JPS63184871A - Mechanical translating device - Google Patents

Abstract

PURPOSE:To perform conversion in one search process independently of the size of a network structure by retrieving a dictionary corresponding to a searched network and starting a network converting rule on the basis of information of this dictionary. CONSTITUTION:Networks consisting of nodes, which have words as constituting elements of sentences as names, and arcs, which have relations among respective nodes as names and indicating directions are stored in a network storage means 12, and rules in accordance which some network is converted to another are stored in a network converting means 14. A network is searched by a control means 11, and a dictionary corresponding to the searched network is retrieved from a dictionary storage means 13 by the header word of the dictionary. The network converting rule in the network converting rule storage means 14 is started on the basis of information in the retrieved dictionary.

Description

[Detailed description of the invention]

[Industrial Application Field] The present invention relates to a machine translation device, and in particular aims to improve a network conversion method for converting a network as a semantic structure obtained as a result of language analysis to obtain a new network. [Prior Art] Conventionally, in m-machine translation, as shown in FIGS. 9(a), (b), and (C), parse trees, dependency structures, and conceptual structures, respectively, obtained as a result of language analysis are used. Networks called semantic networks have been considered. These networks are composed of nodes whose names are words that are constituent elements of a language, and arcs whose names and directions indicate relationships between nodes. In the network conversion method for converting these networks into a new network, a conversion rule for network conversion is written along with the information in a dictionary that describes information about the node using the node name as a headword. There was a method in which when a certain node was searched for in a dictionary, the corresponding conversion rule was executed. Also, a method has been considered in which network conversion rules are stored separately in a conversion rule storage means and conversion is executed. FIG. 1O is an overall block diagram for implementing this method. Here, 21 is an overall control means for controlling the process flow of network conversion, 22 is a network storage means for storing various networks obtained as a result of language analysis, and 23 is a network storage means for storing various networks obtained as a result of language analysis. Dictionary storage means 24 stores information, and network conversion rule storage means 24 stores network conversion rules. Under the above configuration, according to the overall control means 21, first, a network to be converted stored in the network storage means 22 is searched. Next, the dictionary stored in the dictionary storage means 23 is searched using a certain node obtained by searching the network as a heading, and the structure of the network is obtained based on information regarding the nodes in the dictionary. The structure of the obtained network is inspected by sequentially using the network conversion rules stored in the network conversion rule storage means 24,
Determine which conversion rules may be applicable. If there is an applicable conversion rule here, it will be designated as the conversion rule. These methods are, so to speak, top-down methods in which searches are performed one by one for each network conversion rule based on the direct management of the overall control means 21.

[Problems to be solved by the invention]

By the way, the former method of describing network conversion rules in a dictionary is used to convert small structures that depend on individual node names, such as co-occurrence processing of two nodes directly connected by an arc as shown in Figure 9. This is a method suitable for The rules for executing are described in the dictionary. However, in the method described above in which network conversion rules are written in a dictionary, the conversion rules depend on the individual node names, so similar network conversion rules that could normally be combined into a single rule may be different rules. There was a problem in that the number of network conversion rules had to be written one by one, resulting in an increase in the number of network conversion rules. Furthermore, the latter method of storing network conversion rules in the network conversion rule storage means 24 separate from the dictionary is as follows.
By being able to inspect the global network structure, it is possible to convert large structures, but it is difficult to convert large structures if only the network search process for processing individual node names such as the co-occurrence processing described above is sufficient. Therefore, there is a problem in that a separate search process is required to check whether the network has a structure to which the network conversion rule can be applied. Additionally, this method has the problem that the network must be searched one by one for each network conversion rule in one search process. An example of setting up a separate search process is Chapter 7 (9,
1287). Here, for large structural transformations that are difficult to handle during the process of searching and converting networks to perform processes such as co-occurrence processing, separate network search processes are set up before and after the above process, and individual network conversion rules are It is stated that the network is searched independently for each. By the way, in the latter method, similar to the method of writing network conversion rules in a dictionary, for large structural transformations that depend to some extent on the node names on each side, the rules are written as separate network conversion rules in the dictionary. There were times when I had to. SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems and to provide a network conversion method that enables conversion regardless of the size of the network structure in the - search process.

[Means to solve the problem]

Therefore, in the present invention, as shown in FIG. 1, nodes that are obtained as a result of analyzing a sentence and have words that are constituent elements of the sentence as names, and relationships between each of the nodes as names and pointing directions. network storage means 12 for storing a network consisting of arcs; network conversion rule storage means 14 for storing rules for converting any network in the network storage means 12 into another network; The area is composed of an area in which information about a node is described as a headword, and the area contains conditions that should be met by a network structure including a node having a node name that is a headword.
Dictionary storage means 13 for storing a dictionary having information for specifying network conversion rules to be activated in response to conditions; searching for networks; searching by headwords of dictionaries stored in dictionary storage means 13; means for controlling a network, searching for a network, searching a dictionary corresponding to the searched network from the dictionary storage means 13, and converting the network in the network conversion rule storage means based on the information in the searched dictionary. It is characterized by comprising a control means 11 for activating the rules. [Action 1] According to the above configuration, there is a network search process for inspecting local structure and small structural transformation based on it, and inspection of global structure such as transformation of structurally distributed representations and large scale transformation based on it. A large structural transformation is possible through just one network searching process without making any distinction between the network searching process for structural transformation and the network searching process for structural transformation. [Example] Hereinafter, the present invention will be described in detail with reference to the drawings. FIG. 2 is a configuration diagram of an apparatus showing an embodiment of the present invention. Here, 31 is a ROM that stores the processing procedure shown in FIG. 4 and other network subsidiary search and inspection processing procedures.
, 32 is a CPU that controls the entire apparatus such as processing stored in the ROM 31. Reference numerals 33 and 34 denote a conversion rule storage unit that stores network conversion rules and a dictionary storage unit that stores a dictionary that uses node names as headwords, respectively, and these are configured by storage means such as a ROM, for example. 35 is a network storage unit that stores the network obtained as a result of language analysis, and is composed of a converted network area 35^ that stores the network to be converted, and a converted network area 35B that stores the converted network. Ru. These storage units are composed of RAM and the like. 36 is an input device, such as a keyboard, which can input the language to be converted. Reference numeral 37 denotes a display device such as a CRT for displaying sentences inputted by the input device 36 or sentences converted. 38 is a text storage unit that stores sentences before being analyzed by the network and sentences restored from the network form to the normal form; these sentences are inputted through the input device 36 and displayed on the display 37; It can be displayed on Figure 3 (1) shows the semantic network (a) obtained as a result of analyzing the Japanese sentence ``Invention brings about a change in society.''
This is an example of conversion into a semantic network (b) corresponding to zessociety J. In addition, network (a)
and (b) are stored in the network storage section 35. FIG. 3(2) is a network conversion rule for executing the network conversion shown in FIG. 3(1), which is stored in the network conversion rule storage unit 33. In FIG. 3, arcs indicate relationships between nodes by their arc salts, and indicate the direction of the relationships by their pointing directions. Further, the double circle around the node indicates that the node name is a predicate (verb, adjective, adjectival verb), and O indicates the verb "r-transform". This means that, for example, RO
As a storage format in M, etc., two node names or dictionary headwords are used: "transformation: noun" and "transformation: verb." The former indicates the noun "transformation" and the latter indicates the verb "transformation." You can think of it as representing. Network (a) in Figure 3 (1) shows that “invention” is the “cause” and “change,” which is the “subject,” “happens” in “society,” which is the “recipient.” There is. Network (b) in Figure 3 (1) shows that the "subject" of "invention" is the "object" for the act of "transforming" is "society." In this way, the network conversion shown in FIG. 3(1) converts a Japanese expression into a corresponding English expression. Next, the processing procedure of the network search and conversion process in the network conversion shown in FIG. 3 will be explained based on the flowchart shown in FIG. 4. In the process of searching for the network (a) to be converted stored in the network storage unit 35, first
``occurrence'' is searched first, and then its lower nodes ``society'', ``change'', and ``invention'' are searched (step St). Next, referring to each searched node, the dictionary storage unit 34
“Transformation” is the key node of network (a) in
Searching the dictionary, if "change" is indicated by "happen", it is expected that the conversion rule shown in FIG. 3 (2) in the network conversion rule storage section 33 will be applied. , the process jumps to the network conversion rule shown in FIG. 3(2). Note that other nodes "Society" and "
It is not necessary for "invention" to be indicated by "happens," and the relationship in which "happens" indicates "transformation" does not need to be a "subject." In other words, here we are not looking at the overall structure of the network. (Step S2
). Furthermore, we investigated the information surrounding “occurrence” and “transformation” in network (a), that is, the relationship between nodes and arcs, with reference to the information in the dictionary, and obtained the predicted information in Figure 3 (2).
) If it has the same structure as the network structure (C) shown in (C), then the network conversion shown in FIG. 3 (1) is executed for the first time (step S3). Here, to specifically explain the network conversion rule in Figure 3 (2), if the network to be converted has a structure like network structure (C) in Figure 3 (2),
Change the arc salt "cause" of the arc between node 1 and node 4 to "subject", change the arc salt "receiver is hand" of the arc between node 1 and node 2 to "object", and remove node 1. Then, instead of node 1, the node name of node 3 is converted into a phrase (verb). In addition, here, the operation of converting the noun "transformation" into a verb to "transform" is done by changing the node name "transformation: noun" to the node name "transformation: verb" as described above.
This is done by changing to Further, to add more information here, although each node name in network (b) of FIG. 3(1) is not in English, this will be changed to English in the process of searching the dictionary in the dictionary storage unit 34 later. Of course, it is also possible to write a rule so that when executing the network conversion rule (2) in FIG. 3, a dictionary search is performed in the dictionary storage section 34 and a translated word is selected. FIG. 5 shows an embodiment in which a structural transformation larger than the network transformation shown in FIG. 2 is performed. Figure 5 (1) is the Japanese sentence 'Central Processes'.
The semantic network (a) obtained as a result of analyzing "sing IJnit is abbreviated as CPU" is called rCPI in English.
J 5tands forCentral Process
Semantic network corresponding to “ssing Unit” (
This is an example converted to b). FIG. 5(2) is a network conversion rule for executing the network conversion of FIG. 5(1). In Figure 5 (1), "and" means rcentral Processing.
"Unit" is "before" and "after" is r.
It shows that the "cpu" are connected. The network conversion rule in Figure 5 (2) indicates that if the network to be converted has a structure like network structure (C) in Figure 5 (2), it will be converted to network structure (d). ing. However, ^° indicates that it is the same thing as A. To determine whether it is the same thing,
For example, an ID section is provided in the dictionary of A in the dictionary storage section 34 (
Write ^° there), search the ID part of A's dictionary in the dictionary storage section 34 in the network conversion rule,
This can be done by writing a rule that pulls the content and compares it with Ao. In order to jump to the network conversion rules in the network conversion rule storage 33, we will supplement the conditions that the network structure described in the dictionary in the dictionary storage 34 must satisfy. This condition is that in Figure 3, ``change'' is pointed to by an arc from ``happen'', and in Figure 5, ``omitted'' is pointed to by an arc from ``and'' to ``before''. It was said that In other words, the structure between two nodes is described as a condition, but needless to say, the description of the condition is not limited to this. An example thereof will be explained using FIG. Network (a) in Figure 6 (1) is a semantic network obtained by analyzing the Japanese sentence ``I don't think you're right.'' When translating this into English, to make it more familiar, change the position of the negation and use rI do
Don’t think you are right.”
Must be. The semantic network corresponding to this English sentence is network (b) in FIG. 6(1), and the network conversion rule for executing the network conversion in FIG. 6(1) is shown in FIG. 6(2). The conditions that the network structure must satisfy in order for the process to jump to the network conversion rule shown in Figure 6 (2) are described in the dictionary of ``koto'', and the content is that ``koto'' is changed from ``think'' to arc salt. It is pointed by the arc of "object", and the arc salt is pointed by the arc of "apposition" from the node where "thing" is the predicate. As is clear from the above explanation, the present invention enables the inspection of local structures such as co-occurrence processing and the network search process for small structural transformations based on the inspection, and the global search process such as the transformation of structurally distributed representations. Inspection of the structural structure and network search process for major structural transformation based on the inspection
This has the effect of making it possible to perform large structural transformations through a single network search process without distinction. In addition, in Figure 3, the network conversion rule is jumped to only when ``change'' is specified by ``occurrence'', but in this way, the network conversion rule is jumped to only when it is predicted to be necessary. This is more efficient than checking top-down one by one whether the conversion rules are applicable. Furthermore, in the conventional method of writing network conversion rules in a dictionary or in a top-down method, it is difficult to understand that "inventions bring about changes in society" as shown in Figure 3, and as shown in Figure 7 (1). “Change will be brought about” and Figure 8 (1)
``Improvements in society occur through inventions'' and Figure 8 (2)
It is necessary to prepare network conversion rules corresponding to each of the four sentences (even if they are similar rules) for the four sentences ``Inventions cause destruction in society.''
This is because traditional network transformation relies on verbs such as ``happen'' and ``bring about'' and nouns such as ``transform,''``improve,'' and ``disrupt.'' Also, as shown in Figure 7 (2), ``Inventions begin to transform society.''
In the case of the sentence ``The invention causes controversy in society'' as in Figure 8 (3), network transformation as in Figure 3 cannot be performed. Therefore, in any case, it is necessary to prepare rules corresponding to individual verbs and nouns as network conversion rules. On the other hand, according to the present invention, which network conversion rule to select is determined only by determining which conversion rule to activate based on the information in the dictionary. It is sufficient to prepare one network conversion rule for the four sentences (1), FIG. 8(l), and FIG. 8(2). Therefore, there is an effect that the number of network conversion rules can be reduced. [Effects of the Invention] As is clear from the above explanation, the network search process for inspecting local structures and small structural transformations based thereon, and the inspection of global structures such as transforming structurally distributed representations. A large structural transformation is possible through a single network searching process without distinguishing between the network searching process and the network searching process for a large structural transformation based on this. Therefore, it is more efficient than inspecting the structure using network transformation rules individually. In addition, an effect was obtained in that the number of network conversion rules could be reduced.

[Brief explanation of the drawing]

FIG. 1 is an overall configuration diagram of the present invention; FIG. 2 is a configuration diagram of an apparatus showing an embodiment of the present invention; FIG. 3 is an explanatory diagram of an embodiment of the present invention; Fig. 4 is a flowchart of the processing procedure in one embodiment of the present invention, Fig. 5 is an explanatory diagram of another embodiment of the present invention, Fig. 6 is a flowchart of the processing procedure in one embodiment of the present invention, Fig. 6 Furthermore, FIGS. 7 and 8 are explanatory diagrams for explaining the present invention in detail, and FIGS. 9(a), (b), and (C) are obtained as a result of language analysis. FIG. 1O is a configuration diagram of a conventional network conversion method, and FIG. 11 is an explanatory diagram of another conventional network conversion method. Figure 1: The last part of the 71st sentence, Figure 1. Go to Figure 2. N -ノ
N- Torakata Seiri ni Suro Karu J Satote no Su O Flow Chiet Fig. 4 Y... The invention of the ward will bring about changes in the 7 societies.'' An explanatory diagram to explain the effect. Figure 7: With the invention, 7 o's needles are changed (but A is 2. For e = yo, te ne 2 needles are broken and Adan's 2 hearts. ``It Akira, 11.) and Neo * Figure 8

Claims (1)

[Claims] Obtained as a result of analyzing a sentence, it is composed of nodes whose names are words and phrases that are constituent elements of the sentence, and arcs whose names and pointing directions indicate relationships between each of the nodes. network storage means for storing networks; network conversion rule storage means for storing rules for converting any network in the network storage means into another network; It consists of an area where information is described, and this area specifies the conditions that should be satisfied by the network structure including the node having the node name that is the headword, and the network conversion rule that should be activated in response to the condition. dictionary storage means for storing a dictionary having information for searching the network; and means for controlling the search of the network and the search of the dictionary stored in the dictionary storage means using the headword; , control means for searching the dictionary storage means for a dictionary corresponding to the searched network and activating the network conversion rule in the network conversion rule storage means based on the information in the searched dictionary. A machine translation device characterized by: