CN113360624B

CN113360624B - Training method, response device, electronic device and storage medium

Info

Publication number: CN113360624B
Application number: CN202110747050.6A
Authority: CN
Inventors: 庞超; 王硕寰; 孙宇; 李芝
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2021-06-30
Filing date: 2021-06-30
Publication date: 2022-06-21
Anticipated expiration: 2041-06-30
Also published as: CN113360624A

Abstract

The disclosure discloses a training method, a response device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence, in particular to the field of natural language processing. The specific implementation scheme is as follows: constructing a question text based on a ring path, wherein the question text comprises reasoning content, and the ring path comprises at least three triples; and training the pre-training language model by using the problem text to obtain an inference model.

Description

Training method, response device, electronic device and storage medium

Technical Field

The present disclosure relates to the field of artificial intelligence technology, and more particularly, to the field of natural language processing. And more particularly, to a training method, a response method, an apparatus, an electronic device, and a storage medium.

Background

With the development of Natural Language Processing (NLP), the tasks are more finely divided, for example, the Natural Language Processing tasks may include knowledge inference, part-of-speech tagging, text classification, emotion analysis, machine translation, coreference resolution, and the like. The development of pre-training techniques based on pre-training models plays a crucial role in the above task. The core idea of the pre-training model is to pre-train a deep neural network model on a large data set to obtain model parameters, and then apply the trained model to various natural language processing tasks to avoid training from the beginning and reduce the need for labeled data.

With the continuous and deep research in the field of natural language processing, it is proved that the pre-training language model obtained by pre-training with large-scale non-labeled corpora can be helpful to the natural language processing task.

Disclosure of Invention

The disclosure provides a training method, a response device, an electronic device and a storage medium.

According to an aspect of the present disclosure, there is provided a method for training an inference model, including: constructing a question text based on a ring-shaped path, wherein the question text comprises reasoning content, and the ring-shaped path comprises at least three triples; and training the pre-training language model by using the problem text to obtain the inference model.

According to another aspect of the present disclosure, there is provided an answering method including: acquiring a target problem text; and inputting the target question text into an inference model to obtain an answer to the target question text, wherein the inference model is trained by the method.

According to another aspect of the present disclosure, there is provided a training apparatus of an inference model, including: the system comprises a construction module, a query module and a query module, wherein the construction module is used for constructing a question text based on a ring path, the question text comprises reasoning content, and the ring path comprises at least three triples; and the training module is used for training the pre-training language model by using the problem text to obtain the reasoning model.

According to another aspect of the present disclosure, there is provided a response apparatus including: the second acquisition module is used for acquiring a target question text; and an obtaining module, configured to input the target question text into an inference model, and obtain an answer to the target question text, where the inference model is trained by using the apparatus.

According to another aspect of the present disclosure, there is provided an electronic device including: at least one processor; and a memory communicatively coupled to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method.

According to another aspect of the present disclosure, there is provided a non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method as described above.

According to another aspect of the present disclosure, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the method as described above.

It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.

Drawings

The drawings are included to provide a better understanding of the present solution and are not to be construed as limiting the present disclosure. Wherein:

FIG. 1 schematically illustrates an exemplary system architecture for a training method, a response method, and an apparatus to which inference models may be applied, according to an embodiment of the disclosure;

FIG. 2 schematically illustrates a flow chart of a method of training an inference model according to an embodiment of the present disclosure;

FIG. 3 schematically illustrates a flow chart for building a question text based on a looped path according to an embodiment of the present disclosure;

FIG. 4 schematically illustrates a flow diagram for obtaining a looped path from a relational network, according to an embodiment of the disclosure;

FIG. 5 is a schematic diagram schematically illustrating a process of acquiring a looped path from a relational network according to an embodiment of the present disclosure;

FIG. 6 schematically illustrates a schematic diagram of a training process of an inference model according to an embodiment of the present disclosure;

FIG. 7 schematically illustrates a flow chart of a method of responding according to an embodiment of the present disclosure;

FIG. 8 schematically illustrates a block diagram of a training apparatus for inference models in accordance with an embodiment of the present disclosure;

FIG. 9 schematically shows a block diagram of a responding device according to an embodiment of the present disclosure; and

FIG. 10 illustrates a block diagram of an electronic device suitable for a training method or a response method of an inference model, according to an embodiment of the disclosure.

Detailed Description

Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, in which various details of the embodiments of the disclosure are included to assist understanding, and which are to be considered as merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

In the pre-training process, self-supervision training is completed by using large-scale unmarked corpora to obtain a pre-training language model after training. In the fine tuning process, supervised training is completed by using the labeled corpus to obtain a model which can be used for realizing a natural language processing task.

In the process of implementing the disclosed concept, it is found that the pre-trained language model has a poor effect in solving knowledge reasoning, which is caused by the lack of co-occurrence information between the cause and the result in the reasoning process. Knowledge reasoning can be understood as reasoning about unknown knowledge from existing knowledge. The knowledge may be characterized by triplets. The triples may be in the form of (head entity, relationship, tail entity), and the relationship characterizes the relationship between the head entity and the tail entity. Knowledge may include entities and/or relationships.

For example, if B is the father of A and C is the mother of A, then it can be inferred from the above knowledge that B and C are a couple. But if the mother who C is A is missing or the father of A is B, it is difficult to reason that B and C are a couple.

In order to improve the effect of the pre-training language model in knowledge inference, a mode of introducing external knowledge into the pre-training language model can be utilized, that is, an entity obtained based on the processing of the prior pre-training model can be embedded and introduced into the pre-training language model to improve the knowledge learning capability of the pre-training language model, and then the pre-training language model introduced with the entity embedding is trained to obtain an inference model capable of performing knowledge inference. External knowledge may include relationships and/or entities. For example, for the above question of reasoning that B and C are a couple relationship, the "mother" or "father" characterizing the relationship in the mother who lacks C or B is the father of a may refer to external knowledge.

A plurality of single triads can be used for carrying out early pre-training on the representation learning model to obtain an early pre-training model. And processing the single triple by utilizing an advanced pre-training model to obtain external knowledge represented by entity embedding. The representation learning model may include a TransE model, a TransR model, or a TransH model. After external knowledge represented by entity embedding is obtained, the pre-training language model is trained by combining with the training text, and the inference model is obtained.

In the process of realizing the concept disclosed by the invention, the fact that the training sample used in the process of obtaining the early pre-training model through early pre-training is a single triple is found, so that the early pre-training model has poor prediction effect on cross-triple groups, and the knowledge reasoning can be understood as the prediction of cross-triple questions, so that the prediction accuracy of the finally obtained inference model is low. In addition, it is found that, since the inference model is obtained by obtaining external knowledge first, introducing the external knowledge into the pre-trained language model, and training the pre-trained language model by using the training text and the external knowledge, the inference model is obtained through two processes, and errors may accumulate in the two processes and are easily affected by the quality of the external knowledge, so that the prediction accuracy of the inference model is low. Meanwhile, the introduction of external knowledge enables the reasoning ability of the reasoning model to be indirectly learned rather than directly learned, so that the training complexity is increased.

To this end, it was found that the prediction accuracy of inference models can be improved and the training complexity reduced by enabling prediction across tuples and ways to enable direct learning of the models. To achieve prediction across triplets, a circular path may be used for discovery, since a circular path may include at least three triplets, and thus prediction across triplets, i.e., multi-hop inference, may be achieved. In order to realize direct learning, the training text can carry external knowledge, and the pre-training language model can be directly trained by using the training text. Therefore, the embodiment of the disclosure provides a scheme for constructing a problem text including reasoning content based on a ring path, and training a pre-training model by using the problem text to obtain a reasoning model capable of being used for knowledge reasoning.

The disclosed embodiments provide a training method, a response device, an electronic device, a non-transitory computer readable storage medium storing computer instructions, and a computer program product for an inference model. The training method of the data generation model comprises the following steps: and constructing a problem text based on the annular path, wherein the problem text comprises reasoning content, the annular path comprises at least three triples, and the problem text is utilized to train the pre-training language model to obtain a reasoning model.

Fig. 1 schematically illustrates an exemplary system architecture of a training method, a response method, and an apparatus to which inference models may be applied, according to embodiments of the disclosure.

It should be noted that fig. 1 is only an example of a system architecture to which the embodiments of the present disclosure may be applied to help those skilled in the art understand the technical content of the present disclosure, and does not mean that the embodiments of the present disclosure may not be applied to other devices, systems, environments or scenarios. For example, in another embodiment, an exemplary system architecture to which the training method, the response method, and the apparatus of the inference model can be applied may include a terminal device, but the terminal device may implement the training method, the response method, and the apparatus of the inference model provided in the embodiments of the present disclosure without interacting with a server.

As shown in fig. 1, the system architecture 100 according to this embodiment may include

terminal devices

101, 102, 103, a network 104 and a server 105. The network 104 serves as a medium for providing communication links between the

terminal devices

101, 102, 103 and the server 105. Network 104 may include various connection types, such as wired and/or wireless communication links, and so forth.

The user may use the

terminal devices

101, 102, 103 to interact with the server 105 via the network 104 to receive or send messages or the like. The

terminal devices

101, 102, 103 may have installed thereon various communication client applications, such as a knowledge reading application, a web browser application, a search application, an instant messaging tool, a mailbox client, and/or social platform software, etc. (by way of example only).

The

terminal devices

101, 102, 103 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop portable computers, desktop computers, and the like.

The server 105 may be a server providing various services, such as a background management server (for example only) providing support for content browsed by the user using the

terminal devices

101, 102, 103. The backend management server may analyze and process the received data such as the user request, and feed back a processing result (for example, a web page, information, or data obtained or generated according to the user request) to the terminal device.

The Server 105 may be a cloud Server, which is also called a cloud computing Server or a cloud host, and is a host product in a cloud computing service system, so as to solve the defects of high management difficulty and weak service extensibility in a conventional physical host and a VPS (Virtual Private Server). Server 105 may also be a server of a distributed system or a server that incorporates a blockchain.

It should be noted that the training method and the response method of the inference model provided by the embodiments of the present disclosure may be generally executed by the

terminal device

101, 102, or 103. Correspondingly, the training device and the response device of the inference model provided by the embodiment of the present disclosure may also be disposed in the

terminal device

101, 102, or 103.

Alternatively, the training method and the response method of the inference model provided by the embodiments of the present disclosure may be generally performed by the server 105. Accordingly, the training device and the response device of the inference model provided by the embodiment of the present disclosure may be generally disposed in the server 105. The training method and the response method of the inference model provided by the embodiments of the present disclosure may also be performed by a server or a server cluster that is different from the server 105 and is capable of communicating with the

terminal devices

101, 102, 103 and/or the server 105. Correspondingly, the training device and the responding device of the inference model provided by the embodiment of the present disclosure may also be disposed in a server or a server cluster different from the server 105 and capable of communicating with the

terminal devices

101, 102, 103 and/or the server 105.

For example, the server 105 constructs a question text based on the loop path, wherein the question text includes reasoning content, and trains the pre-trained language model with the question text to obtain a reasoning model. Or a server cluster capable of communicating with the

terminal devices

101, 102, 103 and/or the server 105 constructs a question text based on a ring path, and trains the pre-trained language model by using the question text to obtain an inference model.

The server 105 obtains the target question text, inputs the target question text into the inference model, and obtains an answer for the target question text. Or a server cluster capable of communicating with the

terminal devices

101, 102, 103 and/or the server 105 acquires the target question text, inputs the target question text into the inference model, and obtains an answer to the target question text.

It should be understood that the number of terminal devices, networks, and servers in fig. 1 is merely illustrative. There may be any number of terminal devices, networks, and servers, as desired for an implementation.

Fig. 2 schematically shows a flow chart of a method of training an inference model according to an embodiment of the present disclosure.

As shown in FIG. 2, the method 200 includes operations S210-S220.

In operation S210, a question text is constructed based on a circular path, wherein the question text includes inferential content and the circular path includes at least three triples.

In operation S220, the pre-trained language model is trained using the question text to obtain an inference model.

According to an embodiment of the present disclosure, the looped path may comprise at least three triplets and each entity appears in two adjacent triplets, i.e. two adjacent triplets comprise one and the same entity. For example, the three triplets are (a, father, B), (a, mother, C) and (B, wife, C), respectively. Entity A appears in (A, father, B) and (A, mother, C), respectively. Entity B appears in (A, father, B) and (B, wife, C), respectively. Entity C appears in (A, mother, C) and (B, wife, C), respectively. The three triplets form a circular path, and each two triplets in the circular path are adjacent triplets. The number of triplets included in the ring path may be configured according to actual service requirements, and is not limited herein.

According to embodiments of the present disclosure, the question text may include inferential content, which may refer to knowledge that requires inference. For example, the inferential content may include inference questions and inference conditions needed to reason about the inference questions. Since the circular path includes at least three triples, and the question text is constructed based on the circular path, multi-hop knowledge reasoning, i.e., prediction across triples, can be realized based on the question text.

According to an embodiment of the present disclosure, the pre-training language model may include an autoregressive pre-training language model of unidirectional feature characterization, an autoregressive pre-training language model of bidirectional feature characterization, or an autoregressive pre-training language model of bidirectional feature characterization. The bi-directional feature characterized self-encoding pre-training language model may include a BERT (i.e., Bidjective Encode expressions from transformations) model, an ERNIE model, or a MASS (i.e., MAsked Sequence to Sequence pre-training) model. The pre-training language model may be selected according to actual business requirements, and is not limited herein. For example, the pre-trained language model may be a BERT model. The structure of the BERT model is a multi-layer bidirectional transform encoder. Because the BERT model aims to pretrain deep bidirectional representation by jointly adjusting left and right contexts in all layers, an additional output layer is needed, the representation of the pretrained BERT model can be finely adjusted, and then the model is created for different tasks without modifying a large number of model structures for specific tasks. Meanwhile, since the BERT model can convert the distance between two objects at arbitrary positions by an attention mechanism, it has a better Natural Language Understanding (NLU) capability. Because the BERT model is obtained by training a large number of training samples, the BERT model has stronger semantic representation capability and universality.

According to the embodiment of the disclosure, after the circular path is obtained, the question text can be constructed based on the circular path, that is, the triplets included in the circular path can be constructed into the question text including the speculative content, the pre-training language model is trained by using the question text, the inference model capable of responding to the question in the question text is obtained, and the result obtained according to the inference model can be understood as the answer to the question. The method realizes the conversion of the multi-hop knowledge reasoning into the training task aiming at the pre-training language model, and is simple and effective.

According to the embodiment of the disclosure, the problem text including the reasoning content is constructed based on the annular path, the annular path includes at least three triples, and the problem text is used for training the pre-training language model to obtain the reasoning model. Since the circular path may comprise at least three triplets, a prediction across triplets is achieved, i.e. a multi-hop inference is achieved. Moreover, the problem text is constructed based on the annular path, the annular path comprises the triples, and the triples can provide external knowledge, so that the problem text carries the external knowledge, and the pre-training language model can be directly trained by using the training text. Therefore, the prediction accuracy of the inference model is improved, and the training complexity is reduced.

According to an embodiment of the present disclosure, the method for training the inference model may further include the following operations.

A ring path is obtained from a relational network.

According to embodiments of the present disclosure, a relational network may comprise a plurality of nodes and a plurality of edges, each edge may be used to characterize a relationship between two nodes that are connected, i.e. each edge may be used to characterize a relationship in each triplet. A node may characterize an entity in a triplet.

According to the embodiment of the disclosure, the ring path can be acquired from the relational network according to a preset wandering strategy. The preset walk policy may include a policy that makes the start node and the end node in the migrated relationship network the same node. A policy may also be included to traverse the target node, which may refer to the node associated with the wandering-to node in performing the wandering operation. According to the preset migration policy, acquiring the ring path from the relationship network may include: and executing the wandering operation in the relational network according to a preset wandering strategy, and obtaining the annular path according to the relationship between the wandering node and the node in the wandering operation executing process.

According to an embodiment of the present disclosure, operation S220 may include the following operations.

And inputting the question text into a pre-training language model to obtain an answer aiming at the question text. And obtaining an output value by using the label and the answer corresponding to the question text based on the loss function. And adjusting the model parameters of the pre-training language model according to the output value until the output value is converged. And determining the pre-training language model obtained under the condition that the output value is converged as an inference model.

According to an embodiment of the present disclosure, the number of question texts may include a plurality. The label corresponding to the question text may characterize the real answer.

According to the embodiment of the disclosure, the pre-training language model may be trained in a supervised manner, that is, for each question text of a plurality of question texts, the question text is input into the pre-training language model, and an answer corresponding to the question text is obtained. The answer corresponding to the question text is the predicted answer. And inputting the answer and the label corresponding to the question text into a loss function to obtain an output value. Based on the above, an output value corresponding to each of all the question texts is obtained. And determining whether the output value is converged, and in the case that the output value is determined not to be converged, adjusting model parameters of the pre-training language model according to the output value, and repeatedly executing the operations of determining the output value and determining whether the output value is converged until the output value is converged.

Referring to fig. 3 to fig. 6, a training method of a recommendation model according to an embodiment of the present disclosure is further described with reference to specific embodiments.

FIG. 3 schematically illustrates a flow chart for building a question text based on a looped path according to an embodiment of the present disclosure.

As shown in fig. 3, the method 300 includes operations S311 to S312.

In operation S311, a problem triplet and a condition triplet are determined from the looped path.

In operation S312, a question text is constructed from the question triples and the condition triples.

According to embodiments of the present disclosure, a problem triplet may refer to a triplet that includes a problem to be inferred. A conditional triplet may refer to a triplet that includes conditions for reasoning about a problem to be inferred. For each triplet, the triplet may be either a problem triplet or a conditional triplet. The number of problem triplets may include one or more, and the number of conditional triplets may include one or more. For each problem triplet, the problem triplet may correspond to one or more conditional triplets.

For example, a circular path may include three triplets, namely, (a, father, B), (a, mother, C) and (B, wife, C). The problem and condition triplets may be one of: (a, father, B) is a problem triplet, (a, mother, C) and/or (B, wife, C) is a conditional triplet. (a, mother, C) is a problem triplet, (a, father, B) and/or (B, wife, C) is a conditional triplet. (B, wife, C) is a problem triplet, (A, father, B) and/or (A, mother, C) is a conditional triplet.

According to the embodiment of the disclosure, the triples included in the loop path can be divided into the problem triples and the condition triples according to the triples division strategy, and then the problem triples and the condition triples are constructed into the problem text. The triple partitioning policy may include a policy on how to determine problem triples and condition triples. For example, the triple partitioning policy may include a policy of determining that a triple is a problem triple or a condition triple according to whether a problem to be inferred is included and a condition required to determine the problem to be inferred. The above construction method of the question text is only an exemplary embodiment, but is not limited thereto, and may include construction methods known in the art as long as the construction of the question text can be implemented.

According to an embodiment of the present disclosure, operation S311 may include the following operations.

The problem object is determined from the looped path. And determining the triples corresponding to the problem objects as problem triples. And determining part or all of the triples included by the loop path except the problem triples as the conditional triples.

According to an embodiment of the disclosure, the problem object may refer to a problem to be inferred, and the problem to be inferred may include entities or relationships in the triplets, that is, the entities or relationships in the triplets may all become the problem to be inferred. For example, the entity "a", the entity "B", or the relationship "mother" in the triplet "(a, mother, C) included in the circular path may be the question to be inferred.

According to an embodiment of the present disclosure, after determining the problem object from the looped path, the triplet corresponding to the problem object may be determined as a problem triplet, i.e., the triplet including the problem object may be determined as a problem triplet. After the problem triplets are obtained, other triplets included in the loop path besides the problem triplets may be determined as conditional triplets, and the other triplets may include all or part of the triplets except the problem triplets.

For example, a circular path may include four triplets, namely, (E, Google, F), (F, sister, G), (G, Google, H), and (H, sister, E). The problem triplet is (H, sister, E), and the conditional triplet may include at least one of: (E, Google, F), (F, sister, G), and (G, Google, H).

Operation S312 may include the following operations according to an embodiment of the present disclosure.

And constructing an inference problem according to the problem triples. And constructing an inference condition according to the condition triples, wherein the inference condition is a condition for reasoning the inference problem. And constructing a question text according to the reasoning question and the reasoning condition.

According to an embodiment of the present disclosure, an inference problem may refer to a problem characterized according to a first preset format. The inference condition may refer to a condition characterized according to a second preset format. Inference conditions can be used to reason about inference problems. The first preset format and the second preset format may be configured according to actual service requirements, and are not limited herein. For example, the first preset format may be "x is x", "x of" is x ", or" x of "is x. The second preset format may be "x × is ×". "x" may characterize what is known in the triplet, and "may characterize the problem object. It should be noted that for ease of understanding, the symbol "x" is used merely to indicate that information is contained at this location and should not be considered a limitation on the first and second preset formats.

According to embodiments of the present disclosure, after the inference question and the inference condition are constructed, the inference question and the inference condition may be constructed into a question text, i.e., the question text may include the inference question and the inference condition.

For example, the inference question is "B is C", the inference conditions include "A's father is B" and "A's mother is C", the question text may be "if" A's father is B "and" A's mother is C ", then" B ' is C ".

According to an embodiment of the present disclosure, constructing inference questions according to the question triplets may include: and according to a preset problem template, representing the problem triple into an inference problem.

Constructing inference conditions from the conditional triples may include: and representing the condition triples into inference conditions according to a preset condition template.

According to an embodiment of the present disclosure, a preset problem template may refer to a template of an inference problem that is capable of characterizing a problem triplet into a preset format. The preset condition template may refer to a template of the inference condition capable of characterizing the condition triplet into a preset format.

According to the embodiment of the disclosure, the problem triple can be characterized as the inference problem according to the form of the preset problem template, that is, each part included in the problem triple can be converted according to the template corresponding to the part in the preset problem template, so as to obtain the inference problem characterized according to the form of the preset problem template.

According to the embodiment of the disclosure, the condition triple can be characterized as the inference condition according to the form of the preset condition template, that is, each part included in the condition triple can be converted according to the template corresponding to the part in the preset condition template, so as to obtain the inference condition characterized according to the form of the preset condition template.

For example, the preset question template may be "x is x", or "x is x". The preset condition template may be "×" × is × ". "x" may characterize what is known in the triplet, and "may characterize the problem object. It should be noted that the symbol "x" is used only to indicate that information is contained at this position for ease of understanding, and should not be considered as a limitation to the preset question template and the preset condition template.

Fig. 4 schematically illustrates a flow chart of acquiring a looped path from a relational network according to an embodiment of the present disclosure.

As shown in fig. 4, the method 400 includes operations S401 to S403.

In operation S401, a preset node is selected from the relational network as a start node.

In operation S402, starting from the start node, a walk operation is performed in the relational network based on the random walk policy until the end node is the start node.

In operation S403, a path formed by performing the walk completion operation is determined as a circular path.

According to an embodiment of the present disclosure, the relationship network may include a plurality of nodes, and the plurality of nodes may include a preset node. The preset node may be randomly selected, or may be determined according to a preset selection condition. The preset selection condition may be a condition determined according to a degree of complexity of a relationship of the node. Relationship complexity may refer to the complexity of the relationship between a node and other nodes.

According to an embodiment of the present disclosure, the random walk policy may refer to a policy that enables a start node and a stop node that are walked to be the same node and a node associated with a node that has been walked to also be walked to in a process of performing random walk in a relational network.

According to the embodiment of the present disclosure, a walk operation may be performed in the relationship network based on a random walk policy starting from the start node until the walk is again to the start node, in which case, the start node is the termination node, that is, until the termination node is the start node. And in the case that the termination node is the starting node, representing that the execution of the walk operation is completed. And determining a path formed by executing the finish walking operation as a ring path.

Fig. 5 schematically shows a schematic diagram of a process of acquiring a ring path from a relational network according to an embodiment of the present disclosure.

As shown in fig. 5, a process 500 of obtaining a ring path from a relational network may include a relational network 501 and a ring path 502. The relationship network 501 includes

nodes

5010, 5011, and 5012. The node 5010 is a pre-set node, i.e., the node 5010 is an initial node. Starting from the start node, the walk operation is performed in the relationship network 501 based on the random walk policy until the end node is the start node, that is, the walk operation is performed in the direction indicated by the arrow in fig. 5. A path formed by performing the finish walking operation, that is, a path formed by the node 5010 (i.e., the start node) → the node 5011 → the node 5012 → the end node (i.e., the node 5010) is determined as the circular path 502. The ring path 502 includes a node 5010, a node 5011, a node 5012, and a relationship between each two nodes.

Fig. 6 schematically shows a schematic diagram of a training process of an inference model according to an embodiment of the present disclosure.

As shown in fig. 6, the training process 600 of the inference model is as follows: a ring path 602 is obtained from the relationship network 601, and the ring path 602 may include a plurality. Relationship network 601 may include node 6010, node 6011, node 6012. The ring path 602 may include a node 6010, a node 6011, a node 6012, and a relationship between each two nodes. The loop path 602 can be obtained from 601 in the same manner as in fig. 5, and will not be described herein. Node 6010, node 6011, and the relationship parent therebetween may constitute a triple (node 6010, parent, node 6011). Node 6010, node 6012, and the relationship mother between the two may constitute a triplet (node 6010, mother, node 6012). Node 6011, node 6012 and the relationship between the two wives may constitute a triple (node 6011, wife, node 6012)

Question text 603 may be constructed from the looped path 602. Problem text 603 may be characterized as "parent of node 6010 is node 6011 and mother of node 6010 is node 6012", then "parent of node 6011 is node 6012". The question text 603 may include a plurality. The pre-trained language model 604 is trained using the question text 603, resulting in an inference model 605.

Fig. 7 schematically shows a flow chart of a reply method according to an embodiment of the present disclosure.

As shown in fig. 7, the method 700 includes operations S710 to S720.

In operation S710, a target question text is acquired.

In operation S720, the target question text is input into an inference model, which is trained by using a training method of the inference model according to the embodiments of the present disclosure, to obtain an answer to the target question text.

According to the embodiment of the disclosure, the target problem text can be constructed according to the target inference problem and the target inference condition. The target reasoning problem can be obtained by representing a target problem triple according to a preset problem template. The target inference condition may be obtained by characterizing the target condition triple according to a preset condition template. The target issue triplets may be triplets corresponding to target issue objects. The target condition triplets may be some or all of the triplets in the target looped path other than the target problem triplets.

According to the embodiment of the disclosure, the target question text can be input into the inference model obtained by training according to the training method of the inference model described in the embodiment of the disclosure, so as to obtain an answer corresponding to the target question text.

According to the embodiment of the disclosure, the answer to the target question text is obtained by inputting the target question text into the reasoning model, and the reasoning model is obtained by training the pre-training language model by constructing the question text including the reasoning content based on the annular path. Since the circular path may comprise at least three triplets, a prediction across triplets is achieved, i.e. a multi-hop inference is achieved. In addition, the problem text is constructed based on the annular path, the annular path comprises the triples, and the triples can provide external knowledge, so that the problem text carries the external knowledge, and the pre-training language model can be directly trained by using the training text. Therefore, the prediction accuracy of the inference model is improved, and the training complexity is reduced.

Fig. 8 schematically shows a block diagram of a training apparatus of an inference model according to an embodiment of the present disclosure.

As shown in fig. 8, the training apparatus 800 for inference model may include a building model 810 and a training module 820.

A building module 810, configured to build a question text based on a circular path, where the question text includes reasoning content and the circular path includes at least three triples.

And the training module 820 is used for training the pre-training language model by using the question text to obtain an inference model.

According to an embodiment of the disclosure, the building module 810 may include a first determining submodule and a building submodule.

And the first determining submodule is used for determining the problem triples and the condition triples from the annular path.

And the construction sub-module is used for constructing the problem text according to the problem triples and the condition triples.

According to an embodiment of the present disclosure, the first determination submodule may include a first determination unit, a second determination unit, and a third determination unit.

A first determining unit for determining the problem object from the looped path.

And the second determining unit is used for determining the triples corresponding to the problem objects as the problem triples.

And the third determining unit is used for determining part or all triples, except the problem triples, included in the loop path as the conditional triples.

According to an embodiment of the present disclosure, the building submodule may include a first building element, a second building element, and a third building element.

And the first construction unit is used for constructing the reasoning problem according to the problem triplet.

And the second construction unit is used for constructing inference conditions according to the condition triples, wherein the inference conditions are conditions for reasoning the inference problem.

And the third construction unit is used for constructing the question text according to the reasoning question and the reasoning condition.

According to an embodiment of the present disclosure, the first building unit may comprise a first building subunit.

And the first construction subunit is used for representing the problem triple into the inference problem according to a preset problem template.

The second building element may comprise a second building subunit.

And the second construction subunit is used for representing the condition triples into inference conditions according to the preset condition template.

According to an embodiment of the present disclosure, the training apparatus 800 of the inference model may further include a first obtaining module.

The first acquisition module is used for acquiring the annular path from the relational network.

According to an embodiment of the present disclosure, the first obtaining module may include a selecting sub-module, an executing sub-module, and a second determining sub-module.

And the selection submodule is used for selecting a preset node from the relational network as an initial node.

And the execution submodule is used for executing the walk operation in the relational network based on the random walk strategy from the starting node until the termination node is the starting node.

And the second determining submodule is used for determining the path formed by executing the walk-away operation as a ring-shaped path.

According to an embodiment of the present disclosure, the training module 820 may include a first obtaining sub-module, a second obtaining sub-module, an adjusting sub-module, and a third determining sub-module.

And the first obtaining submodule is used for inputting the question text into the pre-training language model to obtain an answer aiming at the question text.

And the second obtaining submodule is used for obtaining an output value by using the label and the answer corresponding to the question text based on the loss function.

And the adjusting sub-module is used for adjusting the model parameters of the pre-training language model according to the output value until the output value is converged.

And the third determining submodule is used for determining the pre-training language model obtained under the condition that the output value is converged as an inference model.

Fig. 9 schematically shows a block diagram of a responding device according to an embodiment of the present disclosure.

As shown in fig. 9, the responding device 900 may include a second obtaining module 910 and obtaining 920.

And a second obtaining module 910, configured to obtain the target question text.

An obtaining module 920, configured to input the target question text into a reasoning model, and obtain an answer to the target question text, where the reasoning model is trained by using a training apparatus of the reasoning model according to the embodiment of the present disclosure.

The present disclosure also provides an electronic device, a readable storage medium, and a computer program product according to embodiments of the present disclosure.

According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method as described above.

According to an embodiment of the present disclosure, a non-transitory computer readable storage medium having stored thereon computer instructions for causing a computer to perform the method as described above.

According to an embodiment of the disclosure, a computer program product comprising a computer program which, when executed by a processor, implements the method as described above.

FIG. 10 illustrates a block diagram of an electronic device suitable for a training method or a response method of an inference model, according to an embodiment of the disclosure. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the disclosure described and/or claimed herein.

As shown in fig. 10, the electronic device 1000 includes a computing unit 1001 that can perform various appropriate actions and processes according to a computer program stored in a Read Only Memory (ROM)1002 or a computer program loaded from a storage unit 1008 into a Random Access Memory (RAM) 1003. In the RAM 1003, various programs and data necessary for the operation of the electronic apparatus 1000 can also be stored. The calculation unit 1001, the ROM 1002, and the RAM 1003 are connected to each other by a bus 1004. An input/output (I/O) interface 1005 is also connected to bus 1004.

A number of components in the electronic device 1000 are connected to the I/O interface 1005, including: an input unit 1006 such as a keyboard, a mouse, and the like; an output unit 1007 such as various types of displays, speakers, and the like; a storage unit 1008 such as a magnetic disk, an optical disk, or the like; and a communication unit 1009 such as a network card, a modem, a wireless communication transceiver, or the like. The communication unit 1009 allows the electronic device 1000 to exchange information/data with other devices through a computer network such as the internet and/or various telecommunication networks.

Computing unit 1001 may be a variety of general and/or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various dedicated Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, and so forth. The calculation unit 1001 executes the respective methods and processes described above, such as the training method or the response method of the inference model. For example, in some embodiments, the training method or the response method of the inference model may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 1008. In some embodiments, part or all of the computer program may be loaded and/or installed onto electronic device 1000 via ROM 1002 and/or communications unit 1009. When the computer program is loaded into the RAM 1003 and executed by the computing unit 1001, one or more steps of the training method or the answering method of the inference model described above may be performed. Alternatively, in other embodiments, the computing unit 1001 may be configured in any other suitable way (e.g. by means of firmware) to perform a training method or a response method of the inference model.

Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuitry, Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), system on a chip (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.

Program code for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program codes, when executed by the processor or controller, cause the functions/operations specified in the flowchart and/or block diagram to be performed. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.

In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), Wide Area Networks (WANs), and the Internet.

The computer system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server may be a cloud server, a server of a distributed system, or a server combining a blockchain.

It should be understood that various forms of the flows shown above may be used, with steps reordered, added, or deleted. For example, the steps described in the present disclosure may be executed in parallel, sequentially, or in different orders, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, and the present disclosure is not limited herein.

The above detailed description should not be construed as limiting the scope of the disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present disclosure should be included in the scope of protection of the present disclosure.

Claims

1. A method for training an inference model comprises the following steps:

constructing a question text based on a circular path, wherein the question text comprises reasoning content, and the circular path comprises at least three triples; and

training a pre-training language model by using the problem text to obtain the inference model;

wherein, the problem text is constructed based on the ring path, and the method comprises the following steps:

determining a problem triplet and a condition triplet from the looped path; and

and constructing the question text according to the question triples and the condition triples.

2. The method of claim 1, wherein the determining a problem triplet and a condition triplet from the looped path comprises:

determining a problem object from the looped path;

determining a triple corresponding to the problem object as the problem triple; and

and determining part or all triples except the problem triples included by the annular path as the condition triples.

3. The method of claim 1 or 2, wherein the constructing the question text from the question triplets and the condition triplets comprises:

constructing an inference problem according to the problem triplet;

constructing an inference condition according to the condition triplet, wherein the inference condition is a condition for inferring the inference problem; and

and constructing the question text according to the reasoning question and the reasoning condition.

4. The method of claim 3, wherein said constructing inference questions from said question triples comprises:

according to a preset problem template, representing the problem triple into the inference problem;

the method for constructing the inference condition according to the condition triple comprises the following steps:

and characterizing the condition triplets into the reasoning conditions according to a preset condition template.

5. The method of claim 1 or 2, further comprising:

and acquiring the ring path from the relation network.

6. The method of claim 5, wherein said obtaining the looped path from a relationship network comprises:

selecting a preset node from the relational network as an initial node;

starting from the starting node, executing a walking operation in the relational network based on a random walking strategy until a termination node is the starting node; and

and determining a path formed by executing the walk operation as the annular path.

7. The method of claim 1 or 2, wherein the training of a pre-trained language model with the question text to obtain the inference model comprises:

inputting the question text into the pre-training language model to obtain an answer aiming at the question text;

obtaining an output value by using the label and the answer corresponding to the question text based on a loss function;

adjusting model parameters of the pre-training language model according to the output value until the output value is converged; and

and determining the pre-training language model obtained under the condition that the output value is converged as the inference model.

8. A method of responding, comprising:

acquiring a target problem text; and

inputting the target question text into an inference model to obtain an answer to the target question text,

wherein the inference model is trained using the method according to any one of claims 1 to 7.

9. An apparatus for training inference models, comprising:

the system comprises a construction module, a query module and a query module, wherein the construction module is used for constructing a question text based on a ring path, the question text comprises reasoning content, and the ring path comprises at least three triples; and

the training module is used for training a pre-training language model by using the question text to obtain the inference model;

wherein the building block comprises:

the first determining submodule is used for determining a problem triplet and a condition triplet from the annular path; and

and the construction sub-module is used for constructing the question text according to the question triples and the condition triples.

10. The apparatus of claim 9, wherein the first determination submodule comprises:

a first determination unit configured to determine a problem object from the loop path;

a second determining unit, configured to determine a triple corresponding to the problem object as the problem triple; and

and a third determining unit, configured to determine, as the conditional triplet, some or all triplets included in the looped path except the problem triplet.

11. The apparatus of claim 9 or 10, wherein the building module comprises:

the first construction unit is used for constructing an inference problem according to the problem triplet;

the second construction unit is used for constructing inference conditions according to the condition triples, wherein the inference conditions are conditions for inferring the inference problem; and

12. The apparatus of claim 11, wherein the first building unit comprises:

the first construction subunit is used for representing the problem triple into the inference problem according to a preset problem template;

the second building element comprising:

and the second construction subunit is used for representing the condition triple into the inference condition according to a preset condition template.

13. The apparatus of claim 9 or 10, further comprising:

and the first acquisition module is used for acquiring the annular path from the relational network.

14. The apparatus of claim 13, wherein the first obtaining means comprises:

the selection submodule is used for selecting a preset node from the relational network as an initial node;

the execution submodule is used for executing the walk operation in the relational network based on a random walk strategy from the starting node until the ending node is the starting node; and

and the second determining submodule is used for determining a path formed by executing the walking operation as the annular path.

15. A transponder device comprising:

the second acquisition module is used for acquiring a target question text; and

an obtaining module, configured to input the target question text into an inference model to obtain an answer to the target question text,

wherein the inference model is trained using an apparatus according to any one of claims 9 to 14.

16. An electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein the content of the first and second substances,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 7 or claim 8.

17. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-7 or 8.

18. A computer program product comprising a computer program which, when executed by a processor, implements the method of any one of claims 1 to 7 or claim 8.