Summarization training method, summarization method, summarization training device, and program

Abstract

A computer implements a training process to train a machine learning model such that the output from the machine learning model, to which a text including a conversation and a text string indicating a viewpoint for summarization of the conversation have been inputted, approaches a summary sentence serving as a correct answer with regard to the text and the text string, thereby assisting with conversational organization that is matched to an intention of a user.

Description

Summary learning method, summary method, summary learning device and program

[0001] The present invention relates to a summarization learning method, a summarization method, a summarization learning device, and a program.

[0002] Contact centers and other facilities face the challenge of utilizing the large amounts of conversation history that accumulate daily to improve the quality of their agents' responses. To do this, the conversation history must be organized in a format that allows it to be utilized.

[0003] A technique for expressing and organizing a dialogue history in a tree structure has been disclosed (Non-Patent Document 1).

[0004] Ayako Hoshino, Hosomikaku, "Dialogue History Summarization Using Phrase Structure Analysis and Clustering," The 32nd Annual Conference of the Japanese Society for Artificial Intelligence, 2018

[0005] In Non-Patent Document 1, dialogues are organized according to clustering and phrase structure, but they are not necessarily organized in the format intended by the user, and there is a problem in that it is difficult to summarize and label long dialogues according to the user's intentions.

[0006] The present invention has been made in view of the above points, and has an object to support the organization of a dialogue in accordance with the user's intentions.

[0007] To solve the above problem, a computer executes a learning procedure to train a machine learning model to which text including a dialogue and a string indicating a viewpoint of a summary of the dialogue are input so that the output from the model approaches a summary sentence that is the correct answer for the text and the string.

[0008] It is possible to assist in organizing dialogue in accordance with the user's intentions.

[0009] FIG. 1 is a diagram showing an example of a hardware configuration of a summarization device 10 according to an embodiment of the present invention; FIG. 2 is a diagram showing an example of a functional configuration of the summarization device 10 according to an embodiment of the present invention; FIG. 3 is a diagram showing an example of dialogue data; FIG. 4 is a diagram showing an example of a source text; FIG. 5 is a diagram showing an example of a target text; FIG. 6 is a diagram for explaining learning of a summary model 12; and FIG. 7 is a diagram for explaining generation of a summary sentence using a trained summary model 12.

[0010] In this embodiment, a technology is disclosed in which, when summarizing text including a dialogue, a summary is generated in accordance with the viewpoint of the summary (i.e., the user's intention) by adding the viewpoint as a label.

[0011] An embodiment of the present invention will now be described with reference to the accompanying drawings. Fig. 1 shows an example of the hardware configuration of a summarizing device 10 according to an embodiment of the present invention. The summarizing device 10 shown in Fig. 1 includes a drive device 100, an auxiliary storage device 102, a memory device 103, a processor 104, and an interface device 105, all of which are interconnected via a bus B.

[0012] The program that realizes the processing in the summarization device 10 is provided by a recording medium 101 such as a CD-ROM. When the recording medium 101 storing the program is set in the drive device 100, the program is installed from the recording medium 101 to the auxiliary storage device 102 via the drive device 100. However, the program does not necessarily have to be installed from the recording medium 101, but may be downloaded from another computer via a network. The auxiliary storage device 102 stores the installed program as well as necessary files, data, etc.

[0013] When an instruction to start the program is received, the memory device 103 reads and stores the program from the auxiliary storage device 102. The processor 104 is a CPU or a GPU (Graphics Processing Unit), or a CPU and a GPU, and executes functions related to the summarization device 10 in accordance with the program stored in the memory device 103. The interface device 105 is used as an interface for connecting to a network.

[0014] Fig. 2 is a diagram showing an example of the functional configuration of a summarizing device 10 according to an embodiment of the present invention. In Fig. 2, the summarizing device 10 includes a learning unit 11, a summary model 12, and a summary sentence generation unit 13. Each of these units is realized by a process executed by a processor 104 in accordance with one or more programs installed in the summarizing device 10. The summarizing device 10 also uses a training data storage unit 14. The training data storage unit 14 can be realized using, for example, an auxiliary storage device 102 or a storage device connectable to the summarizing device 10 via a network.

[0015] The summary model 12 is a machine learning model (e.g., a neural network) that receives source text generated from dialogue data and document labels as input and outputs a summary of the source text according to the document labels.

[0016] Dialogue data refers to data in which a dialogue between multiple speakers is transcribed into text. FIG. 3 is a diagram illustrating an example of dialogue data. FIG. 3 illustrates an example of dialogue data related to a dialogue between an operator and a customer at a contact center. As illustrated in FIG. 3, the dialogue data includes, for each utterance, identification information of the speaker of the utterance (in the example of FIG. 3, "operator" or "customer") and text indicating the content of the utterance. The speaker identification information is hereinafter referred to as a "speaker ID." The dialogue data may be data generated using voice recognition from audio data of a recorded dialogue, or may be data transcribed by a human. The dialogue data may also be data in which the content of a dialogue conducted in text, such as in a chat session, is recorded. The speaker information in the dialogue data may be any information that can distinguish the speaker of each utterance. The speaker information may be a name, a chat user name, or the like.

[0017] The source text is text data obtained by removing speaker information from dialogue data. In other words, the source text is text that includes dialogue and is the target of summarization. Figure 4 shows an example of source text generated based on the dialogue data of Figure 3.

[0018] A document label is a character string that indicates the perspective of a dialogue summary (a character string for organizing the summary). For example, if you want to generate a summary about a question from a dialogue, the document label may be "question," and if you want to generate a summary about an answer from a dialogue, the document label may be "answer." Note that the label only needs to distinguish the perspective of the summary, and the character string used as the label can be determined arbitrarily.

[0019] The training data storage unit 14 stores a set of training data used for training the summary model 12. A single training data set is a set of three pieces of data: a source text, a document label, and a target text.

[0020] The target text of a certain training data refers to text data that indicates the correct summary of the source text of the training data according to (corresponding to) the document label of the training data. In other words, the target text is a summary to be generated according to the source text and document label. For example, if the document label is "question," a summary focusing on the question portion of the source text (a summary of (only) the part of the source text that relates to the question) is created in advance as the target text. An example of the source text in Figure 4 and the target text for the document label "question" is shown in Figure 5.

[0021] The learning unit 11 uses the learning data stored in the learning data storage unit 14 to learn the summary model 12. Learning the summary model 12 means updating the parameters of the summary model 12.

[0022] Fig. 6 is a diagram illustrating the training of the summary model 12. As shown in Fig. 6, for each training data, the training unit 11 trains the summary model 12 (updates the parameters of the summary model 12) so that text data based on the output from the summary model 12 to which the source text and document label of the training data have been inputted approaches the target text of the training data.

[0023] The learning method may be a learning method for generating summaries, such as the learning method disclosed in Japanese Patent Laid-Open No. 2020-140629.

[0024] As a result of the learning, the summary model 12 becomes able to generate an output sentence (summary sentence) according to the document label when the document label and the source text are given at the same time.

[0025] The summary generation unit 13 uses the trained summary model 12 to generate a summary corresponding to the input source text and document label.

[0026] 7 is a diagram illustrating generation of a summary using the trained summary model 12. As shown in FIG. 7, the summary generation unit 13 inputs, for example, source text and document labels input by a user into the trained summary model 12. The summary generation unit 13 acquires a summary output by the summary model 12 in response to this input, and outputs the summary to the user (for example, displays it on a display device). Note that the document label may be input by having the user directly input a character string as the document label, or the summary generation unit 13 may present a choice of document labels and have the user select a document label from the choices.

[0027] Alternatively, the user may be allowed to input two document labels ("question" and "answer" in this embodiment). In this case, for each input document label, the summary generation unit 13 inputs the label and the source text into the summary model 12 to obtain a summary. The summary generation unit 13 outputs the pair of summary sentences obtained for each document label.

[0028] As described above, according to this embodiment, the summary model 12 receives text containing a dialogue and a character string indicating a viewpoint for summarizing the dialogue as input, and can generate a summary of the text according to the viewpoint. Therefore, a summary that reflects the user's desired intent can be generated, and the organization of the dialogue in line with the user's intent can be supported. Specifically, each summary can be used as knowledge directly because the content of the dialogue is organized by viewpoint. For example, if a summary is generated by dividing it into questions and answers, the questions and answers can be used as a set of knowledge.

[0029] In this embodiment, the summarization device 10 is an example of a summarization learning device.

[0030] Although the embodiments of the present invention have been described in detail above, the present invention is not limited to such specific embodiments, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims.

[0031] REFERENCE SIGNS LIST 10 Summarization device 11 Learning unit 12 Summary model 13 Summary sentence generation unit 14 Learning data storage unit 100 Drive device 101 Recording medium 102 Auxiliary storage device 103 Memory device 104 Processor 105 Interface device B Bus

Claims

1. A summary learning method, characterized by being executed by a computer, comprising the steps of: a learning procedure for training a machine learning model to which text including a dialogue and a character string indicating a viewpoint of a summary of the dialogue are input, so that the output from the machine learning model approaches a summary sentence that is a correct answer for the text and the character string.

2. A summarization method characterized by being executed by a computer: a summarization procedure for generating a summary sentence corresponding to input text and a character string indicating a viewpoint of summarizing the dialogue contained in the text, using a machine learning model trained so that when input text including a dialogue and a character string indicating a viewpoint of summarizing the dialogue are used, the output approaches a summary sentence that is a correct answer for the text and the character string.

3. A summary learning device comprising: a learning unit configured to learn a machine learning model to which text including a dialogue and a character string indicating a viewpoint of a summary of the dialogue are input so that the output from the machine learning model approaches a summary sentence that is a correct answer for the text and the character string.

4. A program for causing a computer to execute a learning procedure for training a machine learning model to which text including a dialogue and a character string indicating a viewpoint of a summary of the dialogue are input so that the output from the machine learning model approaches a summary sentence that is a correct answer for the text and the character string.

Images