JPH0950435A - Translation device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To retrieve bilingual examples substantially exceeding the range of sentence examples. SOLUTION: This device is provided with an example base 1 for storing plural examples of bilingual sentences, an index storage means 2 for storing indexes for indicating the correspondence of example vectors for indicating the features of the respective examples in the example base 1 and the respective examples, an input means 3 for inputting a translation request sentence including words, a word dictionary 4 for holding word vectors for indicating the features of the respective words, a vector generation means 5 for analyzing the inputted translation request sentence and generating the context vector of the translation request sentence obtained by using the word dictionary 4, an example candidate retrieval means 6 for retrieving example candidates based on a distance between the context vector of the translation request sentence and the example vector, a display means 7 for displaying the example candidates and a bilingual sentence display means 8 for displaying the bilingual sentence corresponding to the example selected from the example candidates by a user.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a translation device for displaying a bilingual translation example such as a travel conversation sentence when a translation request including a keyword or a sentence is made.

[0002]

2. Description of the Related Art Conventionally, case-based translation apparatuses that perform translation by searching bilingual cases can be roughly classified into a classification display method and a keyword search method.

The classification display system is a system in which bilingual translation examples are classified into various genres according to places, scenes and the like in advance, and an appropriate bilingual translation example is searched based on the classification table. For example, in the case of travel conversation sentences, the major categories are hotels, airports, restaurants, roads ,. . . Etc., subway, bus, taxi, guidance ,. . .
Etc., and various cases are stored under it. For example, the example "What stop do you get off at to get to City Park?"
Stored under.

On the other hand, the keyword search method is a method in which a keyword expressing the contents is added to each bilingual case, or an appropriate bilingual case is searched for by targeting all the words included in an example sentence. For example, in the above example, “city park”, “go”, “what”, “stop”, “get off”, etc. are candidates for keywords. Further, as an extension of the keyword search method, there is an "example-driven machine translation method" of Japanese Patent Laid-Open No. 3-276367. This is a method for searching a case using a word dictionary (thesaurus dictionary) hierarchized in a tree structure based on the similarity of word meanings.

[0005]

However, the case-based translation apparatus based on the above classification display method is effective while the number of cases is small, but when the number of cases becomes enormous, it meets the needs of many people. Classification is difficult, and it is not possible to easily determine to which genre the cases satisfying the user's request are classified. Further, there is a drawback that it takes time to confirm because many cases are classified into the same genre.

In the case search by the keyword search method, only words added to the bilingual sentence as keywords can be searched, and all words supposed to be input by the user are added in advance to the bilingual case as keywords. It is virtually impossible to set. Therefore, it is only possible to search for input sentences within the limited range added as keywords. Similarly, even if a full-text search that searches all the character strings of the case is used, only the words or sentences that appear in the case can be searched. In other words, in both search methods, search beyond the range of the natural language explicitly expressed in the search target case is impossible.

Further, the example-driven machine translation system proposed in Japanese Patent Laid-Open No. 3-276367 realizes more flexible retrieval of parallel translation examples by using a thesaurus in which the meanings of words are classified one-dimensionally. There is. Here, by using linguistic analysis of morphological analysis, an accurate correspondence between the case and the words that make up the translation request is obtained, and similar cases are searched by calculating the distance on the thesaurus between each word.
When the range of translation target is expanded, there are problems that the translation request context is not met, unnecessary translation examples are searched, and large-scale translation examples need to be collected.

SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a translation apparatus which can solve the above-mentioned problems by searching for a bilingual case most suitable for a translation request even in a small number of cases.

[0009]

The translation device of the present invention includes a case base in which a set of a case and its bilingual sentence is stored as a bilingual case, a case vector representing the characteristics of the case, the case vector, and the case vector. Index storage means for storing an index indicating the correspondence with the bilingual case, input means for inputting a translation request including a word, and a word dictionary holding a characteristic word vector indicating the degree of association between the translation request word and the characteristic word A vector generation means for generating a translation request vector based on the addition result of the characteristic word vectors corresponding to the translation request; and a case candidate search means for searching for a case candidate based on the distance between the translation request vector and the case vector. And a bilingual sentence display means for displaying a bilingual sentence corresponding to the retrieved case candidate using an index.

In the translation apparatus of the present invention, for example, assuming a Japanese user on the basis of parallel translation sentences in Japanese and English, when the user inputs a translation request sentence in Japanese, translation is performed using a word dictionary. A request vector is created, the distance to each case vector in the case base is calculated, and a bilingual sentence (English)
Is displayed.

In the present invention, the parallel translation example and the translation request vector are created based on the feature word vector, and the search is realized by vector operation, so that the linguistic analysis of the morphological analysis is not required. In addition, parallel translation examples and translation request vectors are created based on characteristic word vectors that are characterized according to human knowledge and common sense. It is possible to search accordingly.

Further, since the case candidate display means is provided, the case candidates similar to the translation request are displayed in Japanese in order, and when the user selects one case from a plurality of similar case candidates. The degree of freedom is improved and the operability is improved.

Furthermore, it is possible to register a new bilingual translation example consisting of a Japanese original text translation and an English bilingual translation, so that the range of the target bilingual translation can be expanded.

[0014]

FIG. 1 shows a functional block diagram of a case-based translation apparatus according to an embodiment of the present invention. In FIG. 1, it comprises a case base 1, an index storage means 2, an input means 3, a word dictionary 4, a vector generation means 5, a case candidate search means 6, a display means 7, a translated text display means 8 and a case registration means 9. There is.

Next, each component of the translation apparatus will be described. In the case base 1, a plurality of parallel translation cases are stored as a set of a translation request language and a translation text language. The index storage unit 2 stores a case vector indicating the strength of association between each case in the translation request language in the case base 1 and the characteristic word (the characteristic word will be described later), the case vector, and the bilingual case. The index and the correspondence are stored. The case vector may be given in advance,
You may use the sum of the characteristic word vector created by registration of the new case mentioned later.

The user uses the input means 3 including a keyboard.
Use to enter a translation request. The input may be made by recognizing voice or handwritten content online. The word dictionary 4 includes an automaton (transition of characters) for extracting a word from a natural language sentence and a feature vector of each word. As the feature vector, a feature word vector described later is stored. FIG. 2 shows a part of the automaton forming the word dictionary 4. Here, "restaurant",
In this example, only three words "rest room" and "strike" are registered. Here, {Les, s} is all the characters except Les and Sus, the solid line is the goto function, and the broken line is the failure function (the failure function from all other states to the initial state is omitted). The vector value of each word is the output in that state, and indicates the characteristic word vector of the extracted word (here, 3 words). The vector generation means 5 sends the content of the translation request to the automaton of the word dictionary 4, extracts each word of the translation request and its characteristic word vector, and generates a translation request vector from the sum of the characteristic word vectors. The case candidate search means 6 calculates the distance between the translation request vector and each case vector stored in the index storage means 2, and displays the case candidates in the translation request language on the display means 7 which is a display in order from similar cases. To do. Although the display means 7 is not always necessary, even if the display means 7 has a slightly lower priority, the display means 7 can be included in the selection range of the case, so that the usability for the user is improved. The bilingual sentence display means 8 displays the bilingual sentence corresponding to the case candidate selected by the user from the case candidates. The display means 7 and the translated text display means 8 can be realized by the same display, and a printer or the like can be used as the output means.

Further, the case registration means 9 creates a case vector for the new case from the word dictionary 4 and the vector generation means 5 for the new parallel translation case input to the case base 1 using the input means 3. , The index for correlating the case vector and the bilingual case is added to the index storage means 2, and a new bilingual case (case and its bilingual sentence) is registered in the case base 1. When registering a new case, the case vector is the sum of the characteristic word vectors of the words extracted from the case.

FIG. 3 is a block diagram showing a case where the above translation device is realized by electric hardware using a CPU. In FIG. 2, an auxiliary storage device 21, a CPU 22 for performing various processes, a main memory device 23 for storing processing results, and an input / output C for connecting various input / output devices to the CPU.
The translation processing unit 25 includes an h (channel) 24, a display device 26 including a CRT, and a keyboard 27.

The case base 1 and the word dictionary 4 shown in FIG.
The index storage means 2, the vector generation means 5, the case search means 6, and the case registration means 9 stored in the auxiliary storage device 21 of FIG. 1 correspond to the translation processing unit 25 of FIG. 3, and the input means 3 of FIG. Corresponds to the keyboard 27 shown in FIG. 3, and the display means 7 and the translated text display means 8 shown in FIG.
Corresponding to.

Next, the characteristic word vector will be described. The specific configurations of the word dictionary 4, the vector generation means 5, and the case search means 6 in this embodiment are as follows: −1) ”context vector is used.

That is, the context vector indicates the degree of the relationship between the concept of the word in the sentence and the context,
It is a vector expression of the degree of semantic connection with many characteristic words. If n concept classifications are characteristic words,
Each dimension is represented by one point on the n-dimensional vector space corresponding to one feature word. The value of each element of the context vector Xi = (xi1, xi2, ... Xin) of word i is defined as follows.

Xij = 0 if word i is not related to characteristic word j xij = 1 if if i is related to characteristic word j For example, the following six characteristic words (human, sad, art, science, excitement, politics) ) Is selected, the 6-dimensional context vector X of the word “pilot” is expressed as a binary value as follows.

X = (1,0,0,1,1,0) Further, the relationship between the word i and the characteristic word j may be expressed by multiple values as follows according to its strength.

X = (2,0,0,3,1,0) In this embodiment, the context vector is used as a multivalued feature word vector.

The vector generating means 5 generates a translation request vector by normalizing the sum of the characteristic word vectors extracted from the translation request, which is a natural language text, so that the length becomes constant. Since the lengths of the translation request vector and each case vector are normalized to be constant, the case search means 6 can use the inner product as the distance.

The processing flow of the present invention will be described with reference to FIG.

In step S1, the input means 3 causes
The translation request sentence "Where should I get off to go to Central Park?" Is entered. Here, the input is performed in sentences, but only words may be input.

In step S2, the vector generating means 5
Uses the word dictionary 4 to extract four words "central park", "go", "where" and "down" from the input translation request sentence, and obtains a corresponding feature word vector. Of course, if the corresponding characteristic word vector is not stored in the word dictionary 4 in advance, the characteristic word vector is not obtained for a part of the translation request sentence.

The characteristic word vector of each word extracted from the above is the central park (0, 2, 0, 0, 1, 0, 2, 0, 1) going (0, 1, 0, 0, 0, Where (1,1,0,0) where (1,0,0,0,0,1,0,0,1) get down (0,2,0,0,0,2,1,0,0) .

In step S3, the vector CV0 = (1, 5, 0, 0,
1, 4, 4, 0, 2), and then the translation request vector CV1 = which is obtained by normalizing the vector CV0 to a length of 10.
The vector generating means 5 generates (1, 6, 6, 0, 1, 5, 5, 0, 3).

FIG. 5 shows how the case base corresponding to the bilingual case is stored. That is, the bilingual case is stored in the case base 1, and the case vector and the index associating the case vector with the bilingual case are stored in the index storage unit 2. For example, Case 1 “Just push the shutter: Just push the butto”
n. Is stored in the case base 1, the case vector (3,1,1, ...) and the index I1 indicating the storage location in the case base 1 are stored in the index storage means 2. The parallel translation examples of are stored.

In step S4, the case retrieval means 6 calculates the distance between the translation request vector CV1 and all the case vectors stored in the index storage means 2. In the example of FIG.
The distance between the case 1 “just press the shutter” and the translation request sentence is the case vector of the case 1 ((3, 1, 1, ... 0)
Becomes the inner product of the translation request vector CV1 (1, 6, 0, ... 3) and 3 × 1 + 1 × 6 +. . . 3 × 0 = 23 points. Since the length is normalized to 10, the maximum score is 100 points. As shown below, all cases and distance calculations are performed in the same manner.

Case 1: 23 points Case 2: 78 points • Case 12000: 35 points In step S5, several cases are selected in order from a case close to the translation request sentence (large inner product value), and case base 1 as case candidates. It is displayed on the display means 7 together with the score (inner product value).

1. (95 points) What stop do you get off at to go to City Park? 2. (85 points) Is this the road to Central Park? 3. (80 points) Where is Hyde Park? In step S6, the translated text of the case candidate selected by the user from the input means 3 is displayed on the translated text display means 8.

1. (95 points) What stop do you get off at to go to City Park? What bus stop do I get of
f at City Park? Through the above steps, the user can obtain an appropriate bilingual sentence.

Further, when registering a new bilingual case, a case vector is generated for the case in the same manner as the translation request vector generation, and the new case vector is added to the index storage means as shown in FIG. For example, when registering the case 1 of FIG. 5, the characteristic word vectors of “shutter” and “press” are obtained from the word dictionary 4, and a vector obtained by normalizing the sum is used as a case vector, The case vector is stored in the index storage means 2 together with the index I1 indicating the storage location.

As described above, in the embodiment of the present invention, the sentence forming the case is used to create the case vector, but other than that, for example, the case vector is prepared from several words representing the contents of the case given by the user. May be. Further, in the embodiment of the present invention, the Japanese and English bilingual sentences are used, but the present invention is not limited thereto. For example, the case is composed in multiple languages such as Japanese and English, French, Chinese, etc., the translation request sentence is input in the language specified by the user, and the case candidates are displayed using the language in which the translation request sentence is input. However, a configuration in which the bilingual text in the language selected by the user is displayed is also conceivable.

[0038]

EFFECTS OF THE INVENTION The translation device of the present invention is not capable of searching beyond the range of the natural language explicitly expressed in the search target case, whereas the translation device of the present invention is human It is possible to search in natural language that reflects common sense of, and it becomes possible to search for parallel translation examples that greatly exceed the range of natural language, and translation support becomes efficient.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a translation device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a configuration of a word dictionary according to an embodiment of the present invention.

FIG. 3 is an electrical block diagram of a translation device according to an embodiment of the present invention.

FIG. 4 is a flowchart showing a processing flow of the translation device according to the embodiment of the present invention.

FIG. 5 is a diagram showing a storage state of parallel translation cases and corresponding case vectors of the translation apparatus according to the embodiment of the present invention.

[Explanation of symbols]

 1 case base 2 index storage means 3 input means 4 word dictionary 5 vector generation means 6 case search means 7 display means 8 bilingual sentence display means 9 case registration means

Claims (3)

[Claims] 1. A case base in which a set of a case and its bilingual sentence is stored as a bilingual case, a case vector representing a characteristic of the case, and an index indicating a correspondence between the case vector and the bilingual case are stored. Index storage means, an input means for inputting a translation request including words, a word dictionary holding a characteristic word vector indicating the degree of association between the translation request word and the characteristic word, and the characteristic word vector corresponding to the translation request Vector generation means for generating a translation request vector based on the result of addition, a case candidate search means for searching a case candidate based on the distance between the translation request vector and the case vector, and a bilingual sentence corresponding to the searched case candidate. And a translated text display means for displaying by using an index. 2. A case candidate display means for displaying the retrieved case candidates is further provided, and the bilingual sentence corresponding to the case selected by the user from the case candidates displayed on the case candidate displaying means is displayed on the bilingual sentence displaying means. The translation device according to claim 1, which is displayed. 3. A case registration means is further provided, and when a new bilingual case is input, the vector generation means creates a new case vector from the new case using the word dictionary, and the case registration means. 3. The translation device according to claim 1, wherein the translation storage device registers a new index and a new case vector in the index storage means, and also registers the new parallel translation case in the case base.