JPH0614377B2 - Phrase candidate selection device - Google Patents

Description

The present invention relates to a bunsetsu candidate selecting device, and more particularly to a bunsetsu candidate selecting device that can select an optimum bunsetsu candidate even for a bunsetsu that constitutes a sentence that is not grammatically accurate. .

[Prior Art and Problems to be Solved by the Invention] Conventionally, various methods have been proposed and put to practical use as a method of selecting an optimum phrase from a plurality of given phrases to compose a sentence. It's starting. As one of the methods, an optimal phrase candidate selection method is known in which an optimal dependency relation is obtained from a large number of dependency relations and a phrase candidate is selected. Generally, there are multiple clauses in one sentence. When recognition processing is performed using a voice recognition device or a character recognition device, a plurality of bunsetsu candidates are obtained for each bunsetsu as the recognition result. Therefore, there are an extremely large number of bunsetsu candidates in the entire sentence. In the conventional method, a combination of bunsetsu candidates that may form a sentence is found out of these many bunsetsu candidates using a dependency relation, but a sentence that is not grammatically correct is given to the recognition device. In this case, there is a problem that it is not possible to select a correct phrase candidate that can be established as a sentence.

The sentence that is not grammatically correct includes, for example, a sentence including two or more subjects for one predicate, and a sentence in which two or more objects that have already appeared appear. Such sentences often appear in the spoken language, which poses a problem when performing speech recognition processing on them. In the conventional optimal phrase candidate selection method, there was a premise that the phrases that make up a grammatically correct sentence were given, so if the above grammatically incorrect sentence is given, the optimal phrase candidate is selected. I couldn't choose.

The present invention has been made to solve the above problems, and provides a bunsetsu candidate selection device that can select an optimum bunsetsu candidate even when a bunsetsu that constitutes a sentence that is not grammatically accurate is given. With the goal.

[Means for Solving the Problem] The bunsetsu candidate selecting device according to the present invention includes means for giving possibility data indicating the possibility of being selected for each of a plurality of given bunsetsu, and a bunsetsu pair for the bunsetsu pair. The frequency data is obtained by referring to the storage unit for storing the frequency data indicating the appearance frequency, the selection unit for selecting a phrase pair from a plurality of given phrases, and the storage unit for the phrase pair selected by the selection unit. Reference means, distance calculation means for calculating the distance between clauses selected by the selection means, given probability data, frequency data obtained by the reference means, and distance calculation for the clause pair selected by the selection means The evaluation means calculates the distance calculated by the means according to a predetermined evaluation expression, and evaluates the dependency relationship.

[Operation] In the bunsetsu candidate selection device according to the present invention, the evaluation means is calculated by the given possibility data, the frequency data obtained by the reference means, and the distance calculation means for each bunsetsu pair selected by the selection means. The distance is calculated according to a predetermined evaluation formula, and the dependency relationship is evaluated. Therefore, even if a plurality of phrases that form a sentence that is not grammatically accurate are given, the dependency relation is evaluated based on the evaluation result by the evaluation means, so that the optimum phrase can be selected.

[Embodiment of the Invention] FIG. 1 is a block diagram of a recognition system including a phrase candidate selection device 40 showing an embodiment of the present invention. Referring to FIG. 1, the recognition system includes an input device 20 for inputting a voice or a character, a recognition device 30 for performing a recognition process on the input voice or a character, and a recognition device 30.
Phrase candidate selecting device 40 for selecting an optimum phrase candidate from a plurality of phrases obtained by the above, and a modification database 60 in which information on modification relationships is stored in advance.
And a display device 50 for displaying the phrase selected by the phrase candidate selection device 40.

The bunsetsu candidate selection device 40 receives a plurality of bunsetsu candidates, which are recognition results, and the recognition probabilities of each recognized bunsetsu from the recognition device 30, and holds the holding unit 2 and the plurality of held bunsetsu candidates. A bunsetsu candidate selecting unit 3 for selecting an arbitrary pair from the bunsetsu, a susceptibility relationship reference unit 4 for referring to the susceptibility relationship 60 regarding the selected bunsetsu pair, and a susceptibility for the selected bunsetsu pair. The appearance frequency reference unit 5 that refers to the appearance frequency of the database 60, the distance calculation unit 9 that calculates the distance between the selected phrases, the recognition probability from the holding unit 2 and the appearance frequency of the phrase pair, and the phrase interval. In the evaluation value calculation unit 10 that evaluates the dependency on the basis of the distance, the evaluation value holding unit 11 that holds the evaluation values of all the bunsetsu candidates, and the evaluation value holding unit 11. From the lifting evaluation value and a determination unit 12 the highest phrase candidate evaluation value. The dependency database 60 includes dependency relationship data 7 indicating whether a bunsetsu pair has a dependency relationship and appearance frequency data 8 indicating an appearance frequency of the dependency relationship of a bunsetsu pair.

FIG. 2 is a flow chart for explaining the operation of the recognition system shown in FIG. Next, the operation will be described with reference to FIGS. 1 and 2. In the following description, as an example of a sentence that is not grammatically accurate, the input device 20 is given a case where "I will send it as soon as possible to the registration form."

The above sentence is given to the recognition device 30 via the input device 20. First, in step (indicated by SP in the figure) 1, the recognition device 30 performs recognition processing of an input sentence. As a result of the recognition processing, it is assumed that three phrase candidates and recognition probabilities are obtained for each phrase of the input sentence.

FIG. 3 is a correspondence diagram showing the phrase candidates and their recognition probabilities for each phrase of the input sentence. For example, as an example of the phrase candidates corresponding to "urgent", the phrase candidates such as "70", "start and finish", and "urgent" are recognized, and the recognition probabilities 3, 2, and 1 are obtained from the recognition device 30 for each. . The phrase candidates and the recognition probabilities shown in FIG. 3 are held in the holding unit 2.

Next, in step 2, the phrase candidate selection unit 3 selects a phrase pair. This choice is made for a combination of two clauses that have not yet been evaluated for dependency.

In step 3, the distance calculation unit 9 calculates the distance between clauses for the selected clause pair. The calculated distance is given to the evaluation value calculation unit 10.

The following Table 1 shows the contents of data prepared in advance in the dependency database 60.

As shown in Table 1, in the dependency database 60, bunsetsu pairs that may have a dependency relationship and their appearance frequencies are stored in association with each other.

Next, in step 4, the existence of the dependency relationship is checked for the clause pair selected in the dependency database 60 by the dependency relationship reference unit 4.

In step 5, the appearance frequency is referred to from the dependency database 60 for the clause pairs having the dependency relationship. Therefore, in the example of the above sentence, the appearance frequency data for the phrase pair is obtained as shown in Table 1.

In step 6, the evaluation value calculation unit 10 calculates an evaluation value based on the following evaluation formula.

Evaluation value = recognition probability + W1 × appearance frequency−W2 × distance (1) Here, W1 and W2 are predetermined coefficients.

For example, when the phrase candidate selection unit 3 selects the phrase pair “registered” and “paper”, the appearance frequency 12 is obtained from the dependency database shown in Table 1. On the other hand, the third
As shown in the figure, the recognition device 30 has already given the recognition probability 4 for each of the phrases “registered” and “paper”. When the distance between phrases obtained by the distance calculation unit 9 is 4, and the values of the coefficients W1 and W2 in the equation (1) are 1, the evaluation value 16 is obtained from the equation (1). Similarly, when the bunsetsu candidate selecting unit 3 selects a bunsetsu pair “to that” and “send to”, an appearance frequency of 8 is obtained from Table 1, and a recognition probability of 5 is obtained from the recognition result shown in FIG.
And 7 are obtained. Furthermore, assuming that the distance between phrases in this case is 1, the evaluation value 19 is obtained from Expression (1).

In step 7, the evaluation value obtained by the evaluation value calculation unit 10 is held in the evaluation value holding unit 11.

In step 8, it is judged whether or not the evaluation values have been obtained for all clause pairs. If there is a clause whose dependency relationship is not yet evaluated, the procedure returns to step 2 and the above operation is repeated.

FIG. 4 is a schematic diagram showing evaluation values and dependency relationships obtained for each phrase pair. In this way, an evaluation value is obtained for each clause pair having a dependency relationship.

In step 9, the judgment unit 12 selects the clause having the largest evaluation value.

FIG. 5 is a correspondence diagram showing the evaluation value for each phrase candidate. The determination unit 12 determines the bunsetsu that has the largest evaluation value for each bunsetsu candidate, so that the optimum bunsetsu that may form a sentence, that is, "urgent", "send", and "answer" , “Paper”, and “There” are selected.

The optimum phrase candidate obtained in this way is displayed on the display device 5, for example.
It is given to 0 and displayed.

[Effects of the Invention] As described above, according to the present invention, the possibility data of a given bunsetsu, the frequency data of a bunsetsu pair, and the distance between bunsetsus are calculated for a selected bunsetsu pair according to a predetermined evaluation formula. Since the evaluation means for evaluating the dependency relation is provided, a bunsetsu candidate selecting device is provided which can select an optimum bunsetsu candidate even if a plurality of bunsetsus that constitute a sentence that is not grammatically accurate are given.

[Brief description of drawings]

FIG. 1 is a block diagram of a recognition system including a phrase candidate selection device showing an embodiment of the present invention. FIG. 2 is a flow chart for explaining the operation of the recognition system shown in FIG. FIG. 3 is a correspondence diagram showing the correspondence between the phrase candidates output from the recognition device shown in FIG. 1 and the recognition probabilities. FIG. 4 is a schematic diagram showing evaluation values and dependency relationships obtained for each phrase pair. FIG. 5 is a correspondence diagram showing the correspondence between each phrase candidate and its evaluation value. In the figure, 2 is a storage unit, 3 is a phrase candidate selection unit, 4 is a dependency relation reference unit, 5 is an appearance frequency reference unit, 9 is a distance calculation unit, 10 is an evaluation calculation unit, 11 is an evaluation value storage unit, 12 Is a determination unit, 20 is an input device, 30 is a recognition device, 40 is a phrase candidate selection device, 50 is a display device, and 60 is a dependency database.

Claims (1)

[Claims] 1. A bunsetsu candidate selecting device for selecting a bunsetsu optimal for constructing a sentence from a plurality of given bunsetsu, showing a possibility that each of the plurality of bunsetsus should be selected. Possibility data giving means, with respect to a clause pair constituting the plurality of clauses, storage means for storing frequency data indicating the appearance frequency in advance, selecting means for selecting a clause pair from the plurality of clauses, A reference unit for obtaining frequency data by referring to the storage unit for the phrase pair selected by the selection unit, a distance calculation unit for calculating a distance between the phrases selected by the selection unit, and a selection unit selected by the selection unit. The given probability data, the frequency data obtained by the reference means, and the distance calculated by the distance calculation means in a predetermined evaluation formula, Calculated I, and a evaluation means for evaluating a dependency relationship, phrase candidate selection device.