JP2023093349A - Information processing device and information processing method - Google Patents

Abstract

To provide an information processing device and an information processing method.SOLUTION: An information processing device includes: a fifth text generation unit configured to generate a plurality of fifth texts corresponding to a processing waiting text using a text generation model; a candidate text selection unit configured to select, from among the plurality of fifth texts, a fifth text having a semantic matching degree with the processing waiting text equal to or greater than a predetermined matching degree as a candidate fifth text; a text similarity calculation unit configured to calculate, for each candidate fifth text, a text similarity between the candidate fifth text and each of other candidate fifth texts; and a target text selection unit configured to select, from among the candidate fifth texts, a fourth predetermined number of candidate fifth texts having the lowest text similarity between each other as target texts for subsequent processing.SELECTED DRAWING: Figure 6

Description

The present invention relates to the field of information processing, and more particularly to an information processing apparatus and information processing method.

Text is used in many applications in the field of information processing. There are currently several techniques for processing text.

An object of the present invention is to provide an information processing apparatus and an information processing method for processing text.

According to one aspect of the present invention, an information processing device is provided, which comprises:
a second text generation unit configured to, for each of one or more predetermined first texts, generate a plurality of second texts corresponding to the first text using a plurality of text generation models;
selecting a first predetermined number of text generation models from among the plurality of text generation models as candidate models based on the degree of semantic matching between the plurality of second texts and the corresponding first text; a composed candidate model selection unit;
For each of the one or more predetermined first texts, calculate text similarity between each of a plurality of second texts corresponding to the first text generated using the candidate model. and a similarity computation unit configured to compute model similarities between said candidate models based on text similarities between each other of second texts;
a target model selection unit configured to select, as target models, from among said candidate models a second predetermined number of candidate models having the lowest model similarity between each other; and, using said target models, pending processing. A fourth text generation unit configured to generate a second predetermined number of fourth texts corresponding to the text for subsequent processing.

According to another aspect of the present invention, an information processing device is provided, which comprises:
a fifth text generation unit configured to generate a plurality of fifth texts corresponding to the pending text using the text generation model;
a candidate text selection unit configured to select, from among the plurality of fifth texts, a fifth text having a semantic matching degree with the awaiting text equal to or greater than a predetermined matching degree as a candidate fifth text;
a text similarity computation unit configured to compute, for each candidate fifth text, a text similarity between said candidate fifth text and each of the other candidate fifth texts; and said candidate fifth text. a target text selection unit configured to select, from among the texts, a fourth predetermined number of candidate fifth texts having the lowest text similarity between each other as target texts for subsequent processing;

According to another aspect of the present invention, an information processing method is provided, comprising:
for each of one or more predetermined first texts, using a plurality of text generation models to generate a plurality of second texts corresponding to the first text;
selecting a first predetermined number of text generation models from among the plurality of text generation models as candidate models based on the degree of semantic match between the plurality of second texts and the corresponding first text;
For each of the one or more predetermined first texts, calculate text similarity between each of a plurality of second texts corresponding to the first text generated using the candidate model. and calculating the model similarity between the candidate models based on the text similarity between the second texts;
selecting a second predetermined number of candidate models having the lowest model similarity between each other as target models from the candidate models; and using the target models, a second predetermined number corresponding to pending text. Generating a fourth text for subsequent processing.

According to other aspects of the invention, a computer program code and a computer program product for implementing the above method according to the invention and a computer storing the computer program code for implementing the above method according to the invention. A readable storage medium is further provided.

1 is a block diagram of a functional configuration example of an information processing apparatus according to a first embodiment of the present invention; FIG. Fig. 10 shows an example of generating a second text using the back translation method; FIG. 5 is a block diagram of a functional configuration example of an information processing apparatus according to a second embodiment of the present invention; For the Charades-STA dataset, the similarity sequence for the pending text, the average similarity sequence obtained by the video timing positioning unit, the median similarity sequence obtained by the video timing positioning unit, and the manual average FIG. 10 is a diagram showing an example of a result of target frame identification using a similarity sequence; FIG. 11 is a block diagram of an example of the functional configuration of an information processing device according to the third embodiment of the present invention; FIG. 12 is a block diagram of a functional configuration example of an information processing apparatus according to a fourth embodiment of the present invention; For the Charades-STA dataset, the results of target frame identification using the similarity sequence for the pending text, the average similarity sequence obtained by the video timing positioning unit, and the manual average similarity sequence, respectively. FIG. 4 is a diagram showing an example; FIG. 5 is an exemplary flow chart of an information processing method according to the fifth embodiment of the present invention; FIG. FIG. 7 is an exemplary flow chart of an information processing method according to the sixth embodiment of the present invention; FIG. 1 is a block diagram of an exemplary configuration of a personal computer that may be employed in embodiments of the invention; FIG.

Preferred embodiments for carrying out the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that these examples are merely illustrative and do not limit the present invention.

First, an implementation example of an information processing apparatus 100 according to the first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram of a functional configuration example of an information processing apparatus 100 according to the first embodiment of the present invention.

As shown in FIG. 1, the information processing apparatus 100 according to the first embodiment of the present invention includes a second text generation unit 102, a candidate model selection unit 104, a similarity calculation unit 106, a target model selection unit 108 and a fourth text generation unit. 110 may be included.

A second text generation unit 102, for each of one or more predetermined first texts, using a plurality of text generation models to generate a plurality of second texts corresponding to the first text. may be configured. For example, each text generation model may be a single text generation model. Also, for example, each of the multiple text generation models described above may correspond to a different sub-module in the same text generation model. For example, each of the plurality of second texts generated for the same first text may belong to the same word type (language type).

Candidate model selection unit 104, based on the degree of semantic matching between the plurality of second texts generated by the second text generation unit 102 and the corresponding first texts, selects from the plurality of text generation models , a first predetermined number of text generation models may be selected as candidate models. For example, the first predetermined number may be set according to actual needs. For example, existing models of Sentence-BERT may be used to determine the degree of semantic match between each secondary text and the corresponding primary text.

As an example, the candidate model selection unit 104 may select a text generation model with a relatively high semantic match between the generated second text and the corresponding first text as the candidate model.

For example, the candidate model selection unit 104 calculates, for each text generation model, an average semantic match between each second text generated using the text generation model and the corresponding first text, and , a first predetermined number of text generation models having the highest average value of corresponding word sense matching degrees may be selected as candidate models from among the plurality of text generation models described above.

Similarity calculation unit 106, for each of said one or more predetermined first texts, corresponding to said first text generated by the candidate model selected using candidate model selection unit 104 configured to calculate text similarities between the plurality of second texts to each other, and to calculate model similarities between the candidate models to each other based on the text similarities between the second texts to each other. May be.

The target model selection unit 108 may be configured to select, among the candidate models, a second predetermined number of candidate models having the lowest model similarity between each other as target models. For example, the second predetermined number may be set according to actual needs.

The fourth text generation unit 110 may be configured to utilize the target model to generate a second predetermined number of fourth texts corresponding to the pending texts for subsequent processing. If there are multiple pending texts, the fourth text generating unit 110 can generate, for each pending text, a second predetermined number of fourth texts corresponding to the pending text. For example, each of the plurality of fourth texts generated for the same pending text may belong to the same word class.

As described above, the information processing apparatus 100 according to the first embodiment of the present invention determines the degree of semantic matching between the first text and the second text generated using a plurality of text generation models, A target model can be selected from multiple text generation models by considering the text similarity between two texts. Therefore, by using the target model, it is possible to generate a fourth text having an appropriate semantic matching degree with the text to be processed, for example, a fourth text relatively close to the meaning of the text to be processed. For example, a fourth text that is relatively close in meaning to the pending text may be referred to as a high quality fourth text. Also, the text similarity between the fourth texts may be relatively low, resulting in relatively high diversity. In other words, by using the information processing apparatus 100, a fourth text having high quality and high diversity can be generated.

Further, by selecting the target model as described above and generating the fourth text by using the target model, the information processing apparatus 100 can reduce noise caused by inaccurate word sense matching. Here, "noise" refers to fourth text that is not very close in meaning to pending text.

For example, the text to be processed and the first text may belong to the same language, such as Chinese, English, Japanese, etc., but not limited to this.

As another example, the pending text and the first text may belong to different word types. In such a case, for example, the fourth text generation unit 110 first converts the text to be processed into the same vocabulary as the first text, and then uses the target model to generate a second predetermined number of fourth texts. can be

For example, the second text and the fourth text may belong to the same word class, such as Chinese, English, Japanese, etc., but is not limited to this.

As an example, the first text, the second text and the fourth text may belong to the same word type. In such a case, for example, the second text generation unit 102 can use a back-translation method to generate the second text. For example, the second text generation unit 102 generates the first text in a word type (hereinafter referred to as a "second word type") that is different from the word type of the first text (hereinafter may be referred to as a "first word type"). ), and then convert the converted text to the first word type to obtain the second text. For example, FIG. 2 shows an example of generating a second text using the back translation method, wherein the first language type is English and the second language type is Chinese. As shown in FIG. 2, the second text generation unit 102 converts the English first text "The weather is good" into the Chinese text "good weather", and then converts the Chinese text "good weather" into English. can be converted to the second text "Good weather".

Although FIG. 2 shows an example in which the second text generation unit 102 generates one second text, the second text generation unit 102 can generate a plurality of different first texts according to actual needs. The text may be converted to a second language type (eg, Japanese, German, Spanish, etc.), and then the converted text may be converted to the first language type to obtain a plurality of second texts.

Also, for example, the fourth text generation unit 110 may use a back translation method to generate the fourth text. For example, the fourth text generation unit 110 may generate one or more fourth texts using methods similar to those of the second text generation unit 102 described above.

For example, each text generation model may be a single text translation model, e.g., one text generation model is a text generation model for conversion between a first language text and a second language text. It may correspond to a translation model.

As an example, the similarity calculation unit 106 can calculate the text similarity between any two second texts in the following manner: each of the two second texts is a set of words and/or phrases. dividing and obtaining intersections and unions between the obtained sets, and the number of words and phrases contained in the obtained intersections (i.e., the sum of the number of words and the number of phrases); A ratio of the number of words and phrases (i.e., the sum of the number of words and the number of phrases) contained in the obtained union set may be taken as the text similarity between any two second texts. can.

It should be noted that "word" as used herein may represent a single word, such as a single English word, a single Chinese character, a single Japanese word, and the like. Also, as used herein, a "phrase" may refer to a combination of two or more words.

The above text similarity calculation method will be further described below based on an example where the second text is an English text. For example, if the second text generated by the second text generation unit 102 contains the first English phrase "a man was standing in the bathroom holding glasses" and the second English phrase "a person is standing in the bathroom holding a glasses". including The similarity computation unit 106 divides the first English phrase into a first word set {a, man, was, standing, in, the, bathroom, holding, glasses} and the second English phrase into a second word set { a, person, is, standing, in, the, bathroom, holding, a, glass}. The intersection of the first word set and the second word set is {a, man, standing, in, the, bathroom, holding}, containing seven words, and the union of the first word set and the second word set The set is {a, man, was, is, standing, in, the, bathroom, holding, glasses, glasses} and contains 11 words. The similarity calculation unit 106 calculates the ratio (i.e., 7/11) of the number of words in the intersection (i.e., 7) to the number of words in the union (i.e., 11) as the ratio of the first English phrase to the second English phrase. can be used as a text similarity measure between

For example, if the second text is English text and contains uppercase and lowercase letters, the similarity calculation unit 106 can convert the second text to uppercase or lowercase letters, then split the second text, and of course similarity calculation After unit 106 splits the second text, it may convert the split words to upper or lower case. As those skilled in the art will appreciate, such letter conversions for English text are equally applicable to text in other language languages, such as German text, Spanish text, French text, and the like.

As an example, the similarity calculation unit 106 can calculate the model similarity between any two candidate models in the following manner: for each of the one or more predetermined first texts, obtain the text similarity between the second text corresponding to the first text obtained by any two candidate models of , and obtain the text similarity between the second text corresponding to said first text, and the text corresponding to said one or more given first texts The average similarity is calculated as the model similarity between any two candidate models above. For example, if one or more predetermined first texts include text ₁ , text ₂ , _. , the text similarity between the second text of text ₁ obtained using candidate model A and the second text of text ₁ obtained using candidate model B is _s1 , and candidate model A is The text similarity between the second text of text ₂ obtained using candidate model B and the second text of text ₂ obtained using candidate model B is s ₂ and obtained using candidate model A Let the text similarity between the second text of text _m and the second text of text _m obtained using candidate model B be s _m . In such a case, the similarity calculation unit 106 calculates the mean values _of the text similarities _{s 1} , s ₂ , . . . , _{s m corresponding} to text 1 , text 2 , . It can be the model similarity with model B. Of course, the similarity calculation unit 106 may calculate the model similarity between candidate model A and candidate model _B in other ways based on the text similarities s ₁ , s ₂ , . .

As an example, the target model selection unit 108 uses a determinant point process (e.g., Chen L, Zhang G, Zhou E. Fast greedy map influence for determinant point process to improve recommendation Diversity [J] Advances in Neural Information Processing Systems, 2018, 31))), a second predetermined number of candidate models with the lowest model similarity between each other among the candidate models can be selected as target models. For example, the target model selection unit 108 can build an N*N-dimensional matrix SS based on the model similarity between candidate models, where each element in the N*N-dimensional matrix SS is the corresponding candidate model's For example, SS[i,j] represents the similarity between the ith candidate model and the jth candidate model. Wherein N is a natural number greater than 0, which indicates the number of candidate models (ie, the first predetermined number), and i and j are natural numbers greater than 0 and less than or equal to N. The target model selection unit 108 can then use the determinant point process to determine the M*M dimensional maximal determinant submatrix of the N*N dimensional matrix SS, wherein the M*M dimensional maximal determinant submatrix is , corresponds to a second predetermined number of candidate models having the lowest model similarity between each other among the candidate models. Here, M is a natural number greater than 0 and less than N, which represents the second predetermined number.

The information processing apparatus 300 according to the second embodiment of the present invention will be described below with reference to FIG. FIG. 3 is a block diagram of a functional configuration example of an information processing apparatus 300 according to the second embodiment of the present invention.

As shown in FIG. 3, the information processing apparatus 300 according to the second embodiment of the present invention includes a second text generation unit 302, a candidate model selection unit 304, a similarity calculation unit 306, a target model selection unit 308, and a fourth text generation unit. 310 and a video timing positioning unit (also called a video positioning unit) 312 may be included. The second text generation unit 302, the candidate model selection unit 304, the similarity calculation unit 306, the target model selection unit 308, and the fourth text generation unit 310 are included in the information processing apparatus 100 described with reference to FIGS. Since it is similar to the included second text generation unit 102, candidate model selection unit 104, similarity calculation unit 106, target model selection unit 108 and fourth text generation unit 110, a detailed description thereof will be omitted here.

For example, the pending text may be user-inputted text, or text obtained by converting user-inputted speech sounds or images, for example, the pending text may be an object that the user wishes to identify from a given video. , events, etc. Based on the pending text and the second predetermined number of fourth texts generated by the fourth text generating unit 310, the video timing positioning unit 312 generates a frame corresponding to the pending text from the given video (hereinafter "target frame"). (which may be referred to as a "frame") can be identified. Because the video timing positioning unit 312 can utilize the enhanced text (ie, the pending text and the quaternary text) to perform identification for a given video, the identification accuracy can be improved.

For example, video timing positioning unit 312 can use a trained multimodal model to identify the position of frames corresponding to pending text from a given video. For example, the video timing positioning unit 312 uses a trained multimodal model for each of the pending text and the quaternary text to calculate the similarity between each frame in a given video and the text. , we can obtain M+1 similarity sequences, and locate the frame corresponding to the pending text in a given video based on the obtained M+1 similarity sequences. Here, M is a natural number greater than 0, which represents the number of fourth texts. For example, the video timing positioning unit 312 obtains an average similarity sequence by averaging over the obtained M+1 similarity sequences, and calculates pending text in a given video based on the average similarity sequence. can recognize the position of the corresponding frame. Also, for example, the video timing positioning unit 312 determines the median of the peak values of the obtained M+1 similarity sequences, and determines the similarity sequence corresponding to the median (hereinafter "median similarity sequence"). ), the position of the frame corresponding to pending text in a given video can be identified.

FIG. 4 shows the similarity sequence for the pending text, the average similarity sequence obtained by the video timing positioning unit 312, and the median similarity sequence obtained by the video timing positioning unit 312, respectively, for the Charades-STA dataset. , and a manual averaged similarity sequence to perform target frame identification. Among them, the manual averaged similarity sequence consists of M texts corresponding to pending texts generated by M manually selected text generation models (M=10 in the example shown in FIG. 4) and , represents the average similarity sequence obtained by averaging over the similarity sequence with a given video and the similarity sequence for the pending text.

In FIG. 4, IoU0.5 R@1 indicates that if the IoU (Intersection over Union) between the optimal identification result and the true value is greater than 0.5, the recall rate when the identification result is correct is determined. show. As shown in FIG. 4, the recall rate based on the medium similarity sequence obtained by the video timing positioning unit 312 is about 3.01% higher than the recall rate based on the similarity sequence for pending text. , an improvement of about 2.01% compared to the recall rate based on manual average similarity sequences. Also, the recall rate based on the average similarity sequence obtained by the video timing positioning unit 312 improved by about 1.83% compared to the recall rate based on the similarity sequence for pending text, and the manual average similarity It is about 0.83% better than the recall rate based on sequencing.

Also, as shown in FIG. 4, the recall rate based on the median similarity sequences obtained by the video timing positioning unit 312 is about 1 more than the recall rate based on the average similarity sequences obtained by the video timing positioning unit 312. .18% improvement, because the medium similarity sequence can further reduce the effect of noise compared to the average similarity sequence.

As an example, the information processing apparatus 300 may further include a candidate text selection unit 314 and a target text selection unit 316, as shown in FIG.

The candidate text selection unit 314 is configured to select, from among the second predetermined number of fourth texts, a plurality of fourth texts having a semantic matching degree with the text to be processed equal to or greater than a predetermined matching degree as candidate texts. Also good.

The target text selection unit 316 may be configured to select, from among the candidate texts selected by the candidate text selection unit 314, a third predetermined number of candidate texts having the lowest text similarity between each other as target texts. good. The third predetermined number may be set according to actual needs.

The operation of the candidate text selection unit 314 and the target text selection unit 316 described above may, for example, cause further selection of target texts among the fourth texts that have a lower degree of similarity between each other, whereby , can further improve the diversity of target texts.

If the information processing apparatus 300 includes a candidate text selection unit 314 and a target text selection unit 316, the video timing positioning unit 312, based on the pending text and the target text, locates frames corresponding to the pending text from a given video. By identifying, for example, the identification accuracy can be further improved.

For example, target text selection unit 316 may utilize a determinant point process to select a third predetermined number of candidate texts that have the lowest text similarity between each other among the candidate texts as target texts. For example, target text selection unit 316 constructs a matrix in a manner similar to that described for target model selection unit 108 above, and then uses a determinant point process to use the L*L dimensional maxima of the constructed matrix. A determinant sub-matrix can be determined, of which the L*L dimensional maximal determinant sub-matrix corresponds to a third predetermined number of candidate texts with the lowest text similarity between each other among the candidate texts. Here, L is a natural number greater than 0, which represents the third predetermined number.

Note that FIG. 3 uses dashed boxes to denote candidate text selection unit 314 and target text selection unit 316, which in some embodiments information processing apparatus 300 may select candidate text selection unit 314 and target text selection unit 316. means that it does not have to contain

The information processing apparatus 400 according to the third embodiment of the present invention will be described below with reference to FIG. FIG. 5 is a block diagram of a functional configuration example of an information processing apparatus 400 according to the third embodiment of the present invention.

As shown in FIG. 5, the information processing apparatus 400 according to the third embodiment of the present invention includes a second text generation unit 402, a candidate model selection unit 404, a similarity calculation unit 406, a target model selection unit 408, and a fourth text generation unit. 410 and a multimodal model training unit 420. The second text generation unit 402, the candidate model selection unit 404, the similarity calculation unit 406, the target model selection unit 408 and the fourth text generation unit 410 are similar to the first information processing device 100 described above with reference to FIGS. are similar to the second text generation unit 102, the candidate model selection unit 104, the similarity calculation unit 106, the target model selection unit 108 and the fourth text generation unit 110 included in .

For example, the multimodal model training unit 420 trains a multimodal model for video timing positioning based on the pending text and the second predetermined number of fourth texts to obtain a trained multimodal model. may be configured. Thereby, for example, the discrimination accuracy, robustness, etc. of the trained multimodal model can be improved.

As an example, the information processing apparatus 400 may further include a candidate text selection unit 414 and a target text selection unit 416, as shown in FIG. The candidate text selection unit 414 and the target text selection unit 416 are similar to the candidate text selection unit 314 and the target text selection unit 316 described with reference to FIG. 3 above, so a detailed description thereof is omitted here.

For example, multimodal model training unit 420 trains a multimodal model for video timing positioning based on the pending text and the target text selected by candidate text selection unit 414, resulting in a trained multimodal model. , which can further improve, for example, discriminative accuracy, robustness, etc. of the trained multimodal model.

It should be noted that FIG. 5 shows candidate text selection unit 414 and target text selection unit 416 in dashed boxes, which indicates that information processor 400 includes candidate text selection unit 414 and target text selection unit 416 in some embodiments. It means that you can do without it.

The information processing apparatus 500 according to the fourth embodiment of the present invention will be described below with reference to FIG. FIG. 6 is a block diagram of a functional configuration example of an information processing apparatus 500 according to the fourth embodiment of the present invention.

As shown in FIG. 6, the information processing device 500 according to the fourth embodiment of the present invention can include a fifth text generation unit 502, a candidate text selection unit 504, a text similarity calculation unit 506 and a target text selection unit 508.

The fifth text generation unit 502 may be configured to utilize the text generation model to generate a plurality of fifth texts corresponding to the pending text. For example, the fifth text generation unit 502 can generate multiple fifth texts by back translation method. For example, the fifth text generation unit 502 can use multiple text generation models (eg, multiple text translation models) to generate multiple fifth texts. Also, for example, the fifth text generation unit 502 can use one text generation model (eg, one text translation model) to generate a plurality of fifth texts by beam search.

A candidate text selection unit 504 selects, from among the plurality of fifth texts generated by the fifth text generation unit 502, fifth texts having a degree of semantic matching with the text awaiting processing equal to or greater than a predetermined degree of matching as candidate fifth texts. It may be configured to

The text similarity calculation unit 506 may be configured to calculate, for each candidate fifth text, the text similarity between the candidate fifth text and each of the other candidate fifth texts. For example, the text similarity calculation unit 506 calculates the text similarity between the candidate fifth texts using a method similar to the method by which the similarity calculation unit 106 in the first embodiment described above calculates the text similarity. can be calculated.

A target text selection unit 508 is configured to select, from among the candidate fifth texts, a fourth predetermined number of candidate fifth texts having the lowest text similarity between each other as target texts for subsequent processing. may be configured. For example, target text selection unit 508 may utilize a determinant point process to select, from among candidate fifth texts, a fourth predetermined number of candidate fifth texts with the lowest text similarity between each other for subsequent processing. can be selected as the target text for For example, the fourth predetermined number can be set according to actual needs.

As described above, the information processing apparatus 500 according to the fourth embodiment of the present invention determines the degree of word sense matching between the pending text and the fifth text generated using the text generation model, and the degree of matching of the fifth text. A target text can be selected from among a plurality of fifth texts by considering the text similarity between them. Therefore, there is an appropriate degree of semantic match between the target text and the pending text, eg, the semantics of the target text and the pending text may be close to each other. Also, the text similarity between the target texts may be relatively low, resulting in relatively high diversity. In other words, with the information processing device 500, a target text with high quality and high diversity can be obtained.

For example, the text to be processed may include user-inputted text, or text obtained by converting user-inputted speech sounds or images.

As an example, the information processing device 500 may further include a video timing positioning unit 512, as shown in FIG. Based on the pending text and the target text selected by the target text selection unit 508, the video timing positioning unit 512 can recognize the position of the frame corresponding to the pending text from a given video. Because the video timing positioning unit 512 can perform identification for a given video with enhanced text (ie, pending text and target text), identification accuracy can be improved. For example, the video timing positioning unit 512 may have settings similar to those of the video timing positioning unit 312 in the second embodiment described above, so a detailed description thereof will be omitted here.

FIG. 7 shows target frame identification using the similarity sequence for the pending text, the average similarity sequence obtained by the video timing positioning unit 512, and the manual average similarity sequence, respectively, for the Charades-STA dataset. shows an example of the results of In the example shown in FIG. 7, the number of target texts corresponding to one pending text is ten. As shown in FIG. 7, the recall rate based on the average similarity sequence obtained by the video timing positioning unit 512 is about 1.62% better than the recall rate based on the similarity sequence for the pending text, and the manual It is about 0.62% better than the recall rate based on the average similarity sequence.

As another example, the text to be processed may include training text, and as shown in FIG. 6, information processing apparatus 500 may further include multimodal model training unit 520 . For example, multimodal model training unit 520 trains a multimodal model for video timing positioning based on pending text and target text selected by target text selection unit 508, resulting in trained multimodal A model may be obtained, which can, for example, improve the discriminative accuracy, robustness, etc. of the trained multimodal model. For example, the multimodal model training unit 520 may have settings similar to those of the multimodal model training unit 420 in the third embodiment described above, so a detailed description thereof is omitted here.

It should be noted that in FIG. 6 the video timing positioning unit 512 and the multimodal model training unit 520 are shown in dashed boxes, indicating that in some embodiments the information processing apparatus 500 may perform the video timing positioning unit 512 and/or the multimodal model training. This indicates that the unit 520 may not be included.

Although the information processing apparatus according to the embodiments of the present invention has been described above, the present invention further provides the following information processing method embodiments corresponding to the above-described information processing apparatus embodiments.

FIG. 8 is an exemplary flowchart of an information processing method 600 in accordance with an embodiment of the present invention. As shown in FIG. 8, the information processing method 600 according to an embodiment of the present invention may start with a start step S601 and end with an end step S612, and also includes a second text generation step S602, a candidate model selection step S604, A similarity calculation step S606, a target model selection step S608 and a fourth text generation step S610 may be included.

In the second text generation step S602, for each of one or more predetermined first texts, multiple text generation models can be used to generate multiple second texts corresponding to the first text. For example, the second text generation step S602 can be performed by the second text generation units 102, 302 and 402 in the above-described apparatus embodiments, so the specific details are the second text generation units 102, 302 and 402 described above. can be referred to, and the detailed description thereof is omitted here.

In the candidate model selection step S604, the plurality of text generation models are selected based on the degree of semantic matching between the plurality of second texts generated in the second text generation step S602 and the corresponding first texts. From among them, a first predetermined number of text generation models can be selected as candidate models. For example, the first predetermined number can be set according to actual needs. Also, for example, candidate model selection step S604 may be performed by candidate model selection units 104, 304, and 404 in the apparatus embodiments described above, so specific details will be provided for candidate model selection units 104, 304, and 404 described above. , and the detailed description thereof is omitted here.

In a similarity calculation step S606, for each of said one or more predetermined first texts, a plurality of corresponding first text generated by the candidate model selected in the candidate model selection step S604 A text similarity between the second texts can be calculated, and a model similarity between the candidate models can be calculated based on the text similarity between the second texts. For example, the similarity calculation step S606 can be performed by the similarity calculation units 106, 306 and 406 in the above-described apparatus embodiments, so the specific details are given in the description of the similarity calculation units 106, 306 and 406 above. can be referred to, and the detailed description thereof is omitted here.

In the target model selection step S608, from among the candidate models, a second predetermined number of candidate models having the lowest model similarity between each other can be selected as target models. For example, the second predetermined number can be set according to actual needs. For example, target model selection step S608 may be performed by target model selection units 108, 308, and 408 in the above-described apparatus embodiments, so specific details are provided in the descriptions of target model selection units 108, 308, and 408 above. can be referred to, and the detailed description thereof is omitted here.

In a fourth text generation step S610, the target model can be used to generate a second predetermined number of fourth texts corresponding to the pending texts for subsequent processing. For example, the fourth text generation step S610 can be performed by the fourth text generation units 110, 310 and 410 in the above-described apparatus embodiments, so the specific details are described in the fourth text generation units 110, 310 and 410 above. can be referred to, and the detailed description thereof is omitted here.

Similar to the information processing apparatus according to the embodiments of the present invention, the information processing method 600 according to the fifth embodiment of the present invention also uses the semantic definition between the first text and the second text generated by a plurality of text generation models. A target model can be selected from a plurality of text generation models considering the degree of match and the degree of text similarity between the second texts. Therefore, using the target model, it is possible to generate a fourth text having an appropriate degree of semantic matching with the text to be processed, for example, to generate a fourth text close to the meaning of the text to be processed. Also, the text similarity between the fourth texts may be relatively low, resulting in relatively high diversity. In other words, the information processing method 600 allows the generation of fourth texts with high quality and high diversity.

For example, in the similarity calculation step S606, the text similarity between any two second texts can be calculated in the following way: each of the two second texts is a set of words and/or phrases. dividing and obtaining intersections and unions between the obtained sets, and the number of words and phrases contained in the obtained intersections (i.e., the sum of the number of words and the number of phrases); A ratio of the number of words and phrases (ie, the sum of the number of words and the number of phrases) contained in the obtained merged set is calculated as the text similarity between the two arbitrary second texts.

For example, the similarity calculation step S606 can calculate the model similarity between any two candidate models in the following manner: For each of the one or more predetermined first texts, the above obtaining the text similarity between the second texts corresponding to the first texts obtained by any two candidate models; Calculate the mean of the degrees as the model similarity between any two candidate models above.

For example, the target model selection step S608 may use a determinant point process to select as target models a second predetermined number of candidate models that have the lowest model similarity between each other among the candidate models.

As an example, information processing method 600 may further include a video timing positioning step (not shown). The video timing positioning step may identify the position of the frame corresponding to the pending text from the given video based on the pending text and the second predetermined number of fourth texts generated in the fourth text generating step S610. The identification accuracy can be improved because the video timing positioning step can use the enhanced text (ie, the pending text and the quaternary text) to perform the identification for a given video.

For example, the video timing positioning step can be performed by the video timing positioning unit 312 in the second embodiment described above, so for specific details, please refer to the description of the video timing positioning unit 312 above, which is hereby referred to. A detailed explanation is omitted.

As another example, information processing method 600 may further include a multimodal model training step (not shown). In a multimodal model training step, training a multimodal model for video timing positioning based on the pending text and the second predetermined number of fourth texts to obtain a trained multimodal model; Thereby, for example, the discrimination accuracy, robustness, etc. of the trained multimodal model can be improved. For example, the multimodal model training step can be performed by the multimodal model training unit 420 in the third embodiment above, so for specific details see the description of the multimodal model training unit 420 above, here. The detailed explanation is omitted here.

As an example, the information processing method 600 may further include a candidate text selection step and a target text selection step (not shown).

In the candidate text selection step, from among the second predetermined number of fourth texts, a plurality of fourth texts having a word sense matching degree with the text awaiting processing equal to or greater than a predetermined matching degree can be selected as candidate texts. For example, the candidate text selection step may be performed by the candidate text selection units 314 and 414 in the apparatus embodiments described above, so for specific details see the description of the candidate text selection units 314 and 414 above, here. The detailed explanation is omitted here.

In the target text selection step, from among the candidate texts selected in the candidate text selection step, a third predetermined number of candidate texts having the lowest text similarity between each other may be selected as target texts. For example, since the target text selection step can be performed by the target text selection units 316 and 416 in the apparatus embodiments described above, the specific details can be referred to the description of the target text selection units 316 and 416 above, here. The detailed explanation is omitted here.

For example, if the information processing method 600 includes a candidate text selection step and a target text selection step, the video timing positioning step, based on the pending text and the target text, determines the position of the frame corresponding to the pending text from the given video. can be identified, for example, the identification accuracy can be further improved.

Also, for example, if the information processing method 600 includes a candidate text selection step and a target text selection step, the multimodal model training step trains a multimodal model for video timing positioning based on the pending text and the target text. By doing so, a trained multimodal model can be obtained, so that, for example, the discrimination accuracy, robustness, etc. of the trained multimodal model can be further improved.

The information processing method 700 according to the sixth embodiment of the present invention will now be described with reference to FIG. FIG. 9 is an exemplary flowchart of an information processing method 700 in accordance with an embodiment of the present invention. As shown in FIG. 9, the information processing method 700 in an embodiment of the present invention may start with a start step S701 and end with an end step S712, and also include a fifth text generation step S702, a candidate text selection step S704, A text similarity calculation step S706 and a target text selection step S708 may be included.

In the fifth text generation step S702, the text generation model can be used to generate a plurality of fifth texts corresponding to the pending texts. For example, the fifth text generation step S702 can be implemented by the fifth text generation unit 502 in the above fourth embodiment, so the specific details can refer to the description of the fifth text generation unit 502 above, A detailed description thereof is omitted here.

In the candidate text selection step S704, from among the plurality of fifth texts generated in the fifth text generation step S702, a fifth text having a word meaning matching degree with the text awaiting processing equal to or higher than a predetermined matching degree is selected as a candidate fifth text. You can choose. For example, since the candidate text selection step S704 can be implemented by the candidate text selection unit 504 in the fourth embodiment described above, the specific details can be referred to the above description of the candidate text selection unit 504, and here the details are described. Description is omitted.

In a text similarity calculation step S706, for each candidate fifth text, the text similarity between the candidate fifth text and each of the other candidate fifth texts can be calculated. For example, the text similarity calculation step S706 can be implemented by the text similarity calculation unit 506 in the above fourth embodiment, so the specific details can refer to the above description of the text similarity calculation unit 506, A detailed description thereof is omitted here.

In a target text selection step S708, from among the candidate fifth texts, a fourth predetermined number of candidate fifth texts having the lowest text similarity between each other may be selected as target texts for subsequent processing. can. For example, the target text selection step S708 can be implemented by the target text selection unit 508 in the above-described fourth embodiment, so the specific details can refer to the above description of the target text selection unit 508, and here A detailed description thereof is omitted.

As described above, the information processing method 700 according to the sixth embodiment of the present invention provides the degree of semantic matching between the text to be processed and the fifth text generated using the text generation model, and A target text can be selected from among a plurality of fifth texts by considering the text similarity of . Therefore, there is an appropriate degree of semantic matching between the target text and the to-be-processed text, for example, the target text and the to-be-processed text may be close in meaning. Also, the text similarity between the target texts may be relatively low, resulting in relatively high diversity. In other words, the information processing method 700 makes it possible to obtain target texts with high quality and high diversity.

As an example, information processing method 700 may further include a video timing positioning step (not shown). Based on the pending text and the target text, the video timing positioning step can recognize, from a given video, the position of the frame corresponding to the pending text, thus improving the identification accuracy, for example. For example, the video timing positioning step can be performed by the video timing positioning unit 512 in the fourth embodiment described above, so the specific details can be referred to the description of the video timing positioning unit 512 above, and here the details are described. Description is omitted.

As another example, information processing method 700 may further include a multimodal model training step (not shown). For example, the multimodal model training step may train a multimodal model for video timing positioning based on the pending text and the target text to obtain a trained multimodal model, thereby e.g. , can improve the discrimination accuracy, robustness, etc. of trained multimodal models. For example, the multimodal model training step can be performed by the multimodal model training unit 520 in the fourth embodiment above, so for specific details see the description of the multimodal model training unit 520 above, here. The detailed explanation is omitted here.

Although the function settings and operations of the information processing apparatus and the information processing method according to the embodiments of the present invention have been described above, these are merely examples, and those skilled in the art will be able to modify the above-described embodiments based on the principles of the present invention. For example, functional modules and operations in each embodiment may be added, reduced, combined, etc., and all such modifications are within the scope of the present invention.

In addition, since the method embodiments herein correspond to the apparatus embodiments described above, the corresponding descriptions in the apparatus embodiments can be referred to for details not described in the method embodiments. , the detailed description of which is omitted here.

Also, the present invention further provides a storage medium and a program product. It should be noted that the machine-executable instructions in the storage media and program products according to embodiments of the present invention may be further configured to perform the information processing methods described above. Therefore, the contents not described in detail here can be referred to the description of the previous corresponding part, and the detailed description thereof will be omitted here.

Furthermore, the series of processes and devices described above may be implemented by software and/or firmware. When realized by software and/or firmware, a program that constitutes the software is installed from a storage medium or network to a computer having a dedicated hardware structure, for example, a general-purpose personal computer 1000 shown in FIG. Various functions can be executed when various programs are installed.

Correspondingly, storage media carrying program products containing machine-executable instructions as described above are also included in the present disclosure. The storage medium may include, but is not limited to, floppy disks, optical disks, magnetic disks, memory cards, and the like.

Each constituent component, unit, etc. in the above-described apparatus may be configured in the form of software, firmware, hardware, or a combination thereof. Since specific means and methods used for the configuration are well known to those skilled in the art, detailed description thereof will be omitted here. When realized by software or firmware, a program that constitutes the software is installed from a storage medium or a network to a computer having a dedicated hardware structure (for example, a general-purpose computer 1000 shown in FIG. 10), and the computer installs various programs. When installed, various functions can be realized.

FIG. 10 is a block diagram of a hardware configuration (general-purpose computer) 1000 that can implement the methods and apparatus of embodiments of the present invention.

General purpose computer 1000 may be, for example, a computer system. It should be noted that general purpose computer 1000 is exemplary only and does not limit the scope or functionality of the methods and apparatus according to the present invention. Nor does general-purpose computer 1000 rely on any of the modules, assemblies, etc., or combinations thereof in the methods and apparatus described above.

In FIG. 10, a central processing unit (CPU) 1001 performs various processes based on programs stored in a ROM 1002 or programs loaded from a storage unit 1008 to a RAM 1003 . The RAM 1003 can also store data necessary for the CPU 1001 to perform various processes according to needs. The CPU 1001 , ROM 1002 and RAM 1003 are interconnected via a bus 1004 . Input/output interface 1005 is also connected to bus 1004 .

The input/output interface 1005 is further connected with the following components: an input unit 1006 including a keyboard, etc., an output unit 1007 including a display such as a liquid crystal display (LCD) and a speaker. , a storage unit 1008 including a hard disk, etc., and a communication unit 1009 including a network interface card such as a LAN card, a modem, and the like. A communication unit 1009 performs communication processing via a network such as the Internet or a LAN, for example. Drives 1010 may be connected to input/output interfaces 1005 as desired. A removable medium 1011 such as a semiconductor memory can be set in the drive 1010 as necessary, and a computer program read from the medium can be installed in the storage unit 1008 .

Additionally, the present invention further provides a program product including machine-readable instruction code. Such instruction codes, when read and executed by a machine, are capable of performing the methods in the embodiments of the present invention described above. Correspondingly, for carrying such program products, for example magnetic disks (including floppy disks), optical disks (including CD-ROMs and DVDs), magneto-optical disks (MD®) ), and various storage media such as semiconductor storage devices are also included in the present invention.

The storage medium described above may include, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor storage device, etc., but is not limited to these.

Each operation (process) in the above-described method can also be implemented in the form of a computer-executable program stored on various machine-readable storage media.

In addition, the above examples and the like are further disclosed as supplementary notes as follows.

(Appendix 1)
An information processing device,
A second text generation unit configured to generate, for each of one or more predetermined first texts, a plurality of second texts corresponding to the first text using a plurality of text generation models. ;
selecting a first predetermined number of text generation models from among the plurality of text generation models as candidate models based on the degree of semantic matching between the plurality of second texts and the corresponding first text; a composed candidate model selection unit;
For each of the one or more predetermined first texts, calculate text similarity between each of a plurality of second texts corresponding to the first text generated using the candidate model. and a similarity computation unit configured to compute model similarities between said candidate models based on text similarities between each other of second texts;
a target model selection unit configured to select, as target models, from among said candidate models a second predetermined number of candidate models having the lowest model similarity between each other; and, using said target models, pending processing. An apparatus comprising a fourth text generation unit configured to generate a second predetermined number of fourth texts corresponding to the text for subsequent processing.

(Appendix 2)
The information processing device according to Supplementary Note 1,
The text awaiting processing includes text input by the user or text obtained by converting speech sounds or images input by the user;
The information processing apparatus further includes a video timing positioning unit, which, based on the pending text and the second predetermined number of fourth texts, identifies from a given video the position of the frame corresponding to the pending text. A device configured to.

(Appendix 3)
The information processing device according to Supplementary Note 1,
the pending text includes training text;
The information processing device further includes:
a multimodal model training unit configured to train a multimodal model for video timing positioning based on the pending text and the second predetermined number of fourth texts to obtain a trained multimodal model; including, equipment.

(Appendix 4)
The information processing device according to appendix 2, further comprising:
a candidate text selection unit configured to select, as candidate texts, a plurality of fourth texts having a semantic matching degree with the awaiting text equal to or greater than a predetermined matching degree from among the second predetermined number of fourth texts; and a target text selection unit configured to select, as target texts, from among the candidate texts, a third predetermined number of candidate texts having the lowest text similarity between each other;
wherein said video timing positioning unit is further configured to identify, from said predetermined video, a position of a frame corresponding to said pending text based on said pending text and said target text.

(Appendix 5)
The information processing device according to any one of Appendices 1 to 4,
the first text, the second text and the fourth text belong to the same word class;
the second text generation unit is configured to generate the plurality of second texts using a back translation method;
The apparatus, wherein the fourth text generation unit is configured to generate the fourth text using a back-translation method.

(Appendix 6)
The information processing device according to any one of Appendices 1 to 4,
The similarity calculation unit is configured to calculate the text similarity between each other of the plurality of second texts in the following manner:
dividing each secondary text into a set of words and/or phrases; and obtaining and obtaining intersections and unions between the two sets of words and/or phrases corresponding to any two secondary texts. calculating the ratio of the number of words and phrases in the obtained intersection to the number of words and phrases in the obtained union as the text similarity between any two second texts; Device.

(Appendix 7)
The information processing device according to appendix 6,
The similarity computation unit is configured to compute model similarities between the candidate models in the following manner:
For each of the one or more predetermined first texts, obtain text similarity between second texts corresponding to the first text obtained by any two candidate models; and Apparatus for calculating an average value of text similarities corresponding to one or more predetermined first texts as a model similarity between any two candidate models mentioned above.

(Appendix 8)
The information processing device according to any one of Appendices 1 to 4,
The target model selection unit is configured to select the target model in the following manner:
constructing an N*N-dimensional matrix based on the model similarities between the candidate models each other, each element in the N*N-dimensional matrix representing the model similarity between corresponding candidate models, where N is representing said first predetermined number;
determining an M*M dimensional maximal determinant sub-matrix of said N*N dimensional matrix using a determinant point process, wherein M represents said second predetermined number; Apparatus for selecting a corresponding candidate model as said target model.

(Appendix 9)
The information processing device according to appendix 4,
The apparatus, wherein the video timing positioning unit is further configured to determine a final identification result based on a median value of the pending text-based identification result and the target text-based identification result.

(Appendix 10)
An information processing device,
a fifth text generation unit configured to generate a plurality of fifth texts corresponding to the pending text using the text generation model;
a candidate text selection unit configured to select, from among the plurality of fifth texts, a fifth text having a semantic matching degree with the awaiting text equal to or greater than a predetermined matching degree as a candidate fifth text;
a text similarity computation unit configured to compute, for each candidate fifth text, a text similarity between said candidate fifth text and each of the other candidate fifth texts; and said candidate fifth text. a target text selection unit configured to select, from among the texts, a fourth predetermined number of candidate fifth texts having the lowest text similarity between each other as target texts for subsequent processing; Device.

(Appendix 11)
The information processing device according to Appendix 10,
The text awaiting processing includes text input by the user or text obtained by converting speech sounds or images input by the user;
The information processing device further comprises a video timing positioning unit, which is configured to locate a frame corresponding to the pending text from a given video based on the pending text and the target text.

(Appendix 12)
The information processing device according to Appendix 10,
the pending text includes training text;
The information processing device further includes a multimodal model training unit, which trains a multimodal model for video timing positioning based on the pending text and the target text to obtain a trained multimodal model. Constructed, device.

(Appendix 13)
An information processing method,
for each of one or more predetermined first texts, generating a plurality of second texts corresponding to the first text using a plurality of text generation models;
selecting a first predetermined number of text generation models from among the plurality of text generation models as candidate models based on the degree of semantic match between the plurality of second texts and the corresponding first text;
For each of the one or more predetermined first texts, calculate text similarity between each of a plurality of second texts corresponding to the first text generated using the candidate model. and calculating the model similarity between the candidate models based on the text similarity between the second texts;
selecting a second predetermined number of candidate models having the lowest model similarity between each other as target models from the candidate models; and using the target models, a second predetermined number corresponding to pending text. A method comprising generating a fourth text for subsequent processing.

(Appendix 14)
The information processing method according to Appendix 13,
The text awaiting processing includes text input by the user or text obtained by converting speech sounds or images input by the user;
The information processing method further comprises recognizing, from a given video, a position of a frame corresponding to the pending text based on the pending text and the second predetermined number of fourth texts.

(Appendix 15)
The information processing method according to Appendix 13,
the pending text includes training text;
The information processing method further comprises training a multimodal model for video timing positioning based on the pending text and the second predetermined number of fourth texts to obtain a trained multimodal model. .

(Appendix 16)
The information processing method according to appendix 14, further comprising:
selecting, as candidate texts, a plurality of fourth texts having a degree of semantic matching with the text awaiting processing equal to or greater than a predetermined degree of matching, from among the second predetermined number of fourth texts; selecting as target text a third predetermined number of candidate texts having the lowest text similarity between
wherein recognizing a position of a frame corresponding to the pending text includes recognizing a position of a frame corresponding to the pending text from the predetermined video based on the pending text and the target text; Method.

(Appendix 17)
17. The information processing method according to any one of Appendices 13 to 16,
the first text, the second text and the fourth text belong to the same word class;
generating said plurality of second texts and said fourth texts using a back-translation method.

(Appendix 18)
17. The information processing method according to any one of Appendices 13 to 16,
calculating the text similarity between the plurality of second texts in the following manner:
dividing each secondary text into a set of words and/or phrases; and obtaining and obtaining intersections and unions between the two sets of words and/or phrases corresponding to any two secondary texts. calculating the ratio of the number of words and phrases in the obtained intersection to the number of words and phrases in the obtained union as the text similarity between any two second texts; Method.

(Appendix 19)
The information processing method according to Appendix 18,
Compute the model similarity between the candidate models in the following manner:
For each of the one or more predetermined first texts, obtain text similarity between second texts corresponding to the first text obtained by any two candidate models; and A method of calculating an average value of text similarities corresponding to one or more given first texts as the model similarity between any two candidate models mentioned above.

(Appendix 20)
The information processing method according to appendix 16,
determining a final identification result based on a median value of the pending text-based identification result and the target text-based identification result.

Although the preferred embodiment of the present invention has been described above, the present invention is not limited to this embodiment, and all modifications to the present invention fall within the technical scope of the present invention as long as they do not depart from the gist of the present invention.

Claims (10)

An information processing device,
A second text generation unit configured to generate, for each of one or more predetermined first texts, a plurality of second texts corresponding to the first text using a plurality of text generation models. ;
selecting a first predetermined number of text generation models from among the plurality of text generation models as candidate models based on the degree of semantic matching between the plurality of second texts and the corresponding first text; a composed candidate model selection unit;
For each of the one or more predetermined first texts, calculate text similarity between each of a plurality of second texts corresponding to the first text generated using the candidate model. , a similarity computation unit configured to compute model similarities between said candidate models based on text similarities between each other of second texts;
a target model selection unit configured to select, as target models, from among said candidate models a second predetermined number of candidate models having the lowest model similarity between each other; and, using said target models, pending processing. Information processing apparatus comprising a fourth text generation unit configured to generate a second predetermined number of fourth texts corresponding to the text for subsequent processing. The information processing device according to claim 1,
The text awaiting processing includes text input by the user or text obtained by converting speech sounds or images input by the user;
The information processing device further comprises a video timing positioning unit, the video timing positioning unit corresponding to the pending text from a predetermined video based on the pending text and the second predetermined number of fourth texts. An information processing device configured to identify the position of a frame. The information processing device according to claim 1,
The information processing apparatus further includes a multimodal model training unit, wherein the multimodal model training unit develops a multimodal model for video timing positioning based on the pending text and the second predetermined number of fourth texts. An information processor configured to train and obtain a trained multimodal model. The information processing device according to claim 2,
a candidate text selection unit configured to select, as candidate texts, a plurality of fourth texts having a semantic matching degree with the awaiting text equal to or greater than a predetermined matching degree from among the second predetermined number of fourth texts; and further comprising a target text selection unit configured to select, as target texts, a third predetermined number of candidate texts having the lowest text similarity between each other from among the candidate texts;
The information processing apparatus, wherein the video timing positioning unit is further configured to identify, from the predetermined video, a position of a frame corresponding to the pending text based on the pending text and the target text. The information processing device according to any one of claims 1 to 4,
the first text, the plurality of second texts and the fourth text belong to the same word class;
the second text generation unit is configured to generate the plurality of second texts using a back translation method;
Information processing apparatus, wherein the fourth text generation unit is configured to generate the fourth text using a back-translation method. The information processing device according to any one of claims 1 to 4,
The similarity calculation unit is
dividing each secondary text into a set of words and/or phrases; by calculating the ratio of the number of words and phrases contained in the obtained intersection to the number of words and phrases contained in the obtained union set as the text similarity between any two second texts , an information processing device configured to calculate a text similarity between said plurality of second texts. The information processing device according to claim 6,
The similarity calculation unit is
for each of the one or more predetermined first texts, obtaining text similarity between second texts corresponding to the first texts obtained by any two candidate models; model similarity between each other of said candidate models by calculating the mean value of the text similarities corresponding to one or more given first texts as the model similarity between any two candidate models mentioned above; An information processing device configured to calculate degrees. The information processing device according to any one of claims 1 to 4,
The target model selection unit comprises:
constructing an N*N-dimensional matrix based on the model similarities between the candidate models, each element in the N*N-dimensional matrix representing the model similarity between the corresponding candidate models, N being the representing a first predetermined number;
determining an M*M dimensional maximal determinant sub-matrix of said N*N dimensional matrix using a determinant point process, where M represents said second predetermined number; and corresponding to said M*M dimensional maximal determinant sub-matrix. An information processing apparatus configured to select the target model by selecting a candidate model as the target model. An information processing device,
a fifth text generation unit configured to generate a plurality of fifth texts corresponding to the pending text using the text generation model;
a candidate text selection unit configured to select, from among the plurality of fifth texts, a fifth text having a semantic matching degree with the awaiting text equal to or greater than a predetermined matching degree as a candidate fifth text;
a text similarity computation unit configured to compute, for each candidate fifth text, a text similarity between said candidate fifth text and each of the other candidate fifth texts; and said candidate fifth text. a target text selection unit configured to select, from among the texts, a fourth predetermined number of candidate fifth texts having the lowest text similarity between each other as target texts for subsequent processing; Information processing equipment. An information processing method,
for each of one or more predetermined first texts, generating a plurality of second texts corresponding to the first text using a plurality of text generation models;
selecting a first predetermined number of text generation models from among the plurality of text generation models as candidate models based on the degree of semantic match between the plurality of second texts and the corresponding first text;
For each of the one or more predetermined first texts, calculate text similarity between each of a plurality of second texts corresponding to the first text generated using the candidate model. , calculating the model similarity between the candidate models based on the text similarity between the second texts;
selecting a second predetermined number of candidate models having the lowest model similarity between each other as target models from the candidate models; and using the target models, a second predetermined number corresponding to pending text. A method of information processing, comprising generating a fourth text for subsequent processing.

Images