Method and apparatus for training a structured abstract model

Abstract

The embodiments of this specification provide a method and apparatus for training a structured summary model. In this method, fields are extracted from the form to which the structured summary model is applied; the extracted fields are classified according to the filling difficulty of each field in the form to obtain field categories with different filling difficulties; the structured summary model is trained in the order of the field categories with increasing filling difficulty until the training for all field categories in the field category order is completed: for the target field category that is the current training target, the target field category and other field categories with a lower filling difficulty corresponding to the target field category are used as training objects, and the structured summary model is trained using the dialogue sample data and the labels corresponding to each training object; and when the training for the target field category is completed, the next field category in the field category order is determined as the target field category in the next round of training.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of artificial intelligence technology. Specifically, they relate to methods and devices for training a structured summary model. Background Art

[0002] Video face-to-face interviews are conducted between interviewers and interviewees in a video manner for face-to-face access and surveys. During a video face-to-face interview, the interviewer asks questions and the interviewee answers the corresponding questions. Given the convenience of video face-to-face interviews, they are applied in multiple fields, such as job interviews, insurance claims, etc.

[0003] Based on different tasks or purposes, during a video face-to-face interview, the interviewer will ask some questions of concern, such as names, dates, addresses, etc. Then, the interviewer records the relevant information in the corresponding form according to the interviewee's answers. In the traditional recording method, the interviewer needs to record manually, and this recording method is inefficient.

[0004] With the development of artificial intelligence, machine learning models are widely used. Applying a machine learning model during a video face-to-face interview can automatically extract the information required by the form from the conversation and fill the information into the corresponding fields. This recording method based on a machine learning model improves the recording efficiency compared with manual recording. Summary of the Invention

[0005] In view of the above, the embodiments of this specification provide methods and devices for training a structured summary model. Through the technical solutions of the embodiments of this specification, the structured summary model is gradually trained in the order of field categories with increasing filling difficulty. Compared with the method of training all fields together, the training effect is improved, and the reduction of the training effect caused by possible learning conflicts between multiple fields when training all fields together is avoided.

[0006] According to one aspect of the embodiments of the present specification, there is provided a method for training a structured summary model, including: extracting fields from a form applied by the structured summary model; classifying the extracted fields according to the filling difficulty of each field in the form to obtain field categories with different filling difficulties, where the field categories include at least two categories, and each field category includes at least one field; training the structured summary model in the order of field categories with increasing filling difficulty until the training for all field categories in the field category order is completed: for the target field category that is the current training target, using the target field category and other field categories with lower filling difficulty corresponding to the target field category as training objects, and training the structured summary model using dialogue sample data and the labels corresponding to each training object, where the target field category is sequentially determined according to the field category order; and when the training for the target field category is completed, determining the next field category in the field category order as the target field category in the next round of training.

[0007] According to another aspect of the embodiments of the present specification, there is also provided a device for training a structured summary model, including: a field extraction unit that extracts fields from a form applied by the structured summary model; a field classification unit that classifies the extracted fields according to the filling difficulty of each field in the form to obtain field categories with different filling difficulties, where the field categories include at least two categories, and each field category includes at least one field; a model training unit that, for the target field category that is the current training target, uses the target field category and other field categories with lower filling difficulty corresponding to the target field category as training objects, and trains the structured summary model using dialogue sample data and the labels corresponding to each training object, where the target field category is sequentially determined according to the field category order; and a target field category determination unit that, when the training for the target field category is completed, determines the next field category in the field category order as the target field category in the next round of training, where the model training unit and the target field category determination unit train the structured summary model according to the field category order until the training for all field categories in the field category order is completed.

[0008] According to another aspect of the embodiments of the present specification, there is also provided an electronic device, including: at least one processor, a memory coupled to the at least one processor, and a computer program stored on the memory, where the at least one processor executes the computer program to implement the method for training a structured summary model as described in any one of the above.

[0009] According to another aspect of the embodiments of the present specification, there is also provided a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the method for training a structured summary model as described above.

[0010] According to another aspect of the embodiments of the present specification, there is also provided a computer program product including a computer program, and when the computer program is executed by a processor, it implements the method for training a structured summary model as described in any one of the above. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] By referring to the following drawings, a further understanding of the essence and advantages of the content of the embodiments of the present specification can be achieved. In the drawings, similar components or features may have the same reference numerals.

[0012] Figure 1 FIG. shows a schematic diagram of an example of a method for training a structured summary model according to an embodiment of the present specification.

[0013] Figure 2 FIG. shows a flowchart of an example of training a structured summary model according to an embodiment of the present specification.

[0014] Figure 3 FIG. shows a flowchart of an example 300 of training a structured summary model according to an embodiment of the present specification.

[0015] Figure 4 FIG. shows a block diagram of an example of an apparatus for training a structured summary model according to an embodiment of the present specification.

[0016] Figure 5 FIG. shows a block diagram of an electronic device for implementing a method for training a structured summary model according to an embodiment of the present specification. DETAILED DESCRIPTION

[0017] The subject matter described herein will be discussed with reference to example embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thereby implement the subject matter described herein, and is not a limitation on the scope of protection, applicability, or examples set forth in the claims. The functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the content of the embodiments of the present specification. Each example may omit, substitute, or add various processes or components as needed. Additionally, the features described relative to some examples may be combined in other examples.

[0018] As used herein, the term "comprising" and variations thereof are open-ended terms meaning "including but not limited to". The term "based on" means "at least partially based on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc. may refer to different or the same objects. Other definitions may be included below, whether explicit or implicit. Unless explicitly specified in the context, the definition of a term is consistent throughout the specification.

[0019] Video face-to-face interviews are conducted between the interviewer and the interviewee in a video face-to-face manner. During the video face-to-face interview, the interviewer asks questions and the interviewee answers the corresponding questions. Given the convenience of video face-to-face interviews, they are applied in multiple fields, such as job interviews, insurance claims, etc.

[0020] Based on different tasks or purposes, during the video face-to-face interview, the interviewer will ask some questions of concern, such as name, date, address, etc. Then, the interviewer records the relevant information in the corresponding form according to the interviewee's answers. In the traditional recording method, the interviewer needs to record manually, which is inefficient.

[0021] With the development of artificial intelligence, machine learning models are widely used. Applying machine learning models during video face-to-face interviews can automatically extract the information required by the form from the conversation and fill in the corresponding fields with the information. This recording method based on machine learning models improves the recording efficiency compared to manual recording.

[0022] However, currently for the training of the model, all sample data are input into the model for training together, or only the sample data are simply divided and input into the model in batches for several rounds of training. This training method may have learning conflicts between multiple fields, making it difficult to train the model and thus difficult to achieve good training results.

[0023] In view of the above, the embodiments of the present specification provide a method and an apparatus for training a structured summary model. In this method, fields are extracted from the form applied by the structured summary model; the extracted fields are classified according to the filling difficulty of each field in the form to obtain field categories with different filling difficulties; the structured summary model is trained in the order of the field categories with increasing filling difficulty until the training for all field categories in the field category order is completed: for the target field category that is the current training target, the target field category and other field categories with lower filling difficulty corresponding to the target field category are used as training objects, and the structured summary model is trained using the dialogue sample data and the labels corresponding to each training object; and when the training for the target field category is completed, the next field category in the field category order is determined as the target field category in the next round of training. Through the technical solution of the embodiments of the present specification, the structured summary model is gradually trained in the order of increasing filling difficulty of the fields. Compared with the method of training all fields together, the training effect is improved, and the reduction of the training effect caused by the learning conflict between multiple fields that may exist when training all fields together is avoided.

[0024] The following will describe in detail the method and apparatus for training a structured summary model provided by the embodiments of the present specification with reference to the accompanying drawings.

[0025] Figure 1 FIG. 7 shows a schematic diagram of an example 100 of a method for training a structured summary model according to an embodiment of the present specification.

[0026] As Figure 1 shown, at 110, fields can be extracted from the form applied by the structured summary model.

[0027] In the embodiments of the present specification, the structured summary model is a model to be trained. The trained structured summary model can obtain the content required for each field in the form from the input dialogue text and generate a summary that conforms to the format of each field. During the application process of the structured summary model, the form to be filled and the dialogue text can be used as inputs, and the structured summary model outputs the filled form.

[0028] In the embodiments of the present specification, the form may include multiple fields, and the set of all fields forms a form. Each field corresponds to a field filling type as a slot, and the field filling type corresponding to each field is the type of the value to be filled in the field.

[0029] The object targeted by the embodiments of this specification can be a form, or multiple forms. When the targeted object is a single form, all fields to be filled can be extracted from the form. During the subsequent application process of the structured summary model, the object to which the structured summary model is applied is this form. That is, during the application process, the form to be filled can be input into the structured summary model, and the structured summary model outputs the filled form.

[0030] When the targeted object is multiple forms, all fields to be filled can be extracted from these multiple forms. During the subsequent application process of the structured summary model, the object to which the structured summary model is applied is some or all of these multiple forms.

[0031] In one example, the forms in the embodiments of this specification can include various types of forms such as schema. Here, schema will be used as an example for illustration.

[0032] In one example, the trained structured summary model can be applied to the scenario of insurance loss assessment. In the scenario of insurance loss assessment, the loss assessor communicates with the policyholder in the form of a video. During the communication between the loss assessor and the policyholder, the applied structured summary model can obtain the conversation text, and use the conversation text and the form to be filled as inputs, and then output the filled form. It should be noted that the structured summary model can also be applied to other application scenarios, such as job interviews, etc.

[0033] At 120, the extracted fields can be classified according to the filling difficulty of each field in the form to obtain different field categories with different filling difficulties.

[0034] In the form, the content to be filled in each field is different, and the field filling types of each field can also be different. The field filling type of each field can be used to determine the content filling method of the field, that is, the content filling methods of each field can be different.

[0035] The field filling type can include classification type, key information extraction type, information abstraction type, etc. Among them, the classification type can include binary classification type and multi-classification type. Each field corresponds to one field filling type, and the query conditions (i.e., queries) corresponding to different fields belonging to the same field filling type can be different, and the corresponding filled content can also be different. For example, both fields are of the binary classification type, and the value range is "yes" and "no", but the query of one field is "whether male", and the query of the other field is "whether local".

[0036] In a field of the binary classification type, the preset value range of the field only includes two choices, and one of the preset two choices is selected as the assignment of the field. For example, the value range of the binary classification field includes two choices, "yes" and "no", and one of "yes" and "no" is selected as the assignment of the binary classification field.

[0037] In a field of the multi-classification type, the preset value range of the field can include more than two multiple choices, and one of the preset multiple choices is selected as the assignment of the field. For example, the content to be filled in for the field of the multi-classification type is a date, and the preset candidate dates in the field include several dates, and one of the several dates can be selected as the assignment of the field.

[0038] In a field of the key information extraction type, the assignment of the field can be extracted from the dialogue text input to the model. The key information extraction type can be further divided into three sub-types: the direct extraction type, the extraction classification type, and the extraction standardization type. Of course, the above three sub-types are only taken as an example, and the key information extraction type can also include other sub-types.

[0039] In a field of the direct extraction type, the key information required for the query condition corresponding to the field can be directly extracted from the dialogue text input to the model, and the extracted key information can be directly used as the assignment of the field.

[0040] In a field of the extraction classification type, after the key information required for the query condition corresponding to the field is extracted from the dialogue text, the key information can be classified to determine the category to which the key information belongs. Then, the category can be used as the assignment of the field. For example, the content to be filled in a field is a province, and the key information extracted from the dialogue text is Shijiazhuang. Classifying the key information can determine that the category to which it belongs is Hebei Province, so Hebei Province can be used as the assignment of the field.

[0041] In a field of the extraction standardization type, after the key information required for the query condition corresponding to the field is extracted from the dialogue text, the key information can be standardized to convert the key information into standardized format information that conforms to the field format, and the obtained standardized format information is used as the assignment of the field. For example, the content to be filled in a field is time, and the standardized time format set for the field is: yyyy-mm-dd, where yyyy represents the year, mm represents the month, and dd represents the date. The key information extracted from the dialogue text is March 8th. Then, standardizing the key information can convert the key information "March 8th" into the standardized format information "2022-03-08" that conforms to the standardized time format, and this standardized format information can be used as the assignment of the field.

[0042] In the field of information abstraction type, the assignment of the field can be obtained by abstracting according to the meaning expressed in the dialogue text of the input model. Through the abstraction process, the meaning expressed by one or more segments in the dialogue text can be summarized into a short expression to form an abstract output. The meaning expressed by the abstract is the same as or similar to the meaning expressed by the one or more segments in the dialogue text. The formed abstract can be used as the assignment in this field.

[0043] In the embodiments of this specification, the content to be filled in each field is different, which can make the filling difficulty of each field different. For example, when there is more content to be filled in a field, the filling difficulty of this field will be higher; while when there is less content to be filled in, it is easier to fill in. The content to be filled in each field is determined from the dialogue text according to the query condition corresponding to this field, and the filling method of the content to be filled in is determined according to the field filling type. Therefore, the content to be filled in each field is associated with the query condition and the field filling type.

[0044] In one example, the filling difficulty of each field can be determined according to the query condition corresponding to the field and / or the field filling type.

[0045] In the method of determining the filling difficulty according to the query condition, the query condition can exist in the form of a query question or a phrase, etc., and the filling difficulty can be determined according to the length of the query condition. That is, when the text corresponding to the query condition is longer, the filling difficulty corresponding to this query condition is higher; when the text corresponding to the query condition is shorter, the filling difficulty corresponding to this query condition is lower. In addition, the filling difficulty can also be determined according to the sentence structure of the query condition. That is, when the sentence structure corresponding to the query condition is complex, the filling difficulty corresponding to this query condition is higher; when the sentence structure corresponding to the query condition is simple, the filling difficulty corresponding to this query condition is lower.

[0046] In the method of determining the filling difficulty according to the field filling type, different field filling types correspond to different filling difficulties. Among the binary classification type, multi-classification type, key information extraction type, and information abstraction type included in the field filling type, the corresponding filling difficulties increase in sequence. Among them, the extraction operation corresponding to the key information extraction type only needs to extract information from the dialogue text, which is simpler than the abstraction operation corresponding to the information abstraction type. Therefore, the filling difficulty corresponding to the key information extraction type is lower, while the filling difficulty corresponding to the information abstraction type is higher.

[0047] Furthermore, among the direct extraction type, extraction classification type, and extraction standardization type included in the key information extraction type, the corresponding filling difficulty increases in turn. The operation corresponding to the direct extraction type is simpler than the operations corresponding to the extraction classification type and the extraction standardization type, so the filling difficulty is the lowest. The standardization operation in the operation corresponding to the extraction standardization type requires converting the extracted information, which is more complex than the classification operation in the operation corresponding to the extraction classification type. Therefore, the filling difficulty corresponding to the extraction standardization type is higher.

[0048] When the filling difficulty is determined according to the field filling type, when determining the filling difficulty for each field, the field filling type corresponding to each field in the form can be determined first, and then the filling difficulty corresponding to each known field filling type can be used to obtain the filling difficulty corresponding to each field filling type in the form. For example, the field filling types corresponding to the fields in the form include the binary classification type, the key information extraction type, and the information abstraction type. Thus, the increasing order of the filling difficulty can be determined as: the fields of the binary classification type, the fields of the key information extraction type, and the fields of the information abstraction type.

[0049] When the filling difficulty is determined according to the query condition and the field filling type, the filling difficulty can be determined first according to the field filling type, and then the filling difficulty can be further determined according to the query condition on the basis of the determined filling difficulty.

[0050] In an example, when determining the filling difficulty for each field, the first filling difficulty corresponding to each field in the form can be determined according to the field filling type first. Then, for each field with the same first filling difficulty, the field filling types corresponding to these fields are the same. Thus, the filling difficulty of each field with the same first filling difficulty can be further differentiated according to the query condition to update the filling difficulty of each field to the second filling difficulty, and the second filling difficulties corresponding to each field are different. Finally, the first filling difficulty and the second filling difficulty corresponding to each field can be integrated to obtain the filling difficulty of each field.

[0051] For example, when determining the first filling difficulty according to the field filling type, the first filling difficulties corresponding to multiple fields belonging to the binary classification type are the same. Then, the filling difficulty of these multiple fields is differentiated according to the query condition. For example, the longer and more complex the query condition of a field is, the higher the corresponding second filling difficulty is, and the shorter and simpler the query condition of a field is, the lower the corresponding second filling difficulty is. Thus, the second filling difficulties corresponding to these multiple fields can be obtained. The second filling difficulties corresponding to these multiple fields belonging to the binary classification type are all lower than the first filling difficulties corresponding to the fields of other types such as the multi-classification type, the key information extraction type, and the information abstraction type.

[0052] In the embodiments of this specification, the fields can be classified according to the filling difficulty of each field, and field categories with different filling difficulties can be obtained. The obtained field categories include at least two categories, and each field category includes at least one field. For the filling difficulty, the granularity of the filling difficulty can be set. When the granularity is different, the classification of the fields can also be different.

[0053] In one example, in the case of a coarse granularity, the field filling types can include: classification type, key information extraction type, and information abstraction type. At this time, the corresponding filling difficulties of the classification type, key information extraction type, and information abstraction type increase in sequence, and the classification type, key information extraction type, and information abstraction type respectively correspond to one filling difficulty. Thus, the fields can be classified according to these three types: classification type, key information extraction type, and information abstraction type, and each obtained field category corresponds to one type.

[0054] In another example, in the case of a fine granularity, the field filling types can include: binary classification type, multi-classification type, direct extraction type, extraction classification type, extraction standardization type, and information abstraction type. At this time, the corresponding filling difficulties of the binary classification type, multi-classification type, direct extraction type, extraction classification type, extraction standardization type, and information abstraction type increase in sequence, and each type corresponds to one filling difficulty. Thus, the fields can be classified according to these six types: binary classification type, multi-classification type, direct extraction type, extraction classification type, extraction standardization type, and information abstraction type, and each obtained field category corresponds to one type.

[0055] In one example, when classifying the fields according to the filling difficulty of each field, the filling difficulty of each field can be determined first according to the query conditions and field filling types corresponding to each field, and the obtained filling difficulties of each field can be different. Then, the fields are sorted in ascending order of filling difficulty to obtain a field sequence. Next, the field sequence can be divided according to the custom number of classification categories to obtain multiple subsequences. Among them, each subsequence can be used as a classification category. Each subsequence includes at least one field, and the number of fields included in each subsequence can be the same or different. The division method for the field sequence can be random division or division according to a specified rule. For example, average division, and the number of fields included in each obtained subsequence is the same.

[0056] In one example, each field category may include only one field. In this example, the filling difficulty of each field can be determined first according to the query conditions and field filling types corresponding to each field, and the filling difficulties of the obtained fields may be different. Then, sort each field in ascending order of filling difficulty to obtain a field sequence. Each field in the field sequence can be used as a field category, and the order presented by the field sequence is the order in which the filling difficulties of each field category increase.

[0057] In one example, after obtaining the field categories, weights can be configured for each field category, and the weights corresponding to different field categories may be different. The weight of a field category can be used to represent the degree of attention paid by the structured summary model to this field category when training with sample data. The greater the weight, the higher the degree of attention paid by the structured summary model to the corresponding field category; the smaller the weight, the lower the degree of attention paid by the structured summary model to the corresponding field category.

[0058] The weights corresponding to each field category can be determined according to the corresponding filling difficulty. In one example, the smaller the filling difficulty corresponding to each field category, the greater the corresponding weight can be set; the greater the filling difficulty corresponding to each field category, the smaller the corresponding weight can be set. In this way, during the training process of the structured summary model, the structured summary model pays more attention to the field categories with smaller filling difficulties. In another example, the smaller the filling difficulty corresponding to each field category, the smaller the corresponding weight can be set; the greater the filling difficulty corresponding to each field category, the greater the corresponding weight can be set. In this way, during the training process of the structured summary model, the structured summary model pays more attention to the field categories with greater filling difficulties.

[0059] In another example, the weights configured for each field category can be the same, or each field category has no corresponding weight.

[0060] Back to Figure 1 , after obtaining the field categories, the structured summary model can be trained in the order of field categories with increasing filling difficulty until the training for all field categories in the field category order is completed. Among them, the field category order is the order formed by each field category in ascending order of filling difficulty. In the case where each field category includes only one field, the field category order is the order formed by each field in ascending order of filling difficulty. The training process of the structured summary model is described in detail below through operations 130 - 150.

[0061] At 130, for the target field category that is the current training objective, the target field category and other field categories with a lower filling difficulty than that corresponding to the target field category can be used as training objects, and the structured summary model is trained using the sample data and the labels corresponding to each training object.

[0062] In the embodiments of this specification, the target field category can be sequentially determined according to the order of field categories with increasing filling difficulty. Each time, a field category is determined as the target field category, and the determined target field category each time can be used as the current training objective.

[0063] For example, the order of field categories with increasing filling difficulty is: field categories of binary classification type, field categories of key information extraction type, and field categories of information abstraction type. Then, according to this field category order, the field category corresponding to the binary classification type is first determined as the first target field category. Then, the field category corresponding to the key information extraction type is determined as the second target field category. Finally, the field category corresponding to the information abstraction type is determined as the second target field category. Another example is that in the case where each field category includes only one field, the order of fields with increasing filling difficulty is: P1, P2,..., P i …, where P i is a field marker used to represent the field. According to this field order, the field corresponding to P1 is first determined as the first target field category, the field corresponding to P2 is determined as the target field category second, and so on, until the last field in this field order is determined as the target field category.

[0064] After each target field category as the training objective is determined, a round of training for this target field category can be correspondingly executed. In the present disclosure, a round of training can be a training process including multiple sub-loops, and the training objective targeted in a round of training is the same. The end condition for each round of training can be that the specified training end condition is reached for the training objective. For example, the loss for this training objective is less than the specified loss value, or the number of sub-loops reaches the specified number. When there are multiple rounds of training, the training objectives targeted in each round of training are different.

[0065] For each target field category as the current training objective, the target field category and other field categories with a lower filling difficulty than that corresponding to the target field category can be used as training objects. Here, the training object refers to the object that the structured summary model focuses on in a round of training for this target field category. The structured summary model trains for the fields used as training objects, so that the structured summary model after this round of training can better perform the structured summary operation on the fields of this training object.

[0066] For example, in the first round of training, the target field category is the field category with the lowest filling difficulty, and there is no field category with a lower filling difficulty than the target field category. Therefore, only the target field category is used as the training object. In the second round of training, the target field category is the field category ranked second in the field category order. Therefore, the target field category and the field category with the lowest filling difficulty can be used as the training objects.

[0067] After determining the training objects for the current round of training, the structured summary model can be trained using the dialogue sample data and the labels corresponding to each training object.

[0068] In one example, the dialogue sample data and the labels corresponding to each training object can be used as the input to the structured summary model. The structured summary model performs a structured summary task on the dialogue sample data and the form to obtain the summary information of all fields in the form. Then, based on the labels corresponding to each field as the training object and the summary information of each generated field, the loss for each field is calculated. The structured summary model is adjusted according to the calculated loss, and the training of the next cycle continues until the cycle end condition is met.

[0069] Figure 2 FIG. 200 shows a flowchart of an example of training a structured summary model according to an embodiment of the present specification.

[0070] As Figure 2 shown in 131-1, for the target field category as the current training target, the field markers corresponding to each training object field in the target field category and other field categories with a lower filling difficulty than the target field category can be added to the dialogue sample data as a prefix.

[0071] In this example, the training object field is the field used as the training object, and the obtained training object fields are the fields in the target field category and other field categories. Each field is set with a corresponding field marker, and there is a one-to-one correspondence between each field and the field marker. The field marker can represent the corresponding field.

[0072] Adding the field markers corresponding to each training object field to the dialogue sample data makes the dialogue sample data carry the field markers. In one example, the field markers can be added to the dialogue sample data as a prefix to obtain the dialogue sample data with a prefix. The field markers in the dialogue sample data can be used as the prompt information for the corresponding fields to prompt the existence of the corresponding fields in the dialogue sample data, so that the structured summary model pays attention to the fields corresponding to the field markers during the training process.

[0073] The dialogue sample data with a prefix can be expressed as: Q1 Q2 … Qi … X, where Q1, Q2, … Q i respectively represent the set of field markers corresponding to the field categories, and the subscript numbers represent the sorting of the corresponding field categories in the field category order. Each set of field markers corresponding to a field category includes the field markers corresponding to each field included in that field category.

[0074] In one example, when each field category includes only one field, the conversation sample data with prefixes can be expressed as: P1 P2 … P i … X, where P1, P2, … P i respectively represent the field markers corresponding to the fields.

[0075] In 131 - 3, the structured summary model can be trained using the conversation sample data with prefixes and the labels corresponding to each training object field.

[0076] In one example, the data form composed of the conversation sample data with prefixes and the labels corresponding to each training object field can be expressed as (Q1 Q2 … Q i X, V1 S1 V2 S2 … S i-1 V i ), where V1, V2, … V i respectively represent the corresponding label sets, and S1, S2, … S i respectively represent delimiters. Each label set corresponding to a field category includes the labels of each field included in that field category, and the fields included in each field category corresponding to the set of field markers as the prefix are training object fields.

[0077] In one example, when each field category includes only one field, the data form composed of the conversation sample data with prefixes and the labels corresponding to each training object field can be expressed as (P1 P2 … P i X, V1S1V2 S2 … S i-1 V i ), and at this time, V1, V2, … V i respectively represent the labels corresponding to each training object field.

[0078] The conversation sample data with field markers and the labels corresponding to each training object field can be used as the input of the structure summary model, and the output of the structure summary model is the summary for each training object field obtained from the conversation sample data.

[0079] During the operation of the structured summary model, the field tags in the dialogue sample data are involved to prompt the fields corresponding to each field tag existing in the dialogue sample data, enabling the structured summary model to focus on the fields corresponding to each field tag. Thus, during the model training process, through each field tag, the structured summary model can be guided to adapt to the learning tasks for the fields corresponding to each field tag.

[0080] The training using the dialogue sample data with prefixes and the labels corresponding to each training object field is one round of training. In each sub-loop included in this round of training, the loss is calculated based on the summary of each training object field output by the structured summary model and the label corresponding to each training object field, and the structured summary model is adjusted according to the loss. After the adjustment, the structured summary model executes the next sub-loop until the end of this round of training. At the end of this round of training, the operation of 140 is executed.

[0081] It should be noted that when the current target field category is the first field category in the field category order, there is no other field category with a lower filling difficulty than the field category corresponding to it. Thus, all fields in the target field category can be used as training object fields, and the field tags corresponding to each training object field are added to the dialogue sample data as prefixes.

[0082] Figure 3 A flowchart of an example 300 of training a structured summary model according to an embodiment of the present specification is shown.

[0083] As Figure 3 shown, in 132-1, for the target field category that is the current training target, the previous dialogue sample data with prefixes used for training can be obtained by taking the previous field category of the target field category in the field category order as the training target.

[0084] In this example, the previous field category is determined according to the field category order. In the field category order, the previous field category is adjacent to the target field category and is in the previous position of the target field category. The prefix in the previous dialogue sample data includes the field tags corresponding to the fields included in each field category with a lower filling difficulty than the field category corresponding to the current target field category. For example, in the case where each field category includes only one field, the fields in the order of increasing filling difficulty are: P1, P2, P3,... If the current target field category is P3, then the prefix in the previous dialogue sample data includes the field tags corresponding to fields P1 and P2.

[0085] In one example, the set of field identifiers corresponding to the current target field category is Q iWhen, the data form of the previous dialogue sample data with a prefix can be expressed as Q1 Q2 … Q i-1 X.

[0086] In 132-3, in the prefix of the previous dialogue sample data, the field markers corresponding to each training object field in the target field category can be added to obtain the dialogue sample data with a prefix corresponding to the target field category.

[0087] The training object fields in the target field category include all fields in the target field category. When there are at least two fields in the target field category, the set of field markers corresponding to the target field category can be added as a whole to the prefix in the previous dialogue sample data, and the set of field markers includes the field markers corresponding to each training object field in the target field category. When the target field category only includes one field, the field marker corresponding to the field can be added to the prefix in the previous dialogue sample data.

[0088] In an example, the data form of the previous dialogue sample data with a prefix can be expressed as: Q1 Q2 …Q i-1 X, and the set of field markers corresponding to the current target field category is Q i , then after adding the set of field markers corresponding to the target field category to the prefix in the previous dialogue sample data, the dialogue sample data with a prefix corresponding to the target field category obtained is expressed as: Q1 Q2 … Q i-1 Q i X.

[0089] In another example, in the case where each field category only includes one field, the data form of the previous dialogue sample data with a prefix can be expressed as: P1 P2 … P i-1 X, and the field marker corresponding to the current target field is P i , then after adding the field marker corresponding to the target field to the prefix in the previous dialogue sample data, the dialogue sample data with a prefix corresponding to the target field obtained is expressed as: P1 P2 … P i-1 P i X.

[0090] It should be noted that when the current target field category is the first field category in the field category order, there is no previous field category, so there is no previous dialogue sample data with a prefix. Therefore, for the first field category, on the basis of the original dialogue sample data without a prefix, the field markers corresponding to each training object field in the target field category can be added as a prefix.

[0091] At 132-5, the obtained prefixed dialogue sample data and the labels corresponding to the respective training object fields indicated by the respective field tags in the prefix can be used to train the structured summary model.

[0092] In this example, the prefix of the dialogue sample data includes the target field category and the field tags corresponding to the respective fields in other field categories with a lower filling difficulty than the target field category.

[0093] In each sub-loop included in one round of training for the target field category, the loss is calculated based on the summaries for the respective training object fields output by the structured summary model and the labels corresponding to the respective training object fields, and the structured summary model is adjusted according to the loss. After the adjustment, the structured summary model performs the next sub-loop until the end of this round of training. At the end of this round of training, the operation of 140 is performed.

[0094] At 140, it is determined whether the training for all field categories in the field category order is completed. If not, the operation of 150 is performed; if so, the training can be ended to obtain the trained structured summary model.

[0095] At 150, when the training for the target field category is completed, the next field category in the field category order is determined as the target field category in the next round of training, and the operation of 130 is continued.

[0096] Through the technical solution of the embodiments of this specification, the structured summary model is gradually trained in the order of field categories with increasing filling difficulty. Compared with the method of training all fields together, the training effect is improved, and the reduction in the training effect caused by the learning conflicts between multiple fields that may exist when training all fields together is avoided.

[0097] Figure 4 The block diagram of an example of a device for training a structured summary model (hereinafter referred to as the structured summary model training device 400) according to the embodiments of this specification is shown.

[0098] As Figure 4 shown, the structured summary model training device 400 includes: a field extraction unit 410, a field classification unit 420, a model training unit 430, and a target field category determination unit 440.

[0099] The field extraction unit 410 can be configured to extract fields from the form to which the structured summary model is applied.

[0100] The field classification unit 420 can be configured to classify the extracted fields according to the filling difficulty of each field in the form, so as to obtain field categories with different filling difficulties. The field categories include at least two categories, and each field category includes at least one field.

[0101] In one example, the filling difficulty is determined according to the query condition corresponding to the field and / or the field filling type.

[0102] In one example, the field filling type includes binary classification type, multi-classification type, key information extraction type, and information abstraction type. Among them, the filling difficulties corresponding to the binary classification type, multi-classification type, key information extraction type, and information abstraction type increase in turn.

[0103] In one example, each field category includes only one field.

[0104] In one example, each field category is configured with a weight, and the weights corresponding to each field category are determined according to the corresponding filling difficulty.

[0105] The model training unit 430 can be configured to, for the target field category that is the current training target, use the target field category and other field categories with a lower filling difficulty than the target field category as training objects, and train the structured summary model using the dialogue sample data and the labels corresponding to each training object. Among them, the target field category is determined in sequence according to the field category order.

[0106] In one example, the model training unit 430 can also be configured to: for the target field category that is the current training target, add the field markers corresponding to each training object field in the target field category and other field categories with a lower filling difficulty than the target field category to the dialogue sample data as a prefix; and train the structured summary model using the dialogue sample data with the prefix and the labels corresponding to each training object field.

[0107] In one example, the model training unit 430 can also be configured to: for the target field category that is the current training target, obtain the previous dialogue sample data with a prefix used when training the previous field category in the field category order as the training target; add the field markers corresponding to each training object field in the target field category to the prefix in the previous dialogue sample data to obtain the dialogue sample data with a prefix corresponding to the target field category; and train the structured summary model using the obtained dialogue sample data with a prefix and the labels corresponding to each training object field indicated by each field marker in the prefix.

[0108] The target field category determination unit 440 may be configured to, when the training for the target field category is completed, determine the next field category in the field category order as the target field category for the next round of training, where the model training unit and the target field category determination unit train the structured summary model in the field category order until the training for all field categories in the field category order is completed.

[0109] As described above with reference to Figures 1 to 4 , embodiments of a method and an apparatus for training a structured summary model according to embodiments of this specification have been described.

[0110] The apparatus for training a structured summary model according to embodiments of this specification may be implemented in hardware, or may be implemented using software or a combination of hardware and software. Taking software implementation as an example, as a logically meaningful apparatus, it is formed by a processor of the device where it is located reading the corresponding computer program instructions in the memory into the memory for running. In embodiments of this specification, the apparatus for training a structured summary model may be implemented using an electronic device, for example.

[0111] Figure 5 A block diagram of an electronic device 500 for implementing a structured summary model training method according to embodiments of this specification is shown.

[0112] As Figure 5 shown, the electronic device 500 may include at least one processor 510, a memory (e.g., a non-volatile memory) 520, a memory 530, and a communication interface 540, and at least one processor 510, the memory 520, the memory 530, and the communication interface 540 are connected together via a bus 550. At least one processor 510 executes at least one computer-readable instruction stored or encoded in the memory (i.e., the above elements implemented in software form).

[0113] In one embodiment, computer-executable instructions are stored in the memory, which when executed cause at least one processor 510 to: extract fields from the form to which the structured summary model is applied; classify the extracted fields according to the filling difficulty of each field in the form to obtain field categories with different filling difficulties; train the structured summary model in the order of field categories with increasing filling difficulty until the training for all field categories in the field category order is completed: for the target field category as the current training target, use the target field category and other field categories with a lower filling difficulty corresponding to the target field category as training objects, and train the structured summary model using the dialogue sample data and the labels corresponding to each training object; and when the training for the target field category is completed, determine the next field category in the field category order as the target field category for the next round of training.

[0114] It should be understood that computer-executable instructions stored in a memory, when executed, cause at least one processor 510 to perform the various operations and functions described above in connection with the various embodiments of this specification. Figures 1 - 4 described operations and functions.

[0115] According to one embodiment, a program product such as a machine-readable medium is provided. The machine-readable medium may have instructions (i.e., the elements implemented in software as described above), which when executed by a machine, cause the machine to perform the various operations and functions described above in connection with Figures 1 - 4 described operations and functions.

[0116] Specifically, a system or device equipped with a readable storage medium may be provided, on which software program code for implementing the functions of any one of the above-described embodiments is stored, and the computer or processor of the system or device is caused to read and execute the instructions stored in the readable storage medium.

[0117] In this case, the program code read from the readable medium itself can implement the functions of any one of the above-described embodiments, so the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.

[0118] The computer program code required for the operations of the various parts of this specification can be written in any one or more programming languages, including object-oriented programming languages such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB, NET, and Python, conventional procedural programming languages such as C, Visual Basic 2003, Perl, COBOL2002, PHP, and ABAP, dynamic programming languages such as Python, Ruby, and Groovy, or other programming languages. The program code can run on a user's computer, or as a stand-alone software package on a user's computer, or partially on a user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer in any network form, such as a local area network (LAN) or a wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service, such as software as a service (SaaS).

[0119] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROM. Optionally, program code can be downloaded from a server computer or the cloud via a communication network.

[0120] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the figures do not necessarily require the particular order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0121] Not all steps and units in the above-mentioned processes and system structure diagrams are necessary, and some steps or units can be omitted according to actual needs. The execution order of each step is not fixed and can be determined as needed. The device structures described in the above embodiments can be physical structures or logical structures, that is, some units may be implemented by the same physical entity, or some units may be implemented by multiple physical entities respectively, or some components in multiple independent devices may be jointly implemented.

[0122] The term "exemplary" used throughout this specification means "serving as an example, instance, or illustration" and does not mean "preferred" or "advantageous" over other embodiments. For the purpose of providing an understanding of the described technology, the specific embodiments include specific details. However, these technologies can be implemented without these specific details. In some instances, well-known structures and devices are shown in block diagram form to avoid obscuring the concepts of the described embodiments.

[0123] The above has described in detail the optional embodiments of the embodiments of this specification in conjunction with the accompanying drawings. However, the embodiments of this specification are not limited to the specific details in the above embodiments. Within the technical concept scope of the embodiments of this specification, various simple modifications can be made to the technical solutions of the embodiments of this specification, and these simple modifications all fall within the protection scope of the embodiments of this specification.

[0124] The foregoing description of the content of this specification is provided to enable any person of ordinary skill in the art to make or use the content of this specification. Various modifications to the content of this specification will be apparent to those of ordinary skill in the art, and the general principles defined herein can also be applied to other variations without departing from the scope of protection of the content of this specification. Therefore, the content of this specification is not limited to the examples and designs described herein, but is consistent with the broadest scope that conforms to the principles and novel features disclosed herein.

Claims

1. A method for training a structured summary model, comprising: extracting fields from a form to which the structured summary model is applied; classifying the extracted fields according to the filling difficulty of each field in the form to obtain field categories with different filling difficulties, where the field categories include at least two categories, and each field category includes at least one field; training the structured summary model in the order of field categories with increasing filling difficulty until the training for all field categories in the field category order is completed: for a target field category that is the current training target, using the target field category and other field categories with a lower filling difficulty than the target field category as training objects, and training the structured summary model using dialogue sample data and labels corresponding to each training object, where the target field category is determined sequentially according to the field category order; and when the training for the target field category is completed, determining the next field category in the field category order as the target field category for the next round of training.

2. The method according to claim 1, wherein The filling difficulty is determined according to the query conditions corresponding to the field and / or the field filling type.

3. The method according to claim 2, wherein The field filling type includes binary classification type, multi-classification type, key information extraction type, and information abstraction type, where the filling difficulties corresponding to the binary classification type, the multi-classification type, the key information extraction type, and the information abstraction type increase in sequence.

4. The method according to claim 1, wherein, Each field category includes only one field.

5. The method according to claim 1, wherein, Each field category is configured with a weight, and the weights corresponding to each field category are determined according to the corresponding filling difficulty.

6. The method according to claim 1, wherein For a target field category that is the current training target, using the target field category and other field categories with a lower filling difficulty than the target field category as training objects, and training the structured summary model using dialogue sample data and labels corresponding to each training object includes: for a target field category that is the current training target, adding the field markers corresponding to each training object field in the target field category and other field categories with a lower filling difficulty than the target field category to the dialogue sample data as prefixes; and training the structured summary model using the dialogue sample data with prefixes and the labels corresponding to each training object field.

7. The method according to claim 1, wherein, For a target field category that is the current training target, using the target field category and other field categories with a lower filling difficulty than the target field category as training objects, and training the structured summary model using dialogue sample data and labels corresponding to each training object includes: for a target field category that is the current training target, obtaining the previous dialogue sample data with prefixes used when training the previous field category of the target field category in the field category order as the training target; adding the field markers corresponding to each training object field in the target field category to the prefix in the previous dialogue sample data to obtain the dialogue sample data with prefixes corresponding to the target field category; and The obtained prefixed dialogue sample data and the labels corresponding to the respective training object fields indicated by the respective field tags in the prefix are used to train the structured summary model.

8. The method according to claim 1, wherein, The trained structured summary model is applied to the scenario of insurance loss assessment.

9. An apparatus for training a structured summary model, comprising: a field extraction unit that extracts fields from a form to which the structured summary model is applied; a field classification unit that classifies the extracted fields according to the filling difficulty of each field in the form to obtain field categories with different filling difficulties, where the field categories include at least two categories, and each field category includes at least one field; a model training unit that, for a target field category as the current training target, uses the target field category and other field categories with a lower filling difficulty than the target field category as training objects, and uses the dialogue sample data and the labels corresponding to the respective training objects to train the structured summary model, where the target field category is determined sequentially according to the field category order; and a target field category determination unit that, when the training for the target field category is completed, determines the next field category in the field category order as the target field category in the next round of training, wherein the model training unit and the target field category determination unit train the structured summary model according to the field category order until the training for all field categories in the field category order is completed.

10. The apparatus according to claim 9, wherein, The model training unit is configured to: For a target field category as the current training target, add the field tags corresponding to the respective training object fields in the target field category and other field categories with a lower filling difficulty than the target field category to the dialogue sample data as a prefix; and Use the prefixed dialogue sample data and the labels corresponding to the respective training object fields to train the structured summary model.

11. The apparatus according to claim 9, wherein, The model training unit is configured to: For a target field category as the current training target, obtain the previous prefixed dialogue sample data used when training the previous field category of the target field category in the field category order as the training target; Add the field tags corresponding to the respective training object fields in the target field category to the prefix in the previous prefixed dialogue sample data to obtain the prefixed dialogue sample data corresponding to the target field category; and Use the obtained prefixed dialogue sample data and the labels corresponding to the respective training object fields indicated by the respective field tags in the prefix to train the structured summary model.

12. An electronic device, comprising: At least one processor, a memory coupled to the at least one processor, and a computer program stored on the memory, where the at least one processor executes the computer program to implement the method according to any one of claims 1-8.

13. A computer-readable storage medium storing a computer program, where the computer program, when executed by a processor, implements the method according to any one of claims 1-8.

14. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 8.

Images