JPH01156870A - Retrieving system for translation equivalent - Google Patents

Abstract

PURPOSE:To automatically acquire a translation equivalent or synonym corresponding to an input word by copying the state of a node point of a primary network on a node point of a secondary network for decision of the initial state of the secondary network and extracting the translation equivalent out of the secondary network. CONSTITUTION:A secondary network generating means 330 extracts the translation equivalent and its related words to a due word of a primary network 371 by reference to a word dictionary 220 and a meaning relation dictionary 400. Then a secondary network 372 where the relation is defined as an arc with the words set at the side of the extracted translation equivalent as node points is produced by connecting the node points set at the side of said translation equivalent via a related arc. A secondary network state setting means 340 copies the state of the node point corresponding on the network 371 on a node point of the network 372. Thus the initial state of the network 372 is decided. Then the states of the peripheral node points are set or corrected with the node point whose state is set defined as a starting point. In such a way, a translation equivalent or a synonym corresponding to the input word is automatically obtained.

Description

[Detailed Description of the Invention] [Summary] The present invention relates to a bilingual word search system that can output translations and synonyms corresponding to input words, and an object thereof is to automatically acquire translations or synonyms corresponding to input words. , find words related to the input word, generate a first-order network with the found words on the source language side as nodes and relationships as arcs, set states at the nodes of the first-order network, and Find the corresponding translated words and related translated words, generate a secondary network with the found translated words as nodes and the relationships as arcs, and connect the nodes of the primary network corresponding to the nodes on the secondary network. The initial state of the secondary network is determined by copying the state to the nodes of the secondary network, and the state of the nodes of the secondary network is determined based on the state of the nodes of the secondary network and the value of the arc, and the state is determined. The component is to extract translated words from the secondary network. In addition, a constituent requirement is to extract synonyms from the primary network whose state has been determined.

[Industrial application field]

The present invention relates to an improved bilingual word search system that can automatically obtain translations and synonyms corresponding to input words.

[Conventional technology]

Machine translation has various functions, but one of the main ones is a bilingual word search function. Bilingual word search requires a function that not only shows the translation of a given word in the target language, as in a Japanese-English dictionary, but also selects an appropriate translation according to the context of the word. The context referred to here includes grammatical elements related to sentence structure and semantic elements similar to the compatibility between words.There are various ways to handle this, as shown in Figure 11. Broadly speaking, there are two methods: word direct method and method using intermediate representation. The word direct method is a method in which pairs of translated words and target words are individually and directly defined, and appropriate words are selected while generating translated sentences. The method using intermediate representations is a method in which the original text is converted into an intermediate representation representing a concept through semantic and grammatical analysis, and a translated word that matches the intermediate representation is selected.

Methods for converting into intermediate representations are classified into two representative methods: transfer methods and pivot methods, depending on how the intermediate representation is taken. The transfer method involves setting a unique intermediate representation for each language, converting (transferring) the intermediate representation of the target language to the intermediate representation of the translation language, and then
It extracts the translated words. The pivot method sets an intermediate expression that is common to all languages, and uses this to select translated words. In both methods, intermediate representations that systematize sentence structures using linguistic knowledge such as generative grammar theory and knowledge engineering methods such as frame theory are often used.

FIG. 12 is a diagram showing an example of the operation of each method in "Airplane Flying". As shown in the figure, the translation of “plane” is defined for “airplane”, and the translation of “f” is defined for “flying”.
In the word direct method, translations of "ly" and "jump" are defined, and "plane" and "ny" are selected from the linguistic information and sentence structure defined for each word.

Prepare an intermediate representation A of the object concept corresponding to "airplane" or "plane" and intermediate representation B of the concept "flying in the sky", and create an intermediate representation of the original sentence as "A does B". A method using an intermediate representation is to convert "plane" and "fly" into "plane" and "fly" from this intermediate representation.

[Problem that the invention seeks to solve]

A common issue with the above methods is how to set the language information used for translation.

In the word direct method, the information necessary for translation must be given directly to each word or as rules for generating translated sentences, so the information contained in words and the rules for generating sentences tend to be complex. be.

In methods that use intermediate representations, the information used for translation is divided into information that is unique to each language and information that is common to all languages, such as information that words in each language have individually, information that intermediate expressions have, and information that is provided as conversion rules. While it is possible to provide information in a distributed manner to the extent that it becomes a reality, there is the difficulty of having to set up intermediate representations for all possible concepts.

The difficulty and complexity of setting linguistic information becomes apparent when processing unknown words in the original text.

In the word direct method, it is necessary to set new translations for unknown words and provide complex linguistic information and rules. In a method that uses intermediate representations, it is necessary to either fit it into an existing concept (intermediate representation) or set an intermediate representation that matches the unknown word (concept). In both cases, it is quite difficult to perform this work automatically without manual intervention. Particularly in the latter case, the more complex the intermediate representation is based on deep language analysis, the more difficult it becomes to set the intermediate representation.

Acquiring linguistic information (semantics, grammar, etc.) related to the quality of translation is an important issue for general-purpose translation machines that handle almost infinite information, but the majority of this work currently has to be done manually. It is. In order to improve the quality of translation and make the system more efficient, promoting automation of linguistic information acquisition and improving acquisition efficiency are current major challenges.

The present invention was created in view of this point, and aims to provide a bilingual word search system that can automatically and efficiently acquire translations or synonyms corresponding to input words. .

[Means for solving problems]

FIG. 1 is a diagram showing the principle of the present invention. The bilingual word search system of the present invention includes an input unit 100 for inputting an input language, a word dictionary 210 on the source language side, a word dictionary 220 on the translation side, a semantic relation dictionary 400, and a meaning association network of the input language. a primary network generating means 310 for generating a primary network 371; a primary network state setting means 320 for setting the state of the primary network 371;
Secondary network generation means 33 for generating a secondary network 372 which is a meaning association network of the translated language
0, a secondary network status setting means 340 that sets the status of the secondary network 372, a translated word extracting unit 350 that extracts a translated word from the secondary network 372 whose status has been set, and a primary network 37 whose status has been set.
1, and an output unit 500 that outputs the translated words extracted by the translated word extracting means 350 or the similar synonyms extracted by the similar word extracting means 360.

The word dictionary 210 on the source language side contains word headings, information for searching the semantic relationship dictionary (400), and the translation dictionary (210).
20) and grammatical information. The word dictionary 220 on the translation side includes word headings, information for searching the semantic relationship dictionary (400), and grammatical information. The semantic relationship dictionary 400 has information indicating the semantic relationship between words.

The primary network generation means 310 extracts words related to the input language by referring to the word dictionary 210 and the semantic relationship dictionary 400, and creates a primary network 371 in which the extracted words in the original language are nodes and relationships are arcs. is generated by connecting the nodes with arcs according to the relationships.

The primary network state setting means 320 sets an initial state to the node of the input word on the primary network 371,
The state of adjacent nodes is set according to the arc connected to the node of the input word and the state of that node, and then the state of surrounding nodes is set by performing the same process starting from the node where the state has been set. Or fix it.

The secondary network generation means 330 extracts the translated words corresponding to the words on the primary network 371 and words related to the translated words by referring to the word dictionary 220 and the semantic relationship dictionary 400, and extracts the extracted translated words. A quadratic network 372 with nodes and relationships as arcs is generated by connecting nodes on the translation side with arcs corresponding to the relationships.

The secondary network state setting means 340 determines the initial state of the secondary network by copying the state of the corresponding node on the primary network to the node of the secondary network,
Next, starting from the node whose state has been set, the states of surrounding nodes are set or modified.

〔Example〕

The bilingual word search system of the present invention has a function of extracting candidates for words, word pairs, and word strings in the translation language based on the input of words, word pairs, and word strings in the language to be translated, and synonyms from the words consisting of the words and given information. It has a function to output candidates for
These are controlled by inputting control statements. The former includes a function for guessing candidate words even if there is insufficient information for searching for translations.

In the process of realizing both functions, by using a semantic association network as an intermediate representation and simplifying its structure, it is possible to automatically acquire semantic information of linguistic information to a certain extent, and at the same time, it is possible to automatically acquire semantic information among linguistic information. By adding this function, it is now possible to divide language information settings into a translated language and a translated language, promoting more efficient language information settings.

The idea underlying the present invention is that the meaning of human language is defined by the concept of an object (hereinafter referred to as the object side).
It is said to be a chain of concepts (hereinafter referred to as modal words) that express the state, action, and properties of an object. In other words, the meaning of a word that represents a thing (object side) is an adjective that modifies that word, a verb that can be related to the word (the word can be the subject, it can be the object, etc.), etc. The idea is that it is a collection. For example, the word ``apple'' can co-occur with words such as ``red'', ``sweet'', ``eatable'', ``cut'', ``smelt'', ``skin'', etc. (in the same sentence). In other words, they are considered to have a ``meaning'' that means they can exist as a group.

This idea is based on, for example, "one color for each concrete object."

Attributes (slots) such as "taste", etc. are set in advance, and for "apple", the "color" is "red", the "taste" is "sweet", etc., and the values are set in advance. Data handled (conceptual structure) compared to the frame theory approach of matching data.
It is possible to generate a simple intermediate representation in that it has no internal structure.

FIG. 2 is a diagram showing an example of the system configuration. In the figure, 100 is an input section, 210 is a word dictionary on the original language side, and 22
0 is a word dictionary on the translation side, 300 is a processor, 371 is a primary network, 372 is a secondary network, 400
5 indicates a semantic relational dictionary, and 500 indicates an output unit.

The bilingual word search system of the present invention has a word dictionary 21 on the source language side.
0, the word dictionary 220 on the translation side, and the word (synonymous with the concept)
a semantic relationship dictionary 400 that describes the relationships between the two; a processor 300 that generates a semantic association network for each language based on the dictionary and extracts (detects) translations and synonyms; and an input unit 100.
and an output section 500. The word dictionary 210 on the source language side includes words (headings), information for searching the corresponding meaning-related dictionary 400 and translation dictionary 220, and grammatical information. The word dictionary 220 on the translation side includes words (headings), information for searching the semantic relationship dictionary (400) corresponding to the words, and grammatical information.

The semantic relationship dictionary 400 has information such as synonyms, similar meanings, counter-answers, anti-synonyms (synonyms of counter-answers), superior (concept), subordinate (concept), possible co-occurrence, non-co-occurrence, etc. . In addition, similar meanings, anti-synonyms,
Possible co-occurrence and non-co-occurrence are also graded according to the degree of co-occurrence. Synonyms have the strongest degree of similarity;
The counter-answer is considered to be the one with the strongest degree of anti-synonymy.

The semantic relationship dictionary 400 is a dictionary that describes the relationship between concepts regarding the concepts of natural language words. The content of the description is as follows: ■ A superior-subordinate relationship between concepts such as humans being a subordinate concept of animals (included in humans and animals).

■ Synonyms, such as "running" and "running" are synonyms.
Similarity relationship.

■ “Hot” and “cold” have an anti-answer/anti-synonym relationship.

■ The degree of co-occurrence, such that "cold" and "winter" are more likely to co-occur in a sentence than "cool" and "winter".

and so on.

The description format is one that expresses multiple concepts, the relationships established between them, and the certainty (weight) of those relationships based on a set of "two concepts, a relationship, and a weight." It is possible to search for both concepts and relationships.In addition, relationships that do not depend on the type of language such as Japanese and English, and relationships that differ depending on each language are stored separately, and the degree of co-occurrence is calculated. It is also tuned according to the target field and is stored in sections so that it can be easily modified as needed.

Searches are based on concept markers written in words in each word dictionary, and are usually used in combination with word dictionaries.

Counter-answers and counter-synonyms are used in almost the same way as synonyms and similar meanings in the manuscript when calculating the likelihood of translated words (
By treating them as synonyms/synonyms with negative weights, it is possible to calculate the certainty of translations from co-occurrence information). However, unlike synonyms and similar meanings, words that are considered counter-answers and anti-synonyms are not candidates for translation.

Superordinate and subordinate are used in the same way as synonyms when (anti)synonymous/similar synonymous relations or co-occurrence relations do not exist or are lacking. For example, when a translated word does not exist, a translated word of a word that is higher or lower than the original word may be searched.

FIG. 3 is a diagram showing an outline of the process. These processes are performed by the processor 300. The outline of the process is as follows. However, generation of the meaning association network (primary network) of the translated language, setting of the state of the primary network, generation of the meaning association network (secondary network) of the translated language, and setting of the state of the secondary network are almost done all at once. It will be done as follows. In addition, generation of the primary network → 1
Next Network status settings → Secondary network generation →
Translation search processing proceeds as follows: secondary network state setting -> translation extraction, and synonym search processing proceeds as primary network generation -> primary network state setting -> synonym extraction.

The generation of a semantic association network on the translated language side will be explained. The meaning association network on the side of the translated language is called a primary network. Words (pairs or sequences) in the target language
When input, words related to it (all the same notations as the input string, synonyms, synonyms, counter answers, anti-synonyms, possible co-occurrence, non-co-occurrence, superiors, inferiors, etc.) are displayed. is extracted using the word dictionary 210 and the semantic relationship dictionary 400, and by connecting the words (nodes) with arcs according to the relationships, 1
Generate the following network.

The state setting of the primary network will be explained.

By setting the initial state of the word (node) corresponding to the input character string on the primary network, and sending a signal according to the arc connected to that node and the state of the node to the adjacent node through the arc, the adjacent node Set the state of . After that, by using the node whose state has been set as the starting point and performing similar operations to set and modify the state of the surrounding nodes,
− Set the state to all nodes of the next network. The "state" here refers to a set of values that indicate how the network (graph) is connected, and is given to each node according to the following policy: ■ The word (node) corresponding to the input string has the highest state. .

■ Words similar to the word corresponding to the input string have a high status.

■ Words that have a high degree of co-occurrence with words corresponding to the input character string have a high degree of co-occurrence.

■ Words that co-occur in a similar manner to the co-occurrence of words corresponding to the input character string have a higher status.

The generation of a semantic association network on the translation language side will be explained. Translations corresponding to words on the primary association network created by the processing of primary network generation and primary network state setting processing and words related to the translations are extracted using the word dictionary 220 and the semantic relationship dictionary 400. Then, a meaning association network on the translation language side is generated in the same way as the first-order network is generated. The meaning association network on the translation language side is called a secondary network.

Setting the state of the secondary network will be explained. The state of a word (node) on the secondary network that corresponds to a word on the secondary network (defined as a translated word) is copied from the primary network to determine the initial state of the secondary network. Next, starting from the node where the state has been set, the state on the secondary network is determined by performing the same work as setting the state of the primary network.

Explaining the extraction of translated words. Words are extracted from the secondary network based on the state on the secondary network. That is, words, word pairs, and word strings with high status are selected.

Extracting synonyms will be explained. Words on the primary network are extracted based on the state on the primary network.

In order to make it easier to understand the present invention, a simplified method of processing will be shown. Here, the relationships are limited to four types: synonymous, similar, strong co-occurrence, and co-occurrence. Each state has a value for setting the state of the semantic network, and here the values are set as shown in FIG. Arc is 1, arc type and arc value,
or have a value of arc. Nodes have states such as whether they are on the object side or between modalities, degree of similarity, degree of co-occurrence, and estimated likelihood. The degree of similarity is a value that indicates the degree of resemblance to an input word (a word with the same notation as the input string), and the degree of co-occurrence is the degree to which it co-occurs with the input word (a word with the same notation as the input string). The estimated likelihood is a value that indicates how close the connection of nodes is to the input word. Regarding counter-answers, anti-synonyms, and anti-co-occurrences, the explanation here can be used by considering negative synonyms, synonyms, and co-occurrences.

Note that the processing described below is mainly described for the case where the object side and the aspect are given in pairs, and corresponding translated word pairs are output. Also, although not shown in Figure 4,
Words with the same notation as the input character string are treated in the same way as synonyms, but a - symbol is used to distinguish them.

The network consists of nodes corresponding to words and arcs corresponding to relationships.
It has states of similarity, co-occurrence, and estimated luminosity, and each arc also has values as shown in Figure 4.

The network generation method and state settings will be explained. First, extraction of the same notation and synonyms (including synonyms) will be explained. Use a dictionary to calculate the degree of synonymy between synonyms and connect them with synonymous arcs.

In this example, the same notation word is set to 1, and each time a similar meaning arc is exceeded, the value of the similar meaning arc is multiplied.

Note that when multiple degrees of similarity are given, the maximum value is taken. This work ends when the degree of synonymy of the newly extracted word becomes smaller than a certain value.

Furthermore, only when one route connected by a synonymous arc is greater than the existing degree of synonymy, work continues in the form of modifying the degree of synonymy. FIG. 5 is a diagram showing an example of similar synonym extraction. In the figure, the numbers attached to words indicate the degree of synonymy. Word A is connected to the input word by a synonymous arc with value 1,
Word B is connected to word A by a synonymous arc with a value of 0.5, word C is connected to the input word by a synonymous arc with a value of 0.5, and words B and C are connected by a synonymous arc with a value of 0.5. ing. In the illustrated example, if the degree of similarity is set along route a, the degree of similarity for A will be 1.0, the degree of similarity for B will be 0.5, and the degree of similarity for C will be 0.25. Also, if you set the degree of similarity along route b, the degree of similarity to is 0.125, and the degree of similarity to B is 0.
゜25, the degree of similarity of C becomes 0.5. In this case, if the maximum value is taken, the degree of similarity will be as shown in the figure. If you extract words along route b with the synonymy set up to B on route a, the synonymy given to B is 0.25, which is smaller than the synonymy 0.5 that B has. Therefore, we finish processing this route.

The extraction of co-occurring words will be explained. Although this is not done in the case of a translation search, in the case of a synonym search, words that have a degree of co-occurrence greater than the value for the input word and that meet the conditions of the control sentence are extracted, and words specified by the control sentence are extracted. Set similarity.

A control statement specifies options related to system processing, and is a string of commands given to the system.

A set of degrees of co-occurrence will be explained. The value of the co-occurrence arc is set as the degree of co-occurrence at the node (word) beyond the co-occurrence arc originating from the input word or the same written word. When there are multiple values, the sum is calculated. Setting the value to O, etc.
Depends on the control statement specification. FIG. 6 is a diagram showing an example of a co-occurrence degree set. Since the input word S is connected to the input word a by a co-occurrence arc of 0.5, the degree of co-occurrence of the input word S is 0.5. Also, since word B is connected to input word a with a co-occurrence arc of 0.5, the degree of co-occurrence of word B is 0.
．． It becomes 5.

A set of estimated luminosity will be explained. The estimated likelihood (a value indicating the similarity with the co-occurrence situation of the input word) of the words connected by the co-occurrence arc from the words for which the degree of synonymy and the degree of co-occurrence are set (denoted as A) is Set according to the value of degree of similarity, degree of co-occurrence, and co-occurrence arc. Here, it is given by the following formula.

(1,0 - Co-occurrence degree of IA - Co-occurrence arc value 1) * Degree of similarity of A * Degree of co-occurrence of A (1) Ghost-written Figure 7 is a diagram showing an example of a set of estimated likelihoods.

A is on the object side, has a degree of similarity of 1, and a degree of co-occurrence of 1.

Since B' is connected to A with a co-occurrence arc of 0.5,
Equation (1) gives an estimated likelihood of 0.5 to B'. Similarly, for A, the estimated likelihood is 0.5,
A' is given an estimated likelihood of 1, and A' is given an estimated likelihood of 0.25.

The bilingual word search system of the present invention has a translation word search function and a synonym search function, and one or two examples of each will be given below.

FIG. 8 shows a translation search for words without co-occurrence information.

Main dictionary information on the words ``to'' and ``aku'': Information on the original language ○ ``Aku'' is the same written word, and it means opening something that was closed, and when saying ``the store is open.'' There are three things: ``A seat is available'' and ``A seat is available''.

○ “Door” and “gate” strongly co-occur with “akuj”, which means to open.

○ “Aku” can be used as “open” or “go” depending on its meaning.
It corresponds to pen2"dovacant".

○ "Door" and "gate" correspond to "door" and "Bate'".

Information on the translation side ○ “door” + “gate” is “ope” meaning to open
Strongly co-occurs with “n”.

○"gate" weakly co-occurs with "open" which means to begin.

1st candidate "door", 2nd "open"
Candidates “gate”, “open”” In this example, the translation for “door”, which has no co-occurrence information or translation information, is searched using the information of the synonymous/synonymous words “door” and “gate”. .

FIG. 9 is a diagram showing an example of a translation search for a word without translation information.

Main dictionary information for the word ``Cessna'' Source information O ``Cessna'' and ``Airplane'' are synonyms ○ ``Cessna'' and ``Airplane'' are strongly co-occurring with ``fly'' ○ ``Cessna'' is ``small'' Information on the translation word that weakly co-occurs with ○ "plane", "fly", "small" are "plane",
Compatible with 'fly' and 'small'O'Ce5sn
a" and "plane" have similar meanings O"Ce5sna",
“plane” has strong co-occurrence with “rty”○”Ce5sna
” is weakly co-occurring with “small” The first candidate “plan”
e"" Second candidate "Ce5sna" This example searches for a translation word corresponding to "Cessna" for which no translation word has been given. Ce5sna'' is also output as a candidate. This feature helps you find matching pairs among independently registered words in the dictionary.

FIG. 10 is a diagram showing an example of synonym search.

Entering crab "Taijiquan" (! i Word to search for synonyms) and "Karate", "Aikido", "Kendo" (range to search for synonyms) Main dictionary information O "Taijiquan" means "to hit", ""Kicking" and strong co-occurrence O "Karate"
is strongly co-occurring with "hitting" and "kicking" ○ "Aikido" is "hitting"
, "Throw" and strong co-occurrence O "Kendo" strongly co-occur with "Strike" First candidate "Karate" Second candidate "Aikido" Third candidate "Kendo" This example is within the search range of synonyms. There is a system that infers similarities between words based on how they co-occur and ranks synonyms. In addition, there is also a function to search for similar synonyms from co-occurrence information without limiting the scope of the synonym search, although this will be omitted since the network will be large.

〔Effect of the invention〕

The bilingual word search system of the present invention has a search function for translated words (pairs of two columns) as a main function, and a search function for synonyms as a secondary function. The translation search function is realized using translation information, synonym information, and co-occurrence information described in the dictionary, but even if any of these information is missing or missing, it cannot be performed using other information. It also has an analogy function that enables it to operate.

For example, for an input word for which co-occurrence information is not described, it is possible to select a polysemous word using description information of another word from synonym information, etc., and the same can be said for translation word information. Furthermore, even for words for which only co-occurrence information is defined, translations can be searched by using the synonym search function, which is a sub-function.

This analogy function and the main part of network processing are divided into the source language side (primary network) and the target language side (secondary network).
This separation allows language information settings to be performed independently on the source language side and the target language side. As an extreme example, although it is undesirable from the viewpoint of search efficiency, the words that provide information for translation searches are narrowed down to a small number of words, and other words are given synonym information, co-occurrence information, and other words that provide translation information. By providing upper and lower information, it becomes possible to set independent language information for almost each language.

Furthermore, since the setting of synonym information and co-occurrence information can be supported, the setting of linguistic information can be streamlined.

Furthermore, since the analogy function uses related information described in the dictionary all together, the reliability of the analogy function is high. The reason for this is that even if there is slightly strange information, it is difficult to show up in the results due to other information.

Apart from the above, most of the processing of the semantic network is performed by repeating local processing that refers only to the values (states) of the champions of adjacent nodes and the arcs between them.
It is becoming easier to parallelize and distribute processing. multiple C
When a PU is used and each CPU is assigned a part of the network to perform processing, each CPU is able to perform most of the processing independently. This is a very effective feature in improving search speed.

To summarize the above, (a) Since there is a function of analogy of linguistic information, setting of linguistic information has become easier.

0) Since most of the language information used is statistical information, it is easy to build a system that (automatically) supports the setting of language information.

(C) By using a network with a simple structure and making the processing a collection of local operations, it is easy to distribute and parallelize the processing.

That's what I would say. However, since the system of the present invention is a bilingual word search system, it should be noted that the linguistic information referred to here is part of the information used by the translation system.

[Brief explanation of the drawing]

Figure 1 is a diagram of the principle of the present invention, Figure 2 is an example of the system configuration,
Figure 3 is a diagram showing an overview of the process, Figure 4 is a diagram showing arc values and symbols, Figure 5 is a diagram showing an example of synonym extraction, and Figure 6 is a diagram showing an example of synonym extraction.
The figure shows an example of a co-occurrence degree set, Figure 7 shows an example of a set of estimated luminosity, Figure 8 shows an example of a translated word search, and Figure 9 shows the other side of a translated word search. , FIG. 10 is a diagram showing an example of synonym search, FIG. 11 is a diagram showing a conventional translation search method, and FIG. 12 is a diagram showing an example of the operation of each conventional method for "Airplane flies." 100... Input unit, 210... Word dictionary on the original language side,
220... Word dictionary on the translation side, 300... Processor, 371... Primary network, 372... Secondary network, 400... Semantic relationship dictionary, 500...
・Output section. Patent Applicant Fujitsu Ltd. Representative Patent Attorney Kyotani 4 part 9#/) Stirring Hayabusa 3m Ryo-fu's i-nao and symbol Takuki Keike...2t's 4grihatot(2) (Object 71. θ, 0) Estimated r degree name 0 ita 11 DR-I (2) zu again a (thing, 1.chi15) (like, 1.θ, 5.θ
5) (0.25, 0.D of 1 go) (7
Jshi, 1,0.5,0.5n Figure q PltLfLL
C Gunhada (Shakumachisen, (q hand 7 joint road, yu 1 road)) r7z'8f-
ri==:)-r'f tsu (□・1・0・1) (sama, 1.θ, 5.7) karate pickles kendo yoryo proverb 倹子/)4ri′1 Mizui/ρ diagram ? Journalist 1

Claims (1)

[Claims] An input unit (100) for inputting an input language, a word dictionary (210) on the source language side, a word dictionary (220) on the translation side, a semantic relationship dictionary (400), and a meaning of the input language. A primary network generating means (310) for generating a primary network (371) which is an associative network; a primary network state setting means (320) for setting the state of the primary network (371); A secondary network generating means (330) for generating a secondary network (372) which is a semantic association network; a secondary network state setting means (340) for setting the state of the secondary network (372); and a state setting means. A translation word extraction means (350) for extracting translation words from the secondary network (372) for which the status setting has been completed; 350
) is the translated word or synonym extraction means (360
) outputs the synonyms extracted by
), and the word dictionary (210) on the source side contains information for searching word headings and the semantic relationship dictionary (400), and information for searching the word dictionary (220) on the translation side. The word dictionary (220) on the translation side has word headings, information for searching the semantic relationship dictionary (400), and grammatical information. ) has information indicating the semantic relationship between words, and the primary network generation means (310) stores words related to the input language in a word dictionary (210) and a semantic relationship dictionary (40).
0), and the extracted words in the original language are nodes and the relationships are arcs (371)
The primary network state setting means (320) sets an initial state to the nodes of the input word on the primary network (371). , set the state of adjacent nodes according to the arc connected to the node of the input word and the state of the node, and then perform the same process starting from the node where the state has been set, to determine the state of the surrounding nodes. The secondary network generation means (330) stores translations corresponding to words on the primary network (371) and words related to the translations in a word dictionary (220) and a semantic relationship dictionary (220).
A secondary network (372
) by connecting the nodes on the translation side with arcs according to the relationship, and the secondary network state setting means (340) connects the nodes of the secondary network with the corresponding nodes on the primary network. Determine the initial state of the secondary network by copying the state of the node, then start from the node whose state is set,
A bilingual word search system configured to set or modify states of surrounding nodes.