail, according to the language processing apparatus of the first aspect of the present invention, the language processing apparatus comprises a character string in a predetermined language, and information for resolving the ambiguity of the sentence is supplemented in advance. Based on the supplemented information-completed text data and the attribute list of multiple attributes including language attributes for resolving language ambiguities, each sentence of each text data and each attribute included in the attribute list are To clarify whether each attribute in each attribute set exists for each sentence, and to resolve language ambiguity by checking all combinations with each attribute set when divided into multiple sets A plurality of generating means for generating an attribute table indicating the correct answer information of each attribute set, and all the attributes in each attribute set based on the plurality of attribute tables generated by the plurality of generating means. Yo Select the attribute with the largest difference between the entropy after division and the entropy before division, and update the decision tree to generate a node divided by the selected attribute value until it reaches a leaf node By learning, a decision tree having a tree structure of a binary tree format that is divided depending on the attribute value of each attribute, and a decision tree for determining correct information for resolving language ambiguity,
Based on a plurality of learning means generated for each attribute set, and referring to the attribute list based on a character string input in a predetermined language, whether or not each attribute exists by performing a matching for each attribute Extracting means for extracting and outputting an attribute value indicating whether
For each of the attribute values output from the extracting means, each of the decision trees generated by the plurality of learning means is used to determine correct answer information for resolving language ambiguity in the input character string. A plurality of complementing means to output together with the frequency, and, based on the correct answer information output from the plurality of complementing means and the frequency, select a predetermined number of correct answer information having a maximum frequency or a higher frequency, and And complement selection means for outputting a character string in which information including correct answer information has been complemented. Therefore, the present invention has the following specific effects. (A) A decision tree for determining correct information for resolving language ambiguity is automatically created without manually selecting an optimal feature pattern, and an appropriate decision tree is created by using the created decision tree. Correct answer information is determined. (B) A decision tree for determining correct answer information for resolving language ambiguity is more accurately and automatically created without manually adjusting parameters or determining priorities, and the created decision tree The appropriate correct information is determined by using. (C) A decision tree for determining correct information is automatically created for addition of a new feature, and appropriate correct information is determined by using the created decision tree. (D) Compared to the single decision tree learning in the conventional example, even if there is an error in the input of the process, the deterioration of reliability can be suppressed. Therefore, it is possible to determine appropriate correct information with higher accuracy than in the conventional example. (E) Even if an input symbol string that may contain noise, such as a speech recognition result or an OCR character input result, is used as an input sentence, deterioration in accuracy can be suppressed. Therefore, it is possible to determine appropriate correct information with higher accuracy than in the conventional example.

Further, in the language processing device according to the second aspect, the language processing device according to the first aspect is a Japanese omitted element complementing device, and the information-completed text data is
Omitted element-completed text data consisting of a character string of a Japanese natural utterance sentence, in which the correct person of the omitted element of the sentence has been previously complemented. It is an attribute list of a plurality of attributes including an end-of-sentence expression and speaker information as out-of-language information. The attribute table indicates whether each attribute exists for each sentence, and what abbreviation element is the first person. Is an attribute table indicating correct correct person names, each of the decision trees is a decision tree for determining correct correct person names of omitted elements, and each of the complementing means is a correct correct person name of omitted elements in the input character string. Is determined, and the complement selecting means selects a predetermined number of correct personal names having the highest frequency or higher frequency, and outputs a character string in which the omitted element including the correct personal name is complemented. Here, the omitted element is preferably a pronoun of the subject or the object in the Japanese natural utterance sentence. Therefore, the present invention has the following specific effects. (A) A decision tree enabling omission complementation is automatically created without manually selecting an optimal feature pattern, and an appropriate omitted element is complemented by using the created decision tree. (B) A decision tree enabling omission completion can be created more accurately and automatically without manually adjusting parameters or determining priorities, and appropriate omission elements can be obtained by using the created decision tree. Is complemented. (C) A decision tree capable of performing omission complementation even when a new feature is added is automatically created, and an appropriate omitted element is complemented by using the created decision tree. (D) Compared to the single decision tree learning in the conventional example, even if there is an error in the input of the process, the deterioration of reliability can be suppressed. Therefore, it is possible to complement appropriate omitted elements with higher accuracy than in the conventional example. (E) Even if an input symbol string that may contain noise, such as a speech recognition result or an OCR character input result, is used as an input sentence, deterioration in accuracy can be suppressed. Therefore, it is possible to complement appropriate omitted elements with higher accuracy than in the conventional example.

Further, in the language processing apparatus according to the fourth aspect, the language processing apparatus according to the first aspect is a semantic determination apparatus for a translation apparatus that translates from a first language to a second language. The information-completed text data is composed of a character string of a sentence in the first language, and is semantically determined text data in which the meaning of the sentence is complemented in advance. An attribute list of a plurality of attributes including word information. The attribute table is an attribute table indicating whether or not each attribute exists for each sentence, and the correct meaning. Each of the decision trees is a correct word meaning. Is a decision tree for determining the correct word meaning in the input character string, and the complement selecting means selects a predetermined number of correct word meanings having the maximum frequency or the higher frequency. do it,
A character string with the correct word meaning complemented is output. here,
The correct word meaning is preferably a bilingual element of the second language. Therefore, the present invention has the following specific effects. (A) A decision tree capable of determining meaning is automatically created without manually selecting an optimal feature pattern, and by using the created decision tree, an appropriate meaning for a translation device ( Or a bilingual sentence) is determined. (B) Decision trees that enable semantic determination can be created more accurately and automatically without manually adjusting parameters and determining priorities, and using the created decision trees to provide appropriate translation. The meaning (or bilingual sentence) for the device is determined. (C) A decision tree that enables the meaning to be determined even when a new feature is added is automatically created, and by using the created decision tree, an appropriate meaning (or a translated sentence) for the translation device can be obtained. It is determined. (D) Compared to the single decision tree learning in the conventional example, even if there is an error in the input of the process, the deterioration of reliability can be suppressed. Therefore, it is possible to determine an appropriate meaning (or a translated sentence) for the translation device with higher accuracy than the conventional example. (E) Even if an input symbol string that may contain noise, such as a speech recognition result or an OCR character input result, is used as an input sentence, deterioration in accuracy can be suppressed. Therefore, it is possible to determine an appropriate meaning (or a translated sentence) for the translation device with higher accuracy than the conventional example.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the configuration of a Japanese language elimination element complementing device according to a first embodiment of the present invention.

FIG. 2 is a flowchart illustrating an attribute table generation process executed by an attribute table generation unit 11 of FIG. 1;

FIG. 3 is a flowchart showing a decision tree learning process executed by a decision tree learning unit 12 in FIG. 1;

FIG. 4 is a flowchart showing a decision tree generation process which is a subroutine of FIG. 3;

FIG. 5 is a flowchart showing an attribute value extraction process executed by the attribute value extraction unit 13 of FIG.

FIG. 6 is a flowchart illustrating an omitted element complementing process executed by the omitted element complementing unit 14 of FIG. 1;

FIG. 7 is a diagram illustrating a difference between the conventional SDT model having one decision tree and the MDT model processing having a plurality of decision trees according to the present embodiment.

FIG. 8 is a tree structure diagram showing an example of a decision tree based on an attribute set AS1 used in the first embodiment.

FIG. 9 is a tree structure diagram showing an example of a decision tree based on an attribute set AS2 used in the first embodiment.

FIG. 10 is a tree structure diagram showing an example of a decision tree based on an attribute set AS3 used in the first embodiment.

FIG. 11 is a tree structure diagram showing an example of a decision tree based on an attribute set AS4 used in the first embodiment.

FIG. 12 is a block diagram illustrating a configuration of a meaning determining device for a translation device according to a second embodiment of the present invention.

FIG. 13 shows an attribute set AS11 used in the second embodiment.
FIG. 3 is a tree structure diagram showing an example of a decision tree based on the above.

FIG. 14 shows an attribute set AS12 used in the second embodiment.
FIG. 3 is a tree structure diagram showing an example of a decision tree based on the above.

FIG. 15 shows an attribute set AS13 used in the second embodiment.
FIG. 3 is a tree structure diagram showing an example of a decision tree based on the above.

FIG. 16 shows an attribute set AS14 used in the second embodiment.
FIG. 3 is a tree structure diagram showing an example of a decision tree based on the above.

FIG. 17 is a graph showing experimental results of the language processing device according to the first embodiment, showing performance with respect to insertion errors;

FIG. 18 is a graph showing experimental results of the language processing apparatus according to the first embodiment, showing performance against missing errors.

FIG. 19 is a graph showing experimental results of the language processing apparatus according to the first embodiment, showing performance with respect to replacement errors.

FIG. 20 is a graph showing experimental results of the language processing apparatus according to the first embodiment, showing performance against a mixing error;

[Explanation of symbols]

11: Attribute table generator, 12, 12-1 to 12-M: decision tree learning unit, 13: attribute value extractor, 14, 14-1 to 14-M: omitted element complementer, 14a, 14a-1 to 14a-M: Meaning determination unit, 15: Omission element complementing selection unit, 15a: Meaning determination selection unit, 20: Attribute list memory, 20a: Attribute list memory 21: Text data memory for which omitted elements have been supplemented 21a: Text data memory for which meaning has been determined 22: Attribute table memory 22a: Attribute table memory 23, 23-1 to 23-M: Decision tree file memory 23A, 23a- 1 to 23a-M: decision tree file memory; 24: buffer memory

────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] July 6, 1999 (1999.7.6)

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Name of invention

[Correction method] Change

[Correction contents]

Patent application title: Language processing device and meaning determining device

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] Claims

[Correction method] Change

[Correction contents]

[Claims]

[Procedure amendment 3]

[Document name to be amended] Statement

[Correction target item name] 0010

[Correction method] Change

[Correction contents]

[0010]

According to a first aspect of the present invention, there is provided a language processing apparatus comprising a character string of a predetermined language, wherein information for eliminating ambiguity of a sentence is complemented in advance. Each sentence of each text data and each attribute included in the attribute list are divided into a plurality of sets based on the completed text data and the attribute list of multiple attributes including the language attribute for resolving language ambiguity. For all combinations with each attribute set at the time of division, collate them and determine whether each attribute in each attribute set exists for each sentence,
And a plurality of generating means for generating, for each attribute set, attribute tables indicating correct answer information for resolving language ambiguity, and storing an attribute table for each attribute set generated by the plurality of generating means. A first storage device that performs
Based on the plurality of attribute tables stored in the storage device, the attribute having the largest difference between the entropy after division and the entropy before division by all the attributes in each attribute set is selected, and the selected attribute is selected. By learning to update the decision tree to generate nodes divided by value, until reaching the leaf nodes, each leaf node has a frequency and is split depending on the attribute value of each attribute A plurality of learning means for generating, for each attribute set, a decision tree having a tree structure in a binary tree format for determining correct answer information for resolving language ambiguity, and the plurality of learning means Collating each attribute by referring to the attribute list based on a second storage device that stores a decision tree for each attribute set generated by the above and a character string in a predetermined language that is input. By Extracting means for extracting and outputting an attribute value indicating whether or not an attribute exists; and using the respective decision trees stored in the second storage device for the attribute value output from the extracting means. Each of the plurality of complementing means for determining and outputting the correct answer information for resolving the ambiguity of the language in the input character string and the frequency thereof, and the correct answer information output from the plurality of complementary means and the frequency thereof. And a selection means for selecting a predetermined number of correct answer information having the highest frequency or a higher frequency based on the selected information and outputting a character string supplemented with information including the correct answer information.

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] 0011

[Correction method] Change

[Correction contents]

Further, in the language processing apparatus according to the second aspect, the language processing apparatus according to the first aspect is a Japanese abbreviation element complementing apparatus, and the information-completed text data is:
Omitted element-completed text data consisting of a character string of a Japanese natural utterance sentence, in which the correct person of the omitted element of the sentence has been previously complemented. It is an attribute list of a plurality of attributes including an end-of-sentence expression and speaker information as out-of-language information. Is an attribute table indicating correct correct person names, each of the decision trees is a decision tree for determining correct correct person names of omitted elements, and each of the complementing means is a correct correct person name of omitted elements in the input character string. Is determined, and the complement selection means selects a predetermined number of correct personal names having the highest frequency or higher frequency, and outputs a character string in which the omitted element including the correct personal name is complemented.

[Procedure amendment 5]

[Document name to be amended] Statement

[Correction target item name] 0012

[Correction method] Change

[Correction contents]

Further, the language processing device according to claim 3 is
3. The language processing apparatus according to claim 2, wherein the omitted element is a subject of a subject or an object in a Japanese natural utterance sentence.

[Procedure amendment 6]

[Document name to be amended] Statement

[Correction target item name] 0013

[Correction method] Change

[Correction contents]

A semantic determination device according to a fourth aspect of the present invention is a semantic determination device for a translation device for translating a first language into a second language. Semantics-determined text data consisting of a character string and the meaning of a given word of the sentence is determined in advance, and attributes of a plurality of attributes including information on the sentence type, tense, subject, and object for determining the meaning. On the basis of the list, each sentence of each text data and each attribute set when each attribute included in the attribute list is divided into a plurality of sets are checked against all combinations, and each sentence is checked for each sentence. An attribute table indicating whether or not each attribute in the attribute set exists, and an attribute table indicating correct word meaning, respectively, a plurality of generating means for each attribute set, and a plurality of attribute tables for each attribute set generated by the plurality of generating means. First storage device for storing an attribute table And selecting the attribute having the largest difference between the entropy after division and the entropy before division by all the attributes in each attribute set based on the plurality of attribute tables stored in the first storage device. By learning to update the decision tree to generate nodes divided by the selected attribute value until the leaf node is reached, each leaf node has a frequency and the attribute value of each attribute A plurality of learning means for generating, for each attribute set, a decision tree having a tree structure of a binary tree format that is dependently divided, for each attribute set; and a plurality of learning trees generated by the plurality of learning means. A second storage device for storing a decision tree for each attribute set, and a matching for each attribute by referring to the attribute list based on the input character string of the sentence in the first language, Whether the attribute exists Extracting means for extracting and outputting an attribute value indicating whether or not the attribute value is output from the extracting means, using the respective decision trees stored in the second storage device. A plurality of meaning determining means for determining the meaning representing the meaning of the character string and outputting the meaning along with the frequency; and a predetermined meaning having a maximum frequency or a higher frequency based on the meaning and the frequency output from the plurality of meaning determining means. Means for selecting a number of meanings and outputting a character string of the second language in which the meanings are determined.

[Procedure amendment 7]

[Document name to be amended] Statement

[Correction target item name] 0014

[Correction method] Change

[Correction contents]

Further, the meaning determination device according to claim 5 is
5. The meaning determining device according to claim 4, wherein the correct meaning is a bilingual element in a second language.

[Procedure amendment 8]

[Document name to be amended] Statement

[Correction target item name]

[Correction method] Change

[Correction contents]

[0116]

As described above in detail, according to the language processing apparatus of the first aspect of the present invention, the language processing apparatus comprises a character string in a predetermined language, and information for resolving the ambiguity of the sentence is supplemented in advance. Based on the supplemented information-completed text data and the attribute list of multiple attributes including language attributes for resolving language ambiguities, each sentence of each text data and each attribute included in the attribute list are To clarify whether each attribute in each attribute set exists for each sentence, and to resolve language ambiguity by checking all combinations with each attribute set when divided into multiple sets A plurality of generating means for generating an attribute table indicating correct answer information for each attribute set, a first storage device for storing an attribute table for each attribute set generated by the plurality of generating means, First memory Based on a plurality of attribute tables stored in each attribute set, the attribute having the largest difference between the entropy after division by all the attributes in each attribute set and the entropy before division is selected, and the selected attribute value By learning to update the decision tree to generate a split node until it reaches a leaf node, each leaf node has a frequency and is split depending on the attribute value of each attribute A plurality of learning means for generating, for each attribute set, a decision tree for determining correct information for resolving language ambiguity having a tree structure of a simple binary tree, and the plurality of learning means; The second storage device that stores the determined decision tree for each attribute set, and the above-described attribute list based on the input character string in a predetermined language, performs matching for each attribute, attribute Extracting means for extracting and outputting an attribute value indicating whether or not the attribute value exists; and using the respective decision trees stored in the second storage device for the attribute value output from the extracting means. A plurality of complementing means for determining the correct answer information for resolving the ambiguity of the language in the input character string and outputting the information together with the frequency, and based on the correct answer information output from the plurality of complementary means and the frequency. And complement selection means for selecting a predetermined number of correct answer information having a maximum frequency or a higher frequency and outputting a character string in which information including the correct answer information is complemented. Therefore, the present invention has the following specific effects. (A) A decision tree for determining correct information for resolving language ambiguity is automatically created without manually selecting an optimal feature pattern, and an appropriate decision tree is created by using the created decision tree. Correct answer information is determined. (B) A decision tree for determining correct answer information for resolving language ambiguity is more accurately and automatically created without manually adjusting parameters or determining priorities, and the created decision tree The appropriate correct information is determined by using. (C) A decision tree for determining correct information is automatically created for addition of a new feature, and appropriate correct information is determined by using the created decision tree. (D) Compared to the single decision tree learning in the conventional example, even if there is an error in the input of the process, the deterioration of reliability can be suppressed. Therefore, it is possible to determine appropriate correct information with higher accuracy than in the conventional example. (E) Even if an input symbol string that may contain noise, such as a speech recognition result or an OCR character input result, is used as an input sentence, deterioration in accuracy can be suppressed. Therefore, it is possible to determine appropriate correct information with higher accuracy than in the conventional example.

[Procedure amendment 9]

[Document name to be amended] Statement

[Correction target item name]

[Correction method] Change

[Correction contents]

According to a fourth aspect of the present invention, there is provided a semantic determination device for a translation device for translating from a first language to a second language, wherein A sentence character string in which the meaning of a given word in the sentence is determined in advance, and a plurality of attributes including information on the sentence type, tense, subject, and object for determining the meaning. Based on the attribute list, each sentence of each text data is checked against all combinations of each attribute set when each attribute included in the attribute list is divided into a plurality of sets. A plurality of generating means for generating, for each attribute set, whether or not each attribute in each attribute set exists and the correct word meaning, and each attribute set generated by the plurality of generating means. The first to store the attribute table for each A storage device and an attribute having the largest difference between entropy after division and entropy before division by all attributes in each attribute set are selected based on the plurality of attribute tables stored in the first storage device. Then, by learning to update the decision tree to generate a node divided by the selected attribute value until reaching the leaf node, each leaf node has a frequency and the attribute of each attribute A plurality of learning means for generating, for each attribute set, a decision tree having a tree structure of a binary tree format that is divided depending on a value, for each attribute set; A second storage device for storing a decision tree for each attribute set, and performing matching for each attribute by referring to the attribute list based on an input character string of a sentence in the first language. Each attribute exists by Extracting means for extracting and outputting an attribute value indicating whether or not the attribute value is output from the extracting means; A plurality of meaning determining means for determining the meaning representing the meaning of the input character string and outputting the meaning along with the frequency; and based on the meaning and the frequency output from the plurality of meaning determining means, the maximum frequency or the frequency is determined. Means for selecting a predetermined larger number of meanings and outputting a character string of the second language for which the meaning is determined. Here, the correct word meaning is preferably a bilingual element in the second language. Therefore, the present invention has the following specific effects. (A) A decision tree capable of determining meaning is automatically created without manually selecting an optimal feature pattern, and by using the created decision tree, an appropriate meaning for a translation device ( Or a bilingual sentence) is determined. (B) Decision trees that enable semantic determination can be created more accurately and automatically without manually adjusting parameters and determining priorities, and using the created decision trees to provide appropriate translation. The meaning (or bilingual sentence) for the device is determined. (C) A decision tree that enables the meaning to be determined even when a new feature is added is automatically created, and by using the created decision tree, an appropriate meaning (or a translated sentence) for the translation device can be obtained. It is determined. (D) Compared to the single decision tree learning in the conventional example, even if there is an error in the input of the process, the deterioration of reliability can be suppressed. Therefore, it is possible to determine an appropriate meaning (or a translated sentence) for the translation device with higher accuracy than the conventional example. (E) Even if an input symbol string that may contain noise, such as a speech recognition result or an OCR character input result, is used as an input sentence, deterioration in accuracy can be suppressed. Therefore, it is possible to determine an appropriate meaning (or a translated sentence) for the translation device with higher accuracy than the conventional example.

 ────────────────────────────────────────────────── ─── Continued on the front page (72) Inventor Eiichiro Sumida 5 Shiraya, Inaya, Seika-cho, Soraku-cho, Kyoto Pref. AT R.R. 5 Seira-cho, Seiya-cho, Gunma, Subdivision 5, Sanriya, ATR S.A.R.

Claims (5)

[Claims] 1. An information-completed text data composed of a character string in a predetermined language, in which information for resolving the ambiguity of the sentence is complemented in advance, and a language attribute for resolving the ambiguity of the language. Based on the attribute list of the plurality of attributes included, each sentence of each text data is checked against all combinations of each attribute set when each attribute included in the attribute list is divided into a plurality of sets. A plurality of generation means for generating, for each attribute set, whether or not each attribute in each attribute set exists for each sentence, and an attribute table indicating correct answer information for resolving language ambiguity; Based on the plurality of attribute tables generated by the plurality of generating means, the attribute having the largest difference between the entropy after division and the entropy before division by all the attributes in each attribute set is selected. Then, by learning to update the decision tree so as to generate a node divided by the selected attribute value until it reaches a leaf node, the decision tree can be divided depending on the attribute value of each attribute. A plurality of learning means for generating, for each attribute set, a decision tree for determining correct information for resolving language ambiguity having a tree structure of a simple binary tree; Extracting means for extracting and outputting an attribute value indicating whether or not each attribute exists by performing matching for each attribute by referring to the attribute list based on a character string; and outputting from the extracting means. For each attribute value, the correct answer information for resolving the ambiguity of the language in the input character string is determined by using each of the decision trees generated by the plurality of learning means, and together with the frequency, Out A plurality of complementing means, and based on the correct answer information output from the plurality of complementary means and the frequency thereof, select a predetermined number of correct answer information having a maximum frequency or a higher frequency, and obtain information including the correct answer information. A language processing apparatus comprising: a complement selection unit that outputs a complemented character string. 2. A language processing apparatus according to claim 1, wherein the information-completed text data comprises a character string of a Japanese natural utterance sentence, and the correct answer of the omitted element of the sentence. The abbreviation element supplemented text data in which the person is complemented in advance, and the attribute list includes a plurality of attributes including a normal form of the verb, a semantic attribute of the verb, a sentence end expression, and speaker information as extralingual information. The attribute table is an attribute table indicating whether or not each attribute exists for each sentence, and the correct personal name of the omitted element, and each of the decision trees is omitted. A decision tree for determining a correct person of the element; each of the complementing means determines a correct person of the omitted element in the input character string; and the complement selecting means has a maximum frequency or a higher frequency. Select a certain number of correct personals And outputting a character string in which the omitted element including the correct person is complemented. 3. The language processing apparatus according to claim 2, wherein the omitted element is a subject of a subject or an object in a Japanese natural utterance sentence. 4. The language processing apparatus according to claim 1, wherein the language processing apparatus is a semantic determination apparatus for a translation apparatus for translating from a first language to a second language, wherein the information-completed text data is in a first language. The sentence is text data for which the meaning of the sentence has been determined in advance and the meaning of the sentence is complemented in advance. The attribute table is an attribute table indicating whether each attribute exists for each sentence and the correct meaning, and each of the decision trees is a decision tree for determining the correct meaning. Each complementing means determines the correct word meaning in the input character string, and the complement selecting means selects a predetermined number of correct word meanings having the highest frequency or higher frequency, and the character whose correct word meaning is complemented. A language characterized by outputting columns Processing equipment. 5. The language processing apparatus according to claim 4, wherein the correct meaning is a bilingual element of a second language.