CN110555207A - Sentence recognition method, sentence recognition device, machine equipment and computer-readable storage medium - Google Patents
Sentence recognition method, sentence recognition device, machine equipment and computer-readable storage medium Download PDFInfo
- Publication number
- CN110555207A CN110555207A CN201810558865.8A CN201810558865A CN110555207A CN 110555207 A CN110555207 A CN 110555207A CN 201810558865 A CN201810558865 A CN 201810558865A CN 110555207 A CN110555207 A CN 110555207A
- Authority
- CN
- China
- Prior art keywords
- word
- input
- sentence
- intention
- input sentence
- 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.)
- Pending
Links
Landscapes
- Machine Translation (AREA)
Abstract
the invention discloses a sentence recognition method and device. The sentence recognition method comprises the following steps: taking a word as a unit, and carrying out vectorization on the word in an input statement according to the corresponding context to obtain a distributed expression corresponding to the word; obtaining a hidden vector corresponding to the word through forward and backward traversal of the distributed expression corresponding to the word in the input statement; respectively predicting the attribute labels of the words and the intention labels of the input sentences according to the hidden vectors corresponding to the words in the input sentences to obtain the attribute labels corresponding to the words and the intention labels corresponding to the input sentences; and searching the input sentence according to the attribute tag corresponding to the word in the input sentence and the intention tag corresponding to the input sentence. By adopting the method provided by the invention, the sentence identification can be more accurate.
Description
Technical Field
The present invention relates to the field of natural language processing technologies, and in particular, to a sentence recognition method, apparatus, machine device, and computer-readable storage medium.
Background
Recognizing the semantics expressed by a user input sentence is a core module in voice interaction, and comprises two tasks of intention recognition and attribute labeling. The intention recognition is to judge the real intention of the user according to the spoken text input by the user. In a smart television voice interaction scene, common intention identification includes playing a movie, watching a television play, browsing a variety program, playing music, viewing weather and the like; the attribute labels refer to specific information related to the user intention, and the voice interaction system completes specific tasks related to the user intention according to the extracted information related to the intention, for example, searching a movie played by Liudebua according to the voice of the user.
User intention recognition generally uses a machine learning algorithm to perform semantic classification on spoken texts input by users, and attribute labeling generally uses template matching or conditional random fields to perform task-related specific information extraction. The existing method is to divide the intention classification and the attribute labeling into two subtasks, firstly, the intention classification is carried out to obtain the intention label, then, the attribute labeling is carried out to obtain the attribute label under the recognized intention classification.
Therefore, in the implementation of the prior art, the limitation of low accuracy in user intention identification and attribute labeling due to the transfer of errors needs to be solved.
disclosure of Invention
In order to solve the above technical problems, an object of the present invention is to provide a sentence recognition method, apparatus, machine device and computer-readable storage medium.
the technical scheme adopted by the invention is as follows:
A sentence recognition method, comprising: taking a word as a unit, and carrying out vectorization on the word in an input statement according to the corresponding context to obtain a distributed expression corresponding to the word; obtaining a hidden vector corresponding to the word through forward and backward traversal of the distributed expression corresponding to the word in the input statement; respectively predicting the attribute labels of the words and the intention labels of the input sentences according to the hidden vectors corresponding to the words in the input sentences to obtain the attribute labels corresponding to the words and the intention labels corresponding to the input sentences; and searching the input sentence according to the attribute tag corresponding to the word in the input sentence and the intention tag corresponding to the input sentence.
Further, the obtaining of the distributed expression corresponding to the word by vectorizing the word in the input sentence according to the corresponding context in units of the word includes: calculating the probability of occurrence in the stored text corpus information for each word in the input sentence relative to all peripheral words in the input sentence, and obtaining a probability sequence formed by each word in the input sentence relative to all peripheral words in the input sentence; and mapping the probability sequence of each word in the input statement into a low-dimensional vector to obtain the distributed expression corresponding to the word in the input statement.
Further, the obtaining of the hidden vector corresponding to the word through forward traversal and backward traversal of the distributed expression corresponding to the word in the input sentence includes: respectively performing forward traversal and backward traversal according to the distributed expression corresponding to the word in the input statement to obtain a forward implicit vector and a backward implicit vector of the word; and splicing the forward hidden vector and the backward hidden vector of the word in the input statement to obtain the hidden vector corresponding to the word.
further, the performing attribute tag prediction on the word according to the hidden vector corresponding to the word in the input statement to obtain the attribute tag corresponding to the word includes: carrying out nonlinear mapping on the hidden vector corresponding to the word in the input statement through a target parameter matrix to obtain a weighted vector of the word; carrying out probability normalization on the weighted vectors of the words in the input statement to obtain attribute label probability distribution vectors corresponding to the words; and selecting the attribute label corresponding to the maximum probability from the attribute label probability distribution vector corresponding to the word in the input statement as the attribute label corresponding to the word.
further, the predicting the intention label corresponding to the input statement according to the hidden vector corresponding to the word in the input statement to obtain the intention label of the input statement includes: splicing the hidden vectors corresponding to all the words in the input statement to obtain the hidden vectors corresponding to the input statement; carrying out nonlinear mapping on the hidden vector corresponding to the input statement through a target parameter matrix to obtain a weighted vector of the input statement; carrying out probability normalization on the weighted vector of the input statement to obtain an intention label probability distribution vector corresponding to the input statement; and selecting the intention label corresponding to the maximum probability from the intention label probability distribution vector corresponding to the input statement as the intention label corresponding to the input statement.
Further, the sentence recognition method further comprises: summing the attribute tag prediction deviation and the intention tag prediction deviation to obtain a sentence identification deviation corresponding to the input sentence, wherein the attribute tag prediction deviation is an error value between an intention tag corresponding to a word in the input sentence and a real attribute tag, and the intention tag prediction deviation is an error value between an intention tag corresponding to the input sentence and a real intention tag; minimizing the sentence recognition deviation to obtain an intermediate parameter matrix; and updating the intermediate parameter matrix into a target parameter matrix.
a sentence recognition apparatus comprising: the low-dimensional vector mapping module is used for mapping the words in the input statement into low-dimensional vectors according to the context corresponding to the words in the input statement so as to obtain the distributed expression of the words; a forward and backward traversal module for performing forward and backward traversal according to the distributed expression of the words in the input sentence to obtain the hidden vectors of the words; and the sentence label prediction module is used for respectively predicting the attribute labels of the words and the intention labels of the input sentences by the user according to the hidden vectors corresponding to the words in the input sentences so as to obtain the attribute labels corresponding to the words and the intention labels corresponding to the input sentences.
Further, the apparatus in sentence recognition further comprises: a deviation calculation module, configured to perform summation operation on an attribute tag prediction deviation and an intention tag prediction deviation to obtain a sentence recognition deviation of the input sentence, where the attribute tag prediction deviation is an error value between an intention tag corresponding to a word in the input sentence and a real attribute tag, and the intention tag prediction deviation is an error value between an intention tag corresponding to the input sentence and a real intention tag; the deviation value processing module is used for carrying out minimization processing on the sentence recognition deviation to obtain a middle parameter matrix; and the parameter adjusting module is used for updating the intermediate parameter matrix into a target parameter matrix.
A machine device comprising a processor and a memory, the memory having stored thereon computer readable instructions which, when executed by the processor, implement a sentence recognition method as described above.
a computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out a sentence recognition method as set forth above.
In the technical scheme, firstly, the words in the input sentence are vectorized through the context information of the words in the input sentence to obtain the distributed expression corresponding to the words, then the distributed expression corresponding to the words in the input sentence is traversed forwards and backwards to obtain the hidden vectors of the words, and finally the attribute labels corresponding to the words and the intention labels corresponding to the input sentence are respectively predicted according to the hidden vectors corresponding to the words in the input sentence.
In the technical scheme, the attribute label corresponding to each word in the input statement and the intention label corresponding to the input statement are obtained by respectively performing label prediction on the hidden vector corresponding to each word in the input statement, the relevance of sharing semantic text information between intention classification and attribute labeling is fully considered, and the problems of user intention identification and low accuracy of attribute labeling caused by error transmission in the prior art are solved.
in addition, the technical scheme adopted by the invention carries out vectorization on the words in the input sentence through the context information of the words in the input sentence to obtain the distributed expression corresponding to the words, and carries out forward and backward traversal on the distributed expression corresponding to the words in the input sentence to obtain the hidden vector of the words, so that the hidden vector of each word in the input sentence fully considers the context information in the input sentence, and the sentence recognition is more accurate.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description, serve to explain the principles of the invention.
Fig. 1 is a block diagram of an intelligent television according to the present invention.
FIG. 2 is a flow diagram illustrating a sentence recognition method in accordance with an exemplary embodiment.
FIG. 3 is a flow chart of one embodiment of step 210 in the corresponding embodiment of FIG. 2.
FIG. 4 is a diagram illustrating a process for traversing an input sentence in both forward and backward directions in an exemplary embodiment.
FIG. 5 is a flow chart of one embodiment of step 230 of the corresponding embodiment of FIG. 2.
FIG. 6 is a flow diagram of one embodiment of step 250 of the corresponding embodiment of FIG. 2.
FIG. 7 is a flow chart of step 250 in the corresponding embodiment of FIG. 2 in another embodiment.
FIG. 8 is a schematic diagram illustrating analysis of results obtained according to the method shown in FIG. 2, according to one embodiment.
FIG. 9 is a schematic diagram illustrating analysis of results obtained according to the method shown in FIG. 2 according to another embodiment.
FIG. 10 is a schematic diagram illustrating analysis of results obtained according to the method shown in FIG. 2 according to another embodiment.
FIG. 11 is a schematic diagram illustrating analysis of results obtained according to the method shown in FIG. 2 according to another embodiment.
FIG. 12 is a flow chart illustrating a sentence recognition method in accordance with another exemplary embodiment.
Fig. 13 is a block diagram illustrating a sentence recognition apparatus according to an example embodiment.
Fig. 14 is a block diagram illustrating a sentence recognition apparatus according to another exemplary embodiment.
While specific embodiments of the invention have been shown by way of example in the drawings and will be described in detail hereinafter, such drawings and description are not intended to limit the scope of the inventive concepts in any way, but rather to explain the inventive concepts to those skilled in the art by reference to the particular embodiments.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the invention, as detailed in the appended claims.
Fig. 1 is a block diagram of an intelligent television. It should be noted that the technical solution provided by the present invention can be applied to all devices related to natural language processing and machine learning technology, and is not limited to the smart tv shown in fig. 1, and the smart tv shown in fig. 1 is proposed for facilitating understanding of the content of the present invention, and cannot be considered as providing any limitation to the scope of the present invention. Also, the smart tv shown in fig. 1 is only a block diagram of one type of smart tv shown according to an exemplary embodiment, and the smart tv cannot be interpreted as needing to rely on or have to have one or more components in the exemplary smart tv shown in fig. 1.
As shown in fig. 1, the smart tv includes at least one processor 120 (only one shown in fig. 1), a memory 110, a communication module 130, an audio module 140, and a display screen 150.
The memory 110 is used as a carrier for resource storage to store computer programs and modules, such as computer readable instructions and modules corresponding to the sentence recognition method and apparatus in the exemplary embodiment of the present invention, and the processor 120 executes various functions and data processing by executing the computer readable instructions stored in the memory 110, that is, completes the sentence recognition method.
The memory 110 may be random access memory, e.g., high speed random access memory, non-volatile memory such as one or more magnetic storage devices, flash memory, or other solid state memory. The storage means may be a transient storage or a permanent storage.
The communication module 130 is used for transceiving wireless or wired signals, thereby communicating with other devices through the wireless or wired signals. Communication networks on which wireless or wireline signals are based include, but are not limited to, cellular telephone networks, wireless local or metropolitan area networks, and the like, which may use various communication standards, protocols, and technologies.
The audio module 140 is used for inputting/outputting audio signals, and may include one or more microphones, one or more speakers, and one or more audio interfaces, such as a headphone interface, a headset interface, and the like. The user can input a voice signal to the smart television through the audio module 140 to perform voice control on the smart television.
the display screen 150 is used for providing an output interface, so that the smart television displays and outputs output contents formed by any one form or combination of characters, pictures or videos through the display screen 150.
it is to be understood that the configuration shown in fig. 1 is merely illustrative, and the smart tv may further include more or less components than those shown in fig. 1, or have different components than those shown in fig. 1.
Furthermore, the present invention can be implemented by hardware circuits or by a combination of hardware circuits and software, and thus, the implementation of the present invention is not limited to any specific hardware circuits, software, or a combination of both.
FIG. 2 is a flow diagram illustrating a sentence recognition method in accordance with an exemplary embodiment. The sentence recognition method shown in fig. 2 can be applied to the smart television shown in fig. 1, and can also be applied to other devices related to natural language processing and machine learning technologies. As shown in fig. 2, the sentence recognition method includes the following steps:
In step 210, the vectorization of the word in the input sentence is performed according to the corresponding context to obtain the distributed expression corresponding to the word, taking the word as a unit.
in the example that the sentence recognition method is applied to the smart television shown in fig. 1, the input sentence may be text information obtained by performing voice recognition on a voice signal input by a user by the smart television, or text information input to the smart television by the user through a remote controller or in other manners.
Vectorizing a word in an input sentence is a process of mapping the word in the input sentence into a low-dimensional vector, and the obtained low-dimensional vector is called as distributed expression of the word in the input sentence, so that semantic association between different words is established through the distributed expression of the word in the input sentence.
for a word in an input sentence, mapping the word into a low-dimensional vector is specifically obtained by performing conditional mapping on the word with respect to other peripheral words in the input sentence, and the other peripheral words in the input sentence are also referred to as context information corresponding to the word.
The mapping condition may specifically be that probabilities of the word appearing in the text corpus information stored in the text corpus information respectively corresponding to other peripheral words are respectively and sequentially calculated, and then all obtained probability values are combined to form a low-dimensional vector according to the arrangement order of the peripheral words in the input sentence.
For example, when conditional mapping is performed on "liu" words in an input sentence "i want to see liu de hua movie", probabilities of "liu" appearing in text corpus information with respect to "i", "want", "see", "de", "hua", "e", "shadow" are respectively and sequentially calculated, then the obtained probability values are sequentially combined according to the sequence corresponding to "i", "want", "see", "de", "hua", "e" and "shadow" to form an expression form of a vector, and the obtained vector is used as a low-dimensional vector of the "liu" word.
Specifically, the calculation of the probability of occurrence of a word in the input sentence relative to the text corpus information in which other peripheral words are stored is to calculate the probability of occurrence of the current word simultaneously with the text corpus information in which other peripheral words are stored. In one embodiment, the probability of "liu" and "i" occurring simultaneously in the text corpus information is calculated to be 0.25, the probability of "liu" and "thought" occurring simultaneously in the text corpus information is calculated to be 0.2, the probability of "liu" and "look" occurring simultaneously in the text corpus information is calculated to be 0.12, the probability of "liu" and "de" occurring simultaneously in the text corpus information is calculated to be 0.88, the probability of "liu" and "hua" occurring simultaneously in the text corpus information is calculated to be 0.87, the probability of "liu" and "Ji" occurring simultaneously in the text corpus information is calculated to be 0.11, the probability of "liu" and "Ji" occurring simultaneously in the text corpus information is calculated to be 0.76, and the probability of "liu" and "yu" occurring simultaneously in the text corpus information is calculated to be 0.67, and finally the distributed expression of "liu" is calculated to be (0.25,0.2,0.13,0.88, 0.76,0.67, 76, and 0.67).
in order to obtain the distributed expression of each word in the input sentence, each word in the input sentence needs to be mapped to obtain the corresponding low-dimensional vector according to the method.
fig. 3 is a flow chart illustrating an implementation step 210 in an exemplary embodiment. As shown in fig. 3, the implementing step 210 may include the steps of:
In step 211, the probability of occurrence in the stored text corpus information is calculated for each word in the input sentence with respect to all the peripheral words in the input sentence, and a probability sequence formed by each word in the input sentence with respect to all the peripheral words in the input sentence is obtained.
For one word in the input sentence, the word includes all words left after the word is removed from the first word to the last word in the input sentence relative to all peripheral words in the input sentence.
still take the example of the input sentence "i want to see the movie of liu de hua", wherein the word "liu" includes all the words in "i want to see the movie of liu de hua" with respect to all the peripheral words in the input sentence, and the word "electricity" includes all the words in "i want to see the movie of liu de hua" with respect to all the peripheral words in the input sentence.
The probability appearing in the stored text corpus information is calculated for each character in the input sentence relative to all the peripheral characters in the input sentence, so that the association degree of each character in the input sentence with the peripheral characters is obtained, and the higher the probability value is, the higher the association degree of the current character with the peripheral characters is, for example, a 'movie' two character is used as a common word, and the higher the probability value is obtained by calculating the 'shadow' of an 'electronic' character relative to the peripheral characters.
in one exemplary embodiment, the probability of occurrence in the stored text corpus information for each word in the input sentence separately over all surrounding words in the input sentence is calculated according to the word2vector model.
A large amount of text corpus information is stored in the background corpus and is used for sentence training of the word2vector model. The background corpus may be a corpus downloaded from a multimedia medium, such as a Beijing university CCL corpus, a dog-searching text classification corpus, and the like.
And according to text corpus information stored in the background corpus, using a word2vector model to calculate the common occurrence probability of each word in the input sentence and each peripheral word in the background corpus so as to obtain a probability sequence formed by each word in the input sentence relative to all peripheral words in the input sentence.
The objective function of the word2vector model is as follows:
minimizeJ＝-log P(wt|wt-m,…,wt-1,wt+1,…,wt+m)
Wherein "wt"represents the current word to be subjected to probability calculation," t "represents the position of the current word in the input sentence, for example, in the input sentence" i want to see the movie of liu de hua ", the value of" t "when the probability calculation is performed on the current word" liu "is 4; "wt-m,…,wt-1,wt+1,…,wt+m"represents the current word" wt"all surrounding words in the input sentence," m "indicates the position of the surrounding word in the input sentence relative to the current word.
Taking the input sentence as "i want to see the movie of liu de hua" as an example, when the word2vector model is used to calculate the probability sequence of "liu", the position of the peripheral word "i" in the input sentence is: the 3 rd word ahead relative to the "Liu" word, the position in the input sentence where the peripheral word "shadow" is: the 5 th word backward relative to the Liu word, therefore, the peripheral word I is represented as wt-3", peripheral word"i "indicate as" wt+5". As another example, when calculating the probability sequence of "electricity" using the word2vector model, the peripheral word "I" is represented as "wt-7", the peripheral word" I "is denoted as" wt+1”。
In an exemplary embodiment, it is assumed that the probability sequences of "liu" words appearing in the background corpus are 0.25,0.2,0.13,0.88,0.87,0.11,0.76 and 0.67, respectively, and the probability sequences of "liu" words appearing in the background corpus are: 0.53,0.18,0.24,0.43,0.39,0.33,0.43, 0.58.
In step 213, the probability sequence of each word in the input sentence is mapped to a low-dimensional vector, so as to obtain a distributed expression corresponding to the word in the input sentence.
Wherein, the probability sequences of each word in the input sentence obtained in step 211 are combined into a vector expression form to form a distributed expression of each word in the input sentence. For example, the distributed expression of the word "Liu" in the input sentence "movie I want to see Liu De Hua" is obtained as:
v (Liu) ═ 0.25,0.2,0.13,0.88,0.87,0.11,0.76,0.67)
The distributed expression of the "electric" word is:
v (electric) ═ 0.53,0.18,0.24,0.43,0.39,0.33,0.43,0.8)
Compared with the prior art that a One-Hot expression mode is adopted for words in a sentence, the distributed expression adopted by the invention can be associated with peripheral words in the sentence, so that the vector expressions of semantic similar words are consistent, and the accuracy of semantic understanding of the input sentence is improved.
for ease of understanding, assume that the One-Hot expression of "Liu" in the input sentence "I want to see Liu De Hua movie" is: when v (liu) ═ 0,0,1,0,0,0,0,0, the One-Hot expression of "electric" is: and v (electricity) ═ 0,0,0,1,0,0, and the cosine similarity of the distributed expression of "Liu" and "electricity" is 0.8763 and the cosine similarity of the One-Hot expression of "Liu" and "electricity" is 0 through calculation of a cosine text similarity calculation formula. Therefore, from the semantic similarity, the One-Hot expression is only the position of the identifier appearing in the whole background corpus, and semantic association with the context in the sentence cannot be obtained.
In step 230, the hidden vector corresponding to the word is obtained through the forward and backward traversal of the distributed expression corresponding to the word in the input sentence.
Wherein, the forward and backward traversal of the distributed expression corresponding to the words in the input sentence is performed through the circular neural network model.
Specifically, the distributed expressions corresponding to each word in the input sentence are respectively used as the input of a cyclic neural network, the cyclic neural network is used for respectively traversing the input sentence from front to back and from back to front, the specific traversing process is shown in fig. 4, X1 to X6 respectively represent the distributed expressions corresponding to each word in the input sentence, and after the input sentence is traversed from front to back, the hidden vectors corresponding to each word in the input sentence are respectively Y1 to Y6.
The latent vector of a word in an input sentence expresses the semantic association of the word with all words in the input sentence. Forward traversal establishes semantic association between each word and all words before the word in the process of traversing each word of an input statement; backward traversal in the process of traversing each word, semantic associations are established between each word and all words following it. Through the forward and backward traversal of the distributed expression corresponding to each word in the input statement, the semantic expression of each word combined with the whole situation of the input statement can be established.
In an exemplary embodiment, the forward and backward traversal of the distributed expressions corresponding to the words in the input sentence is performed by a Bi-directional long-short-time memory model (Bi-LSTM) to obtain the implicit vectors corresponding to each word in the input sentence.
Fig. 5 is a flowchart illustrating an implementation step 230 in an exemplary embodiment. As shown in fig. 5, the implementing step 230 may include the steps of:
In step 231, forward and backward traversal is performed according to the distributed expression corresponding to the word in the input sentence, so as to obtain a forward hidden vector and a backward hidden vector of the word.
The method comprises the steps of performing forward and backward traversal on distributed expressions corresponding to words in an input statement by adopting a bidirectional long and short time memory model (Bi-LSTM), and obtaining corresponding hidden vectors of the words, wherein the hidden vectors specifically comprise forward hidden vectors and backward hidden vectors.
specifically, the bidirectional long-short time memory model (Bi-LSTM) is composed of two independent long-short time memory models (LSTM), wherein the first long-short time memory model (LSTM) is used for performing forward traversal on the distributed expression of each word in the input sentence, and the other long-short time memory model (LSTM) is used for performing backward traversal on the distributed expression of each word in the input sentence.
the long and short term memory model can output a vector sequence according to the input vector sequence by considering the context relation. And traversing the input statement from front to back by one long and short time memory model according to the distributed expression of each word in the input statement to obtain a forward implicit vector corresponding to each word in the input statement. And traversing the input statement from back to front by the other long and short time memory model according to the distributed expression of each word in the input statement to obtain a backward implicit vector corresponding to each word in the input statement.
Since the back-and-forth traversal process of the long-time memory model is well known to those skilled in the art, it is not described herein.
In step 233, the forward hidden vectors and the backward hidden vectors of the words in the input sentence are concatenated to obtain the hidden vectors corresponding to the words.
The hidden vector of each word in the input sentence is formed by splicing a forward hidden vector and a backward hidden vector corresponding to each word. The essence of splicing the forward hidden vector and the backward hidden vector is to splice the two vectors, and the obtained spliced vector is the hidden vector of the word in the input statement.
In step 250, the attribute labels of the words and the intention labels of the input sentences are respectively predicted according to the hidden vectors corresponding to the words in the input sentences, so as to obtain the attribute labels corresponding to the words and the intention labels corresponding to the input sentences.
the attribute label prediction process of the hidden vector corresponding to each word in the input statement and the intention label prediction process of the hidden vector corresponding to the input statement are carried out synchronously.
Fig. 6 and 7 each represent a flow chart for implementing step 250 in an exemplary embodiment. Fig. 6 is a flowchart illustrating attribute tag prediction on a hidden vector corresponding to each word in an input sentence according to an exemplary embodiment, and fig. 7 is a flowchart illustrating intent tag prediction on a hidden vector corresponding to an input sentence according to an exemplary embodiment.
As shown in fig. 6, the method for performing attribute tag prediction on hidden vectors corresponding to each word in an input sentence includes the following steps:
In step 251, the hidden vector corresponding to the word in the input sentence is nonlinearly mapped by the target parameter matrix to obtain the weighted vector of the word.
and carrying out nonlinear mapping on the hidden vector corresponding to each word in the input statement through a target parameter matrix so as to map the distributed features corresponding to each word in the input statement to a sample mark space, thereby respectively establishing association between each word in the input statement and the sample mark space. The target parameter matrix represents the mapping relation between the hidden vector of the word in the input statement and the prediction probability of different attribute labels, and the target parameter matrix directly influences the accuracy of the attribute label prediction.
The essence of mapping the distributed features corresponding to the words in the input sentence to the sample mark space is that the hidden vectors corresponding to the words in the input sentence and the target parameter matrix are weighted and operated to obtain the weighted vectors corresponding to the words in the input sentence.
Can be expressed as z by formulaj＝WxjWherein "W" represents the target parameter matrix, "xj"represents a hidden vector corresponding to one word of an input sentence," zj"then correspondingly represents the weighting vector corresponding to the word.
In step 253, probability normalization is performed on the weighted vectors of the words in the input sentence to obtain attribute tag probability distribution vectors corresponding to the words.
In an exemplary embodiment, probability normalization of the weighted vectors of words in the input sentence is performed using a Softmax multi-class prediction function. The definition of the Softmax multiclass prediction function is as follows, where zjRepresents the weight vector corresponding to the word in the input sentence:
And mapping the weighted vector of the words in the input statement to a probability vector formed by the probability value sequence combination between (0,1) through a Softmax multi-classification prediction function so as to obtain attribute label probability distribution vectors corresponding to each word in the input statement relative to the peripheral words in the input statement.
In step 255, the attribute label corresponding to the maximum probability is selected from the attribute label probability distribution vectors corresponding to the words in the input sentence as the attribute label corresponding to the word.
Because the attribute label probability distribution vector obtained by the Softmax multi-class prediction function is obtained by associating the word in the input statement with the peripheral word in the input statement, the attribute label corresponding to the maximum probability is closest to the real attribute label of the word in the input statement.
therefore, the attribute label corresponding to the maximum probability in the attribute label probability distribution vector is selected as the attribute label corresponding to the word in the input statement, and the accuracy of the attribute label is ensured to the greatest extent.
As shown in fig. 7, the method for performing the intention tag prediction of an input sentence according to the hidden vector corresponding to the word in the input sentence includes the following steps:
in step 252, the hidden vectors corresponding to all words in the input sentence are spliced to obtain the hidden vector corresponding to the input sentence.
The hidden vectors corresponding to the input statement are obtained by splicing the hidden vectors corresponding to all the words in the input statement according to the sequence of the words in the input statement, and the hidden vectors corresponding to the input statement are spliced vectors.
Because the hidden vector corresponding to each word in the input sentence fully considers the semantic information of the context in the input sentence, the spliced hidden vector corresponding to the input sentence also fully considers the semantic of each word in the input sentence, so that the obtained intention label is more accurate.
In step 254, the hidden vector corresponding to the input sentence is subjected to nonlinear mapping by the target parameter matrix to obtain a weighted vector of the input sentence.
Similar to the method of performing the intention label prediction on the hidden vector corresponding to the input sentence, the essence of performing the nonlinear mapping on the hidden vector corresponding to the input sentence through the same target parameter matrix is to perform the weighted sum operation on the hidden vector corresponding to the input sentence and the target parameter matrix to obtain the weighted vector of the input sentence. The target parameter matrix here represents the mapping relationship between the hidden vector of the input sentence and the prediction probabilities of different intention labels.
In step 256, probability normalization is performed on the weighted vector of the input sentence to obtain an intention label probability distribution vector corresponding to the input sentence.
And calculating the weighted vector corresponding to the input statement by using the same Softmax multi-class prediction function for performing the intention label prediction on the hidden vector corresponding to the input statement to obtain the intention label probability distribution vector corresponding to the attribute labels of all the words in the input statement.
In step 258, the intention label corresponding to the maximum probability is selected from the intention label probability distribution vector corresponding to the input sentence as the intention label corresponding to the input sentence.
because the probability distribution vector of the intention label corresponding to the input statement obtained by the Softmax multi-class prediction function is predicted by the hidden vectors of all the words in the input statement, the intention label corresponding to the maximum probability is closest to the real intention label of the input statement. Therefore, the accuracy of the intention label corresponding to the input statement obtained by the method is guaranteed to the greatest extent.
In step 270, the input sentence is searched according to the attribute tag corresponding to the word in the input sentence and the intention tag corresponding to the input sentence.
After the attribute tag corresponding to each word in the input sentence and the intention tag corresponding to the input sentence are obtained through the steps, the input sentence is searched according to the obtained tags, and a result of recognizing the input sentence is obtained.
Still taking the example of applying the sentence recognition method to the intelligent television shown in fig. 1, after the sentence recognition method described in the above embodiment is executed to obtain the attribute tag corresponding to each word in the input sentence and the intention tag corresponding to the input sentence, the intelligent television learns the user intention according to all the obtained attribute tags and intention tags and executes corresponding operations. For example, by performing label prediction on the input sentence "i want to watch the movie in liu de hua" through the sentence identification method described in the above embodiment, the label information shown in fig. 8 can be obtained, and the smart television can search the movie information in the movie library of liu de hua according to the information shown in fig. 8, and output the search result to the display screen for display.
The intention labels obtained by the sentence recognition method in the above embodiment of the present invention fully consider semantic information of all words in the input sentence, and the obtained attribute labels of each word also fully consider context information in the input sentence where each word is located, so the method provided by the present invention has the following advantages:
in a first aspect, the sentence recognition method provided by the present invention does not rely on fixed sentence templates in the input sentences. Specifically, the method provided by the invention can carry out context processing on the input sentences with irregular sentence expression so as to obtain accurate attribute labels and intention labels.
Still taking the example of applying the sentence recognition method provided by the present invention to the intelligent television shown in fig. 1, if the speech information input by the intelligent television recognition user is "love of movie liudelwa", since the hidden vector of each word in the input sentence obtained by the sentence recognition method according to the above embodiment fully considers the context information in the input sentence, the attribute tags of the movie and television types corresponding to "love" and "love" can be accurately obtained.
The result obtained by sentence recognition of the input sentence "love of movie Liudebua" by the method provided by the invention can be shown in fig. 9, and the intelligent television searches movies of love types played by Liudebua in the video library thereof according to the information shown in fig. 9 and outputs the searched movie information to the display screen.
in a second aspect, the sentence recognition method provided by the invention can solve the problem of noise existing in the input sentence to a certain extent.
For example, the smart television recognizes that the voice information input by the user is "i want to watch the movie of liud flower", where "liud flower" is a recognition error, but the method provided by the present invention can accurately obtain the attribute tag of "flower" according to the context information of the input sentence, as shown in fig. 10.
In a third aspect, the sentence identification method provided by the invention can solve the problem that the intention identification is influenced by the interference tag existing in the input sentence to a certain extent.
For example, the smart television recognizes that the voice information input by the user is "animation monkey king", where the attribute label of the "monkey king" may be a song name or a movie name, but the method provided by the present invention may accurately predict the attribute label as a movie name according to the context information, as shown in fig. 11.
The sentence recognition method provided by the invention can be used as an off-line training stage and an on-line prediction stage respectively in the process of being applied to equipment. The purpose of the offline training stage is to optimize the target parameter matrix used in the process of predicting the label of the input sentence in step 250, so as to obtain the optimal target parameter matrix, and the accuracy of the label obtained by predicting the optimal target parameter matrix is optimal. And in the online prediction stage, label prediction is carried out on the text information input by the user in real time by directly using the optimal target parameter matrix obtained in the offline training stage, so as to directly obtain the attribute label and the intention label with the highest accuracy.
FIG. 12 is a flow chart illustrating a sentence recognition method in accordance with another exemplary embodiment, the method illustrated in FIG. 12 being adapted for use in an offline training phase. As shown in fig. 12, after the method for sentence recognition obtains the attribute tag and the intention tag by using the method described in the above embodiment, the method further includes the following steps:
In step 310, the attribute tag prediction deviation and the intention tag prediction deviation are summed to obtain a sentence recognition deviation corresponding to the input sentence.
The attribute tag prediction deviation is an error value between an intention tag corresponding to a word in the input sentence and a real attribute tag, and the intention tag prediction deviation is an error value between an intention tag corresponding to the input sentence and a real intention tag.
In one embodiment, the sentence recognition bias value corresponding to the input sentence is calculated according to a cross entropy loss function. The cross entropy loss function defined by the invention is:
Wherein i represents the number of intent types, y is the true intent tag, the value takes 0 or 1,Is the confidence score that predicts the intention label. m represents the number of words in the current sentence, n represents the number of attribute tag types, am,ja tag that represents a property of the real world,representing a confidence score in predicting the attribute tag.
the cross entropy loss function considers deviation values generated by comparing the attribute label and the intention label predicted by the input statement with the real label, and the deviation value obtained by adopting the cross entropy loss function is more suitable for judging the accuracy of label prediction.
If the deviation value obtained by calculation through the cross entropy loss function cannot meet the preset threshold condition, it indicates that the target parameter matrix adopted in the current label prediction process is not optimal, and therefore an optimal target parameter matrix needs to be searched.
In step 330, the sentence recognition bias is minimized, and an intermediate parameter matrix is obtained.
In an exemplary embodiment, the sentence recognition bias value is minimized using a gradient descent method. The specific treatment process comprises the following steps:
Firstly, derivation is carried out on the cross entropy loss function based on the currently adopted target parameter matrix, and a partial derivative of the cross entropy loss function to the adopted target parameter matrix is obtained. The resulting partial derivatives are also referred to as gradient values.
And then, carrying out subtraction operation on the currently adopted target parameter matrix and the partial derivative, wherein the obtained result is an intermediate parameter matrix.
In step 350, the intermediate parameter matrix is updated to the target parameter matrix.
After the intermediate parameter matrix is obtained, the intermediate parameter matrix is updated to the target parameter matrix used in the process of performing label prediction on the input statement in step 250.
In the off-line training stage, if the obtained deviation value cannot meet the preset threshold condition, the input sentence needs to be trained for the second time. In the second training, after the updated target parameter matrix is used to perform label prediction on the input sentence in step 250 to obtain the updated attribute label and intention label of the target parameter, the deviation value calculation is performed again in step 310. If the obtained deviation value still does not meet the preset threshold condition, the method in the embodiment is repeated to train the input sentence for the next time until the obtained deviation value meets the preset condition, the target parameter matrix used at this time is the optimal target parameter matrix, and the off-line training stage is completed at this time.
Fig. 13 is a block diagram illustrating a sentence recognition apparatus according to an example embodiment. The sentence recognition apparatus executes all or part of the steps of the sentence recognition method shown in any one of fig. 2, as shown in fig. 13, and includes, but is not limited to, a low-dimensional vector mapping module 410, a forward and backward traversing module 430, a sentence tag predicting module 450, and a sentence searching module 470.
The low-dimensional vector mapping module 410 is configured to map the words in the input sentence into low-dimensional vectors according to contexts corresponding to the words in the input sentence, so as to obtain a distributed expression of the words in the input sentence.
The forward and backward traversal module 430 is configured to perform forward and backward traversal according to the distributed expression of the words in the input sentence, so as to obtain a hidden vector of the words in the input sentence.
the sentence tag prediction module 450 is configured to perform attribute tag prediction on a word in an input sentence and intent tag prediction on the input sentence according to a hidden vector corresponding to the word in the input sentence, so as to obtain an attribute tag corresponding to the word in the input sentence and an intent tag corresponding to the input sentence;
The sentence searching module 470 is used for searching the input sentence according to the attribute tag corresponding to the word in the input sentence and the intention tag corresponding to the input sentence.
Further, the low-dimensional vector mapping module 410 includes a probability operation unit and a vector mapping unit. The probability operation unit is used for respectively operating the probability of occurrence in the stored text corpus information for each word in the input sentence relative to all the peripheral words in the input sentence, and obtaining the probability sequence formed by each word in the input sentence relative to all the peripheral words in the input sentence. The vector mapping unit is used for mapping the probability sequence of each word in the input statement into a low-dimensional vector to obtain the distributed expression corresponding to the word in the input statement.
Further, the backward-forward traversal module 430 includes a backward-forward traversal unit and a hidden vector stitching unit. The forward and backward traversal unit is used for respectively performing forward traversal and backward traversal according to the distributed expression corresponding to the words in the input statement to obtain forward hidden vectors and backward hidden vectors of the words in the input statement. The hidden vector splicing unit is used for splicing a forward hidden vector and a backward hidden vector of a word in an input sentence to obtain a hidden vector corresponding to the word in the input sentence.
further, the sentence tag prediction module 450 includes a nonlinear mapping unit, a probability normalization unit, and a tag obtaining unit. The nonlinear mapping unit is used for carrying out nonlinear mapping on the hidden vector corresponding to the word in the input statement and the hidden vector corresponding to the input statement through the target parameter matrix to obtain a weighted vector of the word in the input statement and a weighted vector of the input statement. The probability normalization unit is used for carrying out probability normalization on the weighted vector of the word in the input statement and the weighted vector of the input statement to obtain an attribute label probability distribution vector of the word in the input statement and an intention label probability distribution vector of the input statement. The label obtaining unit is used for obtaining an attribute label corresponding to the maximum probability from the attribute label probability distribution vector of the words in the input statement and obtaining an intention label corresponding to the maximum probability from the intention label probability distribution vector of the input statement.
Fig. 14 is a block diagram illustrating a sentence recognition apparatus according to another exemplary embodiment. As shown in fig. 14, the apparatus further includes a deviation value calculating module 510 and a parameter adjusting module 530.
The deviation value calculating module 510 is configured to sum the attribute tag prediction deviation and the intention tag prediction deviation to obtain a sentence recognition deviation corresponding to the input sentence, where the attribute tag prediction deviation is an error value between an intention tag corresponding to a word in the input sentence and a real attribute tag, and the intention tag prediction deviation is an error value between an intention tag corresponding to the input sentence and a real intention tag.
The bias value processing module 530 is configured to perform minimization on the sentence recognition bias to obtain an intermediate parameter matrix.
The parameter adjusting module 550 is configured to update the intermediate parameter matrix to the target parameter matrix.
In one exemplary embodiment, an apparatus comprises:
a processor; and
A memory, on which a computer readable program is stored, the computer readable program, when executed by the processor, implementing the sentence recognition method in the above embodiments.
In an exemplary embodiment, a computer-readable storage medium has a computer program stored thereon, and the computer program, when executed, implements the sentence recognition method in the above embodiments.
The above-mentioned embodiments are merely preferred examples of the present invention, and are not intended to limit the embodiments of the present invention, and those skilled in the art can easily make various changes and modifications according to the main concept and spirit of the present invention, so that the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. A sentence recognition method, comprising:
taking a word as a unit, and carrying out vectorization on the word in an input statement according to the corresponding context to obtain a distributed expression corresponding to the word;
Obtaining a hidden vector corresponding to the word through forward traversal and backward traversal of the distributed expression corresponding to the word in the input statement;
Respectively predicting the attribute labels of the words and the intention labels of the input sentences according to the hidden vectors corresponding to the words in the input sentences to obtain the attribute labels corresponding to the words and the intention labels corresponding to the input sentences;
and searching the input sentence according to the attribute tag corresponding to the word in the input sentence and the intention tag corresponding to the input sentence.
2. The method of claim 1, wherein the vectorizing of a word in an input sentence according to a corresponding context to obtain a distributed expression corresponding to the word in units of the word comprises:
calculating the probability of occurrence in the stored text corpus information for each word in the input sentence relative to all peripheral words in the input sentence, and obtaining a probability sequence formed by each word in the input sentence relative to all peripheral words in the input sentence;
And mapping the probability sequence of each word in the input statement into a low-dimensional vector to obtain the distributed expression corresponding to the word in the input statement.
3. The method of claim 1, wherein obtaining the hidden vector corresponding to the word through a forward traversal and a backward traversal of the distributed representation corresponding to the word in the input sentence comprises:
respectively performing forward traversal and backward traversal according to the distributed expression corresponding to the word in the input statement to obtain a forward implicit vector and a backward implicit vector of the word;
And splicing the forward hidden vector and the backward hidden vector of the word in the input statement to obtain the hidden vector corresponding to the word.
4. The method of claim 1, wherein the performing attribute tag prediction for a word in the input sentence according to a hidden vector corresponding to the word, and obtaining an attribute tag corresponding to the word comprises:
Carrying out nonlinear mapping on the hidden vector corresponding to the word in the input statement through a target parameter matrix to obtain a weighted vector of the word;
carrying out probability normalization on the weighted vectors of the words in the input statement to obtain attribute label probability distribution vectors corresponding to the words;
and selecting the attribute label corresponding to the maximum probability from the attribute label probability distribution vector corresponding to the word in the input statement as the attribute label corresponding to the word.
5. the method of claim 1, wherein the performing the prediction of the intention label corresponding to the input sentence according to the hidden vector corresponding to the word in the input sentence, and obtaining the intention label of the input sentence comprises:
Splicing the hidden vectors corresponding to all the words in the input statement to obtain the hidden vectors corresponding to the input statement;
Carrying out nonlinear mapping on the hidden vector corresponding to the input statement through a target parameter matrix to obtain a weighted vector of the input statement;
carrying out probability normalization on the weighted vector of the input statement to obtain an intention label probability distribution vector corresponding to the input statement;
And selecting the intention label corresponding to the maximum probability from the intention label probability distribution vector corresponding to the input statement as the intention label corresponding to the input statement.
6. The method of claim 4 or 5, further comprising:
Summing the attribute tag prediction deviation and the intention tag prediction deviation to obtain a sentence identification deviation corresponding to the input sentence, wherein the attribute tag prediction deviation is an error value between an intention tag corresponding to a word in the input sentence and a real attribute tag, and the intention tag prediction deviation is an error value between an intention tag corresponding to the input sentence and a real intention tag;
Minimizing the sentence recognition deviation to obtain an intermediate parameter matrix;
And updating the intermediate parameter matrix into the target parameter matrix.
7. a sentence recognition apparatus, the apparatus comprising:
The low-dimensional vector mapping module is used for mapping the words in the input sentences into low-dimensional vectors according to the corresponding contexts to obtain distributed expressions corresponding to the words;
a forward and backward traversal module for performing forward and backward traversal according to the distributed expression of the words in the input sentence to obtain the hidden vectors of the words;
a statement label prediction module, configured to perform attribute label prediction on a word and intent label prediction on the input statement according to a hidden vector corresponding to the word in the input statement, so as to obtain an attribute label corresponding to the word and an intent label corresponding to the input statement;
And the sentence searching module is used for searching the input sentence according to the attribute tag corresponding to the word in the input sentence and the intention tag corresponding to the input sentence.
8. The apparatus of claim 7, wherein the apparatus further comprises:
A deviation calculation module, configured to perform summation operation on an attribute tag prediction deviation and an intention tag prediction deviation to obtain a sentence recognition deviation of the input sentence, where the attribute tag prediction deviation is an error value between an intention tag corresponding to a word in the input sentence and a real attribute tag, and the intention tag prediction deviation is an error value between an intention tag corresponding to the input sentence and a real intention tag;
The deviation value processing module is used for carrying out minimization processing on the sentence recognition deviation to obtain a middle parameter matrix;
And the parameter adjusting module is used for updating the intermediate parameter matrix into a target parameter matrix.
9. A machine device, comprising:
a processor;
Memory storing a computer program that, when executed by the processor, implements the method of any of claims 1 to 6.
10. a computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the method of any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810558865.8A CN110555207A (en) | 2018-06-01 | 2018-06-01 | Sentence recognition method, sentence recognition device, machine equipment and computer-readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810558865.8A CN110555207A (en) | 2018-06-01 | 2018-06-01 | Sentence recognition method, sentence recognition device, machine equipment and computer-readable storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110555207A true CN110555207A (en) | 2019-12-10 |
Family
ID=68734855
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810558865.8A Pending CN110555207A (en) | 2018-06-01 | 2018-06-01 | Sentence recognition method, sentence recognition device, machine equipment and computer-readable storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110555207A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111507108A (en) * | 2020-04-17 | 2020-08-07 | 腾讯科技（深圳）有限公司 | Alias generation method and device, electronic equipment and computer readable storage medium |
| EP4177784A4 (en) * | 2020-08-10 | 2023-11-15 | Beijing Bytedance Network Technology Co., Ltd. | Method and apparatus for generating prediction information, and electronic device and medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106227774A (en) * | 2016-07-15 | 2016-12-14 | 海信集团有限公司 | Information search method and device |
| CN106569998A (en) * | 2016-10-27 | 2017-04-19 | 浙江大学 | Text named entity recognition method based on Bi-LSTM, CNN and CRF |
| CN107526798A (en) * | 2017-08-18 | 2017-12-29 | 武汉红茶数据技术有限公司 | A kind of Entity recognition based on neutral net and standardization integrated processes and model |
| CN108062538A (en) * | 2017-12-29 | 2018-05-22 | 成都智宝大数据科技有限公司 | Face identification method and device |
-
2018
- 2018-06-01 CN CN201810558865.8A patent/CN110555207A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106227774A (en) * | 2016-07-15 | 2016-12-14 | 海信集团有限公司 | Information search method and device |
| CN106569998A (en) * | 2016-10-27 | 2017-04-19 | 浙江大学 | Text named entity recognition method based on Bi-LSTM, CNN and CRF |
| CN107526798A (en) * | 2017-08-18 | 2017-12-29 | 武汉红茶数据技术有限公司 | A kind of Entity recognition based on neutral net and standardization integrated processes and model |
| CN108062538A (en) * | 2017-12-29 | 2018-05-22 | 成都智宝大数据科技有限公司 | Face identification method and device |
Non-Patent Citations (2)
| Title |
|---|
| XIAODONG ZHANG 等: "A Joint Model of Intent Determination and Slot Filling for Spoken Language Understanding", 《PROCEEDINGS OF THE TWENTY-FIFTH INTERNATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE 》, 9 July 2016 (2016-07-09), pages 3 * |
| 韦鹏程 等: "《当代机器深度学习方法与应用研究》", 电子科技大学出版社, pages: 139 - 140 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111507108A (en) * | 2020-04-17 | 2020-08-07 | 腾讯科技（深圳）有限公司 | Alias generation method and device, electronic equipment and computer readable storage medium |
| CN111507108B (en) * | 2020-04-17 | 2021-03-19 | 腾讯科技（深圳）有限公司 | Alias generation method and device, electronic equipment and computer readable storage medium |
| EP4177784A4 (en) * | 2020-08-10 | 2023-11-15 | Beijing Bytedance Network Technology Co., Ltd. | Method and apparatus for generating prediction information, and electronic device and medium |
| US11842163B2 (en) | 2020-08-10 | 2023-12-12 | Beijing Bytedance Network Technology Co., Ltd. | Method and apparatus for generating prediction information, and electronic device and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109543030B (en) | Method, device, equipment and storage medium for classifying session texts of customer service robot | |
| US20180341698A1 (en) | Method and apparatus for parsing query based on artificial intelligence, and storage medium | |
| CN112685565A (en) | Text classification method based on multi-mode information fusion and related equipment thereof | |
| CN111368548A (en) | Semantic recognition method and device, electronic equipment and computer-readable storage medium | |
| CN107491534A (en) | Information processing method and device | |
| CN110162766B (en) | Word vector updating method and device | |
| CN114245203B (en) | Video editing method, device, equipment and medium based on script | |
| CN113421547B (en) | Voice processing method and related equipment | |
| US11036996B2 (en) | Method and apparatus for determining (raw) video materials for news | |
| EP4109324A2 (en) | Method and apparatus for identifying noise samples, electronic device, and storage medium | |
| CN110114765B (en) | Electronic device performing translation by sharing context of utterance and operating method thereof | |
| CN111079418A (en) | Named body recognition method and device, electronic equipment and storage medium | |
| WO2022187063A1 (en) | Cross-modal processing for vision and language | |
| CN111582360A (en) | Method, apparatus, device and medium for labeling data | |
| US20200387676A1 (en) | Electronic device for performing translation by sharing context of utterance and operation method therefor | |
| CN116050496A (en) | Determination method and device, medium and equipment of picture description information generation model | |
| CN114330474A (en) | Data processing method and device, computer equipment and storage medium | |
| CN110827799B (en) | Method, apparatus, device and medium for processing voice signal | |
| JP2022537860A (en) | Voice packet recommendation method, device, electronic device and program | |
| CN110555207A (en) | Sentence recognition method, sentence recognition device, machine equipment and computer-readable storage medium | |
| CN117765132A (en) | Image generation method, device, equipment and storage medium | |
| CN117453949A (en) | Video positioning method and device | |
| CN110851629A (en) | Image retrieval method | |
| CN111222011B (en) | Video vector determining method and device | |
| CN112287159A (en) | Retrieval method, electronic device and computer readable medium |
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 |