Machine reading comprehension method and machine reading comprehension system

Abstract

To provide a machine reading comprehension method.SOLUTION: A machine reading comprehension method comprises: a step of establishing a machine reading comprehension model based on training data and a plurality of first mask problems; a step of inputting a plurality of second mask problems into the machine reading comprehension model to train the machine reading comprehension model; a step of generating a dialogue generation model according to the machine reading comprehension model; and a step of generating a dialogue according to the dialogue generation model.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present application relates to a machine reading method and a machine reading system. In particular, the present application relates to a machine reading method and a machine reading system using mask learning. In this specification, "mask learning" refers to learning by masking phrases in a sentence, and a "mask problem" refers to a sentence in which part of the text is masked. [Background technology]

[0002] Machine reading comprehension (MRC) is a technology that allows a computer to read text and solve related problems. In recent years, a large amount of text data has been generated in various industries. The traditional manual processing methods face problems such as slow processing speed, large overhead, and inability to fully cover question-and-answer pairs, so the demand for machine reading comprehension is gradually increasing. How to effectively read and understand large amounts of text data is one of the problems that need to be solved in this field. Summary of the Invention [Means for solving the problem]

[0003] According to one aspect of the present disclosure, there is provided a machine reading comprehension method, comprising: a step of establishing a machine reading comprehension model based on training data and a plurality of first mask problems; a step of inputting a plurality of second mask problems into the machine reading comprehension model to train the machine reading comprehension model; a step of generating a dialogue generation model according to the machine reading comprehension model; and a step of generating a dialogue according to the dialogue generation model.

[0004] Another aspect of the present disclosure provides a machine reading comprehension system including a machine reading comprehension circuit for establishing a machine reading comprehension model based on training data and a plurality of first mask problems, and inputting a plurality of second mask problems into the machine reading comprehension model to train the machine reading comprehension model, a dialogue generation circuit coupled to the machine reading comprehension circuit for generating a dialogue generation model according to the machine reading comprehension model and for generating a dialogue according to the dialogue generation model, and a memory coupled to the machine reading comprehension circuit and the dialogue generation circuit for storing the machine reading comprehension model and the dialogue generation model. [Brief description of the drawings]

[0005] The contents of this application can be better understood with reference to the embodiments in the following paragraphs and the following drawings. [Figure 1] 1 is a schematic diagram of a machine reading system according to some embodiments of the present application. [Diagram 2] 1 is a flowchart of a machine comprehension method according to some embodiments of the present application. [Diagram 3] 3 is a flowchart of one step in FIG. 2 according to some embodiments of the present application. [Figure 4] 3 is a flowchart of one step in FIG. 2 according to some embodiments of the present application. [Diagram 5] 3 is a flowchart of one step in FIG. 2 according to some embodiments of the present application. [Figure 6] FIG. 2 is a schematic diagram of a creation policy according to some embodiments of the present application. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0006] The spirit of the present application will be clearly explained below with reference to the drawings and detailed description. After understanding the embodiments of the present application, anyone skilled in the art can make changes and modifications to the technology taught by the present application without departing from the spirit and scope of the present application.

[0007] The terminology used herein is for the purpose of describing particular embodiments only and is not meant to be limiting of the present application. As used herein, the singular forms "one," "this," "this," and "the" include the plural forms as well.

[0008] As used herein, "including," "comprising," "having," "containing," and the like are all open-ended terms, meaning including but not limited to.

[0009] For the terms used in this specification, unless otherwise specified, each term generally has its common meaning as used in the art and in the context of this application and in the particular context. Some of the terms used to describe this application are discussed below or elsewhere in this specification to provide those skilled in the art with further guidance in describing this application.

[0010] 1 is a schematic diagram of a machine reading system 100 according to some embodiments of the present application. In some embodiments, referring to FIG 1, the machine reading system 100 includes a machine reading circuit 110, a dialogue generating circuit 130, and a memory 150. In some embodiments, the machine reading circuit 110, the dialogue generating circuit 130, and the memory 150 are coupled to each other.

[0011] Referring to Fig. 2, in order to better understand the present disclosure, the detailed operation of the machine reading system 100 will be described in combination with the embodiment shown in Fig. 2. Fig. 2 is a flowchart of a machine reading method 200 according to some embodiments of the present application. The machine reading method 200 can be applied to a machine reading system 100 having the same or similar structure as shown in Fig. 1, but the present disclosure is not limited to the application of the embodiment in Fig. 1.

[0012] As shown in FIG. 2, the machine-readable method 200 includes steps S210 to S270.

[0013] In step S210, a machine reading comprehension model is established based on the training data and the first plurality of mask problems. In some embodiments, step S210 is performed by the machine reading comprehension circuitry 110 in FIG. 1. In some embodiments, the machine reading comprehension model is stored in memory 150 in FIG. 1.

[0014] Please refer to Figure 3. Figure 3 is a flow chart of step S210 according to some embodiments of the present application. As shown in Figure 3, step S210 includes step S212 and step S214.

[0015] In step S212, at least one character in the training data is randomly masked to generate a first mask problem. In some embodiments, the training data is text data that includes many characters, and the at least one character may be a single word or a sentence. In some embodiments, the machine reading circuit 110 masks the sentence based on punctuation.

[0016] For example, in some embodiments, assuming that the text data includes the character content "Good morning, I want to go to eat breakfast, what do you want to eat?", the machine reading comprehension circuitry 110 masks the characters between the punctuation comma and the punctuation period of the sentence based on the punctuation comma and the question mark to generate a masked question "Good morning, I want to go to eat breakfast, what do you want to eat, [mask]?". In another embodiment, the machine reading comprehension circuitry 110 masks only single words to generate a masked question, such as "Good morning, I want to go to eat [mask], what do you want to eat?".

[0017] In some embodiments, the machine comprehension circuitry 110 is further used to convert the characters in the training data into codes and then mask them in the form of codes to generate masked questions, such as "1.2, 3.5, 1.5, [mask]?"

[0018] In step S214, the training data is searched to obtain a first answer corresponding to the first mask problem. For example, for the mask problem “Good morning, I want to go eat [mask], what do you want to eat?”, the machine reading comprehension circuit 110 obtains “breakfast” as the answer after searching the text content.

[0019] In some embodiments, the machine reading comprehension circuit 110 establishes a mask problem and finds a corresponding answer according to the above steps S212 and S214, and establishes a machine reading comprehension model based on the mask problem and the corresponding answer.

[0020] In some embodiments, the machine-reading model may include a look-up table and a classification model based on a neural network. The classification model based on a neural network includes a plurality of nodes and weights of connection lines between the nodes, and each node corresponds to one mask problem or one answer. The machine-reading circuit 110 establishes a mask problem through the classification model based on a neural network and searches for a corresponding answer in the training data, while establishing a plurality of nodes and weights of connection lines between the nodes. Based on this, the machine-reading circuit 110 can search for an answer corresponding to the mask problem based on the nodes and weights of connection lines between the nodes through the classification model based on a neural network.

[0021] The lookup table includes a plurality of mask problems and a plurality of answers. Each of the plurality of answers includes a weight value or a probability value. The machine-reading circuit 110 can establish the mask problems and the weight values ​​or probability values ​​including the corresponding answers while establishing the mask problems and searching the corresponding answers in the training data. Based on this, the machine-reading circuit 110 searches the corresponding answers and their weight values ​​or probability values ​​in the lookup table based on the mask problems.

[0022] In step S230, the plurality of second mask problems are input to the machine reading comprehension model to train the machine reading comprehension model. In some embodiments, step S230 is performed by the machine reading comprehension circuitry 110 in FIG.

[0023] Please refer to Fig. 4. Fig. 4 is a flow chart of step S230 according to some embodiments of the present application. As shown in Fig. 4, step S230 includes steps S232 to S236.

[0024] In step S232, at least one character in the training data is randomly masked to generate a plurality of second mask problems. In some embodiments, the at least one character may be a single word or a sentence. In some embodiments, the machine reading circuit 110 masks the sentence based on punctuation. For example, the training data includes the character content "Good morning, I want to go to eat breakfast, what do you want to eat?" and the mask problem may be "Good morning, I want to go to eat breakfast, what do you want to eat, [mask]?" or "Good morning, I want to go to eat [mask], what do you want to eat?"

[0025] In step S234, the masked problems are input to a machine reading comprehension model, and the machine reading comprehension model determines a plurality of second answers corresponding to the plurality of second masked problems based on the probabilities of the plurality of putative answers.

[0026] For example, the machine reading comprehension circuitry 110 inputs the masked problem into the machine reading comprehension model for the masked problem "Good morning, I want to go eat [mask], what would you like to eat?"

[0027] In step S234, the machine reading comprehension model generates a putative answer corresponding to the second mask problem in step S234 based on the first mask problem and the corresponding first answer learned in step S210. In step S234, the machine reading comprehension model is not looking for an answer in the training data.

[0028] For example, in some embodiments, the machine reading comprehension model generates multiple putative answers corresponding to the mask problem “Good morning, I want to go eat [mask], what would you like to eat?”, including “hill”, “breakfast”, and “insurance”, based on the first mask problem and the corresponding first answer learned in step S210.

[0029] Each of the putative answers includes a probability. In some embodiments, the machine reading comprehension model generates a list of probabilities corresponding to the mask question "Good morning, I want to go eat [mask], what would you like to eat?" based on the first mask question and the corresponding first answer learned in step S210. In one embodiment, the list of probabilities includes: Hill [0.2] Breakfast [0.9] Insurance [0.1].

[0030] In some embodiments, the machine reading comprehension model determines the highest probability in the list of probabilities, and takes the putative answer corresponding to the highest probability as the answer to the corresponding mask question.

[0031] For example, in the above probability list, the probability of the estimated answer "breakfast" is the highest. The machine reading comprehension model uses the estimated answer "breakfast" as the answer to the mask question "Good morning, I want to go eat [mask]. What would you like to eat?"

[0032] It should be noted that in the above embodiment, a probability list is taken as an example, in some other embodiments, the probability list may be expressed as a weight value list, a score list, or in other ways.

[0033] In step S236, a machine reading comprehension model is trained based on the multiple feedback scores of the multiple second answers. For example, in step S234, if the answer to the mask question "Good morning, I want to go eat [mask], what do you want to eat?" is "breakfast", the feedback score is [+1]. In step S234, if the answer to the mask question "Good morning, I want to go eat [mask], what do you want to eat?" is "insurance", the feedback score is [-1]. In some embodiments, the feedback score [+1] indicates that the answer is correct, while the feedback score [-1] indicates that the answer is incorrect. Any of the various feedback scores are within the embodiments of the present application, and the embodiments of the present application are not limited to the above feedback scores.

[0034] In some embodiments, the machine reading comprehension circuitry 110 is further used to train a machine reading comprehension model based on the mask questions, the answers and the feedback scores thereof to optimize the machine reading comprehension model.

[0035] In step S250, a dialogue generation model is generated according to the machine reading comprehension model. In some embodiments, step S250 is performed by the dialogue generation circuit 130 in Figure 1. In some embodiments, the dialogue generation model is stored in the memory 150 in Figure 1.

[0036] Please also refer to Figure 5. Figure 5 is a flow chart of step S250 according to some embodiments of the present application. As shown in Figure 3, step S250 includes steps S252 to S256.

[0037] In step S252, a plurality of dialogue sentences are input to the dialogue generation model. For example, a user inputs a first dialogue sentence, "Good morning, have you woken up?", to the machine reading comprehension system 100 via an input / output circuit (not shown).

[0038] In step S254, the dialogue generation model generates multiple response sentences according to the machine reading comprehension model. For example, the dialogue generation circuit 130 generates a first response sentence "Good morning, did you wake up?" according to the dialogue generation model based on the first dialogue sentence "Good morning, did you wake up?". In some embodiments, the dialogue sentence and the response sentence are present in the training data in step S210, and after the machine reading comprehension model learns, the dialogue generation model can be directly used to generate the response sentence. In this way, by repeating the input of the dialogue sentence and the generation of the response sentence multiple times, the dialogue generation circuit 130 can generate a dialogue generation model based on the dialogue sentence and the response sentence.

[0039] In step S256, a dialogue generation model is trained based on the multiple feedback scores of the multiple response sentences. For example, when the dialogue generation circuit 130 generates the first response sentence "Good morning, I just woke up" according to the dialogue generation model based on the first dialogue sentence "Good morning, did you wake up?", the feedback score is [+1], whereas when the dialogue generation circuit 130 generates the first response sentence "Special offer 3 for 100 yuan" according to the dialogue generation model based on the first dialogue sentence "Good morning, did you wake up?", the feedback score is [-1]. In some embodiments, the feedback score [+1] indicates that the answer is correct, whereas the feedback score [-1] indicates that the answer is incorrect.

[0040] In some embodiments, the dialogue generation circuitry 130 is further used to train a dialogue generation model based on the dialogue sentences, the responses and their feedback scores, so as to optimize the dialogue generation model.

[0041] In some embodiments, the dialogue generation model may include a dialogue template and corresponding text content, a classification model based on a neural network, and a look-up table.

[0042] The dialogue template may be a dialogue template generated according to a machine reading comprehension model. The mask sentence obtains the sentence following the mask question to establish the dialogue template.

[0043] The classification model based on the neural network includes a plurality of nodes and weights of the connection lines between the nodes, and each node corresponds to one dialogue or one response. The dialogue generation circuit 130 can establish a dialogue through the classification model based on the neural network and search for a corresponding response in the machine reading comprehension model, while at the same time establishing a plurality of nodes and weights of the connection lines between the nodes. On this basis, the dialogue generation circuit 130 can search for a response corresponding to the dialogue based on the weights of the nodes and the connection lines between the nodes through the classification model based on the neural network.

[0044] The lookup table includes a plurality of dialogue sentences and a plurality of response sentences. Each of the plurality of response sentences includes a weight value or a probability value. The dialogue generation circuit 130 can search for a response sentence corresponding to the dialogue sentence through the machine reading comprehension model, and at the same time, establish a weight value or a probability value of the response sentence, and further establish a weight value or a probability value of a dialogue sentence following the response sentence. Based on this, the dialogue generation circuit 130 searches for a corresponding response sentence and its weight value or probability value in the lookup table based on the dialogue sentence.

[0045] In the above embodiment, the machine reading comprehension model and the dialogue generation model are used to generate daily dialogue content, but the present disclosure is not limited thereto. In one embodiment, the machine reading comprehension model and the dialogue generation model can be trained on legal article content or legal document by the above mask, for detailed description, see the following paragraph.

[0046] 3, for example, in step S212, the training data is a legal document. The first mask problem may be, for example, "Patents in this Act are divided into three types: 1. Invention patents; 2. [mask]; and 3. Designs."

[0047] In step S214, the machine reading circuit 110 searches the text content for the mask question "Patents in this Act are divided into the following three types: 1. Invention patents; 2. [mask]; 3. Designs." and obtains the answer "invention".

[0048] Based on this, in step S210 of FIG. 2, the machine reading comprehension circuit 110 randomly masks the legal document to establish a machine reading comprehension model.

[0049] Then, in step S230, the machine reading comprehension circuit 110 trains a reading comprehension model based on the multiple second mask problem machines. In detail, referring also to FIG. 4, in step S232, at least one character in the training data is randomly masked to generate multiple second mask problems. The second mask problem may be, for example, "Patents as referred to in this Act are divided into the following three types: 1. [Mask]. 2. Invention. 3. Design."

[0050] In step S234, the machine reading circuit 110 obtains a list of multiple putative answers and corresponding probabilities based on the second mask problem. For example, Invention patent [0.9] Beef noodles [0.1] Patent requirement [0.2].

[0051] Based on the above probability list, the machine reading comprehension model sets the presumed answer “invention patent” as the answer to the mask question: “Patents as referred to in this Act are divided into the following three types: 1. [mask]. 2. Invention. 3. Design.”

[0052] Then, in step S236, the machine reading comprehension circuit 110 can train a machine reading comprehension model based on a plurality of second answers corresponding to the second mask problem and their feedback scores. For example, in step S234, if the answer to the mask problem "Patents in this law are divided into the following three types: 1. [Mask]. 2. Invention. 3. Design" is "invention patent", the feedback score is [+1]. In step S234, if the answer to the mask problem "Patents in this law are divided into the following three types: 1. [Mask]. 2. Invention. 3. Design" is "beef noodles", the feedback score is [-1]. In some embodiments, the feedback score [+1] indicates that the answer is correct, while the feedback score [-1] indicates that the answer is incorrect. Various feedback scores are all within the embodiments of the present application, and the embodiments of the present application are not limited to the above feedback scores. Then, in step S250, a dialogue generation model is generated according to the re-machine reading comprehension model.

[0053] 5, in step S252, a plurality of dialogue sentences are inputted into the dialogue generation model. The first dialogue sentence may be, for example, "What kinds of patents are there?"

[0054] In step S254, the dialogue generation circuit 130 generates a first response sentence "Patents are divided into three types" according to the machine reading comprehension model based on the first dialogue sentence "Are there several types of patents?" In this way, the dialogue generation circuit 130 can repeat input of a dialogue sentence and generation of a response sentence multiple times, and generate a dialogue generation model based on the dialogue sentence and the response sentence.

[0055] In step S256, a dialogue generation model is trained based on the multiple feedback scores of the multiple response sentences. For example, when the dialogue generation circuit 130 generates the first response sentence "3 types" according to the dialogue generation model based on the first dialogue sentence "Are there several types of patents?", the feedback score is [+1], whereas when the dialogue generation circuit 130 generates the first response sentence "I want to eat beef noodles" according to the dialogue generation model based on the first dialogue sentence "Are there several types of patents?", the feedback score is [-1]. In some embodiments, the feedback score [+1] indicates that the answer is correct, whereas the feedback score [-1] indicates that the answer is incorrect.

[0056] Based on this, in some embodiments, the dialogue generation circuitry 130 is further used to train a dialogue generation model based on the dialogue sentences, the responses and their feedback scores, so as to optimize the dialogue generation model.

[0057] In some embodiments, the plurality of dialogues and responses form a generation policy, see also Figure 6, which is a schematic diagram of a generation policy according to some embodiments of the present application.

[0058] As shown in Fig. 6, in Fig. 6, the response sentences that the dialogue generation circuit 130 may generate based on the dialogue Q1 include a response sentence A1, a response sentence A2, or a response sentence A3. After the dialogue generation circuit 130 returns the response sentence A1 based on the dialogue Q1, the response sentences that the dialogue generation circuit 130 may further generate based on the subsequent dialogue Q2 (not shown) and the response sentence A1 include a response sentence A11 or a response sentence A12. After the dialogue generation circuit 130 returns the response sentence A2 based on the dialogue Q1, the response sentences that the dialogue generation circuit 130 may further generate based on the subsequent dialogue Q2 (not shown) and the response sentence A2 include a response sentence A21 or a response sentence A22. After the dialogue generation circuit 130 returns the response sentence A3 based on the dialogue Q1, the response sentences that the dialogue generation circuit 130 may further generate based on the subsequent dialogue Q2 (not shown) and the response sentence A3 include a response sentence A31 or a response sentence A32. The above dialogue and response sentences are merely for illustrative purposes, and the embodiments of the present application are not limited to the above.

[0059] In some embodiments, response sentences A1 and A11 are generated based on the dialogue Q1 to form a generation policy S1, response sentences A1 and A12 are generated based on the dialogue Q1 to form a generation policy S2, response sentences A2 and A21 are generated based on the dialogue Q1 to form a generation policy S3, response sentences A2 and A22 are generated based on the dialogue Q1 to form a generation policy S4, response sentences A3 and A31 are generated based on the dialogue Q1 to form a generation policy S5, and response sentences A3 and A32 are generated based on the dialogue Q1 to form a generation policy S6.

[0060] In some embodiments, each of the response sentences A1 to A3, A11 to A32 includes a feedback score. In some embodiments, the dialogue generation circuit 130 optimizes the dialogue generation model based on the plurality of response sentences and the feedback scores included in each of them.

[0061] In step S270, a dialogue is generated according to the dialogue generation model. In some embodiments, step S270 is performed by the dialogue generation circuit 130 in Fig. 1. Referring to Fig. 6, in some embodiments, the dialogue generation circuit 130 calculates a total feedback score for each of the multiple generation policies S1 to S6, determines the one with the highest total feedback score for each of the multiple generation policies S1 to S6 as optimal, and generates a dialogue based on the optimal generation policy.

[0062] For example, in some embodiments, the dialogue generation circuitry 130 adds the feedback scores of responses A1 and A11 to arrive at a total feedback score for generated policy S1, and the remainder is inferred in this manner.

[0063] When the generation policy S1 has the highest score among the total feedback scores of the generation policies S1 to S6, the dialogue generation circuit 130 generates a dialogue based on the generation policy S1.

[0064] Based on the above examples, the embodiments of the present application provide a machine reading comprehension method and a machine reading comprehension system, which learns training data by masking, and then adds feedback scores to optimize a machine reading comprehension model. A dialogue generation model is generated according to the machine reading comprehension model, and the dialogue generation model is optimized by the feedback scores to generate dialogue. The embodiments of the present application can effectively read and understand a large amount of text data, and generate dialogue based on the viewed text data, thereby improving the effectiveness of machine learning.

[0065] In some embodiments, the machine reading system 100 as described above can be integrated into an electronic device or system. In some embodiments, the server can be implemented as a cloud server including a central processing unit. In some embodiments, the server can be, but is not limited to, a single processor or a combination of multiple microprocessors, including memory and input / output circuitry.

[0066] In some embodiments, the machine reading circuitry 110 and the dialogue generating circuitry 130 may be a central processor unit (CPU), a microprocessor (MCU), a server, or other computing circuitry or element having data access, data calculation, data storage, data transmission and reception, or similar functions.

[0067] In some embodiments, each of the machine-reading circuitry 110 and the dialogue generating circuitry 130 includes a processor and input / output circuitry. The processor may be a circuit or element having a data access, data calculation, data storage, or similar function. The input / output circuitry may be a circuit or element having a data transmission / reception, or similar function.

[0068] In some embodiments, memory 150 may be any circuit or element having data storage, data access, or similar functionality.

[0069] Although the present application has been disclosed above in detailed examples, it does not exclude other possible embodiments. Therefore, the scope of protection of the present application should not be limited by the above examples, but should be based on what is defined by the following claims.

[0070] Those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Based on the above embodiments, all changes and modifications made to the present application are also included within the protection scope of the present application. [Explanation of symbols]

[0071] 100: Machine Reading Systems 110: Machine reading circuit 130: Dialogue generation circuit 150:Memory 200: Machine Reading Methods S210, S230, S250, S270: Process S212, S214: Process S232, S234, S236: Project S252, S254, S256: Project Q1: Dialogue sentence A1, A2, A3: Response sentence A11, A12, A21, A22, A31, A32: Response sentence S1, S2, S3, S4, S5, S6: Generation policy

Claims

1. establishing a machine reading comprehension model based on the training data and a plurality of first mask problems; training the machine reading comprehension model by inputting a plurality of second mask problems into the machine reading comprehension model; generating a dialogue generation model according to the machine reading comprehension model; generating a dialogue in accordance with the dialogue generation model; A machine reading method comprising:

2. Establishing the machine reading comprehension model based on the training data and the first plurality of mask problems, randomly masking at least one character in the training data to generate the plurality of first mask problems; searching the training data to obtain a plurality of first answers corresponding to the plurality of first mask problems; 2. The machine-readable method of claim 1, comprising:

3. The step of inputting the plurality of second mask problems to the machine reading comprehension model to train the machine reading comprehension model includes: randomly masking at least one character in the training data to generate the plurality of second mask problems; inputting the second mask problems into the machine reading comprehension model, and the machine reading comprehension model determining second answers corresponding to the second mask problems based on probabilities of putative answers; training the machine reading comprehension model based on a plurality of feedback scores of the plurality of second answers; 3. The machine reading method of claim 2, comprising:

4. generating the dialogue generation model according to the machine reading comprehension model, inputting a plurality of dialogue sentences into the dialogue generation model; generating a plurality of response sentences according to the machine reading comprehension model; training the dialogue generation model based on a plurality of feedback scores of a plurality of responses; 2. The machine-readable method of claim 1, comprising:

5. generating a plurality of generation policies based on the plurality of dialogue sentences and the plurality of response sentences; generating the interaction based on an optimal one of the plurality of generation policies; The machine reading method of claim 4 further comprising:

6. a machine reading comprehension circuit for establishing a machine reading comprehension model based on the training data and a plurality of first mask problems, and inputting a plurality of second mask problems to the machine reading comprehension model to train the machine reading comprehension model; a dialogue generation circuit coupled to the machine reading comprehension circuit for generating a dialogue generation model according to the machine reading comprehension model and for generating dialogue according to the dialogue generation model; a memory coupled to said machine reading circuitry and said dialogue generation circuitry for storing said machine reading model and said dialogue generation model; A machine reading system including:

7. 7. The machine reading comprehension system of claim 6, wherein the machine reading comprehension circuitry is further adapted to randomly mask at least one character in the training data to generate the plurality of first masked problems, and to search the training data to obtain a plurality of first answers corresponding to the plurality of first masked problems.

8. 8. The machine reading comprehension system of claim 7, wherein the machine reading comprehension circuitry is further used to randomly mask at least one character in the training data to generate the plurality of second masked problems, input the plurality of second masked problems to the machine reading comprehension model, wherein the machine reading comprehension model determines a plurality of second answers corresponding to the plurality of second masked problems based on a plurality of probabilities of a plurality of putative answers, and is used to train the machine reading comprehension model based on a plurality of feedback scores of the plurality of second answers.

9. 7. The machine reading comprehension system of claim 6, wherein the dialogue generation circuitry is further used to input a plurality of dialogue sentences to the dialogue generation model, the dialogue generation model is used to generate a plurality of response sentences according to the machine reading comprehension model, and is used to train the dialogue generation model based on a plurality of feedback scores of the plurality of response sentences.

10. 10. The machine reading comprehension system of claim 9, wherein the dialogue generation circuitry is further used to generate a plurality of generation policies based on the plurality of dialogue sentences and the plurality of response sentences, and to generate the dialogue based on an optimal one of the plurality of generation policies.

Images