CN108804627B - Information acquisition method and device - Google Patents
Information acquisition method and device Download PDFInfo
- Publication number
- CN108804627B CN108804627B CN201810551681.9A CN201810551681A CN108804627B CN 108804627 B CN108804627 B CN 108804627B CN 201810551681 A CN201810551681 A CN 201810551681A CN 108804627 B CN108804627 B CN 108804627B
- Authority
- CN
- China
- Prior art keywords
- key content
- candidate
- sentence
- optimal
- key
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Abstract
The embodiment of the invention provides an information acquisition method and device, and belongs to the technical field of natural language processing. The method comprises the following steps: respectively inputting the query text and the reply text matched with the query text into the N key content calculation models, and acquiring candidate key content output by each key content calculation model; the key content calculation model is obtained by training based on sample query texts, sample reply texts and sample key contents in the sample reply texts, and each candidate key content is extracted from the reply texts; and acquiring the optimal key content according to each candidate key content, and using the optimal key content as a response corresponding to the query text. Because the output results of the N key content calculation models can be fused, the problem that a single model has deviation in the training process is effectively avoided, the reliability and the accuracy of the reply content can be improved, and the interaction experience of a user in question-answer interaction with equipment is improved.
Description
Technical Field
The embodiment of the invention relates to the technical field of natural language processing, in particular to an information acquisition method and device.
Background
In recent years, with the development of artificial intelligence-related disciplines, particularly computational linguistics, various question answering systems and dialogue robots have come into operation, and people can communicate with devices in a natural language to acquire required information. In the related art, for a question of a user, the content of a response corresponding to the question of the user is determined through a model. Since a single model is inevitably biased in the training process, it is difficult to completely fit the entire training data distribution, resulting in low reliability of the response content.
Disclosure of Invention
In order to solve the above problems, embodiments of the present invention provide an information acquisition method and apparatus that overcome the above problems or at least partially solve the above problems.
According to a first aspect of the embodiments of the present invention, there is provided an information acquisition method, including:
respectively inputting the query text and the reply text matched with the query text into the N key content calculation models, and acquiring candidate key content output by each key content calculation model; the key content calculation model is obtained by training based on sample query texts, sample reply texts and sample key contents in the sample reply texts, each candidate key content is extracted from the reply texts, and N is a positive integer greater than 1;
and acquiring the optimal key content according to each candidate key content, and using the optimal key content as a response corresponding to the query text.
According to the method provided by the embodiment of the invention, the query text and the reply text matched with the query text are respectively input into the N key content calculation models, and the candidate key content output by each key content calculation model is obtained. And acquiring the optimal key content according to each candidate key content, and using the optimal key content as a response corresponding to the query text. Because the output results of the N key content calculation models can be fused, the problem that a single model has deviation in the training process and is difficult to completely fit all training data distribution is effectively avoided, the reliability and the accuracy of reply content can be improved, and the interaction experience of a user in question-answer interaction with equipment is improved.
According to a second aspect of the embodiments of the present invention, there is provided an information acquisition apparatus including:
the candidate key content acquisition module is used for respectively inputting the query text and the reply text matched with the query text into the N key content calculation models and acquiring candidate key content output by each key content calculation model; the key content calculation model is obtained by training based on sample query texts, sample reply texts and sample key contents in the sample reply texts, each candidate key content is extracted from the reply texts, and N is a positive integer greater than 1;
and the optimal key content acquisition module is used for acquiring optimal key content according to each candidate key content and taking the optimal key content as a response corresponding to the query text.
According to a third aspect of embodiments of the present invention, there is provided an information acquisition apparatus including:
at least one processor; and
at least one memory communicatively coupled to the processor, wherein:
the memory stores program instructions executable by the processor, the processor calling the program instructions being capable of performing the information obtaining method provided by any of the various possible implementations of the first aspect.
According to a fourth aspect of the present invention, there is provided a non-transitory computer readable storage medium storing computer instructions for causing a computer to perform the information acquisition method provided in any one of the various possible implementations of the first aspect.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory and are not restrictive of embodiments of the invention.
Drawings
Fig. 1 is a schematic flow chart of an information acquisition method according to an embodiment of the present invention;
FIG. 2 is a schematic flowchart of a candidate key content obtaining method according to an embodiment of the present invention;
FIG. 3 is a flowchart illustrating an optimal key content obtaining method according to an embodiment of the present invention;
fig. 4 is a flowchart illustrating a voting score obtaining method when an initial sentence is used as an initial sentence in optimal key content according to an embodiment of the present invention;
fig. 5 is a flowchart illustrating a voting score obtaining method when an end sentence serves as an end sentence in optimal key content according to an embodiment of the present invention;
FIG. 6 is a flowchart illustrating a method for obtaining N key content calculation models according to an embodiment of the present invention;
FIG. 7 is a block diagram of an information acquisition apparatus according to an embodiment of the present invention;
fig. 8 is a block diagram of an information acquisition apparatus according to an embodiment of the present invention.
Detailed Description
The following describes embodiments of the present invention in further detail with reference to the drawings and examples. The following examples are intended to illustrate the examples of the present invention, but are not intended to limit the scope of the examples of the present invention.
At present, people can communicate with equipment in a natural language mode to acquire required information. In the related art, for a question of a user, the content of a response corresponding to the question of the user is determined through a model. Since a single model is inevitably biased in the training process, it is difficult to completely fit the entire training data distribution, resulting in low reliability of the response content. In view of the above situation, an embodiment of the present invention provides an information acquisition method. The method can be used in an intelligent question and answer scene, and can also be used in other scenes needing an intelligent question and answer function, such as a driving scene, a shopping scene and the like, which are not specifically limited in the embodiment of the invention. The method may be performed by different devices in combination with different usage scenarios, which are not limited in this embodiment of the present invention. For example, if the method is used in a driving scenario, the execution subject of the method may be an in-vehicle device; if the method is used in a shopping scenario, the execution subject of the method may be a mobile terminal. Referring to fig. 1, the method includes:
101, respectively inputting a query text and a reply text matched with the query text into N key content calculation models, and acquiring candidate key content output by each key content calculation model; the key content calculation model is obtained by training based on sample query texts, sample reply texts and sample key contents in the sample reply texts, each candidate key content is extracted from the reply texts, and N is a positive integer greater than 1.
Before the above process is executed, the voice data when the user asks the question can be obtained, and the voice data is subjected to voice recognition to obtain a query text; alternatively, the text input by the user may also be directly obtained and used as the query text, which is not specifically limited in the embodiment of the present invention. In addition, the reply text matching the query text may contain the reply content of the query corresponding to the question. Specifically, if the query text corresponds to a question for inquiring how a certain function in a product is used, the reply text matched with the query text may be a description document of the product; further, considering that a product usually has multiple functions, and the description document of the product usually has usage description information of all functions of the product, if the description document is divided into a plurality of structured texts according to each function in advance, and the query text queries a certain function in the product, the reply text matched with the query text can be the structured text corresponding to the function. If the query text corresponds to a query that is a definition of a technical term, the reply text may be a technical dictionary containing the definition of the technical term. Of course, the reply text may be in other forms besides the above-listed form, and the embodiment of the present invention is not particularly limited thereto.
Since the reply text may contain some redundant information that is not related to the question corresponding to the query text, the redundant information in the reply text can be filtered out by the key content calculation model. Specifically, for any of the key content calculation models, after the reply text is input to the key content calculation model, candidate key content may be output. And each key content calculation model correspondingly outputs one candidate key content. The candidate key content may be a specific text content extracted from the reply text through the key content calculation model, or may be a sentence number in the reply text.
In addition, before the above process is executed, N key content calculation models can be obtained through pre-training. For any one of the N key content calculation models, the key content calculation model can be obtained by training in the following manner, specifically: firstly, collecting a large amount of sample query texts and sample reply texts matched with the sample query texts; wherein, the key content of the sample in the sample reply text is predetermined and is the reply content of the sample query text corresponding to the question. And training the initial model based on the sample query text, the sample reply text and the sample key content to obtain the key content calculation model. The initial model may be a single neural network model or a combination of a plurality of neural network models, and the embodiment of the present invention does not specifically limit the type and structure of the initial model.
It should be noted that, when N key content calculation models are obtained through pre-training, each key content calculation model can be obtained through training the same type of initial model and the same sample. In the actual training process, N key content calculation models with the same function and different output effects can be obtained by adjusting the parameters of the initial model. Taking the example that the initial model comprises a convolutional neural network or comprises a long-short term memory network, the N different key content calculation models can be obtained by adjusting the size and the number of convolutional kernels of the convolutional neural network or adjusting the number of nodes of a hidden layer of the long-short term memory network. Alternatively, N different key content calculation models may be obtained through training by using different training methods, such as dropout or a regularization method. Of course, different parameters may be adjusted and different training methods may be adopted, and the two methods may be combined to obtain N different key content calculation models through training, which is not specifically limited in the embodiment of the present invention.
And 102, acquiring the optimal key content according to each candidate key content, and using the optimal key content as a response corresponding to the query text.
As can be seen from the above steps, N candidate key contents can be output through the N key content calculation models. Different key content calculation models have different output effects, so that candidate key contents output by the N key content calculation models respectively can be fused in the step to obtain the optimal key content. It should be noted that, as can be seen from the content of the above steps, the content format of the candidate key content may be a specific text content or a sentence number. Therefore, in this step, the content format of the optimal key content corresponds to the content format of the candidate key content, which may be a specific text content or a sentence number, and this is not particularly limited in the embodiment of the present invention.
When the optimal key content is obtained according to each candidate key content, each candidate key content can be input into a preset model, so that the optimal key content is obtained; or, the number of times that each clause appears in all candidate key contents in the reply text may also be counted, so that the clause is selected from all candidate key contents to constitute the optimal key content according to the number of times that each clause appears, which is not specifically limited in the embodiment of the present invention. After the optimal key content is obtained, the optimal key content can be used as a response corresponding to the query text. Specifically, the optimal key content may be used as a response to the user in a voice broadcast or text display manner, which is not specifically limited in this embodiment of the present invention.
According to the method provided by the embodiment of the invention, the query text and the reply text matched with the query text are respectively input into the N key content calculation models, and the candidate key content output by each key content calculation model is obtained. And acquiring the optimal key content according to each candidate key content, and using the optimal key content as a response corresponding to the query text. Because the output results of the N key content calculation models can be fused, the problem that a single model has deviation in the training process and is difficult to completely fit all training data distribution is effectively avoided, the reliability and the accuracy of reply content can be improved, and the interaction experience of a user in question-answer interaction with equipment is improved.
Based on the content of the foregoing embodiment, as an optional embodiment, the embodiment of the present invention does not specifically limit a manner in which the query text and the reply text matching the query text are respectively input to the N key content calculation models, and candidate key content output by each key content calculation model is obtained. Referring to fig. 2, including but not limited to:
for candidate key contents output by any key content calculation model, the starting probability of each clause in the reply text as a starting sentence in the candidate key contents is obtained, and the ending probability of each clause in the reply text as an ending sentence in the candidate key contents is obtained 201.
And 202, selecting the clause corresponding to the maximum starting probability from the reply text as the starting sentence of the candidate key content according to the starting probability corresponding to each clause, and selecting the clause corresponding to the maximum ending probability from the reply text as the ending sentence of the candidate key content according to the ending probability corresponding to each clause.
And 203, taking the clauses, the starting sentences and the ending sentences positioned between the starting sentences and the ending sentences in the reply text as candidate key contents.
Taking an example that the reply text includes 8 clauses, for any candidate key content, when determining a starting sentence and an ending sentence in the candidate key content, each clause in the reply text is taken as a starting probability of the starting sentence in the candidate key content, and each clause in the reply text is taken as an ending probability of the ending sentence in the candidate key content, as shown in table 1 below: TABLE 1
As shown in table 1 above, the starting probability of clause 4 is the highest, and the ending probability of clause 6 is the highest. Therefore, clause 4 may be used as the starting sentence of the candidate key content, clause 6 may be used as the ending sentence of the candidate key content, and clauses 4, 5, and 6 may be used as the candidate key content. Through the process in the above example, N candidate key contents may be obtained.
According to the method provided by the embodiment of the invention, each clause in the reply text is obtained to serve as the starting probability of the starting sentence in each candidate key content and the ending probability of the ending sentence, and the candidate key content is determined based on the starting probability and the ending probability, so that the calculation accuracy of the key content calculation model is improved, and the reliability and the accuracy of the candidate key content can be improved.
It can be known from the content of the above embodiments that a single model has a deviation in the training process, and it is difficult to completely fit all the training data distributions, so that the key content output by the single model is not reliable enough and has low accuracy. For this situation, based on the content of the foregoing embodiment, as an alternative embodiment, the embodiment of the present invention does not specifically limit the manner of obtaining the optimal key content according to each candidate key content. Referring to fig. 3, including but not limited to:
301, for each starting sentence in the starting sentences contained in all the candidate key contents, determining the optimal starting sentence in the optimal key contents according to the voting score when each starting sentence is used as the starting sentence in the optimal key contents.
It can be known from the content of the above embodiment that the candidate key contents output by the N key content calculation models may be different, some candidate key contents may have the same start statement, some candidate key contents may have the same end statement, and some candidate key contents may not have an information intersection. In order to fuse N candidate key contents and select an optimal starting sentence from the N candidate key contents, the embodiments of the present invention may adopt a voting manner to vote starting sentences appearing in all candidate key contents, that is, determine voting values of each starting sentence in all the appearing starting sentences in a voting process, so as to select the starting sentence with the largest voting value from all the appearing starting sentences as the optimal starting sentence.
The voting score corresponding to the initial sentence in the candidate key content represents the accuracy and reliability of the initial sentence serving as the initial sentence in the optimal key content. The larger the voting score is, the higher the reliability and accuracy of the initial sentence are. The voting score may be determined according to a start probability corresponding to a start statement in each candidate key content, or may be determined according to the number of times that the start statement appears in all candidate key contents, which is not specifically limited in the embodiment of the present invention.
302, for each ending sentence in the ending sentences included in all the candidate key contents, determining an optimal ending sentence in the optimal key content according to the voting score of the ending sentence when the ending sentence is used as the ending sentence in the optimal key content.
It can be known from the content of the above embodiment that the candidate key contents output by the N key content calculation models may be different, some candidate key contents may have the same start statement, some candidate key contents may have the same end statement, and some candidate key contents may not have an information intersection. In order to fuse the N candidate key contents and select an optimal end sentence from the N candidate key contents, the embodiments of the present invention may adopt a voting manner to vote for the end sentences appearing in all the candidate key contents, that is, determine the voting score of each end sentence in all the appearing end sentences in the voting process, so as to select the end sentence with the largest voting score from all the appearing end sentences as the optimal end sentence.
The voting score corresponding to the ending statement in the candidate key content represents the accuracy and reliability of the ending statement when the ending statement is used as the ending statement in the optimal key content. The greater the voting score is, the higher the reliability and accuracy of the corresponding ending statement are. The voting score may be determined according to a start probability corresponding to an ending statement in each candidate key content, or may be determined according to the number of times that the ending statement appears in all candidate key contents, which is not specifically limited in the embodiment of the present invention.
303, taking the clause, the optimal starting sentence and the optimal ending sentence positioned between the optimal starting sentence and the optimal ending sentence in the reply text as the optimal key content.
According to the method provided by the embodiment of the invention, the optimal starting statement and the optimal ending statement are determined according to the voting score of the starting statement and the voting score of the ending statement, so that the optimal key content is determined, the candidate key contents output by the N key content calculation models are fused, and the reliability and the accuracy of the reply content are improved.
As can be seen from the content of the foregoing embodiment, the voting score corresponding to the starting sentence in the candidate key content can be determined according to the starting probability corresponding to the starting sentence in the candidate key content. Therefore, before determining the optimal starting sentence in the optimal key content according to the vote score when each starting sentence is used as the starting sentence in the optimal key content, as an optional embodiment based on the principle and the contents of the above embodiments, the embodiment of the present invention further provides a manner for determining the vote score corresponding to the starting sentence in the candidate key content. Referring to fig. 4, including but not limited to:
401, regarding any candidate starting sentence in all candidate starting sentences, taking the candidate key content meeting the first preset condition in all candidate key contents as the first target candidate key content, where all candidate starting sentences are the starting sentences included in all candidate key contents, and the first preset condition is that any candidate starting sentence is included and any candidate starting sentence is taken as the first clause.
402, calculating the voting score when any candidate starting sentence is used as the starting sentence in the optimal key content according to the total number of the first target candidate key contents in all the candidate key contents and the starting probability corresponding to the starting sentence in each first target candidate key content.
For ease of understanding, taking N as 5 as an example, 5 key content calculation models may output 5 candidate key contents. For ease of description, the starting sentence is represented by its sentence-order number in the reply text. Taking the candidate key contents as a, b, c, d and e as an example, the starting sentence and the starting probability corresponding to the starting sentence in each candidate key content can be referred to as the following table 2:
TABLE 2
As can be seen from table 2, the initial sentences included in the 5 candidate key contents are the clauses with clause numbers 2, 3 and 4 in the reply text, that is, all the candidate initial sentences are the clauses with clause numbers 2, 3 and 4. For the candidate starting sentence with the sentence sequence number of 2, the candidate key contents including the candidate starting sentence as the starting sentence and the candidate starting sentence as the first sentence (that is, satisfying the first preset condition) are a and d, respectively, that is, the first target candidate key contents. At this time, the total number of the first target candidate key contents is 2, the start probability corresponding to the start sentence in the first target candidate key content a is 80%, and the start probability corresponding to the start sentence in the first target candidate key content d is 94%. And calculating the voting value of the candidate initial sentence with the sentence sequence number of 2 when the candidate initial sentence is used as the initial sentence in the optimal key content according to the parameters. The specific calculation process can refer to the following formula:
scorestart-i＝countstart(index＝i)+sum(pstart-i)/countstart(index＝i)
wherein, scorestart-iAnd taking the candidate starting sentence with the sentence sequence number i as the voting value of the starting sentence in the optimal key content. countstart(index ═ i) is the total number of first target candidate key contents, sum (p)start-i) Representing the sum of the starting probabilities corresponding to the starting sentences in all the first target candidate key contents. It should be noted that, here, the first target candidate key content startsThe sentence is a candidate starting sentence with a sentence dividing sequence number i.
As can be seen from the contents in the above example, for the candidate start sentence with the sentence number of 2, the candidate key contents with the candidate start sentence as the start sentence are a and d, respectively. With the contents of the above example and the above calculation formula, the voting score of the candidate start sentence with sentence number 2 when the candidate start sentence is used as the start sentence in the optimal key content can be calculated to be scorestart-2＝2.87。
Similarly, for the candidate starting sentence with the sentence sequence number of 3, as can be seen from table 2 above, the candidate key contents taking the candidate starting sentence as the starting sentence are c and e, respectively, that is, the candidate key contents are the first target key contents. At this time, the total number of the first target candidate key contents is 2, the start probability corresponding to the start sentence in the first target candidate key contents c is 72%, and the start probability corresponding to the start sentence in the first target candidate key contents e is 76%. Based on the parameters and the calculation formula, the candidate starting sentence with the sentence sequence number of 3 can be calculated, and the voting value when the candidate starting sentence is used as the starting sentence in the optimal key content is scorestart-32.74. For the candidate starting sentence with sentence number 4, based on the above calculation process, the candidate starting sentence with sentence number 4 can be determined, and the voting score when the candidate starting sentence is used as the starting sentence in the optimal key content is scorestart-4＝1.64。
As can be seen from the content of the foregoing embodiment, when determining the optimal starting sentence in the optimal key content, according to the vote score when each candidate starting sentence is used as the starting sentence in the optimal key content, the candidate starting sentence corresponding to the maximum vote score may be selected as the optimal starting sentence in the optimal key content. While in the above example, the maximum vote score was scorestart-22.87, namely the candidate starting sentence with sentence number 2. Therefore, the candidate start sentence with sentence number 2 in the reply text may be used as the optimal start sentence in the optimal key content in the above example.
As can be seen from the content of the foregoing embodiment, the voting score corresponding to the end sentence in the candidate key content can be determined according to the end probability corresponding to the end sentence in the candidate key content. Therefore, before determining the optimal ending sentence in the optimal key content according to the voting score of each ending sentence as the ending sentence in the optimal key content, based on the principle and the contents of the above embodiments, as an optional embodiment, the embodiment of the present invention further provides a manner of determining the voting score corresponding to the ending sentence in the candidate key content. Referring to fig. 5, including but not limited to:
and 501, regarding any candidate ending sentence in all candidate ending sentences, taking the candidate key content meeting a second preset condition in all candidate key contents as a second target candidate key content, wherein all candidate ending sentences are ending sentences contained in all candidate key contents, and the second preset condition is that any candidate ending sentence is contained and any candidate ending sentence is taken as a last clause.
502, according to the total number of the second target candidate key contents in all the candidate key contents and the end probability corresponding to the end sentence in each second target candidate key content, calculating the voting score when any candidate end sentence is used as the end sentence in the optimal key content.
For ease of understanding, also taking N as 5 as an example, 5 key content calculation models may output 5 candidate key contents. For convenience of description, the ending sentence is represented by a sentence number of the ending sentence in the reply text. Taking the candidate key contents as a, b, c, d and e as an example, the ending probability corresponding to the ending sentence and the ending sentence in each candidate key content can be referred to as the following table 3:
TABLE 3
As can be seen from table 3, the end sentences included in the 5 candidate key contents are the clauses with clause numbers 3, 4 and 5 in the reply text, that is, all the candidate end sentences are the clauses with clause numbers 2, 3 and 4. For the candidate ending sentence with the sentence sequence number of 4, the candidate key contents including the candidate ending sentence and the candidate ending sentence as the last sentence (that is, meeting the second preset condition) are b, d and e, respectively, that is, the second target candidate key contents. At this time, the total number of the second target candidate key contents is 3, the end probability corresponding to the end sentence in the second target candidate key content b is 64%, the end probability corresponding to the end sentence in the second target candidate key content d is 94%, and the end probability corresponding to the end sentence in the second target candidate key content e is 76%. And calculating the voting value of the candidate ending sentence with the sentence sequence number of 4 when the candidate ending sentence is used as the optimal key content to end the sentence according to the parameters. The specific calculation process can refer to the following formula:
scoreend-i＝countend(index＝i)+sum(pend-i)/countend(index＝i)
wherein, scoreend-iAnd taking the candidate ending sentence with the sentence sequence number i as the voting score of the ending sentence in the optimal key content. countend(index ═ i) is the total number of second target candidate key contents, sum (p)end-i) And the sum of the end probabilities corresponding to the end sentences in all the second target candidate key contents is represented. It should be noted that, here, the end sentence in the second target candidate key content is the candidate end sentence with the sentence number i.
As can be seen from the contents in the above example, for the candidate end sentence with the sentence number of 4, the candidate key contents with the candidate end sentence as the end sentence are b, d, and e, respectively. With the contents of the above example and the above calculation formula, the voting score of the candidate ending sentence with sentence number 4 when the candidate ending sentence is taken as the ending sentence in the optimal key content can be calculated to be scoreend-4＝3.78。
Similarly, for the candidate ending sentences with the sentence sequence number of 3, as can be seen from table 3 above, the candidate key contents taking the candidate ending sentences as ending sentences are respectively c, that is, the candidate key contents are the second target key contents. At this time, the total number of the second target candidate key contents is 1, and the start corresponding to the end sentence in the second target candidate key content cThe probability is 72%. Based on the parameters and the calculation formula, the candidate ending sentence with sentence number 3 can be calculated, and the voting value when the sentence is ended in the optimal key content is scoreend-31.72. For the candidate end sentence with sentence number 5, based on the above calculation process, the candidate end sentence with sentence number 5 can be determined, and the voting score when the sentence is ended in the optimal key content is scoreend-5＝1.8。
As can be seen from the content of the foregoing embodiment, when determining the optimal end sentence in the optimal key content, according to the vote score when each candidate end sentence is used as an end sentence in the optimal key content, the candidate end sentence corresponding to the maximum vote score may be selected as the optimal end sentence in the optimal key content. While in the above example, the maximum vote score was scoreend-4That is, the candidate end sentence with sentence number 4 is 3.78. Therefore, the candidate end sentence with the sentence number of 4 in the reply text may be used as the optimal end sentence in the optimal key content in the above example.
According to the method provided by the embodiment of the invention, the voting score when the ending statement is used as the ending statement in the optimal key content is determined based on the starting probability corresponding to the starting statement and the ending probability corresponding to the ending statement in the candidate key content, so that the voting score of each starting statement and the voting score of each ending statement are determined in a voting mode, and the optimal starting statement and the optimal ending statement in the optimal key content are elected based on the voting score. Therefore, candidate key contents output by the N key content calculation models can be fused, and reliability and accuracy of the reply contents are improved.
In the actual implementation process, a plurality of key content calculation models to be selected may be obtained through pre-training. In the using process of the candidate key content calculation models, the accuracy and the reliability of the reply content obtained based on some candidate key content calculation models are higher, and the accuracy and the reliability of the reply content obtained based on some candidate key content calculation models are lower. Therefore, for the pre-trained candidate key content calculation model, it is necessary to filter it. Based on the principle, if the "N key content calculation models" involved in the above embodiments are obtained after screening, the query text and the reply text matched with the query text are input to the N key content calculation models, and before candidate key content output by each key content calculation model is obtained, M key content calculation models to be selected obtained through pre-training may be further screened to obtain N key content calculation models.
Based on the content of the above embodiment, as an optional embodiment, the embodiment of the present invention provides a method for obtaining N key content calculation models by screening. Referring to fig. 6, including but not limited to:
601, combining M key content calculation models to be selected to obtain a plurality of target model sets to be selected, wherein each target model set to be selected comprises N key content calculation models to be selected, and M is not less than N.
The to-be-selected key content calculation models related to the embodiment of the invention can be obtained by the same type of initial models and the same sample training. In the actual training process, a plurality of candidate key content calculation models with the same function but different output effects can be obtained by adjusting the parameters of the initial model. Taking the example that the initial model comprises a convolutional neural network or comprises a long-short term memory network, the M different key content calculation models to be selected can be obtained through training by adjusting the size and the number of convolutional kernels of the convolutional neural network or adjusting the number of nodes of a hidden layer of the long-short term memory network. Or, M different candidate key content calculation models may be obtained through training by using different training methods, such as dropout or a regularization method. Of course, different parameters may be adjusted and different training methods may be adopted, and the two methods may be combined to obtain M different candidate key content calculation models through training, which is not specifically limited in the embodiment of the present invention. For a specific training process, reference may be made to the training process of the key content calculation model in the foregoing embodiment, which is not described herein again.
Take M as 5, and A, B, C, D and E, respectively, as an example. If N is 3, 3 candidate key content calculation models can be selected from the 5 candidate key content calculation models in different combination modes, so that a plurality of groups of target candidate model sets including the 3 candidate key content calculation models are obtained. For example, (A, B, C), (A, B, E) and (B, C, D), etc. It should be noted that, in the above example, different candidate key content calculation models are combined to obtain a target candidate model set. In an actual implementation process, the same candidate key content calculation models may also be combined to obtain a target candidate model set, which is not specifically limited in the embodiment of the present invention. For example, the obtained target candidate model set may be (a, a), (B, B), and (C, C).
602, respectively inputting a plurality of sample test cases to each group of target candidate model sets, and acquiring the optimal key content of the samples obtained when each sample test case is used as input in each group of target candidate model sets; each sample test case comprises a sample query text and a sample reply text matched with the sample query text, each sample test case corresponds to the relevant key content, and the key content corresponding to each sample test case is extracted in the sample reply text in advance based on the sample query text.
In this step, for any sample test case and any group of target candidate model sets, where the group of target candidate model sets includes N candidate key content calculation models, the sample test case is input to the target candidate model set, that is, a sample query text and a sample reply text matched with the sample query text are respectively input to the N candidate key content calculation models in the group of target candidate model sets. Similarly to the above embodiment, after the sample query text and the sample reply text matched with the sample query text are respectively input to the N candidate key content calculation models in the set of target candidate models, the optimal key content can be obtained according to the above process of obtaining the optimal key content.
It should be noted that, by inputting a sample test case into a group of target candidate model sets, optimal key content of a sample can be obtained. If the number of the sample test cases is m and the number of the groups of the candidate model sets is n, the optimal key content of m × n samples can be obtained by inputting each sample test case into each group of target candidate model sets respectively.
603, comparing the optimal key content of the sample obtained when each sample test case is input into each group of target candidate model sets with the key content corresponding to each sample test case, determining the information acquisition accuracy rate corresponding to each group of target candidate model sets according to the comparison result, and selecting the target candidate model set corresponding to the maximum information acquisition accuracy rate as N key content calculation models.
For the sample test example, on the premise that both the sample query text and the sample reply text matched with the sample query text are determined, the reply content of the query corresponding to the sample query text, that is, the key content related to the embodiment of the present invention, may also be predetermined. Therefore, in this step, for any sample test case and any group of target candidate model sets, the optimal key content of the sample obtained when the group of target candidate model sets take the sample test case as input is compared with the key content corresponding to the sample test case, so that it can be determined whether the output result of the group of target candidate model sets is correct or wrong when the sample test case takes the sample test case as input.
Specifically, if the optimal key content of the sample output by the set of target candidate model sets is consistent with the key content corresponding to the sample test case, it may be determined that the output result of the set of target candidate model sets is correct when the sample test case is used as an input. If the optimal key content of the sample output by the set of target candidate model sets is inconsistent with the key content corresponding to the sample test case, it can be determined that the output result of the set of target candidate model sets is wrong when the sample test case is used as input.
Therefore, for any group of target candidate model sets, each time a sample test case is input into the group of target candidate model sets, the test case can be regarded as a test process for the group of target candidate model sets. Through the comparison process, whether the output result of each test process is correct or not can be judged. Therefore, for any group of target candidate model sets, the plurality of sample test cases in the steps are respectively input into the group of target candidate model sets, and the information acquisition accuracy of the group of target candidate model sets can be determined according to the output result. For example, if the number of the sample test cases is 100, and the number of times of outputting the correct result is 72 after the 100 sample test cases are respectively input to the set of target candidate model sets, it may be determined that the information acquisition accuracy of the set of target candidate model sets is 72%. Similarly, for each group of target candidate model sets in the plurality of groups of target candidate model sets, the information acquisition accuracy corresponding to each group of target candidate model sets can be determined according to the above process.
After the information acquisition accuracy rate corresponding to each group of target candidate model sets in the plurality of groups of target candidate model sets is determined, the target candidate model set corresponding to the maximum information acquisition accuracy rate can be selected, and N candidate key content calculation models in the group of target candidate model sets are used as N key content calculation models.
According to the method provided by the embodiment of the invention, the target candidate model set corresponding to the maximum information acquisition accuracy rate is selected as N key content calculation models based on the information acquisition accuracy rate of each group of target candidate model sets, so that the reliability and the accuracy of the reply content can be ensured.
Based on the content of the above embodiment, as an optional embodiment, the embodiment of the present invention does not specifically limit the manner in which the M candidate key content calculation models are combined to obtain the plurality of groups of target candidate model sets, including but not limited to: and based on a greedy algorithm, combining M key content calculation models to be selected one by one according to the information acquisition accuracy rate corresponding to the model set to be selected obtained after combination until a plurality of groups of target model sets to be selected are obtained.
The greedy algorithm is to achieve an overall optimal solution of a problem through a series of local optimal selections, namely greedy selection. According to the content of the embodiment, the target candidate model set comprises N candidate key content calculation models. For convenience of understanding, a process of combining A, B, C, D and E to obtain a plurality of sets of target candidate model sets is described by taking M candidate key content calculation models as A, B, C, D and E, N as 3, and a as an initial combination model as an example.
Since A is the initial combinatorial model, A can be combined with A, B, C, D and E, respectively, to yield (A, A), (A, B), (A, C), (A, D) and (A, E). Since the five sets are also sets obtained by combining the key content calculation models to be selected, the information acquisition accuracy rates corresponding to the 5 sets can be calculated respectively based on the calculation process of the information acquisition accuracy rates corresponding to the target model set to be selected in the embodiment. Because the greedy algorithm is based on the local optimal solution, the set corresponding to the maximum information acquisition accuracy can be selected from the 5 sets to continue the subsequent combination process.
Taking the set (a, B) as the set corresponding to the maximum information acquisition rate as an example, the combination of (a, B) with A, B, C, D and E can be continued to obtain (a, B, a), (a, B), (a, B, C), (A, B, D) and (a, B, E). Similarly, the information acquisition accuracy rates corresponding to (a, B, a), (a, B), (a, B, C), (A, B, D) and (a, B, E) can be calculated respectively according to the above calculation process. And selecting a set corresponding to the maximum information acquisition accuracy from the 5 sets, wherein the set can be used as a target model set to be selected, which is obtained by taking A as an initial combination model.
Similarly, B, C, D and E are respectively used as the initial combination models, so that the target candidate model combination using B, C, D and E as the initial combination models can be obtained. And combining all the target candidate models obtained according to the process to obtain a plurality of groups of target candidate model sets.
According to the method provided by the embodiment of the invention, M key content calculation models to be selected are combined one by one through a greedy algorithm according to the information acquisition accuracy rate corresponding to the model set to be selected obtained after combination until a plurality of groups of target model sets to be selected are obtained. The M key content calculation models to be selected can be combined based on a local optimal solution mode to obtain a plurality of groups of target model sets to be selected, so that the reliability and accuracy of the reply content can be guaranteed.
It should be noted that, all the above-mentioned alternative embodiments may be combined arbitrarily to form alternative embodiments of the present invention, and are not described in detail herein.
Based on the content of the above embodiments, an embodiment of the present invention provides an information acquisition apparatus. The device is used for executing the information acquisition method provided in the method embodiment. Referring to fig. 7, the apparatus includes:
a candidate key content obtaining module 701, configured to input the query text and the reply text matched with the query text to the N key content calculation models, respectively, and obtain candidate key content output by each key content calculation model; the key content calculation model is obtained by training based on sample query texts, sample reply texts and sample key contents in the sample reply texts, each candidate key content is extracted from the reply texts, and N is a positive integer greater than 1;
an optimal key content obtaining module 702, configured to obtain optimal key content according to each candidate key content, and use the optimal key content as a response corresponding to the query text.
As an alternative embodiment, the candidate key content obtaining module 701 includes:
the probability obtaining unit is used for obtaining the starting probability of each clause in the reply text as the starting sentence in the candidate key content and obtaining the ending probability of each clause in the reply text as the ending sentence in the candidate key content for the candidate key content output by any key content calculation model;
the sentence acquisition unit is used for selecting the clause corresponding to the maximum starting probability from the reply text as the starting sentence of the candidate key content according to the starting probability corresponding to each clause, and selecting the clause corresponding to the maximum ending probability from the reply text as the ending sentence of the candidate key content according to the ending probability corresponding to each clause;
and the candidate key content determining unit is used for taking the clause, the starting sentence and the ending sentence positioned between the starting sentence and the ending sentence in the reply text as candidate key content.
As an alternative embodiment, the optimal key content obtaining module 702 includes:
the optimal starting sentence determining unit is used for determining the optimal starting sentence in the optimal key content according to the voting score when each starting sentence is used as the starting sentence in the optimal key content for each starting sentence in the starting sentences contained in all the candidate key contents;
an optimal end sentence determining unit, configured to determine, for each end sentence in end sentences included in all candidate key contents, an optimal end sentence in the optimal key content according to a vote score when each end sentence is used as an end sentence in the optimal key content;
and the optimal key content determining unit is used for taking the clauses, the optimal starting sentences and the optimal ending sentences which are positioned between the optimal starting sentences and the optimal ending sentences in the reply text as the optimal key content.
As an alternative embodiment, the optimal key content obtaining module 702 further includes:
a first target candidate key content obtaining unit, configured to, for any candidate starting sentence in all candidate starting sentences, use, as a first target candidate key content, candidate key contents that satisfy a first preset condition in all candidate key contents, where all candidate starting sentences are starting sentences included in all candidate key contents, and the first preset condition is that any candidate starting sentence is included and any candidate starting sentence is used as a first clause;
and the initial sentence voting score acquisition unit is used for calculating the voting score when any candidate initial sentence is taken as the initial sentence in the optimal key content according to the total number of the first target candidate key contents in all the candidate key contents and the initial probability corresponding to the initial sentence in each first target candidate key content.
As an alternative embodiment, the optimal key content obtaining module 702 further includes:
a second target candidate key content obtaining unit, configured to, for any candidate ending statement in all candidate ending statements, use, as a second target candidate key content, a candidate key content that satisfies a second preset condition in all candidate key contents, where all candidate ending statements are ending statements included in all candidate key contents, and the second preset condition is that any candidate ending statement is included and any candidate ending statement is taken as a last clause;
and the end sentence voting score acquisition unit is used for calculating the voting score when any candidate end sentence is taken as the end sentence in the optimal key content according to the total number of the second target candidate key contents in all the candidate key contents and the end probability corresponding to the end sentence in each second target candidate key content.
As an alternative embodiment, the apparatus further comprises:
the target candidate model set acquisition module is used for combining M candidate key content calculation models to obtain a plurality of groups of target candidate model sets, each group of target candidate model sets comprises N candidate key content calculation models, and M is not less than N;
the optimal key content acquisition module is used for respectively inputting a plurality of sample test cases into each group of target candidate model sets and acquiring the optimal key content of the sample obtained when each sample test case is used as input in each group of target candidate model sets; each sample test case comprises a sample query text and a sample reply text matched with the sample query text, each sample test case corresponds to related key content, and the key content corresponding to each sample test case is extracted in the sample reply text in advance based on the sample query text;
and the key content calculation model acquisition module is used for comparing the optimal key content of the sample obtained when each sample test case is input into each group of target model sets to be selected with the key content corresponding to each sample test case, determining the information acquisition accuracy rate corresponding to each group of target model sets to be selected according to the comparison result, and selecting the target model set to be selected corresponding to the maximum information acquisition accuracy rate as the N key content calculation models.
As an optional embodiment, the sample optimal key content obtaining module is configured to combine, based on a greedy algorithm, the M key content calculation models to be selected one by one according to the information obtaining accuracy corresponding to the model set to be selected obtained after combination until a plurality of groups of target model sets to be selected are obtained.
According to the device provided by the embodiment of the invention, the query text and the reply text matched with the query text are respectively input into the N key content calculation models, and the candidate key content output by each key content calculation model is obtained. And acquiring the optimal key content according to each candidate key content, and using the optimal key content as a response corresponding to the query text. Because the output results of the N key content calculation models can be fused, the problem that a single model has deviation in the training process and is difficult to completely fit all training data distribution is effectively avoided, the reliability and the accuracy of reply content can be improved, and the interaction experience of a user in question-answer interaction with equipment is improved.
Secondly, each clause in the reply text is obtained to serve as the starting probability of the starting sentence in each candidate key content and the ending probability of the ending sentence, and then the candidate key content is determined based on the starting probability and the ending probability, so that the calculation accuracy of the key content calculation model is improved, and the reliability and the accuracy of the candidate key content can be improved.
And thirdly, determining the optimal starting statement and the optimal ending statement according to the voting score of the starting statement and the voting score of the ending statement, and further determining the optimal key content, so that the candidate key contents output by the N key content calculation models are fused, and the reliability and the accuracy of the reply content are improved.
And determining the voting value of the ending sentence as the ending sentence in the optimal key content by determining the starting probability corresponding to the starting sentence and the ending probability corresponding to the ending sentence in the candidate key content, thereby determining the voting value of each starting sentence and the voting value of each ending sentence in a voting way, and electing the optimal starting sentence and the optimal ending sentence in the optimal key content on the basis of the voting values. Therefore, candidate key contents output by the N key content calculation models can be fused, and reliability and accuracy of the reply contents are improved.
In addition, the target candidate model set corresponding to the maximum information acquisition accuracy rate is selected as N key content calculation models based on the information acquisition accuracy rate of each group of target candidate model sets, so that the reliability and the accuracy of the reply content can be ensured.
And finally, combining M key content calculation models to be selected one by one according to the information acquisition accuracy rate corresponding to the model set to be selected obtained after combination based on a greedy algorithm until a plurality of groups of target model sets to be selected are obtained. The M key content calculation models to be selected can be combined based on a local optimal solution mode to obtain a plurality of groups of target model sets to be selected, so that the reliability and accuracy of the reply content can be guaranteed.
The embodiment of the invention provides information acquisition equipment. Referring to fig. 8, the apparatus includes: a processor (processor)801, a memory (memory)802, and a bus 803;
the processor 801 and the memory 802 communicate with each other via a bus 803; the processor 801 is configured to call the program instructions in the memory 802 to execute the information obtaining method provided by the foregoing embodiments, for example, including: respectively inputting the query text and the reply text matched with the query text into the N key content calculation models, and acquiring candidate key content output by each key content calculation model; the key content calculation model is obtained by training based on sample query texts, sample reply texts and sample key contents in the sample reply texts, each candidate key content is extracted from the reply texts, and N is a positive integer greater than 1; and acquiring the optimal key content according to each candidate key content, and using the optimal key content as a response corresponding to the query text.
An embodiment of the present invention provides a non-transitory computer-readable storage medium, where the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable a computer to execute the information acquisition method provided in the foregoing embodiment, for example, the method includes: respectively inputting the query text and the reply text matched with the query text into the N key content calculation models, and acquiring candidate key content output by each key content calculation model; the key content calculation model is obtained by training based on sample query texts, sample reply texts and sample key contents in the sample reply texts, each candidate key content is extracted from the reply texts, and N is a positive integer greater than 1; and acquiring the optimal key content according to each candidate key content, and using the optimal key content as a response corresponding to the query text.
Those of ordinary skill in the art will understand that: all or part of the steps for implementing the method embodiments may be implemented by hardware related to program instructions, and the program may be stored in a computer readable storage medium, and when executed, the program performs the steps including the method embodiments; and the aforementioned storage medium includes: various media that can store program codes, such as ROM, RAM, magnetic or optical disks.
The above-described embodiments of the information acquiring apparatus and the like are merely illustrative, and units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute the various embodiments or some parts of the methods of the embodiments.
Finally, the method of the present application is only a preferred embodiment and is not intended to limit the scope of the embodiments of the present invention. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the embodiments of the present invention should be included in the protection scope of the embodiments of the present invention.
Claims (10)
1. An information acquisition method, comprising:
respectively inputting a query text and a reply text matched with the query text into N key content calculation models, and acquiring candidate key content output by each key content calculation model; the key content calculation model is obtained by training based on a sample query text, a sample reply text and sample key contents in the sample reply text, each candidate key content is extracted from the reply text, and N is a positive integer greater than 1; the N key content calculation models have different output effects;
and acquiring optimal key content according to each candidate key content, and using the optimal key content as a response corresponding to the query text.
2. The method of claim 1, wherein the step of inputting the query text and the reply text matching with the query text into N key content calculation models respectively to obtain the candidate key content output by each key content calculation model comprises:
for candidate key contents output by any key content calculation model, acquiring the starting probability of each clause in the reply text as a starting sentence in the candidate key contents, and acquiring the ending probability of each clause in the reply text as an ending sentence in the candidate key contents;
selecting a clause corresponding to the maximum starting probability from the reply text as a starting sentence of the candidate key content according to the starting probability corresponding to each clause, and selecting a clause corresponding to the maximum ending probability from the reply text as an ending sentence of the candidate key content according to the ending probability corresponding to each clause;
and taking the clause, the starting sentence and the ending sentence positioned between the starting sentence and the ending sentence in the reply text as the candidate key content.
3. The method of claim 2, wherein obtaining the optimal key content according to each candidate key content comprises:
for each initial sentence in the initial sentences contained in all candidate key contents, determining an optimal initial sentence in the optimal key contents according to the voting score of each initial sentence when the initial sentence is used as the initial sentence in the optimal key contents;
for each ending statement in ending statements contained in all candidate key contents, determining an optimal ending statement in the optimal key contents according to a voting score when each ending statement is used as an ending statement in the optimal key contents;
and taking the clause, the optimal starting sentence and the optimal ending sentence which are positioned between the optimal starting sentence and the optimal ending sentence in the reply text as the optimal key content.
4. The method of claim 3, wherein before determining the optimal starting sentence in the optimal key content according to the vote score of each starting sentence as the starting sentence in the optimal key content, further comprising:
for any candidate starting sentence in all candidate starting sentences, taking the candidate key content meeting a first preset condition in all candidate key contents as first target candidate key content, wherein all candidate starting sentences are starting sentences contained in all candidate key contents, and the first preset condition is that any candidate starting sentence is contained and any candidate starting sentence is taken as a first clause;
and calculating the voting score when any candidate starting statement is taken as the starting statement in the optimal key content according to the total number of the first target candidate key contents in all the candidate key contents and the starting probability corresponding to the starting statement in each first target candidate key content.
5. The method according to claim 3 or 4, wherein before determining the optimal end sentence in the optimal key content according to the vote score of each end sentence as the end sentence in the optimal key content, further comprising:
for any candidate ending sentence in all candidate ending sentences, taking the candidate key content meeting a second preset condition in all candidate key contents as a second target candidate key content, wherein all candidate ending sentences are ending sentences contained in all candidate key contents, and the second preset condition is that any candidate ending sentence is contained and any candidate ending sentence is taken as a last clause;
and calculating the voting score when any candidate ending statement is taken as the ending statement in the optimal key content according to the total number of the second target candidate key contents in all the candidate key contents and the ending probability corresponding to the ending statement in each second target candidate key content.
6. The method of claim 1, wherein before inputting the query text and the reply text matching the query text into the N key content calculation models and obtaining the candidate key content output by each key content calculation model, further comprising:
combining M key content calculation models to be selected to obtain a plurality of groups of target model sets to be selected, wherein each group of target model sets to be selected comprises N key content calculation models to be selected, and M is not less than N;
respectively inputting a plurality of sample test cases into each group of target candidate model sets, and acquiring the optimal key content of the samples obtained when each sample test case is used as input in each group of target candidate model sets; each sample test case comprises a sample query text and a sample reply text matched with the sample query text, each sample test case corresponds to related key content, and the key content corresponding to each sample test case is extracted in advance from the sample reply text on the basis of the sample query text;
and comparing the optimal key content of the sample obtained when each sample test case is input into each group of target candidate model sets with the key content corresponding to each sample test case, determining the information acquisition accuracy rate corresponding to each group of target candidate model sets according to the comparison result, and selecting the target candidate model set corresponding to the maximum information acquisition accuracy rate as the N key content calculation models.
7. The method according to claim 6, wherein the combining M candidate key content calculation models to obtain a plurality of sets of target candidate models comprises:
and combining the M key content calculation models to be selected one by one according to the information acquisition accuracy rate corresponding to the model set to be selected obtained after combination until a plurality of groups of target model sets to be selected are obtained.
8. An information acquisition apparatus characterized by comprising:
the candidate key content acquisition module is used for respectively inputting a query text and a reply text matched with the query text into the N key content calculation models and acquiring candidate key content output by each key content calculation model; the key content calculation model is obtained by training based on a sample query text, a sample reply text and sample key contents in the sample reply text, each candidate key content is extracted from the reply text, and N is a positive integer greater than 1; the N key content calculation models have different output effects;
and the optimal key content acquisition module is used for acquiring optimal key content according to each candidate key content and taking the optimal key content as a response corresponding to the query text.
9. An information acquisition apparatus characterized by comprising:
at least one processor; and
at least one memory communicatively coupled to the processor, wherein:
the memory stores program instructions executable by the processor, the processor invoking the program instructions to perform the method of any of claims 1 to 7.
10. A non-transitory computer-readable storage medium storing computer instructions that cause a computer to perform the method of any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810551681.9A CN108804627B (en) | 2018-05-31 | 2018-05-31 | Information acquisition method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810551681.9A CN108804627B (en) | 2018-05-31 | 2018-05-31 | Information acquisition method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108804627A CN108804627A (en) | 2018-11-13 |
| CN108804627B true CN108804627B (en) | 2021-04-06 |
Family
ID=64089724
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810551681.9A Active CN108804627B (en) | 2018-05-31 | 2018-05-31 | Information acquisition method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108804627B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103379433A (en) * | 2012-04-27 | 2013-10-30 | 华为终端有限公司 | Method and device for giving alarm through distance measurement |
| CN103902652A (en) * | 2014-02-27 | 2014-07-02 | 深圳市智搜信息技术有限公司 | Automatic question-answering system |
| CN104933097A (en) * | 2015-05-27 | 2015-09-23 | 百度在线网络技术（北京）有限公司 | Data processing method and device for retrieval |
| CN106997376A (en) * | 2017-02-28 | 2017-08-01 | 浙江大学 | The problem of one kind is based on multi-stage characteristics and answer sentence similarity calculating method |
| CN107273416A (en) * | 2017-05-05 | 2017-10-20 | 深信服科技股份有限公司 | The dark chain detection method of webpage, device and computer-readable recording medium |
| CN107908803A (en) * | 2017-12-26 | 2018-04-13 | 上海智臻智能网络科技股份有限公司 | The response method and device, storage medium, terminal of question and answer interaction |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9740985B2 (en) * | 2014-06-04 | 2017-08-22 | International Business Machines Corporation | Rating difficulty of questions |
-
2018
- 2018-05-31 CN CN201810551681.9A patent/CN108804627B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103379433A (en) * | 2012-04-27 | 2013-10-30 | 华为终端有限公司 | Method and device for giving alarm through distance measurement |
| CN103902652A (en) * | 2014-02-27 | 2014-07-02 | 深圳市智搜信息技术有限公司 | Automatic question-answering system |
| CN104933097A (en) * | 2015-05-27 | 2015-09-23 | 百度在线网络技术（北京）有限公司 | Data processing method and device for retrieval |
| CN106997376A (en) * | 2017-02-28 | 2017-08-01 | 浙江大学 | The problem of one kind is based on multi-stage characteristics and answer sentence similarity calculating method |
| CN107273416A (en) * | 2017-05-05 | 2017-10-20 | 深信服科技股份有限公司 | The dark chain detection method of webpage, device and computer-readable recording medium |
| CN107908803A (en) * | 2017-12-26 | 2018-04-13 | 上海智臻智能网络科技股份有限公司 | The response method and device, storage medium, terminal of question and answer interaction |
Non-Patent Citations (1)
| Title |
|---|
| 定义问答检索关键技术研究;龙华;《中国博士学位论文全文数据库 信息科技辑》;20110715(第07期);第I138-133页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108804627A (en) | 2018-11-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108287858B (en) | Semantic extraction method and device for natural language | |
| CN110543552B (en) | Conversation interaction method and device and electronic equipment | |
| CN112270196B (en) | Entity relationship identification method and device and electronic equipment | |
| CN108763535B (en) | Information acquisition method and device | |
| CN110297893B (en) | Natural language question-answering method, device, computer device and storage medium | |
| CN112346567A (en) | Virtual interaction model generation method and device based on AI (Artificial Intelligence) and computer equipment | |
| CN109522397B (en) | Information processing method and device | |
| CN106649739B (en) | Multi-round interactive information inheritance identification method and device and interactive system | |
| CN111310440A (en) | Text error correction method, device and system | |
| CN108959388B (en) | Information generation method and device | |
| CN112507167A (en) | Method and device for identifying video collection, electronic equipment and storage medium | |
| CN112686051A (en) | Semantic recognition model training method, recognition method, electronic device, and storage medium | |
| CN111950267B (en) | Text triplet extraction method and device, electronic equipment and storage medium | |
| CN116738956A (en) | Prompt template generation method and device, computer equipment and storage medium | |
| CN109033082B (en) | Learning training method and device of semantic model and computer readable storage medium | |
| CN108804627B (en) | Information acquisition method and device | |
| CN110852066B (en) | Multi-language entity relation extraction method and system based on confrontation training mechanism | |
| EP4322066A1 (en) | Method and apparatus for generating training data | |
| WO2023040516A1 (en) | Event integration method and apparatus, and electronic device, computer-readable storage medium and computer program product | |
| CN115470328A (en) | Open field question-answering method based on knowledge graph and related equipment | |
| WO2021082570A1 (en) | Artificial intelligence-based semantic identification method, device, and semantic identification apparatus | |
| CN110852074B (en) | Method and device for generating correction statement, storage medium and electronic equipment | |
| CN113254788A (en) | Big data based recommendation method and system and readable storage medium | |
| CN111858863B (en) | Reply recommendation method, reply recommendation device and electronic equipment | |
| CN114117057A (en) | Keyword extraction method of product feedback information and terminal equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |