Summary generation method, summary generation system, and computer program
The summary generation system uses a dual language model approach to efficiently generate concise summaries, addressing inefficiencies in existing models by leveraging generative methods to handle large or repetitive text data.
Patent Information
- Application Number
- JP2023125438
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2023-05-10
- Filing Date
- 2023-08-01
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2043-08-01
AI Technical Summary
Existing language models are not utilized effectively to generate summaries efficiently, particularly when dealing with large amounts of text data or repetitive content.
A summary generation system utilizing both a front-end and back-end language model, where the front-end model processes text data using a generative method to reduce character count and the back-end model generates a summary, expanding the applicability of the back-end model and improving summary generation versatility.
The system effectively generates concise summaries by leveraging both models, enhancing the performance of the back-end model and handling large or repetitive text data more efficiently.
Smart Images

Figure 0007588689000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a method for generating a summary, and in particular to a method for generating a summary applied to text data.The present invention further relates to a system and a computer program for generating a summary applied to text data. [Background technology]
[0002] Natural language is the main language that humans use to express their thoughts when communicating. Therefore, natural language processing has a wide range of applications, including speech recognition, machine translation, text classification, question answering, and other specific applications. With the development of machine learning technology, natural language processing has become a popular research direction in the field of artificial intelligence. Therefore, in recent years, not only have the types of language models become more abundant, but their performance is also continuously improved. However, the performance of a language model is closely related to how it is used, so how to use language models more effectively and improve their usage efficiency has become a new research direction. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] China Patent Application Publication No. 114647720 [Patent Document 2] China Patent Application Publication No. 115577096 Summary of the Invention [Problem to be solved by the invention]
[0004] SUMMARY OF THE PRESENT DISCLOSURE It is therefore an object of the present invention to provide a method and system for generating a summary that utilizes a language model more effectively, and a computer program therefor. [Means for solving the problem]
[0005] The method is performed by a summary generation system electrically connected to a user device, the summary generation system storing a front-end language model and a back-end language model realized by machine learning techniques. The summary generating method includes the steps of: A) performing token splitting on target text data and determining whether the number of split tokens is equal to or greater than a predetermined threshold; B) if it is determined that the number of split tokens is equal to or greater than the predetermined threshold, inputting the target text data into a front-end language model to obtain pre-processed text data expressed in a natural language by a generative technique, the number of characters being less than the number of characters in the target text data, and then inputting the pre-processed text data into a back-end language model to obtain a first summary result expressed in a natural language by a generative technique, the number of characters being less than the number of characters in the pre-processed text data, and transmitting the first summary result to a user-side device for output; and C) if it is determined that the number of split tokens is not equal to or greater than the predetermined threshold, inputting the target text data into the back-end language model to obtain a second summary result expressed in a natural language by a generative technique, the number of characters being less than the number of characters in the target text data, and transmitting the second summary result to a user-side device for output. The summary generating system includes a processing unit electrically connected to the user device, and a storage unit electrically connected to the processing unit. The storage unit stores a front-end language model and a back-end language model that are realized by machine learning techniques. The processing unit is configured to perform token splitting on the target text data, determine whether the number of split tokens is equal to or greater than a predetermined threshold, and if it is determined that the number of split tokens is equal to or greater than the predetermined threshold, input the target text data to a front-end language model, obtain pre-processed text data by a generative technique, which is expressed in a natural language and has a number of characters less than the number of characters in the target text data, and then input the pre-processed text data to a back-end language model, obtain a first summary result by a generative technique, which is expressed in a natural language and has a number of characters less than the number of characters in the pre-processed text data, and send the first summary result to a user-side device for output; and if it is determined that the number of split tokens is not equal to or greater than the predetermined threshold, input the target text data to the back-end language model, obtain a second summary result by a generative technique, which is expressed in a natural language and has a number of characters less than the number of characters in the target text data, and send the second summary result to the user-side device for output. The computer program, when executed by a computer system, causes the computer system to perform the summary generation method described above using a front-end language model and a back-end language model implemented by machine learning. Effect of the Invention
[0006] By executing the summary generation method according to the present invention, when the number of divided tokens of the target text data is determined to be equal to or greater than a predetermined threshold (i.e., the number of characters in the target text data is relatively large), the summary generation system inputs the target text data into a front-end language model to obtain pre-processed text data, and then inputs the input text data into a back-end language model to obtain a summary result. In this way, when the back-end language model has a limit on the number of input characters, the present invention can widen the application range of the back-end language model and provide a more versatile summary generation function. In addition, since the front-end language model generates pre-processed text data by a generative method, when the target text data contains a lot of repetitive content, it is possible to summarize the content of the target text data more effectively than by an extractive method, and generate pre-processed text data with high information density to be input to the back-end language model. Other features and advantages of the present invention will become apparent in the following detailed description of the embodiments which proceeds with reference to the accompanying drawings. [Brief description of the drawings]
[0007] [Figure 1] 1 is a block diagram showing an example of a summary generation system according to an embodiment of the present invention and a user device to which the system is applied; [Diagram 2] 2 is a flowchart illustrating an example of a summary generating method according to the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0008] Before describing the present invention in more detail, unless otherwise specified, the term "electrically connect" in this specification is used to describe a coupling relationship between computer hardware (e.g., electronic systems, equipment, devices, units, parts, etc.), and refers to a "wired electrical connection" in which multiple computer hardware pieces are physically connected via conductor or semiconductor materials, or a "wireless electrical connection" in which wireless data transmission is realized using wireless communication technology (e.g., wireless network, Bluetooth (registered trademark), electrical induction, etc.). Meanwhile, unless otherwise specified, the term "electrically connect" in this specification further refers to a "direct electrical connection" in which multiple computer hardware pieces are directly connected to each other, or an "indirect electrical connection" in which multiple computer hardware pieces are connected to each other via other computer hardware.
[0009] The summary generation system 1 of the present invention is configured to be electrically connected to multiple user side devices 5 via a network. The user side devices 5 are smartphones, tablet computers, notebook computers, or desktop computers used by users. For convenience of explanation, the following describes the summary generation system 1 electrically connected to only one user side device 5 as shown in FIG. 1.
[0010] In this embodiment, the summary generation system 1 is a server equipment, and includes a processing unit 11 and a storage unit 12 electrically connected to the processing unit 11. The processing unit 11 is electrically connected to a user side device 5 via a network.
[0011] More specifically, in this embodiment, the processing unit 11 is a processor realized by an integrated circuit and has the functions of command transmission / reception and data calculation, and the storage unit 12 is a data storage device (e.g., a hard disk, a hard disk array, or other types of computer-readable storage media) that stores digital data. In a similar embodiment, the processing unit 11 may be a processing circuit having a processor, and the storage unit 12 may be a collection of multiple storage devices of the same or different types. Furthermore, in another embodiment, the summary generation system 1 may be multiple server facilities electrically connected to each other, in which case the processing unit 11 is a collection of processors or processing circuits that each of the multiple server facilities has, and the storage unit 12 is a collection of data storage devices that each of the multiple server facilities has. Therefore, the computer hardware realization of the summary generation system 1 is not limited to this embodiment.
[0012] The storage unit 12 includes a speech processing model M0, a front-end language model LM1, and a back-end language model LM2.
[0013] The voice processing model M0 generates text data corresponding to the voice data to be output for the voice data to be input. More specifically, it is determined whether the voice data belongs to a single speaker type representing the voice of a single speaker or a multiple speaker type representing the voices of multiple speakers, and when it is determined that the voice data belongs to the multiple speaker type, it further recognizes the speaker using speaker separation based on voiceprint recognition, separates the voice of the voice data for each speaker, and then generates corresponding text data based on the voice represented by the voice data using speech-to-text. Therefore, when the voice data belongs to the multiple speaker type, the output text data includes multiple speech parts each corresponding to one of the multiple speakers and indicating the content of one utterance of the corresponding speaker, and can indicate the order of utterance of each speaker and the content of each utterance. In addition, the voice processing model M0 is realized by machine learning using voice data, which is, for example, a recording file of the single speaker type and the multiple speaker type, as training data. Note that the training of the voice processing model M0 can be realized by conventional technology and is not the point of the present invention, so it will not be described in detail.
[0014] The front-end language model LM1 generates output text data from input text data using a generative method that utilizes natural language processing. The output text data is expressed in a natural language and is a summary of the input text data with fewer characters than the input text data.
[0015] The back-end language model LM2 generates output text data from input text data using a generative method that uses natural language processing. The output text data is a summary of the input text data, expressed in natural language, with fewer characters than the input text data.
[0016] The front-end language model LM1 and the back-end language model LM2 are pre-trained language models realized by machine learning, using text data, which is a character record of a sentence or a conversation, as training data. More specifically, in this embodiment, the front-end language model LM1 is preferably BLOOMZ, but may be a pre-trained language model capable of generating text expressed in a natural language, such as BLOOM, MT0, GPT-2, or T5. Meanwhile, in this embodiment, the back-end language model LM2 is preferably GPT-3, but may be a pre-trained language model capable of generating text expressed in a natural language, such as GPT-4, GPT-3.5, or GPT-2.
[0017] In this specification, "generative" (also called "abstractive") means that a language model generates output text data based on input text data using natural language generation technology. As a person having ordinary knowledge in the technical field to which the present invention belongs knows, "generative" means that a language model understands input text data and then generates output text data of a new document using natural language generation technology, so that the output text data may contain expressions that are not included in the input text data. For example, the output text data may contain words or sentences that are not included in the input text data, or may express the contents of the input text data more concisely, or may summarize the contents of the input text data in bullet points or in a table. As a result, the "generative" method in this specification is different from the "extractive" method in which words and sentences included in the input text data are extracted and combined to produce output text data.
[0018] Referring to FIG. 2, a summary generation method executed by the summary generation system 1 of the present embodiment is shown.
[0019] In step S1, the processing unit 11 receives voice data from the user-side device 5. Specifically, for example, the user-side device 5 transmits the voice data to the processing unit 11 via a network in response to a user's operation, in other words, uploads the voice data to the summary generation system 1. The voice data may be, for example, the voice portion of a video such as a speech or an interview, or a recording file of a monologue or a meeting involving multiple participants.
[0020] In step S2, the processing unit 11 inputs the received voice data into the voice processing model M0 to obtain corresponding text data. More specifically, the processing unit 11 uses the voice processing model M0 to determine whether the received voice data belongs to a single speaker type or a multi-speaker type, and then generates corresponding text data. The text data is, for example, a verbatim transcript corresponding to the recording file input into the voice processing model M0.
[0021] In step S3, the processing unit 11 performs tokenization on the text data obtained in step S2 (hereinafter referred to as target text data) to obtain a plurality of divided tokens. In this embodiment, each divided token is one character or a combination of multiple characters divided from the target text data, that is, a token. In addition, the processing unit 11 performs tokenization based on a token table pre-stored in the storage unit 12. For example, in this embodiment, the processing unit 11 divides the text data of "natural language" into two tokens, "natural" and "language", based on the token table. In other embodiments, the processing unit 11 may divide the text data of "natural language" into four tokens, "self", "natural", "language", and "language", and is not limited to this embodiment.
[0022] In step S4, the processing unit 11 determines whether the number of divided tokens is equal to or greater than a predetermined threshold. The predetermined threshold may be set to, for example, 2000, but can be freely set and adjusted according to actual situations and needs, and is not limited to a fixed value. If it is determined that the number of divided tokens is equal to or greater than the predetermined value, the flow proceeds to step S5, and if it is determined that the number of divided tokens is smaller than the predetermined value, the flow proceeds to step S8.
[0023] If it is determined that the number of divided tokens is equal to or greater than the predetermined threshold, this indicates that the number of characters in the target text data is relatively large, so in step S5, the processing unit 11 inputs the target text data and the soft prompts into the front-end language model LM1 to obtain output pre-processed text data.
[0024] The soft prompt is predicted in advance by a language model (for example, but not limited to, the front-end language model LM1) using a prompt learning technique of prompt engineering. The prompt learning may be, but is not limited to, prefix tuning, tuning initialized with discrete prompts, or hard-soft prompt hybrid tuning. The soft prompt (also called "continuous prompt") is expressed by a vector or a numerical value in other non-natural languages. The front-end language model LM1 uses the soft prompt to understand the meaning of the input text data, and then uses its own attention (attention mechanism) to generate output text data for parts of the input text data that are highly relevant to the context (i.e., ignoring parts that are less relevant or irrelevant). The attention is realized by the front-end language model LM1 using the steepest descent method in the training phase. Also, attention is a conventional technique, so it will not be described in detail in this specification.
[0025] In addition, when the voice data corresponding to the target text data belongs to a single speaker type (the target text data does not include multiple speech parts), the input soft prompt is a first soft prompt corresponding to the single speaker type, so that the front-end language model LM1 generates pre-processed text data according to the single speaker type. On the other hand, when the voice data corresponding to the target text data belongs to a multiple speaker type (the target text data includes multiple speech parts), the input soft prompt is a second soft prompt corresponding to the multiple speaker type and different from the first soft prompt, so that the front-end language model LM1 generates pre-processed text data according to the multiple speaker type. In addition, unlike hard prompts, soft prompts can effectively avoid a situation in which the output of a language model differs greatly due to a slight difference in input. In other words, soft prompts can improve the stability and reliability of a language model. Therefore, it is better to use different soft prompts according to text data corresponding to single speaker type or multiple speaker type voice data than to use different hard prompts.
[0026] The front-end language model LM1 ignores parts of the target text data that are less relevant to the context and generates pre-processed text data for parts that are highly relevant, so that when the number of characters in the target text data is relatively large, the effect of simplifying the target text data can be obtained. In addition, since the front-end language model LM1 generates pre-processed text data by a generative method, when the target text data contains a lot of repetitive content, the present invention can better summarize the content of the target text data and generate pre-processed text data with high information density compared to an extractive method.
[0027] In step S6, the processing unit 11 inputs the pre-processed text data and the hard prompts into the back-end language model LM2 to obtain an output first summary result. The hard prompts are preset and stored in the summary generation system 1. The back-end language model LM2 uses the hard prompts to generate more fluent text data, which has better human readability than the text data generated by the front-end language model LM1.
[0028] In step S7, the processing unit 11 transmits the first summary result to the user-side device 5, and causes the user-side device 5 to display and output the first summary result.
[0029] If it is determined that the number of divided tokens is not greater than the predetermined threshold, in step S8, the processing unit 11 inputs the target text data and the hard prompt into the back-end language model LM2 to obtain a second summary result as output.
[0030] In step S9, the processing unit 11 transmits the second summary result to the user-side device 5, and causes the user-side device 5 to display and output the second summary result.
[0031] That is, when the number of characters in the target text data is relatively small, the processing unit 11 does not use the front-end language model LM1, but uses only the back-end language model LM2.
[0032] It should be understood that steps S1 to S9 are merely illustrative of the summary generation method of the present invention. If steps S1 to S9 are combined, divided, or the order is changed, and the same effect can be obtained in a manner substantially the same as that of this embodiment, this corresponds to an embodiment of the summary generation method of the present invention and should be included in the scope of the present invention. Therefore, steps S1 to S9 do not limit the present invention.
[0033] In this embodiment, the computer program includes a front-end language model LM1, a back-end language model LM2, and a speech processing model M0. When the computer program of this embodiment is executed by a computer system (e.g., one computer device or a server device, or a combination of multiple computer devices or server devices), the computer system becomes a summary generation system 1, and causes the computer system to execute a summary generation method using the front-end language model LM1, the back-end language model LM2, and the speech processing model M0. In another embodiment, the front-end language model LM1, the back-end language model LM2, and the speech processing model M0 may be stored in a remote server. When the computer program is executed by a computer system, the computer system accesses the front-end language model LM1, the back-end language model LM2, and the speech processing model M0 via a network.
[0034] In summary, by executing the summary generation method of the present invention, when the number of divided tokens of the target text data is determined to be equal to or greater than a predetermined threshold (i.e., the number of characters in the target text data is relatively large), the summary generation system 1 processes the target text data using the front-end language model LM1 and then obtains a summary result using the back-end language model LM2. In this way, when the back-end language model LM2 is limited in the number of input characters, the present invention can widen the application range of the back-end language model LM2 and provide a more versatile summary generation function. In addition, since the front-end language model LM1 generates pre-processed text data using a generative method, when the target text data contains a lot of repetitive content, it can summarize the content of the target text data using an extractive method, and generate pre-processed text data with high information density to be input to the back-end language model LM2. Thus, the present invention uses a language model that utilizes two generative methods to realize a more versatile summary generation system 1, improve the performance of the back-end language model LM2, and reliably achieve the object of the present invention.
[0035] In the above description, for the purpose of explanation, numerous specific details are set forth to provide a thorough understanding of the embodiments. However, it will be apparent to one skilled in the art that one or more other embodiments may be implemented without the specific details. In addition, in the description of "one embodiment" or "one embodiment" in this specification, it should be understood that all descriptions accompanied by ordinal numbers or other indications may be included in the specific implementation of the present invention having specific aspects, structures, and features. Furthermore, in this specification, multiple variations are sometimes incorporated into one embodiment, drawing, or description thereof, but this is for the purpose of streamlining the specification and for the purpose of understanding the multiple aspects of the present invention, and one or more features or specific examples in one embodiment may be implemented together with one or more features or specific examples in other embodiments, where appropriate, in the implementation of the present invention.
[0036] Although the embodiments and modified examples of the present invention have been described above, the present invention is not limited to these, and is intended to encompass all modifications and equivalent configurations as various configurations falling within the spirit and scope of the broadest interpretation. [Explanation of symbols]
[0037] 1 Summary Generation System 11 Processing Unit 12 Storage Unit M0 Audio Processing Model LM1 Front-end Language Model LM2 Backend Language Models 5 User side equipment S1-S9 Steps
Claims
1. A method for generating a summary executed by a summary generation system electrically connected to a user-side device, the method comprising the steps of: A) performing token division on target text data and determining whether the number of divided tokens is equal to or greater than a predetermined threshold; B) if it is determined that the number of the divided tokens is equal to or greater than the predetermined threshold, inputting the target text data into the front-end language model to obtain pre-processed text data, which is expressed in a natural language by a generative technique and has a number of characters less than that of the target text data, and then inputting the pre-processed text data into the back-end language model to obtain a first summary result, which is expressed in a natural language by a generative technique and has a number of characters less than that of the pre-processed text data, and transmitting the first summary result to the user-side device for output; C) if it is determined that the number of the divided tokens is not equal to or greater than the predetermined threshold, inputting the target text data into the back-end language model, and obtaining a second summary result by a generative method, the second summary result being expressed in a natural language and having a number of characters less than the number of characters in the target text data, and transmitting the second summary result to the user-side device for output. Summary generation method.
2. The method of claim 1 , wherein in step B), the summary generation system further inputs soft prompts predicted by a language model together with the target text data into the front-end language model.
3. The method of claim 2 , further comprising, prior to step A), the step of: D) receiving voice data from the user-side device and generating text data of the target based on the voice represented by the voice data.
4. In step D), it is further determined whether the speech data belongs to a single speaker type, which represents the speech of a single speaker, or a multi-speaker type, which represents the speech of multiple speakers; When the voice data is determined to belong to the multiple speaker type, the generated target text data includes a plurality of speech portions each corresponding to one of the plurality of speakers; 4. The summary generating method of claim 3, wherein in step B), the soft prompt input to the front-end language model is a first soft prompt corresponding to the single-speaker type when the voice data belongs to the single-speaker type, and is a second soft prompt corresponding to the multiple-speaker type and different from the first soft prompt when the voice data belongs to the multiple-speaker type.
5. a processing unit electrically connected to a user side device; a storage unit electrically connected to the processing unit; The storage unit stores a front-end language model and a back-end language model realized by machine learning technology; The processing unit includes: Tokenizing the target text data and determining whether the number of tokens obtained by the tokenizing is equal to or greater than a predetermined threshold; if it is determined that the number of divided tokens is equal to or greater than the predetermined threshold, input the target text data into the front-end language model to obtain pre-processed text data, which is expressed in a natural language by a generative technique and has a number of characters less than the number of characters in the target text data, and then input the pre-processed text data into the back-end language model to obtain a first summary result, which is expressed in a natural language by a generative technique and has a number of characters less than the number of characters in the pre-processed text data, and transmit the first summary result to the user-side device for output; and if it is determined that the number of the divided tokens is not equal to or greater than the predetermined threshold, inputting the target text data into the back-end language model, and obtaining a second summary result by a generative method, the second summary result being expressed in a natural language and having a number of characters less than the number of characters in the target text data, and transmitting the second summary result to the user-side device for output. Summary generation system.
6. The summary generation system of claim 5 , wherein the processing unit further inputs soft prompts predicted by a language model together with the target text data into the front-end language model.
7. The summary generation system of claim 6 , wherein the processing unit is further configured to receive voice data from the user-side device and generate text data of the subject based on the speech represented by the voice data.
8. The processing unit is further configured to determine whether the speech data belongs to a single speaker type, which represents a speech of a single speaker, or a multi-speaker type, which represents a speech of multiple speakers; When the voice data is determined to belong to the multiple speaker type, the generated target text data includes a plurality of speech portions each corresponding to one of the plurality of speakers; 8. The summary generation system of claim 7, wherein the soft prompt input to the front-end language model is a first soft prompt corresponding to the single-speaker type when the voice data belongs to the single-speaker type, and is a second soft prompt corresponding to the multiple-speaker type and different from the first soft prompt when the voice data belongs to the multiple-speaker type.
9. A computer program product which, when executed by a computer system, causes the computer system to perform the method of claim 1 using a front-end language model and a back-end language model implemented by machine learning.
Citation Information
Patent Citations
Abstract generation method and device
CN114647720A
Text abstract generation method, system and equipment based on pre-training model and medium
CN115577096A
Knowledge question and answer processing method and device
CN118227771A
Document outputting device and its control method
JP2005258676A
Generation apparatus, generation method, and generation program
JP2018005324A
