CN113139043B

CN113139043B - Question-answer sample generation method and device, electronic equipment and storage medium

Info

Publication number: CN113139043B
Application number: CN202110476855.1A
Authority: CN
Inventors: 张文君; 宋丹丹; 张玉东; 庞海龙
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2021-04-29
Filing date: 2021-04-29
Publication date: 2023-08-04
Anticipated expiration: 2041-04-29
Also published as: CN113139043A

Abstract

The application discloses a question and answer sample generation method, a question and answer sample generation device, electronic equipment and a storage medium, and relates to the technical fields of artificial intelligence, deep learning and big data. The specific implementation scheme is as follows: acquiring a target question text and an auxiliary answer text set; and selecting a target answer text for the target question text according to the similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set so as to obtain a negative question answer sample comprising the target question text and the target answer text. The method and the device can improve the generation efficiency of the negative question-answer samples.

Description

Question-answer sample generation method and device, electronic equipment and storage medium

Technical Field

The application relates to the technical field of data processing, in particular to an artificial intelligence technology, a big data technology and a deep learning technology, and specifically relates to a question-answer sample generation method, a question-answer sample generation device, electronic equipment and a storage medium.

Background

With the development of technology and the continuous progress of internet technology, a search-based interactive community question-answering platform has become an important channel for people to acquire and share knowledge in life and work. Community questions and answers (Community Question Answering, CQA) are direct answers to questions provided by users participating in a web site in combination with an open knowledge sharing web site, utilizing the collective intelligence of the web users.

However, due to the openness of CQAs, the quality of the responses of CQAs varies greatly, some of which can help the questioner to obtain information, and some of which cannot meet the questioner's needs, i.e., answer questions, even contain various irrelevant, low quality, and even malicious information. This difference in content quality is a major problem to be solved in the question-and-answer community.

Disclosure of Invention

The application provides a question and answer sample generation method, a question and answer sample generation device, electronic equipment and a storage medium.

According to an aspect of the present application, there is provided a question-answer sample generation method, including:

acquiring a target question text and an auxiliary answer text set;

and selecting a target answer text for the target question text according to the similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set so as to obtain a negative question answer sample comprising the target question text and the target answer text.

According to another aspect of the present application, there is provided a question-answer sample generation device, including:

the question-answer source data acquisition module is used for acquiring a target question text and an auxiliary answer text set;

and the target answer text screening module is used for selecting a target answer text for the target question text according to the similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set so as to obtain a negative question answer sample comprising the target question text and the target answer text.

According to another aspect of the present application, there is provided an electronic device including:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein, the liquid crystal display device comprises a liquid crystal display device,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the question-answer sample generation method of any embodiment of the present application.

According to another aspect of the present application, there is provided a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the question-answer sample generation method according to any embodiment of the present application.

According to another aspect of the present application, there is provided a computer program product comprising a computer program which, when executed by a processor, implements a question-answer sample generation method as described in any of the embodiments of the present application.

The method and the device can improve the question-answer sample generation efficiency.

It should be understood that the description of this section is not intended to identify key or critical features of the embodiments of the application or to delineate the scope of the application. Other features of the present application will become apparent from the description that follows.

Drawings

The drawings are for better understanding of the present solution and do not constitute a limitation of the present application. Wherein:

FIG. 1 is a schematic diagram of a question and answer sample generation method according to an embodiment of the present application;

FIG. 2 is a schematic diagram of a question and answer sample generation method according to an embodiment of the present application;

FIG. 3 is a schematic diagram of a question and answer sample generation method according to an embodiment of the present application;

FIG. 4 is a schematic diagram of a question and answer sample generation method according to an embodiment of the present application;

FIG. 5 is a schematic diagram of a question and answer sample generation method according to an embodiment of the present application;

FIG. 6 is a schematic diagram of a method for training a question-answer correlation detection model according to an embodiment of the application;

FIG. 7 is a schematic illustration of a question-answer correlation detection model training in accordance with embodiments of the present application;

fig. 8 is a schematic diagram of a question-answer sample generation device according to an embodiment of the present application;

fig. 9 is a block diagram of an electronic device for implementing a question-answer sample generation method according to an embodiment of the present application.

Detailed Description

Exemplary embodiments of the present application are described below in conjunction with the accompanying drawings, which include various details of the embodiments of the present application to facilitate understanding, and should be considered as merely exemplary. Accordingly, one 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 application. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

Fig. 1 is a flowchart of a question-answer sample generation method disclosed in accordance with an embodiment of the present application, which may be applied to the case of generating a negative question-answer sample for training a question-answer correlation detection model. The method of the embodiment can be executed by a question-answer sample generating device, the device can be realized in a software and/or hardware mode, and the device is specifically configured in an electronic device with certain data operation capability, wherein the electronic device can be a client device, a mobile phone, a tablet personal computer, a vehicle-mounted terminal, a desktop computer and the like, and can also be a server device.

S101, acquiring a target question text and an auxiliary answer text set.

The target question text may refer to text containing a question, illustratively, "is apple happy? ". The set of auxiliary answer texts includes at least two auxiliary answer texts, which may refer to texts including answers, and the auxiliary answer texts are different from standard answer texts of the target question text. The standard answer text of the target question text refers to an accurate answer corresponding to the target question text. Illustratively, the target question text is "do apple get good? The standard answer text is "delicious", and the auxiliary answer text is "pineapple delicious". The auxiliary answer text set is used for selecting at least one auxiliary answer text, and at least one question-answer pair is formed by the auxiliary answer text set and the target question text respectively and is used as a negative question-answer sample.

In fact, there is no correlation between the target question text and each of the supplemental answer text in the supplemental answer text set. The target question text can be obtained from the question data collected in each community question-answering platform in the network. The question data is a sentence with the semantic meaning of a question, which is extracted from an interactive text of a question-answer relation, relative to any sentence in the interactive text. And acquiring an auxiliary answer text set from answer data collected in a community question and answer platform in the network. The answer data is a sentence which is extracted from the interactive text of the question-answer relation and has the meaning of an answer relative to any sentence in the interactive text. By way of example, the community question and answer platform may be referred to as an open community question and answer platform. The question and answer data in the community question and answer platform can be Chinese or other foreign characters, such as English. The method comprises the steps of collecting question data and answer data in a random mode, wherein the collecting of the question data and the collecting of the answer data are both carried out in a random mode, and therefore, a correlation does not exist between a target question text collected in the random mode and each auxiliary answer text in an auxiliary answer text set.

S102, selecting a target answer text for the target question text according to the similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set, so as to obtain a negative question-answer sample comprising the target question text and the target answer text.

The question-answer sample is a question-answer pair consisting of one question text and one answer text. The positive question-answer sample refers to a question-answer pair related to a question text and an answer text, and the negative question-answer sample refers to a question-answer pair unrelated to the question text and the answer text, namely a question-answer pair which can be understood as a question-answer not-questions. Illustratively, the question text is "do apple get good? "and the answer text is" pineapple happiness "and the question-answer pair formed by" pineapple happiness "is a negative question-answer sample; the question text is "do apple happy? "and question-answer pairs formed by answer text of" happy "are positive question-answer samples.

The similarity value is used for describing the similarity degree between the target question text and each auxiliary answer text. For example, the auxiliary answer text with the highest similarity value may be selected to be determined as the target answer text, or the first i auxiliary answer texts with high similarity value may be selected to be determined as the target answer text. The similarity value can be calculated by at least one method of word Frequency-inverse text Frequency index (Term Frequency-Inverse Document Frequency, TF-IDF), implicit dirichlet allocation (Latent Dirichlet Allocation, LDA), deep learning and the like. And screening the target answer text according to the similarity between the target question text and each auxiliary answer text, and inquiring the auxiliary answer text with a certain similarity or a plurality of similarities. However, since the question text and the auxiliary answer text set are randomly acquired and have low correlation, answers similar to but not related to the questions can be screened. And forming a negative question-answering sample by the screened target answer text and the target question text with low correlation.

In the prior art, recognition based on relevance or recognition of answers without relevance based on low-quality word list is generally adopted. Based on correlation recognition, the matching degree of questions and answers is measured through correlation technology, and the method is used for solving the non-correlation answer questions. The recognition is based on a low-quality word list, and the low-quality answers are recognized by manually mining the low-quality word list and utilizing a word list matching technology, so that the answer-free questions containing the characteristic words are solved. But for the question text: is apple bad? Answer text 1: apple is a fruit, answer text 2: the pears are delicious. Answer text 1 and answer text 2 are both uncorrelated answer text relative to the question text. The application range of the correlation-based identification method is limited, and the method is only suitable for question and answer content identification scenes with low correlation, and is difficult to accurately identify question and answer questions with certain correlation; the low-quality recognition method based on the vocabulary is only suitable for the low-quality content which can hit the characteristic words, and is easy to detect errors for the missed low-quality content, and meanwhile, the arrangement of the low-quality vocabulary can also have great manual workload.

In view of this, the present application can find out the target answer text which is not related to the specified similarity of the target question text and form a negative question-answer sample by screening the target answer text according to the similarity, reduce the labor cost of generating the negative question-answer sample, improve the generating efficiency, train the question-answer correlation detection model, improve the model detection accuracy, and select a plurality of corresponding target answer texts according to the similarity screening, respectively generate a plurality of negative question-answer samples according to the similarity, thereby generating question-answer pairs with different similarities, increase the similarity range covered by the negative question-answer samples, increase the diversity of the negative question-answer samples, thereby increasing the representativeness of the negative question-answer samples, and simultaneously, accurately detect the negative question-answer samples with different similarities by training the model, i.e. improve the accuracy of the question-answer correlation detection.

According to the technical scheme, the target answer text is selected for the target question text according to the similarity between the target question text and at least two service answer texts in the auxiliary answer text set so as to generate the negative question answer sample, the automatic generation of the negative question answer sample is realized, the labor cost for mining the negative question answer sample is reduced, the generation efficiency of the negative question answer sample is improved, the representativeness of the negative question answer sample is increased, and therefore the accuracy of question answer correlation detection is improved.

Fig. 2 is a flowchart of another question-answer sample generation method disclosed in the embodiment of the present application, which is further optimized and expanded based on the above technical solution, and may be combined with the above various alternative embodiments. In the case that the target question text selects a target answer text, the optimization is as follows: and acquiring a new auxiliary answer text from the auxiliary non-labeling answer set, and replacing the selected target answer text with the new auxiliary answer text to update the auxiliary answer text set for generating a new negative question-answer sample.

S201, acquiring a target question text and an auxiliary answer text set.

S202, selecting a target answer text for the target question text according to the similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set, so as to obtain a negative question-answer sample comprising the target question text and the target answer text.

S203, under the condition that the target question text selects the target answer text, a new auxiliary answer text is obtained from the auxiliary non-labeling answer set, and the selected target answer text is replaced by the new auxiliary answer text so as to update the auxiliary answer text set, so that a new negative question answer sample is generated.

The auxiliary non-annotated answer set may refer to a set of answer text formations without annotated data, wherein the auxiliary non-annotated answer set is used to form and update the auxiliary answer text set to screen out target answer text unrelated to the target question text. The new auxiliary answer text is used for replacing the selected target answer text and is added into the auxiliary answer text set to screen the target answer text of the next round. In fact, under the condition of no labeling answer text, the probability of combining the question text and the answer text with strong correlation into a question-answer sample can be reduced by screening the target answer text, and the correlation between the target question text and the target answer text in the negative question-answer sample is reduced, so that the quality of the negative question-answer sample is improved.

New auxiliary answer text is obtained from the auxiliary non-labeling answer set, the target answer text is replaced in the auxiliary answer text set, and the auxiliary answer text can be supplemented in the auxiliary answer text set after a negative question-answer sample is formed. The auxiliary answer text is updated, so that the problems that the repeated content of the formed negative question-answer sample is too much, the representativeness is reduced, the accuracy of the trained model is reduced and the like caused by the fact that a plurality of target question texts select the same auxiliary answer text as the target answer text can be avoided.

K auxiliary answer texts can be screened from the auxiliary non-labeling answer sets to form auxiliary answer text sets. And each time a negative question-answer sample is generated, the target answer text in the negative question-answer sample in the auxiliary answer text set is popped up, one answer text is selected from the auxiliary non-labeling answer set, and the answer text is added into the auxiliary answer text set, so that the target answer text is replaced, and the auxiliary answer text set is updated. And, the number of the auxiliary answers included in the updated auxiliary answer text set is still K. K is a super parameter, which can be configured by the user. The correlation between the target question text and the target answer text in the negative question-answer sample is determined by the value of K: generally, the larger K, the greater the correlation; the smaller K, the smaller the correlation. It can be understood that the fewer the number of auxiliary answer texts related to the target question text, the more the number of auxiliary answers included in the auxiliary answer text set, the higher the overall correlation between the target question text and the auxiliary answer text set, and the higher the probability of querying the auxiliary answer text related to the target question text, so that the higher the correlation between the target question answer text and the target answer text is obtained through screening; accordingly, the fewer the number of auxiliary answers included in the auxiliary answer text set, the lower the overall correlation between the target question text and the auxiliary answer text set, and the lower the probability of querying the auxiliary answer text related to the target question text, so that the lower the correlation between the target question answer text and the target answer text is obtained through screening.

Optionally, the auxiliary non-labeling answer set is obtained from the original non-labeling answer set.

The original unlabeled answer set may be a set formed by collecting the obtained answer text from at least one community question and answer platform. The auxiliary non-labeling answer set is obtained from the original non-labeling answer set. In practice, the number of answer texts obtained from multiple community question-answering platforms is very large, answer texts are directly selected from original answer texts to form an auxiliary answer text set, random screening is needed in a huge data volume, and screening cost is very high. The auxiliary non-labeling answer set is obtained from the original non-labeling answer set, and the auxiliary answer text set is generated and updated from the auxiliary non-labeling answer set, so that the source data quantity required to be screened can be reduced, the screening cost is reduced, and the generation efficiency of the negative question-answer sample is improved.

In addition, the target question text is obtained from the original unmarked question set. The original unlabeled question set may be a set formed by collecting acquired question text from at least one community question-and-answer platform. The question text can be sequentially selected from the original non-labeling question set and determined to be the target question text. The original unlabeled question set includes a number of question texts that is less than a number of answer texts that the original unlabeled answer set includes.

In a specific example, as shown in fig. 3, the forming step of the negative question-answer sample includes: (1) Randomly extracting m answer texts (answers) and n question texts from a database which is formed by collecting data from a community question-answer platform and respectively serves as an answer data source (namely an auxiliary non-marked answer set) and a question data source (namely an original non-marked question set), wherein m is greater than n, and one question text can be finally screened out to form a negative question-answer sample, wherein the database comprises an original non-marked answer set and an original non-marked question set. (2) From the answer data source (auxiliary non-labeling answer set), K answer texts are randomly popped up and put into an answer matching pool, namely, an auxiliary answer text set is generated. (3) 1 question text A is randomly popped up from the question data source, and similarity calculation is carried out on the question text A and all answer texts (answers) in the answer matching pool. (4) And determining a target answer text (answer) B according to the similarity, popping up the answer matching pool by the B, and forming a question-answer pair (question-answer pair) by the A and the B. (5) the AB question-answer pair is output to the generated data set. (6) From the answer data source (auxiliary non-annotated answer set), 1 answer is randomly popped up to supplement the answer matching pool (auxiliary answer text set), and K answers are kept in the answer matching pool. Repeating (3) - (6) until sufficient production data is obtained.

The answer text with highest similarity can be taken as a target answer text B to pop up, so that the recognition capability of literal similarity can be enhanced, and the recognition of deep semantics can be improved; meanwhile, the degree of k is controlled, the similarity degree of the questions and the answers can be controlled, and the k with too large existence probability is matched with the correct answer; and, the answer data source, the question data source and the answer matching pool popup data are all unreplaced.

According to the technical scheme, the new auxiliary answer text is obtained from the auxiliary non-labeling answer set, the target answer text is replaced, the auxiliary answer text set is updated, the probability of forming a question-answer sample by the question text and the answer text with strong correlation can be reduced, the correlation between the target question text and the target answer text in the negative question-answer sample is reduced, the quality of the negative question-answer sample is improved, the fact that a plurality of target question texts select the same auxiliary answer text as the target answer text can be avoided, the representativeness of the negative question-answer sample is improved, and the accuracy of a trained model is improved.

Fig. 4 is a flowchart of another question-answer sample generation method disclosed in the embodiment of the present application, which is further optimized and expanded based on the above technical solution, and may be combined with the above various alternative embodiments. Selecting a target answer text for the target question text according to the similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set, wherein the selecting comprises the following steps: respectively calculating similarity values between the target question text and at least two auxiliary answer texts in the auxiliary answer text set; wherein the similarity value comprises a literal similarity value and/or a grammar structure similarity value; and selecting a target answer text for the target question text according to the similarity value between the target question text and each auxiliary answer text.

S301, acquiring a target question text and an auxiliary answer text set.

S302, similarity values between the target question text and at least two auxiliary answer texts in the auxiliary answer text set are calculated respectively; wherein the similarity value comprises a literal similarity value and/or a grammatical structure similarity value.

Literal similarity is used to evaluate whether text is alike. Generally, literal similarity refers to text similarity and is semantic independent. Where literal text similarity can solve whatever the semantics, as long as two texts are much longer they can be considered similar. Grammar structure similarity may refer to the similarity of text grammars. In the case where the similarity value includes a literal similarity value and a grammatical structure similarity value, the similarity value may be equal to a weighted sum of the literal similarity value and the grammatical structure similarity value, or the similarity value may be equal to a product between an exponent of the literal similarity value and an exponent of the grammatical structure similarity value, wherein the weights in the weighted sum may be set as desired.

The calculating the similarity value grammar structure similarity value between the target question text and the auxiliary answer text may include: respectively carrying out grammar analysis on the target question text and the auxiliary answer text to generate a grammar structure character sequence corresponding to the target question text and a grammar structure character sequence corresponding to the auxiliary answer text; and calculating the similarity between the grammar structure character sequence corresponding to the target question text and the grammar structure character sequence corresponding to the auxiliary answer text.

The grammar parsing generates grammar structure character sequences, and for example, grammar parsing can be performed by using a grammar parsing generator developed by JAVA. For the two grammar structure character sequences, a TFIDF method and other methods can be adopted to calculate the similarity value, or a grammar structure word embedded (embedding) network can be trained according to the encoder method, and then the cosine similarity value and the like can be calculated. The grammar analysis can comprise analysis of parts of speech, dependency relationship, semantic roles and the like, and fusion is carried out to obtain a grammar structure character sequence. For example, part-of-speech strings, dependency strings, and semantic role strings are concatenated to form a grammar structure string, where concatenation may refer to employing preset spacers, e.g., spaces, short connecting lines, even specified characters, etc. Exemplary, question text is: who is the wife with little brightness. Wherein, "who" is a pronoun, denoted by r; "yes" is a verb, denoted by v; "Xiaoming" is a name, denoted by nr; "typically" occurs after a noun, representing an uncountable noun, denoted by u; "wife" is a noun, denoted by n. The part-of-speech string may be rvnrun. The dependency relationship between "who" and "yes" is a master-called relationship, expressed by SBV; the dependency relationship between "Yes" and "wife" is a moving guest relationship, denoted by VOB; the dependency relationship between "Xiaoming" and "is the" word structure, "denoted by DE; the dependency between "and" wife "is a" word structure, denoted by DE. The corresponding dependency string is sbvvobde. The grammar structure character sequence can be part-of-speech character strings and dependency character strings spliced by short connecting lines, such as rvnrun-sbvvobde. In addition, other representations may be set as needed.

S303, selecting a target answer text for the target question text according to the similarity value between the target question text and each auxiliary answer text, so as to obtain a negative question-answer sample comprising the target question text and the target answer text.

Optionally, the selecting the target answer text for the target question text according to the similarity value between the target question text and each auxiliary answer text includes: dividing the similarity value between the target question text and each auxiliary answer text into at least two similarity intervals; the number of the similarity intervals is the same as the number of the target answer texts to be selected; and selecting at least two target answer texts for the target question texts according to the at least two similarity intervals.

The similarity interval is used for dividing the similarity value so as to screen a plurality of target answer texts. The number of the similarity intervals is the same as the number of the target answer texts to be selected, and the similarity values are divided according to the number of the target answer texts to be selected, so that the similarity intervals with the same number are obtained.

For example, one auxiliary answer text may be selected in each similarity interval, so as to obtain target answer texts with the number of similarity intervals. Or, according to the distribution situation of the similarity values, at least two auxiliary answer texts are selected in the similarity intervals of the similarity values with more distribution, and no auxiliary answer text is selected in the similarity intervals of the similarity values with less distribution, and the setting can be specifically performed according to the needs.

In a specific example, the similarity value range is between 0 and 1, and the similarity interval may be divided into: 0-0.1, 0.1-0.2. 0.2 … … 0.9, 0.9-1.0, or evenly dividing any gear, and can be set according to application requirements. Each similarity interval may be one, more or less, or more.

By dividing a plurality of similarity intervals and selecting a plurality of target answer texts in the plurality of similarity intervals, negative question-answering samples with different similarities can be generated, the diversity of the negative question-answering samples is improved, the representativeness of the negative question-answering samples is improved, and therefore the accuracy of question-answering correlation detection is improved.

According to the technical scheme, the similarity between the text of each auxiliary answer and the text of the question and/or the similarity value of the grammar structure are calculated, the similarity is calculated by adopting a way of combining the words and the grammar, and the similarity between the answer and the question can be calculated in a multi-dimensional manner, so that negative question-answering samples with different similarities are generated, the diversity of the negative question-answering samples is improved, the representativeness of the negative question-answering samples is improved, and the accuracy of the question-answering correlation detection is improved.

Fig. 5 is a flowchart of another question-answer sample generation method disclosed in the embodiment of the present application, which is further optimized and expanded based on the above technical solution, and may be combined with the above various alternative embodiments. In the case of obtaining a negative question-answer sample comprising the target question text and the target answer text, optimizing to be: acquiring a question and answer sample; and training a question-answer correlation detection model by adopting the generated negative question-answer sample and the positive question-answer sample.

S401, acquiring a target question text and an auxiliary answer text set.

And S402, selecting a target answer text for the target question text according to the similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set so as to obtain a negative question-answer sample comprising the target question text and the target answer text.

S403, acquiring a positive question-answer sample.

The positive question-answer samples correspond to the negative question-answer samples. The positive question-answer sample is a question-answer pair formed by the associated question text and answer text.

The question and answer sample can be obtained by manual labeling and collection, or can be obtained by screening in a community question and answer platform according to posterior data. For example, the posterior data may include at least one of authority information of a user of the answer text, an answer to which the user has a challenge, whether the answer text is original text, and interaction statistics of the answer text. The interactive statistical information of the answer text may include statistical information of at least one of comment data, praise data, comment data, and the like. And generating a question-answer pair as a positive question-answer sample according to the posterior data and the reliable answer text and the question text.

S404, training a question-answer correlation detection model by adopting the generated negative question-answer sample and the positive question-answer sample.

The question-answer correlation detection model is used for detecting the correlation between a question and an answer. The question-answer correlation detection model may include a semantic and grammatical parsing layer and a convolutional neural network, wherein the convolutional neural network may be replaced with a recurrent neural network or other neural network.

Optionally, the training question-answer correlation detection model includes: analyzing a question-answer sample to form semantic information and grammar information, wherein the grammar information comprises part-of-speech information and/or dependency information, and the question-answer sample comprises the positive question-answer sample and the negative question-answer sample; and training a question-answer correlation detection model according to the semantic information and the grammar information.

As shown in fig. 6, the question-answer correlation detection model may include an parsing layer (parameter), a coding layer, an embedding layer, a convolution layer (may include a plurality of convolution operations), a pooling layer, a stitching layer, and a neural network (full-layers), where the neural network may employ the aforementioned convolution neural network or cyclic neural network, and so on. And analyzing the question-answer sample through an analysis layer to obtain semantic information and grammar information, wherein the grammar information comprises part-of-speech information and/or dependency information. The three information are respectively encoded, embedded, convolved and pooled to correspondingly obtain three vectors, the three vectors are spliced to form a vector, and the vector is processed through a full connection layer to obtain a correlation detection result.

Accordingly, as shown in fig. 7, the question-answer pair is parsed into semantic information, part-of-speech information and dependency information, and information compression is performed respectively. The information compression is shown in fig. 6, and the operations of an analysis layer, a coding layer, an embedding layer, a convolution layer, a pooling layer and the like are performed to map the information into vectors, the three vectors obtained by compressing the three paths of information are spliced to obtain a vector, and the vector is detected through a neural network to obtain a correlation detection result. The target question text and the target answer text are respectively analyzed, and part-of-speech information (verb, noun, adjective and the like) lists, dependency relationship (moving-guest relationship, main-predicate relationship, association structure and the like) lists and semantic information lists are analyzed. Information compression, convolutional neural networks may be employed. The semantic information list, the part-of-speech information list, and the dependency information list are encoded separately, and one-bit efficient encoding (one-hot) may be used. And carrying out the ebadd operation on the coding result, namely compressing the sparse matrix formed by the one-hot into a dense matrix, namely carrying out data dimension reduction. The convolution and pooling may be compression into semantic vectors, part-of-speech vectors, and dependency vectors via a plurality of one-dimensional convolutions and one-dimensional pooling. And splicing the semantic vector, the part-of-speech vector and the dependency relation vector after information compression to form a new vector. And then, the spliced vector is subjected to a plurality of full connection layers, and finally a 0 and 1 classification is obtained, namely, a final correlation detection result of whether correlation exists, wherein 0 can be uncorrelated, and 1 can be correlated.

By combining grammar features and semantic features to detect question-answer correlation, feature information can be enriched, and correlation detection accuracy is improved.

According to the technical scheme, the positive question-answering sample is obtained, and the generated negative question-answering sample training model with high representativeness and high generation efficiency is combined, so that the training efficiency of the question-answering correlation detection model can be improved, the model training labor cost is reduced, and the question-answering correlation detection accuracy is improved.

Fig. 8 is a block diagram of a question-answer sample generation device in an embodiment of the present application, which is applicable to a case of generating an image sample for target detection of a truncated object according to an embodiment of the present application. The device is realized by software and/or hardware, and is specifically configured in the electronic equipment with certain data operation capability.

A question-answer sample generation device 500 as shown in fig. 8, comprising: a question-answer source data acquisition module 501 and a target answer text screening module 502; wherein, the liquid crystal display device comprises a liquid crystal display device,

a question-answer source data acquisition module 501, configured to acquire a target question text and an auxiliary answer text set;

and a target answer text screening module 502, configured to select a target answer text for the target question text according to a similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set, so as to obtain a negative question-answer sample including the target question text and the target answer text.

Further, the question-answer sample generating device further includes: and the auxiliary answer text updating module is used for acquiring a new auxiliary answer text from the auxiliary non-labeling answer set under the condition that the target answer text is selected for the target question text, and replacing the selected target answer text with the new auxiliary answer text so as to update the auxiliary answer text set and generate a new negative question answer sample.

Further, the auxiliary non-labeling answer set is obtained from the original non-labeling answer set.

Further, the target answer text filtering module 502 includes: a similarity value obtaining unit, configured to calculate similarity values between the target question text and at least two auxiliary answer texts in the auxiliary answer text set, respectively; wherein the similarity value comprises a literal similarity value and/or a grammar structure similarity value; and the target answer text determining unit is used for selecting a target answer text for the target question text according to the similarity value between the target question text and each auxiliary answer text.

Further, the similarity value obtaining unit includes: the similarity interval dividing subunit is used for dividing the similarity value between the target question text and each auxiliary answer text into at least two similarity intervals; the number of the similarity intervals is the same as the number of the target answer texts to be selected; and the target answer text partition acquisition subunit is used for selecting at least two target answer texts for the target question text according to the at least two similarity intervals.

Further, the question-answer sample generating device further includes: the positive question and answer sample acquisition module is used for acquiring a positive question and answer sample under the condition that a negative question and answer sample comprising the target question text and the target answer text is acquired; and the model training module is used for jointly training a question-answer correlation detection model by adopting the generated negative question-answer sample and the positive question-answer sample.

Further, the model training module includes: the question-answer sample analysis unit is used for analyzing a question-answer sample to form semantic information and grammar information, wherein the grammar information comprises part-of-speech information and/or dependency information, and the question-answer sample comprises the positive question-answer sample and the negative question-answer sample; the semantic grammar fusion training unit is used for training a question-answer correlation detection model according to the semantic information and the grammar information.

The object detection device can execute the question-answer sample generation method provided by any embodiment of the application, and has the corresponding functional modules and beneficial effects of executing the question-answer sample generation method.

According to embodiments of the present application, there is also provided an electronic device, a readable storage medium and a computer program product.

Fig. 9 shows a schematic block diagram of an example electronic device 600 that may be used to implement embodiments of the present application. 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 telephones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the application described and/or claimed herein.

As shown in fig. 9, the apparatus 600 includes a computing unit 601 that can perform various appropriate actions and processes according to a computer program stored in a Read Only Memory (ROM) 602 or a computer program loaded from a storage unit 608 into a Random Access Memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the device 600 may also be stored. The computing unit 601, ROM 602, and RAM 603 are connected to each other by a bus 604. An input/output (I/O) interface 605 is also connected to bus 604.

Various components in the device 600 are connected to the I/O interface 605, including: an input unit 606 such as a keyboard, mouse, etc.; an output unit 607 such as various types of displays, speakers, and the like; a storage unit 608, such as a magnetic disk, optical disk, or the like; and a communication unit 609 such as a network card, modem, wireless communication transceiver, etc. The communication unit 609 allows the device 600 to exchange information/data with other devices via a computer network, such as the internet, and/or various telecommunication networks.

The computing unit 601 may be a variety of general and/or special purpose processing components having processing and computing capabilities. Some examples of computing unit 601 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various specialized Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the respective methods and processes described above, such as a question-answer sample generation method. For example, in some embodiments, the question and answer sample generation method may be implemented as a computer software program tangibly embodied on a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program may be loaded and/or installed onto the device 600 via the ROM 602 and/or the communication unit 609. When the computer program is loaded into the RAM 603 and executed by the computing unit 601, one or more steps of the question-answer sample generation method described above may be performed. Alternatively, in other embodiments, the computing unit 601 may be configured to perform the question-answer sample generation method in any other suitable way (e.g. by means of firmware).

Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuit systems, field Programmable Gate Arrays (FPGAs), application Specific Integrated Circuits (ASICs), application Specific Standard Products (ASSPs), systems On 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, the one or more computer programs may be executed and/or interpreted on a programmable system including at least one programmable processor, which may be a special purpose or general-purpose programmable processor, that may receive data and instructions from, and transmit data and instructions to, a storage system, at least one input device, and at least one output device.

Program code for carrying out methods of the present application may be written in any combination of one or more programming languages. These program code 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 code, when executed by the processor or controller, causes the functions/operations specified in the flowchart and/or block diagram to be implemented. 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 application, 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. The 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 pointing device (e.g., a mouse or 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 may 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 input, speech input, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a background 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 background, 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), blockchain networks, and the internet.

The computer system may include a client and a server. The client and server are typically 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 can be a cloud server, also called a cloud computing server or a cloud host, and is a host product in a cloud computing service system, so that the defects of high management difficulty and weak service expansibility in the traditional physical hosts and VPS service are overcome.

It should be appreciated that various forms of the flows shown above may be used to reorder, add, or delete steps. For example, the steps described in the present disclosure may be performed in parallel, sequentially, or in a different order, provided that the desired results of the technical solutions disclosed in the present application are achieved, and are not limited herein.

The above embodiments do not limit the scope of the application. It will be apparent to those skilled in the art that various modifications, combinations, sub-combinations and alternatives are possible, depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application are intended to be included within the scope of the present application.

Claims

1. A question and answer sample generation method comprises the following steps:

acquiring a target question text and an auxiliary answer text set;

selecting a target answer text for the target question text according to the similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set so as to obtain a negative question answer sample comprising the target question text and the target answer text;

The auxiliary answer text refers to a text containing an answer, and the auxiliary answer text is different from a standard answer text of the target question text; the question-answer sample is a question-answer pair consisting of a question text and an answer text; the auxiliary answer text set is used for selecting at least one auxiliary answer text, and forming at least one question-answer pair with the target question text respectively to serve as a negative question-answer sample; the positive question-answer sample refers to a question-answer pair related to a question text and an answer text, and the negative question-answer sample refers to a question-answer pair unrelated to the question text and the answer text.

2. The method of claim 1, further comprising, in the case where the target question text selects a target answer text:

and acquiring a new auxiliary answer text from the auxiliary non-labeling answer set, and replacing the selected target answer text with the new auxiliary answer text to update the auxiliary answer text set for generating a new negative question-answer sample.

3. The method of claim 2, wherein the set of auxiliary unlabeled answers is obtained from an original set of unlabeled answers.

4. The method of claim 1, wherein the selecting the target answer text for the target question text based on a similarity between the target question text and at least two of the set of auxiliary answer texts comprises:

Respectively calculating similarity values between the target question text and at least two auxiliary answer texts in the auxiliary answer text set; wherein the similarity value comprises a literal similarity value and/or a grammar structure similarity value;

and selecting a target answer text for the target question text according to the similarity value between the target question text and each auxiliary answer text.

5. The method of claim 4, wherein the selecting the target answer text for the target question text according to the similarity value between the target question text and each of the auxiliary answer texts comprises:

dividing the similarity value between the target question text and each auxiliary answer text into at least two similarity intervals; the number of the similarity intervals is the same as the number of the target answer texts to be selected;

and selecting at least two target answer texts for the target question texts according to the at least two similarity intervals.

6. The method of claim 1, further comprising, in the case of obtaining a negative question-answer sample comprising the target question text and the target answer text:

acquiring a question and answer sample;

And training a question-answer correlation detection model by adopting the generated negative question-answer sample and the positive question-answer sample.

7. The method of claim 6, wherein the training question-answer correlation detection model comprises:

analyzing a question-answer sample to form semantic information and grammar information, wherein the grammar information comprises part-of-speech information and/or dependency information, and the question-answer sample comprises the positive question-answer sample and the negative question-answer sample;

and training a question-answer correlation detection model according to the semantic information and the grammar information.

8. A question-answer sample generation device, comprising:

the target answer text screening module is used for selecting a target answer text for the target question text according to the similarity between the target question text and at least two auxiliary answer texts in the auxiliary answer text set so as to obtain a negative question answer sample comprising the target question text and the target answer text;

9. The apparatus of claim 8, further comprising:

and the auxiliary answer text updating module is used for acquiring a new auxiliary answer text from the auxiliary non-labeling answer set under the condition that the target answer text is selected for the target question text, and replacing the selected target answer text with the new auxiliary answer text so as to update the auxiliary answer text set and generate a new negative question answer sample.

10. The apparatus of claim 9, wherein the secondary set of unlabeled answers is obtained from an original set of unlabeled answers.

11. The apparatus of claim 8, wherein the target answer text screening module comprises:

a similarity value obtaining unit, configured to calculate similarity values between the target question text and at least two auxiliary answer texts in the auxiliary answer text set, respectively; wherein the similarity value comprises a literal similarity value and/or a grammar structure similarity value;

and the target answer text determining unit is used for selecting a target answer text for the target question text according to the similarity value between the target question text and each auxiliary answer text.

12. The apparatus of claim 11, wherein the similarity value acquisition unit comprises:

the similarity interval dividing subunit is used for dividing the similarity value between the target question text and each auxiliary answer text into at least two similarity intervals; the number of the similarity intervals is the same as the number of the target answer texts to be selected;

and the target answer text partition acquisition subunit is used for selecting at least two target answer texts for the target question text according to the at least two similarity intervals.

13. The apparatus of claim 8, further comprising:

the positive question and answer sample acquisition module is used for acquiring a positive question and answer sample under the condition that a negative question and answer sample comprising the target question text and the target answer text is acquired;

and the model training module is used for jointly training a question-answer correlation detection model by adopting the generated negative question-answer sample and the positive question-answer sample.

14. The apparatus of claim 13, wherein the model training module comprises:

the question-answer sample analysis unit is used for analyzing a question-answer sample to form semantic information and grammar information, wherein the grammar information comprises part-of-speech information and/or dependency information, and the question-answer sample comprises the positive question-answer sample and the negative question-answer sample;

The semantic grammar fusion training unit is used for training a question-answer correlation detection model according to the semantic information and the grammar information.

15. An electronic device, comprising:

at least one processor; and

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the question-answer sample generation method of any one of claims 1-7.

16. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the question-answer sample generation method according to any one of claims 1-7.