JP2960936B2 - Dependency analyzer - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a dependency analysis apparatus for performing a sentence dependency analysis method in natural language processing such as machine translation, natural language understanding, and automatic index extraction. Things. [Related Art] Dependency analysis of a sentence is a pre-process of natural language processing, and the accuracy of the dependency greatly affects the performance of the entire processing. Therefore, a highly accurate dependency analysis method is strongly desired. In general, there is a rule in Japanese language dependency that a received word is placed as close to the word as possible. But,
Not all of the dependency relationships fall under the above rules,
Ambiguous dependency relationships occur. Therefore, various methods have been conventionally proposed in order to accurately determine the dependency relationship. Kinukawa et al. (Kinukawa, Kimura: Automatic indexing method by Japanese sentence structure analysis, IPSJ Transactions, vo1.21, no.
3, 1980) has proposed. FIG. 6 is an example of a sentence pattern table of verb case relations used in the above dependency analysis method. Focus on verbs in sentences,
The noun and case particle connected to the verb and the semantic classification of the noun phrase are defined, and dependency analysis is performed using the relation. However, this method requires a difficult action to describe the case relation between the verb and the noun for all verbs.
Furthermore, in the case of dependency from a noun phrase to a noun phrase or dependency on a noun phrase by modifying a continuation of a word, there is a problem that it is difficult to determine the dependency. Takamatsu, Nishida et al. (Takamatsu, Kusaka, Nishida: Automatic extraction of related information from technical abstracts, IPSJ Transactions vo1.25, no.2,
1984) supplements the problem of the dependency analysis method based on the sentence pattern table of Kinukawa et al. This method combines a case structure pattern of a verb with knowledge of a specialized field described individually in a technical abstract such as a claim sentence, and cannot be analyzed by case structure relation alone. The analysis can be determined. The method will be described using a specific example. First, the case structure of the input sentence is analyzed, and then the access table to the knowledge table in FIG. 7 is referred to. Next, based on the access table, a set of case labels corresponding to the case structure of the sentence is searched from the knowledge table. Then, the case related to the case described in the knowledge table is set as the maximum likelihood dependency candidate, and the dependency is determined. FIG. 8 is an example of a knowledge table regarding a semiconductor device. Ninth
An example of dependency analysis of the above method is shown in the sentence example in the figure. In this example, two interpretations (a) and (b) hold whether the verb “contain” relates to “on the silicon substrate” or “insulating layer”. OBJ and PARTIC are taken as the case of the "contain" type verb,
The seventh is that there is a conceptual relationship of "COMPOSITION" between them.
It is determined from the access table in FIG. Next, a knowledge table of “COMPOSITION” in the semiconductor device of FIG. 8 is searched to check an example of a case relationship between OBJ and PARTIC. Actually, case examples of "channel region" (PARTIC) and "silicon substrate" (OBJ) are extracted from the knowledge table, and the channel region is included in the silicon substrate, but the insulating layer is not included in the silicon substrate. Knowledge can be extracted,
In this example, it is determined that the dependency of (a) is correct. [Problems to be solved by the invention] Dependency analysis on association modification can be performed accurately by this method, but unless various knowledge tables such as constituents and detailed knowledge tables for each specialized field are created, There is a problem that it does not. An object of the present invention is to eliminate the ambiguity of dependency by using information in a sentence without using detailed case relations or specialized knowledge when parsing a sentence having a lot of repeated expressions such as a claim sentence. Another object of the present invention is to provide a dependency analysis method. [Means for Solving the Problems] A dependency analyzing apparatus according to the present invention provides a connection table for recording a connection relationship consisting of a pair of a dependency word and a received word, a meaning category number of the dependency word, and a meaning of the received word. A semantic category number connection relation list that records a semantic category number connection relation consisting of a pair with a category number, and a dependency in which a pair of the connection relation and the semantic category number connection relation is recorded as a dependency relation. It has a dependency candidate storage unit composed of a relation table. Further, a phrase unit dividing unit that divides the input sentence into phrases, a phrase extracting unit that extracts the phrases that are sequentially divided from the beginning of the input sentence, and a phrase that receives the remaining phrases as the related words and receives the remaining phrases as words A dependency candidate extracting unit for extracting inter-phrase dependency candidates composed of pairs of words, and a prefix that divides the compound word into words when the independent word of the extracted phrase is a compound word composed of two or more words. A compound word division unit for removing a word and a suffix, and a word-to-word dependency relationship between the compound words as a connection relationship, and a meaning category of the word and a meaning of the word. A compound word dependency analyzing unit for linking a category pair to a semantic category number connection, and the inter-word dependency relationship is registered in the dependency candidate storage unit, and when the inter-phrase dependency candidate is one, Save the candidate Having a dependency relationship registration unit that registers the new part. When the inter-phrase dependency candidates are two or more, when a candidate that matches the connection relation of the candidates can be retrieved from the connection table, the candidate with the closest distance between the dependency word and the received word is determined to be the correct answer. However, if the correct answer cannot be determined, and a candidate that matches the connection relation of the candidates can be retrieved from the semantic category connection relation list, the candidate with the closest distance between the dependent word and the received word is determined to be the correct answer. A dependency candidate search unit that determines whether a candidate has the same semantic category number as a dependency word in the candidate when the correct answer cannot be determined in the dependency candidate search unit. A dependency determination unit that determines whether the candidate is correct if the word exists, and suspends the determination of inter-phrase dependency if the word does not exist. Dependency relation determination unit that, for the inter-phrase dependency candidates for which the correct answer could not be determined at the time of completion of the dependency analysis of all phrases, accurately determines the inter-phrase dependency candidate with the shortest distance between the dependency word and the received word. And Note that the dependency candidate search unit determines that a candidate whose connection relationship with the candidate for which the determination of the dependency has been suspended matches the connection relationship registered in the above is determined to be a correct answer. Is determined as a correct answer if the connection relationship of the candidate for which the reservation is suspended coincides with the newly registered connection relationship. [Operation] In the dependency analyzing apparatus according to the present invention, the dependency candidate storage unit stores, as a dependency relationship table, a connection table that records a connection relationship composed of a pair of a dependency word and a reception word, a semantic category number of the dependency word, and the like. A semantic category number concatenation relation list that records concatenation relations of semantic category numbers composed of pairs with semantic category numbers of received words, and records pairs of connection relations and semantic category number concatenation relations as dependency relations . When a sentence is input, the input sentence is divided into phrases by a phrase unit dividing unit. The phrase extraction unit extracts phrases that are sequentially divided from the beginning of the input sentence, and the dependency candidate extraction unit extracts inter-phrase dependency candidates consisting of pairs in which the extracted phrases are used as the dependent words and the remaining phrases are used as the received words. Extract. When the extracted independent word of the phrase is a compound word composed of two or more words, the compound word dividing unit divides the compound word into words and removes the prefix and the suffix. The compound dependency analysis unit connects the pair of the dependency word and the received word relating to the immediately succeeding word as the inter-word dependency relationship, and connects the pair of the meaning category of the dependency word and the meaning category of the received word to the semantic category number. Relationship. The dependency relation registration unit newly registers the dependency relation between words in the dependency candidate storage unit, and newly registers the candidate in the dependency candidate storage unit when there is one inter-phrase dependency candidate. When the number of inter-phrase dependency candidates is two or more, the dependency candidate search unit searches the connection table for a candidate that matches the connection relationship of the candidate, and the distance between the dependency word and the received word is the shortest. If the candidate is determined to be correct, and the correct answer cannot be determined, the candidate that has the closest relationship between the dependent word and the received word is found when the candidate that matches the connection relationship of the candidate can be searched from the semantic category connection relationship list. Is determined to be the correct answer. The dependency determination unit determines whether there is a candidate having the same semantic category number of the dependency word and the semantic category number of the received word in the candidate when the correct candidate cannot be determined by the dependency candidate search unit. If it exists, the candidate is regarded as the correct answer, and if it does not exist, the determination of inter-phrase dependency is suspended. The dependency relation determination unit determines that the inter-phrase dependency candidate with the shortest distance between the dependency word and the received word among the inter-clause dependency candidates for which the correct answer could not be determined at the time when the dependency analysis of all the clauses has been completed is determined to be the correct answer. I do. Note that the dependency candidate search unit determines that the candidate whose connection relation has been suspended and whose match is identical to the newly registered connection relation is a correct answer. If the correct answer cannot be determined, the dependency Is determined as a correct answer if the connection relationship of the candidate for which is suspended is coincident with the newly registered connection relationship. Therefore, the dependency analyzing apparatus of the present invention can determine the dependency relationship between the phrases obtained by dividing the input sentence in one way. Here, at the time of the dependency analysis, the dependency relationship between words constituting a compound word and the inter-segment dependency candidate defined as one are registered in the dependency candidate storage unit and used for the subsequent dependency analysis. [Example] Hereinafter, an example of the present invention will be described. In the embodiment of the present invention, both the dependency word and the dependency word in the dependency relationship are in terms of a phrase. As the dependency relationship, a connection relationship between the dependency category and the semantic category of the dependency word and a connection relationship between the dependency word and the dependency word are used. Here, the semantic category represents a common semantic concept of a word, and a semantic category number or the like corresponding to the concept is assigned to each word. The semantic category numbers used in this example are classified vocabulary tables published by the National Institute of Japanese Language and Linguistics (National Institute for Japanese Language Linguistics 6 Classified vocabulary tables, Hideei Shuppan, 1964)
Use the one described in. FIG. 1 is a flowchart of a dependency analysis process used in the dependency analysis device of the present invention, and FIG. 2 is a block diagram showing an embodiment of the dependency analysis device according to the present invention. First,
The configuration shown in FIG. 2 will be described, and then the operation will be described with reference to FIG. In FIG. 2, reference numeral 1 denotes a phrase unit dividing unit, which divides an input sentence into phrases. Reference numeral 2 denotes a phrase extraction unit which extracts phrases sequentially divided from the beginning of the input sentence. Reference numeral 3 denotes a dependency candidate extraction unit that extracts inter-phrase dependency candidates for pairs in which the extracted phrase is used as a dependency word and the remaining phrases are used as receiving words. Reference numeral 4 denotes a compound word division unit that divides an independent word into words when the extracted independent word of the phrase is a compound word composed of two or more words. Reference numeral 5 denotes a compound word dependency analysis unit for analyzing the dependency relationship between words. Reference numeral 6 denotes a dependency candidate storage unit, which stores a connection table that records a connection relationship consisting of a pair of a dependency word and a reception word, and a semantic category number consisting of a pair of a meaning category number of the dependency word and a meaning category number of the reception word. It is composed of a connection list of semantic category numbers for recording connection relations, and is a dependency relation table for recording pairs of connection relations and connection relations of semantic category numbers as dependency relations. 7
Is a dependency relation registration unit, which newly registers the inter-word dependency candidate in the dependency candidate storage unit 6 and, when there is only one inter-phrase dependency candidate, newly registers the candidate in the dependency candidate storage unit 6 I do. Reference numeral 8 denotes a dependency candidate search unit. When there are two or more inter-phrase dependency candidates, a connection table that matches the connection relation of the candidate is searched from the connection table, and the distance between the dependency word and the received word is the shortest. A close candidate is determined as a correct answer. If the correct answer cannot be determined, a candidate that matches the related relationship of the candidate is searched from the semantic category connection relationship list, and a candidate having the closest distance between the dependent word and the received word is determined to be the correct answer. Reference numeral 9 denotes a dependency determination unit. If the correct candidate cannot be determined by the dependency candidate search unit 8, whether or not there is a candidate having the same semantic category number of the dependency word and the semantic category number of the received word in the candidate Is determined. If it exists, the candidate is regarded as the correct answer. If it does not exist, the determination of the inter-phrase dependency relation is suspended. For the candidate for which the determination of the dependency relationship has been suspended, the dependency candidate search unit 8 determines that the candidate whose connection relationship matches the newly registered connection relationship is a correct answer. A candidate whose connection relationship matches the newly registered connection relationship is determined to be correct. When the dependency analysis of all the clauses is completed, the dependency relation determination unit 10 determines the inter-clause dependency candidate in which the distance between the dependency word and the subject word is the shortest among the inter-clause dependency candidates for which the correct answer could not be determined. Correct answer. The numbers (1) to (16) in FIG. 2 indicate the steps of the dependency analysis performed by each component, and correspond to the steps (1) to (16) in FIG. Each step is performed in the following procedure. First, as a step (1), the phrase unit dividing unit 1 divides an input sentence into phrases. In step (2), the phrase extraction unit 2 extracts phrases sequentially divided from the beginning of the input sentence. In step (3), the dependency candidate extraction unit 3 uses the extracted phrase as a dependency, and determines all the phrase candidates that can be received words according to the part of speech of the independent word in the target phrase, the inflected form, and the type of the auxiliary word to be used, ie, Extract inter-phrase dependency candidates. In step (4), the compound word dividing unit 4 determines whether or not the independent word constituting the hang-up word is a compound word composed of two or more words. When there is a certain time, the process proceeds to step (5), and when not, the process proceeds to step (8). In step (5), the compound word is divided into words, and prefixes and suffixes are removed. In step (6), the compound word dependency analysis unit 5 performs a connection relation consisting of a pair of the dependency word and the immediately following word, and a meaning consisting of a pair of the meaning category number of the dependency word and the meaning category number of the received word. The connection relationship between category numbers, that is, the dependency relationship between words is analyzed. In step (7), the dependency relation registration unit 7 stores the dependency relation between words.
Is recorded in the dependency candidate storage unit 6. In step (8), if there is one inter-phrase dependency candidate, step (9) is executed. In step (9), the candidate is newly recorded in the dependency candidate storage unit 6 by the dependency relationship registration unit 7.
Through these processes, in the inter-phrase dependency analysis, not only inter-phrase dependency information but also inter-word dependency information can be used. Step (12) when there are two or more inter-phrase dependency candidates
Execute Step (12) is the dependency candidate search section 8
In the connection table that matches the connection relationship of the candidate is searched from the connection table, and the candidate with the shortest distance between the dependency and the received word is determined as the correct answer. If the correct answer cannot be determined,
A candidate that matches the connection relation of the candidate is searched from the semantic category connection relation list, and a candidate having the shortest distance between the dependency word and the received word is determined to be the correct answer. Step (13) proceeds to step (17) if the correct answer can be determined by the dependency candidate search unit 8, and proceeds to step (14) if the correct answer cannot be determined. In step (14), the dependency determining unit 9 determines whether or not there is a candidate having the same semantic category number of the dependency word and the semantic category number of the received word in the candidate. In step (15), when the candidate exists, the candidate is determined to be correct, and the process proceeds to step (17). If not present, the process proceeds to step (16), where the dependency determination is suspended, and the process proceeds to step (17). In step (17), the target of the hang-up word is advanced to the next clause. Regarding the dependency candidates for which the dependency determination has been suspended, in the step (10), the candidate whose connection relation matches the connection relation newly registered in the step (7) and the step (9) in the dependency candidate search unit 8 is correctly answered. To determine the dependency relationship between phrases. When a correct answer is not determined, a candidate whose connection relation matches the newly registered connection relation is determined as a correct answer, and the inter-phrase dependency relation is determined. When the dependency analysis of all phrases is completed in step (11), the dependency relation determination unit 10 determines, for the dependency candidates for which the correct answer could not be determined, the candidate having the shortest distance between the dependency word and the received word. Determine as the correct answer. Hereinafter, the dependency analysis used in the dependency analysis apparatus of the present invention will be described by taking the analysis of the claims as an example. FIG. 3 shows an example of a claim sentence. When this sentence is subjected to dependency analysis, there are dependency candidates that cannot be determined in one way. For example, in the example of [1] in the example sentence,
Assuming that the phrase “under construction” is a dependency, there are four types of dependency relations with the following “in the sentence”, “conversion result”, “same phonetic”, and “means” as the receiving words. It is possible.
In the dependency relation table shown in FIG. 4, since a connection relation consisting of a pair with a reception word using "under construction" as a dependency word is not recognized, it is not possible to narrow down candidates for the dependency relation using the connection relation. Therefore, the dependency relationship is determined by using the concatenation relationship of the semantic category numbers, which are pairs of the semantic category numbers of the dependency words and the semantic category numbers of the receiving words. FIG. 4 shows an example of a dependency relation table obtained by executing the dependency analysis up to the phrase “under construction”. Here, 12 is a dependency word and a meaning category number of the dependency word, 13 is a receiving word and a meaning category number of the receiving word, and 14 is a concatenation consisting of a pair of the meaning category number of the dependency word and the meaning category number of the receiving word. 6 shows a connection list of semantic category numbers in which the relation is recorded. In particular, since 12 and 13 indicate the connection relationship between the dependent word and the received word in which the dependent word and the received word are paired, 12 and 13 can be regarded collectively as a connection table. Further, since the dependency relation table is configured in such an arrangement, the dependency relation, that is, the connection relation and the connection relation are recorded in association with each other. Here, the relationship between the first two pairs of “kana”-“kanji” and “kanji”-“conversion” is a dependency relationship extracted from the compound word “kana-kanji conversion” shown in FIG. The second relationship between “kana-kanji conversion” and “conversion result” is a dependency relationship obtained by performing dependency analysis on the dependency word “at the time of kana-kanji conversion” and the receiving word “conversion result”. As a semantic category number as a compound,
A semantic category number of a word that clearly indicates the meaning of a compound in a group of words forming the compound is used rather than a semantic category number of the compound. The compound words used in the claims are often expressed by the last word (excluding the suffix). In the example of FIG. 4, the semantic category number of the last word is used as the semantic category number of the compound word. The last three lists in FIG. 4 are dependency relationships extracted from the sentence “With means to be displayed in the sentence being created” (first line in FIG. 3). An example of the dependency analysis for the sentence shown in [1] using the dependency relationship created up to this point will be described. Dependency candidates (c), (d), (e),
The connection relationships in (f) are (1.386, 1.3154) and (1.38, respectively).
6,1.1112), (1.386,1.3112) and (1.386,1.1113). Among them, the connection relation (c) (1.38) that matches the connection relation recorded in the connection relation list of the semantic category number in FIG.
6,1.3154). In other words, semantic category number 1.
The dependency relationship in which the word of the generation relationship indicated by 386 is the dependency word and the word of the sentence relationship indicated by the semantic category number 1.3154 is the recipient word is selected as a candidate. Therefore, it can be seen that there is a dependency relationship in which the generation-related word "creation" is a dependency and the sentence-related word "sentence" is a candidate for a receiving word. Is the correct answer. As shown in this example, the verb “create” (semantic category number) from the connection relationship (1.386,1.3154)
1.386) indicates the possibility of having a word (text, paper, sentence,...) With a semantic category number of 1.3154 as a case. Therefore, even if the same word is not necessarily used for the dependency, it can be determined that the same dependency relationship is established between the words having the same meaning category number. For example, when you make a sentence ...
It can also be used to determine a dependency relationship between the phrase "make" and the phrase "make". In addition, the dependency relationship in which the dependency word "of the created text ..." and the reception word are reversed can be determined using the connection relationship list of the semantic category number without distinguishing the dependency and the dependency by the connection relationship. . In the example of [3] in FIG. 3, a processing example of a dependency candidate for which a matching connection relationship and a connection relationship are not recorded in the dependency relationship table is shown. Here, since the phrase “selected” is a modified form of the adnominal, it is assumed that there is a possibility that the phrase is related to both a noun (phrase), that is, “homophone display selection means” and “homophone output method”. Here, it is assumed that neither the connection relation nor the connection relation of the dependency candidates matches the dependency table. At this time, step (14) is executed. Therefore, the candidates consisting of words having a meaning category number of 1.3063 equal to the word "selection" and words having a meaning category number of 1.3063 equal to the word "selection" are narrowed down from all candidates for the dependency. . In this case, since the semantic category number of “selection” of the compound word “homophone display selection means” is 1.3063, when “choose” is used as a dependency, “homophone display selection means” becomes the dependency. Relationships are specified. Even if the semantic category number of the dependency word and the semantic category number of the receiving word do not match, a candidate for the dependency is shown in FIG. 3 [2].
There may be two or more types as in the example of. Here, when the phrase "of the conversion result" is used as a dependent word, "homophone" (g) and "means" (h) are received words. Therefore, all the dependency candidates having the phrase “conversion result” as the dependency are stored, the determination of the dependency is suspended, and the dependency analysis of the next phrase is executed. As a result of executing the dependency analysis of the other clauses, the dependency relation using the phrase "conversion result" as a dependency is obtained by using the analysis result of the dependency relation in the range shown by the wavy line 11 in the second row of FIG. Can be determined. In other words, the received word “same phonetic word” is determined in one way from the dependency word “of the conversion result” at the wavy line portion 11.
These connection relations are recorded in the connection table, and a connection relation consisting of a pair of the meaning category number 1.1112 of the word “conversion result” and the meaning category number 1.3112 of the word “same word” is recorded in the connection relation list of the semantic category. . Therefore, the dependency relationship of the phrase “of the conversion result” with respect to the phrase “same word” can be determined from the newly recorded connection relationship or connection relationship. That is, in the example of [2] in FIG. 3, the dependency relationship of (g) in which the preceding phrase is a dependency is determined to be correct by performing dependency analysis on the later phrase. In addition, in the example of [4] in FIG. 3, when the phrase “homophone” is used as a dependent word, as shown in the figure, (k) “display”,
(L) "Select", (m) "Prepared",
(N) There are four types of dependency candidates each having "characteristic" as a receiving word. The connection relation of the semantic category numbers corresponding to the connection relation of the semantic category numbers corresponding to these dependency relations is not recognized in the connection relation list of the semantic category numbers in FIG. However, after the compound word "same word display selecting means" is divided into words, a dependency relationship between the words such as a dependency relationship between the words "same word" and the dependent word "display" as a dependent word, etc. By determining the relationship, it is possible to determine that the phrase “homophone” relates to the phrase (k) “display”. [Effect of the Invention] The dependency analyzing apparatus of the present invention sequentially stores a dependency relationship that can be uniquely determined in a dependency candidate storage unit, and determines a dependency relationship in which a dependency cannot be uniquely determined. By referring to the dependency candidate storage unit, the dependency relationship existing in the dependency candidate storage unit is adopted in preference to the dependency relationship not present in the dependency candidate storage unit. By extracting and using information in each sentence without using, a highly accurate dependency analysis can be realized. Then, when the dependency relation is obtained as a concatenation of semantic categories, it is possible to handle the modification based on the case relation of the verb and the modification by the noun phrase in a unified manner, and also to the sentence including the paraphrase. The relationship can be determined uniquely. According to the present invention, it is possible to reduce the cost required for creating specialized knowledge for each field. In addition, even if this method is realized by hardware or software, an extremely small-scale system can be provided because an area for retaining expert knowledge is not required. Therefore, the dependency analyzer according to the present invention is effective for analyzing sentences that repeatedly use the same expression, such as patent sentences and legal sentences. Further, the analysis apparatus can be applied to automatic document processing such as automatic classification of text, automatic index extraction, and automatic summarization. Also, there is an excellent advantage that it can be applied not only to Japanese sentences but also to English sentences.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flowchart of a processing procedure of a dependency analyzing apparatus of the present invention, FIG. 2 is a block diagram showing one embodiment of a dependency analyzing apparatus of the present invention, and FIG. FIG. 4 is a diagram showing a sentence example of a claim sentence used for describing one embodiment, FIG. 4 is a dependency relation table created in the embodiment, and FIG. 5 is a diagram showing an example of compound word analysis in the embodiment. FIG. 6 is a diagram showing an example of a case relation table used in the conventional dependency analysis method,
7 and 8 show an example of an access table to a knowledge table and a knowledge table used in a conventional dependency analysis method, and FIG. 9 is used to explain an example of execution of a conventional dependency analysis. It is a figure showing the example of a sentence of a patent sentence. In the figure, 1 is a phrase unit division unit, 2 is a phrase extraction unit, 3 is a dependency candidate extraction unit, 4 is a compound word division unit, 5 is a compound word dependency analysis unit, 6 is a dependency candidate storage unit, and 7 is a dependency candidate storage unit. A dependency relationship registration unit, 8 is a dependency candidate search unit, 9 is a dependency determination unit, and 10 is a dependency relationship determination unit.

   ────────────────────────────────────────────────── ─── Continuation of front page    (72) Inventor Fumihiko Kobashi               1-2356 Take, Yokosuka-shi, Kanagawa Sun               Inside the Combined Communication Research Laboratories of the Telegraph and Telephone Corporation                (56) References JP-A-61-260366 (JP, A)                 JP-A-62-197076 (JP, A)

Claims (1)

(57) [Claims] A connection table that records a connection relation composed of a pair of a dependency word and a reception word, and a semantic category number that records a connection relation of a semantic category number composed of a pair of a meaning category number of a modification word and a semantic category number of a reception word A dependency candidate storage unit consisting of a connection relationship list, and a dependency relationship table storing a pair of the connection relationship and the connection relationship of the semantic category number as a dependency relationship; and a phrase unit for dividing an input sentence into phrases. A segmentation unit; a segment extraction unit that extracts a segment that is sequentially divided from the beginning of the input sentence; and an inter-segment dependency candidate consisting of a pair of the extracted segment as a dependency word and the remaining segments as a word. A dependency candidate extraction unit that, when the independent word of the extracted phrase is a compound word composed of two or more words, divides the compound word into words to obtain a prefix and a suffix; The compound word division unit to be excluded, and the inter-word dependency relationship of the compound word, a pair of the dependency word and the receiving word relating to the immediately succeeding word is set as a connection relationship, and the pair of the meaning category of the dependency word and the meaning category of the receiving word is determined. A compound word dependency analyzing unit for connecting the semantic category numbers, and the inter-word dependency relationship is newly registered in the dependency candidate storage unit, and if the inter-phrase dependency candidate is one, the candidate And a dependency relation registering section for newly registering the same in the dependency candidate storage section, and when the number of inter-phrase dependency candidates is two or more, a connection that matches the connection relation of the candidate can be searched from the connection table. When the candidate has the closest distance between the dependent word and the received word, the candidate is determined to be the correct answer, and if the correct answer cannot be determined, a candidate that matches the connection relationship of the candidate is searched from the semantic category connection relationship list. A dependency candidate search unit that determines a candidate having the closest distance between the dependency word and the received word as a correct answer, and if the correct answer cannot be determined by the dependency candidate search unit, the dependency word in the candidate Determine whether there is a candidate whose semantic category number is equal to the semantic category number of the receiving word. If there is, determine the candidate as the correct answer. If not, suspend the determination of inter-phrase dependency. Dependency determination When the dependency analysis of all clauses is completed, the inter-clause dependency candidate with the shortest distance between the dependent word and the received word is regarded as the correct answer for the inter-clause dependency candidate for which the correct answer could not be determined. And a dependency relationship determining unit, wherein the dependency candidate search unit determines that a candidate whose connection relationship with the candidate for which the determination of the dependency is suspended matches the newly registered connection relationship is a correct answer, If the solution can not be determined, characterized in that the connection of candidates pending determination of the dependency is determined to correct the candidate that matches the registered connection relationship to the new dependency analyzer.