CN113591462A - Bullet screen reply generation method and device and electronic equipment - Google Patents
Bullet screen reply generation method and device and electronic equipment Download PDFInfo
- Publication number
- CN113591462A CN113591462A CN202110855573.2A CN202110855573A CN113591462A CN 113591462 A CN113591462 A CN 113591462A CN 202110855573 A CN202110855573 A CN 202110855573A CN 113591462 A CN113591462 A CN 113591462A
- Authority
- CN
- China
- Prior art keywords
- bullet screen
- vector
- screen information
- word
- theme
- 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
- 238000000034 method Methods 0.000 title claims abstract description 51
- 239000013598 vector Substances 0.000 claims abstract description 269
- 238000004364 calculation method Methods 0.000 claims abstract description 19
- 238000007781 pre-processing Methods 0.000 claims abstract description 7
- 238000013528 artificial neural network Methods 0.000 claims description 57
- 230000006870 function Effects 0.000 claims description 31
- 238000012549 training Methods 0.000 claims description 31
- 230000001419 dependent effect Effects 0.000 claims description 21
- 238000004891 communication Methods 0.000 claims description 16
- 230000007246 mechanism Effects 0.000 claims description 11
- 238000010606 normalization Methods 0.000 claims description 5
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 238000013473 artificial intelligence Methods 0.000 abstract description 2
- 238000003062 neural network model Methods 0.000 abstract 1
- 238000011084 recovery Methods 0.000 abstract 1
- 238000004590 computer program Methods 0.000 description 6
- 238000001514 detection method Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 6
- 230000008569 process Effects 0.000 description 5
- 238000013459 approach Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 150000001875 compounds Chemical class 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 239000003607 modifier Substances 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 239000002699 waste material Substances 0.000 description 2
- 241000039077 Copula Species 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000007654 immersion Methods 0.000 description 1
- 230000009191 jumping Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
- 230000017105 transposition Effects 0.000 description 1
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
- H04N21/47—End-user applications
- H04N21/488—Data services, e.g. news ticker
- H04N21/4884—Data services, e.g. news ticker for displaying subtitles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/211—Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/258—Heading extraction; Automatic titling; Numbering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- General Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Signal Processing (AREA)
- Multimedia (AREA)
- Machine Translation (AREA)
Abstract
The embodiment of the invention relates to the technical field of artificial intelligence and discloses a bullet screen reply generation method and device and electronic equipment. The method comprises the following steps: acquiring the bullet screen information; preprocessing the bullet screen information to convert the bullet screen information into an initial text vector; determining a subject matter semantic relation distribution of the initial text vector; determining a syntax dependence vector and a position relation vector between words in the bullet screen information according to the initial text vector; performing weighted calculation on the syntactic dependency relationship vector and the position relationship vector to obtain an input vector of each word; and inputting the input vector and the subject semantic relation distribution into a target neural network model for prediction so as to generate a bullet screen information context for replying the bullet screen information context. Through the mode, the bullet screen recovery quality is improved.
Description
Technical Field
The embodiment of the invention relates to the technical field of artificial intelligence, in particular to a bullet screen reply generation method and device and electronic equipment.
Background
A bullet screen is a segment of text that is displayed in synchronization with a video for user interaction with the video or between users. The user can better participate in the video communication and interaction by issuing the barrage related to the video situation.
According to the bullet screen information issued by the user, the bullet screen reply information is generated, so that the participation and immersion of the user can be effectively improved. In the related art, the generation of the bullet screen reply information is mainly based on an extraction method and an understanding generation method. However, the inventors found in the course of implementing embodiments of the present invention that: the quality of the bullet screen reply information generated in the related technology is poor, for example, the interactivity of the bullet screen reply information generated by adopting an extraction method is poor, and the bullet screen reply information generated by adopting an understanding generation method has defects in semantic logic.
Disclosure of Invention
In view of the foregoing problems, embodiments of the present invention provide a bullet screen reply generation method, device and electronic device, so as to solve the problem in the prior art that the bullet screen reply information quality is poor.
According to an aspect of the embodiments of the present invention, there is provided a bullet screen reply generation method, including:
acquiring the bullet screen information;
preprocessing the bullet screen information to convert the bullet screen information into an initial text vector;
determining a subject matter semantic relation distribution of the initial text vector;
determining a syntax dependence vector and a position relation vector between words in the bullet screen information according to the initial text vector;
performing weighted calculation on the syntactic dependency relationship vector and the position relationship vector to obtain an input vector of each word;
and inputting the input vector and the subject semantic relation distribution into a target neural network for prediction so as to generate a bullet screen information context for replying the bullet screen information context.
In an optional manner, the determining, according to the initial text vector, a syntactic dependency vector and a position relationship vector between words in the bullet screen information includes:
determining the position of a current word, wherein the current word is any word vector in the initial text vectors;
traversing the position of the dependent word corresponding to the current word, and ending the traversal when traversing to the position of the core word, wherein the dependent word is the dependent word of the current word;
calculating the absolute value of the difference value between the position of the current word and the position of the dependent word;
and accumulating the absolute value of the difference value to calculate the core weight of the current word to obtain the syntactic dependency vector.
In an alternative approach, the target neural network includes an attention mechanism, a feed-forward network, and a subject word prediction model; the step of inputting the input vector and the subject semantic relation distribution into a target neural network for prediction to generate a bullet screen information context for replying the bullet screen information context comprises:
determining the theme of the initial text vector and the corresponding word distribution under the theme according to the theme semantic relationship distribution;
according to the theme and the corresponding word distribution under the theme, performing attention mechanism calculation on the input vector to obtain an interword relation vector;
after residual connection and normalization are carried out on the inter-word relation vectors, the inter-word relation vectors are input into the feedforward network, and theme semantic vectors of all words in the input vectors are obtained;
residual error connection is carried out on the theme semantic vector and the inter-word relation vector to obtain a bullet screen information sequence vector;
and inputting the bullet screen information sequence vector into the subject word prediction model to generate a bullet screen information context for replying the bullet screen information context.
In an optional manner, the bullet screen information sequence vector includes a theme above the bullet screen information, word distribution under the theme, theme probability and semantic information; the inputting the bullet screen information sequence vector into the subject term prediction model to generate a bullet screen information context for replying the bullet screen information context includes:
determining a last prediction result, wherein the last prediction result is a predicted word predicted before a current predicted word;
taking the last prediction result and the bullet screen information sequence vector as input, and obtaining the probability of the current predicted word according to a preset vocabulary model and the distribution of the subject semantic relation;
weighting and calculating the theme probability and the current predicted word probability to obtain corresponding predicted words under each theme;
and generating a bullet screen information context for replying the bullet screen information context according to the corresponding prediction words under the topics.
In an optional manner, the method further comprises:
calculating a quality value of the bullet screen information according to a preset quality value formula, wherein the quality value is used for representing the logic collocation quality, the content diversity quality and the theme quality of the bullet screen information;
and if the quality value of the bullet screen information context is larger than a preset quality value threshold, outputting the bullet screen information context.
In an alternative manner, calculating the mass value of the bullet screen information according to a preset mass value formula includes:
respectively determining a logic collocation quality value, a content diversity quality value and a theme quality value under the bullet screen information;
calculating a product of the logical collocation quality value and the content diversity quality value, and determining a sum of the product and the topic quality value as the quality value.
In an optional manner, the method further comprises:
inputting a training sample into the target neural network for training to obtain an output result; the training sample is a bullet screen dialogue text with a label; the label comprises a theme of the bullet screen dialogue text and a corresponding bullet screen dialogue text;
calculating a loss value of the target neural network according to the output result, the bullet screen dialog context and a preset loss function;
adjusting parameters of the target neural network according to the loss value, re-executing the step of inputting a training sample into the target neural network for training to obtain an output result, calculating the loss value of the target neural network according to the output result, the bullet screen dialog context and a preset loss function, and adjusting the parameters of the target neural network according to the loss value until the preset loss function is converged or is smaller than a preset threshold value;
the preset loss function is composed of a word prediction likelihood function and a cosine similarity distance of a context semantic vector.
According to another aspect of the embodiments of the present invention, there is provided a bullet screen reply generation apparatus, including:
the acquisition module is used for acquiring the bullet screen information;
the conversion module is used for preprocessing the bullet screen information and converting the bullet screen information into an initial text vector;
the first determining module is used for determining the subject semantic relation distribution of the initial text vector;
the second determining module is used for determining a syntactic dependency vector and a position relation vector between words in the bullet screen information according to the initial text vector;
the calculation module is used for carrying out weighting calculation on the syntactic dependency relationship vector and the position relationship vector to obtain an input vector of each word;
and the prediction module is used for inputting the input vector and the subject semantic relation distribution into a target neural network for prediction so as to generate a bullet screen information context for replying the bullet screen information context.
According to another aspect of the embodiments of the present invention, there is provided an electronic device including: the system comprises a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface complete mutual communication through the communication bus;
the memory is used for storing at least one executable instruction, and the executable instruction enables the processor to execute the operation of the bullet screen reply generation method.
According to another aspect of the embodiments of the present invention, a computer-readable storage medium is provided, where at least one executable instruction is stored in the storage medium, and when the executable instruction is executed on an electronic device, the electronic device executes the operations of the bullet screen reply generation method described above.
In the embodiment of the invention, after converting the bullet screen information into the initial text vector, the theme semantic relationship distribution of the initial text vector, and the syntax dependence vector and the position relationship vector between words in the bullet screen information can be determined; and determining an input vector of each word according to the syntactic dependency vector and the position relation vector, inputting the input vector and the subject semantic relation distribution into a target neural network for prediction, and generating a bullet screen information context for replying the bullet screen information context. The method and the device can determine the input vector of each word through the syntactic dependency vector and the position relation vector, and can more accurately mine semantic roles borne by the predicted words, so that the logic structure of the bullet screen information is more accurate; the theme characteristics above the bullet screen information can be dynamically captured by determining the theme semantic relationship distribution of the initial text vector, so that the theme characteristics below the bullet screen information are consistent with the theme characteristics above the bullet screen information, and the generation quality of the bullet screen information is further improved.
The foregoing description is only an overview of the technical solutions of the embodiments of the present invention, and the embodiments of the present invention can be implemented according to the content of the description in order to make the technical means of the embodiments of the present invention more clearly understood, and the detailed description of the present invention is provided below in order to make the foregoing and other objects, features, and advantages of the embodiments of the present invention more clearly understandable.
Drawings
The drawings are only for purposes of illustrating embodiments and are not to be construed as limiting the invention. Also, like reference numerals are used to refer to like parts throughout the drawings. In the drawings:
fig. 1 is a schematic structural diagram illustrating a content generation network provided by an embodiment of the present invention;
fig. 2 is a schematic flow chart illustrating a bullet screen reply generation method according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram illustrating a bullet screen reply generation apparatus according to an embodiment of the present invention;
fig. 4 shows a schematic structural diagram of an electronic device provided in an embodiment of the present invention.
Detailed Description
Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be embodied in various forms and should not be limited to the embodiments set forth herein.
Fig. 1 is a schematic structural diagram illustrating a content generation network according to an embodiment of the present invention. As shown in fig. 1, the content generation network is a bullet screen theme generation network constructed based on a Transformer deep learning model, and includes a semantic learning module, a theme learning module, and a bullet screen prediction module.
The semantic learning module is used for learning the main semantic logical relationship of the bullet screen based on attention mechanics by adding syntactic relationship learning codes to the bullet screen data, so that noise logical interference is reduced; in the theme learning module, by adding theme relation identification information into the bullet screen data, the theme semantic relation distribution of sentences and words is obtained through context information; and in the bullet screen prediction module, words under the subjects similar to the upper part of the bullet screen are obtained through the context to generate a conversation, so that the generated conversation is flexible and various.
Further, the embodiment of the present invention provides an evaluation feedback method, i.e., an STGE (Semantic Text generation) method, for the quality of automatically generated bullet screen content. The method evaluates the logic collocation quality, the content diversity quality and the theme quality of the bullet screen sentence, provides reference for the defects of the existing automatic evaluation method, and simultaneously provides feedback for the generation of high-quality conversation. The following describes the implementation process of the content generation network in detail with reference to specific embodiments.
Fig. 2 is a flowchart illustrating a bullet screen reply generation method according to an embodiment of the present invention, where the method is executed by an electronic device. The memory of the electronic device is used for storing at least one executable instruction, and the executable instruction enables the processor of the electronic device to execute the operation of the bullet screen reply generation method. As shown in fig. 1 and fig. 2, the method includes the following steps:
step 110: and acquiring the bullet screen information.
Wherein, the bullet screen information is generally edited and released by the user. The barrage information above may be, for example, "teacher happy! "," handset and my shape do not leave ", etc. Barrage information is generally associated with the current video above, e.g., the subject of the current video is "teacher's festival," the barrage information may be, e.g., "teacher's festival happy! When the theme of the current video is "mobile phone", the bullet screen information may be "mobile phone and my movie do not leave" above, for example. Since the theme is dynamically changed during the video playing process, the barrage information published by the user generally has real-time property, namely, is related to the current video theme.
Step 120: and preprocessing the bullet screen information and converting the bullet screen information into an initial text vector.
When the bullet screen information is preprocessed, the bullet screen information can be subjected to word segmentation, then word vectors of all words are calculated, and initial text vectors corresponding to the bullet screen information can be obtained according to the word vectors of all the words. For example, if the bullet screen information is "useless words are not much, three links are worship", and the word segmentation can be performed to obtain "useless words/not much/words/,/three links/is/worship/".
Step 130: and determining the subject semantic relation distribution of the initial text vector.
The theme semantic relation distribution of the bullet screen dialogue sequence can be trained in advance through n-layer coding learning. The theme semantic relationship distribution of the pre-training barrage conversation sequence comprises theme semantic relationship distribution corresponding to each word in the barrage conversation sequence, wherein one word may correspond to a plurality of themes, and each theme corresponds to a theme probability.
The pre-trained topic vector Z may be represented as { Z1, Z2, z3... zm }, the word vector w corresponding to the initial text vector may be represented as { d1, d2, d3... dn }, and the topic semantic relationship distribution of the initial text vector may be determined by the topic semantic relationship distribution corresponding to the word vector w in the initial text vector. Further, the distribution of the semantic relationship of the topics corresponding to the word vector w can be determined by constructing a relationship matrix of the word vector and the topic vector Z. In the initial text vector, the topic probability of the word corresponding to the word vector w can be determined by the following formula.
Wherein, p (z)k| w) represents the topic probability of the word corresponding to the word vector w, zkRepresenting the topic to which the word vector w belongs, m representing the total dimension of the topic vector Z, n representing the total dimension of the word vector w, p (Z) can be obtained by the above calculation formulak| w) is expressed as the semantic distance of the word vector from the corresponding topic vector.
Therefore, topic similarity detection is carried out on each word vector in the initial text vector and topic semantic relation distribution of the pre-training barrage dialogue sequence, so that topic semantic relation distribution corresponding to the initial text vector is obtained, and the topic semantic relation distribution comprises topics and topic probabilities corresponding to each word.
Step 140: and determining a syntactic dependency vector and a position relation vector among the words in the bullet screen information according to the initial text vector.
The syntax dependence vector is used for representing the semantic logic relationship of each word in the bullet screen information, and the position relationship vector is used for representing the position relationship of each word in the bullet screen information. The part of speech analysis can be carried out on the bullet screen information, and the semantic relation of each participle is analyzed. For example, "the word is not much spoken, the word" three links are worship "and after the part of speech analysis:
(ROOT (IP (NP (NP (NR WASTE))) (NP (NN NOT))) (VP (VV so (PU'))) (IP (VP (QP (CD Tri)))) (VP (VC so (NP (NN Jing)))))))))))))))))
Where ROOT denotes a sentence of a text to be processed, IP denotes a simple clause, NP denotes a noun phrase, NR denotes a proper noun, NN denotes a common noun, VP denotes a verb phrase, VV denotes a verb, PU denotes a sentence break such as a period, question mark, exclamation mark, QP denotes a quantifier phrase, and CD denotes a radix word. The result of analyzing the semantic relation of each participle is as follows:
nmod, assmod (no more than-2, waste word-1), nsubj (say-3, no more than-2), ROOT (ROOT-0, say-3), punct (say-3, -4), conj (Jing-7, san Lian-5), cop (Jing-7, is-6), ccomp (say-3, Jing-7).
Wherein nmod (non compound modifier) represents compound noun modification, assmod (associative modifier) represents associated modification, nsubj (non subject) represents noun subject, root represents root node, designation represents punctuation symbol, conj (join) represents connective, cop (copula) represents systematic verb, and ccomp represents subordinate clause complement.
The syntax dependency vector can be calculated by the following equation, for example.
In the above formula, p represents the position of the current word, pkdThe position of the dependent word is determined, r is the position of the core word, i (the value range is 0-d) represents the dimension of the current word, and d represents the total dimension. SE (p, i) represents a syntactic dependency vector of a word at the p-th position in the sequence of words { w1, w 2.. wn }. Further, when determining a syntactic dependency vector between words in the bullet screen information according to the initial text vector, the position of the current word can be determined first, and the current word can be any word vector in the initial text vector; traversing the position of a dependent word corresponding to the current word, and ending the traversal when the position of the core word is traversed, wherein the dependent word is the dependent word of the current word; calculating the absolute difference between the current word position and the dependent word positionA value; and accumulating the absolute values of the difference values to calculate the core weight of the current word so as to obtain the syntactic dependency vector corresponding to the current word. For example, in "do nothing, three links are worship", the core word is "say", the stem of the whole sentence surrounding the core word is "do nothing, worship", and for "do nothing", the pk position is 2, pkdPosition 3. By coding the syntactic relation through the formula, the core relation words such as the chief and the predicate guest can be focused on, so that the influence of other meaningless noises is reduced, the attention learning of the subject characteristics is focused, and the accurate identification of the logic relation among the words is improved.
Further, the positional relationship vector can be calculated by using the following equation, for example.
Where pos represents the location of the current word, dmodel represents the total dimension, and i represents the dimension of the location relationship vector.
Step 150: and performing weighted calculation on the syntactic dependency relationship vector and the position relationship vector to obtain an input vector of each word.
And performing weighted calculation on the syntactic dependency relationship vector and the position relationship vector to obtain an input vector of each word, so that the semantic relationship characteristic and the position relationship characteristic of each word can be represented by the input vector of each word. In the embodiment of the present invention, the syntactic dependency relationship vector and the position relationship vector corresponding to each word may be directly added or added after setting different weights respectively to obtain the input vector. The embodiment of the present invention does not specifically limit how to weight the specific form, and those skilled in the art can set the weight accordingly according to the specific scenario.
Step 160: and inputting the input vector and the subject semantic relation distribution into a target neural network for prediction so as to generate a bullet screen information context for replying the bullet screen information context.
Wherein, the target neural network is a neural network generated through training. The target neural network may be based on, for example, a Transformer model, which is a natural language processing model developed by ***.
In an alternative approach, the target neural network includes an attention mechanism, a feed-forward network, and a subject word prediction model. When the input vector and the distribution of the subject semantic relationship are input into a target neural network for prediction, the subject of an initial text vector and the distribution of words corresponding to the subject can be determined according to the distribution of the subject semantic relationship, attention mechanism calculation is carried out on the input vector according to the subject and the distribution of the words corresponding to the subject to obtain an inter-word relationship vector, the inter-word relationship vector is subjected to residual connection and normalization, and then the inter-word relationship vector is input into a feed-forward network to obtain the subject semantic vector of each word in the input vector; residual error connection is carried out on the theme semantic vector and the interword relation vector to obtain a bullet screen information sequence vector; and inputting the bullet screen information sequence vector into a subject word prediction model to generate a bullet screen information context for replying the bullet screen information context.
The forward propagation may be performed through a feedforward neural network, and the forward propagation process may use, for example, a GELU (Gaussian Error Linear Units) activation function. The GELU activation function may be calculated using the following equation, for example.
In an optional manner, the bullet screen information sequence vector includes a topic above the bullet screen information, a word distribution under the topic, a topic probability, and semantic information. When the bullet screen information sequence vector is input into the subject word prediction model, a last prediction result can be determined firstly, wherein the last prediction result is a predicted word predicted before a current predicted word; and taking the last prediction result and the bullet screen information sequence vector as input, and obtaining the probability of the current predicted word according to the preset vocabulary model and the theme semantic relationship distribution. The preset vocabulary model may be manually preset for encoding words, thereby establishing a unique correspondence between words and corresponding codes. Further, carrying out weighted calculation on the theme probability and the current predicted word probability to obtain corresponding predicted words under each theme; and obtaining a plurality of pieces of predicted bullet screen reply information and corresponding probabilities according to corresponding predicted words under each theme, and determining the predicted bullet screen reply information with the highest probability as the bullet screen information context.
Further, a word prediction function may be used to generate the corresponding predicted word under each topic, and the word prediction function may be calculated using the following formula, for example.
Wherein V represents a preset vocabulary model, K represents topic semantic relation distribution, V and K are subsets of V and K respectively, s represents a topic vector above bullet screen information, yi-1Representing the last prediction, w and b are the network parameters used to represent the coefficients and bias terms, respectively, p (y)i) As a function of word prediction, wTDenotes the transposition of W, WSV、Wsk、Wyv、WykAnd WxKRespectively for representing the components of w, x representing the word vector above the bullet screen information, bvAnd bkRepresenting the respective components of b, respectively.
It should be noted that, after the bullet screen information context for replying to the bullet screen information context is generated according to the above embodiment, the bullet screen information context may be further detected to determine the quality value of the bullet screen information context, and the quality value of the bullet screen information context is compared with the preset quality value threshold.
In an alternative manner, when determining the quality value of the bullet screen information context, the bullet screen information context is first divided into a plurality of logic detection units, for example, a punctuation may be added to the bullet screen information context according to a preset step length, and the logic detection units are divided according to the punctuation. And adopting a jumping mechanism to carry out logic collocation detection on different logic detection units, and determining a logic collocation quality value under the bullet screen information according to a logic collocation detection result. Furthermore, weighting can be carried out on the theme vector of each word in the bullet screen information to determine the theme vector in the bullet screen information, calculating the cosine similarity between the theme vector in the bullet screen information and the theme vector in the bullet screen information, and determining the cosine similarity as the theme quality value in the bullet screen information. Further, jump combination of different sliding windows can be performed on the word sequences under the bullet screen information, for example, word combination (w1, 1, w2) represents word combination w1w2 with distance of 1, and the logical dependency relationship of the word combination is calculated, for example, (w1, nsubj, 1, w2) represents the logical dependency relationship of word combination w1w2, and the logical relationship between words in the training word set C and the word collocation combination are combined to count the logical and specificity of the word combination between different distances in the bullet screen information, so that the logical collocation quality value and the content diversity quality value in the bullet screen information are calculated.
In an optional mode, the quality value of the bullet screen information context can be calculated according to a preset quality value formula, the quality value is used for representing the logic collocation quality, the content diversity quality and the theme quality of the bullet screen information context, and if the quality value of the bullet screen information context is larger than a preset quality value threshold, the bullet screen information context is output.
In an alternative mode, when the quality value of the bullet screen information context is calculated according to the preset quality value formula, the logic collocation quality value, the content diversity quality value and the theme quality value of the bullet screen information context can be respectively determined, the product of the logic collocation quality value and the content diversity quality value is calculated, and the sum of the product and the theme quality value is determined as the quality value of the bullet screen information context.
The preset quality value formula may be, for example, the following formula:
wherein the content of the first and second substances,
a theme quality value representing the context of the bullet screen information,
a topic vector representing the context of the bullet screen information,
the theme vector representing the bullet screen information is used for representing the theme quality value through the cosine similarity between the theme vector in the bullet screen information and the theme vector in the bullet screen information;
a logical collocation quality value is represented below the bullet screen information,
representing the content diversity quality value of the bullet screen information, r representing the set of reply words, C representing the set of training words, (w)m,i,wn) Denotes w at an interval im wnPhrase, count is used to count w without deduplicationm wnNumber of phrases, uniq for statistical deduplication wm wnThe number of phrases, m and n are self-defined parameters.
It should be noted that before inputting the input vector and the subject semantic relationship distribution into the target neural network for prediction, the target neural network needs to be trained in advance.
In an optional mode, inputting a training sample into a target neural network for training to obtain an output result, wherein the training sample is a bullet screen dialogue upper text with a label, and the label comprises a theme of the bullet screen dialogue upper text and a corresponding bullet screen dialogue lower text; calculating a loss value of the target neural network according to an output result of the target neural network, the bullet screen dialogue context and a preset loss function, adjusting parameters of the target neural network according to the loss value of the target neural network, training the target neural network again according to the training sample, and adjusting the parameters of the target neural network according to the loss value of the target neural network until the preset loss function is converged or is smaller than a preset threshold value.
In an alternative, the predetermined loss function may be formed by a cosine similarity distance between the word prediction likelihood function and the context semantic vector. For example, the preset loss function may be, for example:
wherein logP (x)i|xi-k,.....xi-1Theta) is a word prediction likelihood function,
the cosine similarity distance of the context semantic vector,
the above theme vector of the bullet screen is shown,
representing the bullet screen context topic vector. Alpha and beta are adjustable parameters, theta is a parameter to be estimated, and k and m are self-defined parameters.
In the embodiment of the invention, after converting the bullet screen information into the initial text vector, the theme semantic relationship distribution of the initial text vector, and the syntax dependence vector and the position relationship vector between words in the bullet screen information can be determined; and determining an input vector of each word according to the syntactic dependency vector and the position relation vector, inputting the input vector and the subject semantic relation distribution into a target neural network for prediction, and generating a bullet screen information context for replying the bullet screen information context. The method and the device can determine the input vector of each word through the syntactic dependency vector and the position relation vector, and can more accurately mine semantic roles borne by the predicted words, so that the logic structure of the bullet screen information is more accurate; the theme characteristics above the bullet screen information can be dynamically captured by determining the theme semantic relationship distribution of the initial text vector, so that the theme characteristics below the bullet screen information are consistent with the theme characteristics above the bullet screen information, and the generation quality of the bullet screen information is further improved.
Fig. 3 shows a schematic structural diagram of a bullet screen generating device according to an embodiment of the present invention. As shown in fig. 3, the apparatus 300 includes: an acquisition module 310, a conversion module 320, a first determination module 330, a second determination module 340, a calculation module 350, and a prediction module 360.
The obtaining module 310 is configured to obtain the bullet screen information;
the conversion module 320 is configured to preprocess the bullet screen information to convert the bullet screen information into an initial text vector;
a first determining module 330, configured to determine a subject semantic relationship distribution of the initial text vector;
a second determining module 340, configured to determine, according to the initial text vector, a syntactic dependency vector and a position relationship vector between words in the bullet screen information;
a calculating module 350, configured to perform weighted calculation on the syntactic dependency relationship vector and the position relationship vector to obtain an input vector of each word;
and the prediction module 360 is configured to input the input vector and the subject semantic relationship distribution into a target neural network for prediction, so as to generate a bullet screen information context for replying the bullet screen information context.
In an alternative manner, the second determining module 340 is configured to:
determining the position of a current word, wherein the current word is any word vector in the initial text vectors;
traversing the position of the dependent word corresponding to the current word, and ending the traversal when traversing to the position of the core word, wherein the dependent word is the dependent word of the current word;
calculating the absolute value of the difference value between the position of the current word and the position of the dependent word;
and accumulating the absolute value of the difference value to calculate the core weight of the current word to obtain the syntactic dependency vector.
In an alternative approach, the target neural network includes an attention mechanism, a feed-forward network, and a subject word prediction model; the prediction module 360 is to:
determining the theme of the initial text vector and the corresponding word distribution under the theme according to the theme semantic relationship distribution;
according to the theme and the corresponding word distribution under the theme, performing attention mechanism calculation on the input vector to obtain an interword relation vector;
after residual connection and normalization are carried out on the inter-word relation vectors, the inter-word relation vectors are input into the feedforward network, and theme semantic vectors of all words in the input vectors are obtained;
residual error connection is carried out on the theme semantic vector and the inter-word relation vector to obtain a bullet screen information sequence vector;
and inputting the bullet screen information sequence vector into the subject word prediction model to generate a bullet screen information context for replying the bullet screen information context.
In an optional manner, the bullet screen information sequence vector includes a theme above the bullet screen information, word distribution under the theme, theme probability and semantic information; the prediction module 360 is to:
determining a last prediction result, wherein the last prediction result is a predicted word predicted before a current predicted word;
taking the last prediction result and the bullet screen information sequence vector as input, and obtaining the probability of the current predicted word according to a preset vocabulary model and the distribution of the subject semantic relation;
weighting and calculating the theme probability and the current predicted word probability to obtain corresponding predicted words under each theme;
and generating a bullet screen information context for replying the bullet screen information context according to the corresponding prediction words under the topics.
In an optional manner, the apparatus 300 further comprises an output module for:
calculating a quality value of the bullet screen information according to a preset quality value formula, wherein the quality value is used for representing the logic collocation quality, the content diversity quality and the theme quality of the bullet screen information;
and if the quality value of the bullet screen information context is larger than a preset quality value threshold, outputting the bullet screen information context.
In an alternative form, the output module is configured to:
respectively determining a logic collocation quality value, a content diversity quality value and a theme quality value under the bullet screen information;
calculating a product of the logical collocation quality value and the content diversity quality value, and determining a sum of the product and the topic quality value as the quality value.
In an optional manner, the apparatus 300 further comprises a training module for:
inputting a training sample into the target neural network for training to obtain an output result, wherein the training sample is a bullet screen dialogue text with a label, and the label comprises a theme of the bullet screen dialogue text and a corresponding bullet screen dialogue text;
calculating a loss value of the target neural network according to the output result, the bullet screen dialog context and a preset loss function;
adjusting parameters of the target neural network according to the loss value, re-executing the step of inputting a training sample into the target neural network for training to obtain an output result, calculating the loss value of the target neural network according to the output result, the bullet screen dialog context and a preset loss function, and adjusting the parameters of the target neural network according to the loss value until the preset loss function is converged or is smaller than a preset threshold value;
the preset loss function is composed of a word prediction likelihood function and a cosine similarity distance of a context semantic vector.
In the embodiment of the invention, after converting the bullet screen information into the initial text vector, the theme semantic relationship distribution of the initial text vector, and the syntax dependence vector and the position relationship vector between words in the bullet screen information can be determined; and determining an input vector of each word according to the syntactic dependency vector and the position relation vector, inputting the input vector and the subject semantic relation distribution into a target neural network for prediction, and generating a bullet screen information context for replying the bullet screen information context. The method and the device can determine the input vector of each word through the syntactic dependency vector and the position relation vector, and can more accurately mine semantic roles borne by the predicted words, so that the logic structure of the bullet screen information is more accurate; the theme characteristics above the bullet screen information can be dynamically captured by determining the theme semantic relationship distribution of the initial text vector, so that the theme characteristics below the bullet screen information are consistent with the theme characteristics above the bullet screen information, and the generation quality of the bullet screen information is further improved.
Fig. 4 is a schematic structural diagram of an electronic device according to an embodiment of the present invention, and the specific embodiment of the present invention does not limit the specific implementation of the electronic device.
As shown in fig. 4, the electronic device may include: a processor (processor)402, a Communications Interface 404, a memory 406, and a Communications bus 408.
Wherein: the processor 402, communication interface 404, and memory 406 communicate with each other via a communication bus 408. A communication interface 404 for communicating with network elements of other devices, such as clients or other servers. The processor 402 is configured to execute the program 410, and may specifically execute the relevant steps in the foregoing bullet screen generation method embodiment.
In particular, program 410 may include program code comprising computer-executable instructions.
The processor 402 may be a central processing unit CPU or an application Specific Integrated circuit asic or one or more Integrated circuits configured to implement embodiments of the present invention. The electronic device comprises one or more processors, which can be the same type of processor, such as one or more CPUs; or may be different types of processors such as one or more CPUs and one or more ASICs.
And a memory 406 for storing a program 410. Memory 406 may comprise high-speed RAM memory, and may also include non-volatile memory (non-volatile memory), such as at least one disk memory.
The program 410 may specifically be invoked by the processor 402 to cause the electronic device to perform the following operations:
acquiring the bullet screen information;
preprocessing the bullet screen information to convert the bullet screen information into an initial text vector;
determining a subject matter semantic relation distribution of the initial text vector;
determining a syntax dependence vector and a position relation vector between words in the bullet screen information according to the initial text vector;
performing weighted calculation on the syntactic dependency relationship vector and the position relationship vector to obtain an input vector of each word;
and inputting the input vector and the subject semantic relation distribution into a target neural network for prediction so as to generate a bullet screen information context for replying the bullet screen information context.
In an alternative, the program 410 is invoked by the processor 402 to cause the electronic device to perform the following operations:
determining the position of a current word, wherein the current word is any word vector in the initial text vectors;
traversing the position of the dependent word corresponding to the current word, and ending the traversal when traversing to the position of the core word, wherein the dependent word is the dependent word of the current word;
calculating the absolute value of the difference value between the position of the current word and the position of the dependent word;
and accumulating the absolute value of the difference value to calculate the core weight of the current word to obtain the syntactic dependency vector.
In an alternative approach, the target neural network includes an attention mechanism, a feed-forward network, and a subject word prediction model; the program 410 is invoked by the processor 402 to cause the electronic device to perform the following operations:
determining the theme of the initial text vector and the corresponding word distribution under the theme according to the theme semantic relationship distribution;
according to the theme and the corresponding word distribution under the theme, performing attention mechanism calculation on the input vector to obtain an interword relation vector;
after residual connection and normalization are carried out on the inter-word relation vectors, the inter-word relation vectors are input into the feedforward network, and theme semantic vectors of all words in the input vectors are obtained;
residual error connection is carried out on the theme semantic vector and the inter-word relation vector to obtain a bullet screen information sequence vector;
and inputting the bullet screen information sequence vector into the subject word prediction model to generate a bullet screen information context for replying the bullet screen information context.
In an optional manner, the bullet screen information sequence vector includes a theme above the bullet screen information, word distribution under the theme, theme probability and semantic information; the program 410 is invoked by the processor 402 to cause the electronic device to perform the following operations:
determining a last prediction result, wherein the last prediction result is a predicted word predicted before a current predicted word;
taking the last prediction result and the bullet screen information sequence vector as input, and obtaining the probability of the current predicted word according to a preset vocabulary model and the distribution of the subject semantic relation;
weighting and calculating the theme probability and the current predicted word probability to obtain corresponding predicted words under each theme;
and generating a bullet screen information context for replying the bullet screen information context according to the corresponding prediction words under the topics.
In an alternative, the program 410 is invoked by the processor 402 to cause the electronic device to perform the following operations:
calculating a quality value of the bullet screen information according to a preset quality value formula, wherein the quality value is used for representing the logic collocation quality, the content diversity quality and the theme quality of the bullet screen information;
and if the quality value of the bullet screen information context is larger than a preset quality value threshold, outputting the bullet screen information context.
In an alternative, the program 410 is invoked by the processor 402 to cause the electronic device to perform the following operations:
respectively determining a logic collocation quality value, a content diversity quality value and a theme quality value under the bullet screen information;
calculating a product of the logical collocation quality value and the content diversity quality value, and determining a sum of the product and the topic quality value as the quality value.
In an alternative, the program 410 is invoked by the processor 402 to cause the electronic device to perform the following operations:
inputting a training sample into the target neural network for training to obtain an output result; the training sample is a bullet screen dialogue text with a label; the label comprises a theme of the bullet screen dialogue text and a corresponding bullet screen dialogue text;
calculating a loss value of the target neural network according to the output result, the bullet screen dialog context and a preset loss function;
adjusting parameters of the target neural network according to the loss value, re-executing the step of inputting a training sample into the target neural network for training to obtain an output result, calculating the loss value of the target neural network according to the output result, the bullet screen dialog context and a preset loss function, and adjusting the parameters of the target neural network according to the loss value until the preset loss function is converged or is smaller than a preset threshold value;
the preset loss function is composed of a word prediction likelihood function and a cosine similarity distance of a context semantic vector.
In the embodiment of the invention, after converting the bullet screen information into the initial text vector, the theme semantic relationship distribution of the initial text vector, and the syntax dependence vector and the position relationship vector between words in the bullet screen information can be determined; and determining an input vector of each word according to the syntactic dependency vector and the position relation vector, inputting the input vector and the subject semantic relation distribution into a target neural network for prediction, and generating a bullet screen information context for replying the bullet screen information context. The method and the device can determine the input vector of each word through the syntactic dependency vector and the position relation vector, and can more accurately mine semantic roles borne by the predicted words, so that the logic structure of the bullet screen information is more accurate; the theme characteristics above the bullet screen information can be dynamically captured by determining the theme semantic relationship distribution of the initial text vector, so that the theme characteristics below the bullet screen information are consistent with the theme characteristics above the bullet screen information, and the generation quality of the bullet screen information is further improved.
The embodiment of the invention provides a computer-readable storage medium, wherein at least one executable instruction is stored in the storage medium, and when the executable instruction runs on an electronic device, the electronic device is enabled to execute the bullet screen reply generation method in any method embodiment.
The embodiment of the invention provides a bullet screen generating device which is used for executing the bullet screen reply generating method.
Embodiments of the present invention provide a computer program, where the computer program can be called by a processor to enable an electronic device to execute a bullet screen reply generation method in any of the above method embodiments.
An embodiment of the present invention provides a computer program product, where the computer program product includes a computer program stored on a computer-readable storage medium, and the computer program includes program instructions, when the program instructions are run on a computer, the computer is caused to execute the bullet screen reply generation method in any of the above method embodiments.
The algorithms or displays presented herein are not inherently related to any particular computer, virtual system, or other apparatus. Various general purpose systems may also be used with the teachings herein. The required structure for constructing such a system will be apparent from the description above. In addition, embodiments of the present invention are not directed to any particular programming language. It is appreciated that a variety of programming languages may be used to implement the teachings of the present invention as described herein, and any descriptions of specific languages are provided above to disclose the best mode of the invention.
In the description provided herein, numerous specific details are set forth. It is understood, however, that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.
Similarly, it should be appreciated that in the foregoing description of exemplary embodiments of the invention, various features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof for the purpose of streamlining the invention and aiding in the understanding of one or more of the various inventive aspects. However, the disclosed method should not be interpreted as reflecting an intention that: that the invention as claimed requires more features than are expressly recited in each claim.
Those skilled in the art will appreciate that the modules in the device in an embodiment may be adaptively changed and disposed in one or more devices different from the embodiment. The modules or units or components of the embodiments may be combined into one module or unit or component, and may be divided into a plurality of sub-modules or sub-units or sub-components. All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and all of the processes or elements of any method or apparatus so disclosed, may be combined in any combination, except combinations where at least some of such features and/or processes or elements are mutually exclusive. Each feature disclosed in this specification (including any accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise.
It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and that those skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The usage of the words first, second and third, etcetera do not indicate any ordering. These words may be interpreted as names. The steps in the above embodiments should not be construed as limiting the order of execution unless specified otherwise.
Claims (10)
1. A bullet screen reply generation method is characterized by comprising the following steps:
acquiring the bullet screen information;
preprocessing the bullet screen information to convert the bullet screen information into an initial text vector;
determining a subject matter semantic relation distribution of the initial text vector;
determining a syntax dependence vector and a position relation vector between words in the bullet screen information according to the initial text vector;
performing weighted calculation on the syntactic dependency relationship vector and the position relationship vector to obtain an input vector of each word;
and inputting the input vector and the subject semantic relation distribution into a target neural network for prediction so as to generate a bullet screen information context for replying the bullet screen information context.
2. The method of claim 1, wherein determining the syntactic dependency vector and the position relationship vector between words in the bullet screen information according to the initial text vector comprises:
determining the position of a current word, wherein the current word is any word vector in the initial text vectors;
traversing the position of the dependent word corresponding to the current word, and ending the traversal when traversing to the position of the core word, wherein the dependent word is the dependent word of the current word;
calculating the absolute value of the difference value between the position of the current word and the position of the dependent word;
and accumulating the absolute value of the difference value to calculate the core weight of the current word to obtain the syntactic dependency vector.
3. The method of claim 1, wherein the target neural network comprises an attention mechanism, a feed forward network, and a subject term prediction model; the step of inputting the input vector and the subject semantic relation distribution into a target neural network for prediction to generate a bullet screen information context for replying the bullet screen information context comprises:
determining the theme of the initial text vector and the corresponding word distribution under the theme according to the theme semantic relationship distribution;
according to the theme and the corresponding word distribution under the theme, performing attention mechanism calculation on the input vector to obtain an interword relation vector;
after residual connection and normalization are carried out on the inter-word relation vectors, the inter-word relation vectors are input into the feedforward network, and theme semantic vectors of all words in the input vectors are obtained;
residual error connection is carried out on the theme semantic vector and the inter-word relation vector to obtain a bullet screen information sequence vector;
and inputting the bullet screen information sequence vector into the subject word prediction model to generate a bullet screen information context for replying the bullet screen information context.
4. The method of claim 3, wherein the barrage information sequence vector comprises a topic above the barrage information, a word distribution under the topic, a topic probability, and semantic information; the inputting the bullet screen information sequence vector into the subject term prediction model to generate a bullet screen information context for replying the bullet screen information context includes:
determining a last prediction result, wherein the last prediction result is a predicted word predicted before a current predicted word;
taking the last prediction result and the bullet screen information sequence vector as input, and obtaining the probability of the current predicted word according to a preset vocabulary model and the distribution of the subject semantic relation;
weighting and calculating the theme probability and the current predicted word probability to obtain corresponding predicted words under each theme;
and generating a bullet screen information context for replying the bullet screen information context according to the corresponding prediction words under the topics.
5. The method of claim 1, further comprising:
calculating a quality value of the bullet screen information context according to a preset quality value formula, wherein the quality value is used for representing the logic collocation quality, the content diversity quality and the theme quality of the bullet screen information context;
and if the quality value of the bullet screen information context is larger than a preset quality value threshold, outputting the bullet screen information context.
6. The method of claim 5, wherein calculating a quality value for the bullet screen information context according to a preset quality value formula comprises:
respectively determining a logic collocation quality value, a content diversity quality value and a theme quality value under the bullet screen information;
calculating a product of the logical collocation quality value and the content diversity quality value, and determining a sum of the product and the topic quality value as the quality value.
7. The method of claim 1, further comprising:
inputting a training sample into the target neural network for training to obtain an output result; the training sample is a bullet screen dialogue text with a label; the label comprises a theme of the bullet screen dialogue text and a corresponding bullet screen dialogue text;
calculating a loss value of the target neural network according to the output result, the bullet screen dialog context and a preset loss function;
adjusting parameters of the target neural network according to the loss value, re-executing the step of inputting a training sample into the target neural network for training to obtain an output result, calculating the loss value of the target neural network according to the output result, the bullet screen dialog context and a preset loss function, and adjusting the parameters of the target neural network according to the loss value until the preset loss function is converged or is smaller than a preset threshold value;
the preset loss function is composed of a word prediction likelihood function and a cosine similarity distance of a context semantic vector.
8. A bullet screen reply generation apparatus, characterized in that the apparatus comprises:
the acquisition module is used for acquiring the bullet screen information;
the conversion module is used for preprocessing the bullet screen information and converting the bullet screen information into an initial text vector;
the first determining module is used for determining the subject semantic relation distribution of the initial text vector;
the second determining module is used for determining a syntactic dependency vector and a position relation vector between words in the bullet screen information according to the initial text vector;
the calculation module is used for carrying out weighting calculation on the syntactic dependency relationship vector and the position relationship vector to obtain an input vector of each word;
and the prediction module is used for inputting the input vector and the subject semantic relation distribution into a target neural network for prediction so as to generate a bullet screen information context for replying the bullet screen information context.
9. An electronic device, comprising: the system comprises a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface complete mutual communication through the communication bus;
the memory is used for storing at least one executable instruction, and the executable instruction causes the processor to execute the operation of the bullet screen reply generation method according to any one of claims 1 to 7.
10. A computer-readable storage medium, wherein the storage medium has at least one executable instruction stored therein, and when the executable instruction is executed on an electronic device, the electronic device is caused to perform the operations of the bullet screen reply generation method according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110855573.2A CN113591462A (en) | 2021-07-28 | 2021-07-28 | Bullet screen reply generation method and device and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110855573.2A CN113591462A (en) | 2021-07-28 | 2021-07-28 | Bullet screen reply generation method and device and electronic equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113591462A true CN113591462A (en) | 2021-11-02 |
Family
ID=78251177
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110855573.2A Pending CN113591462A (en) | 2021-07-28 | 2021-07-28 | Bullet screen reply generation method and device and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113591462A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114548115A (en) * | 2022-02-23 | 2022-05-27 | 北京三快在线科技有限公司 | Method and device for explaining compound nouns and electronic equipment |
| CN114942986A (en) * | 2022-06-21 | 2022-08-26 | 平安科技（深圳）有限公司 | Text generation method and device, computer equipment and computer readable storage medium |
| WO2023108981A1 (en) * | 2021-12-15 | 2023-06-22 | 平安科技（深圳）有限公司 | Method and apparatus for training text generation model, and storage medium and computer device |
| CN117521680A (en) * | 2024-01-04 | 2024-02-06 | 深圳和成视讯科技有限公司 | Online translation method of recorder and recorder |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106101756A (en) * | 2016-06-07 | 2016-11-09 | 腾讯科技（北京）有限公司 | Barrage display packing, barrage adding method, Apparatus and system |
| CN107229610A (en) * | 2017-03-17 | 2017-10-03 | 咪咕数字传媒有限公司 | The analysis method and device of a kind of affection data |
| CN109408641A (en) * | 2018-11-22 | 2019-03-01 | 山东工商学院 | It is a kind of based on have supervision topic model file classification method and system |
| CN110035325A (en) * | 2019-04-19 | 2019-07-19 | 广州虎牙信息科技有限公司 | Barrage answering method, barrage return mechanism and live streaming equipment |
| CN110850983A (en) * | 2019-11-13 | 2020-02-28 | 腾讯科技（深圳）有限公司 | Virtual object control method and device in video live broadcast and storage medium |
| CN110895656A (en) * | 2018-09-13 | 2020-03-20 | 武汉斗鱼网络科技有限公司 | Text similarity calculation method and device, electronic equipment and storage medium |
| CN111241295A (en) * | 2020-01-03 | 2020-06-05 | 浙江大学 | Knowledge map relation data extraction method based on semantic syntax interactive network |
| US20200294507A1 (en) * | 2019-03-11 | 2020-09-17 | Adobe Inc. | Pose-invariant Visual Speech Recognition Using A Single View Input |
| CN111783474A (en) * | 2020-07-16 | 2020-10-16 | 厦门市美亚柏科信息股份有限公司 | Comment text viewpoint information processing method and device and storage medium |
| US20210200957A1 (en) * | 2019-12-27 | 2021-07-01 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and apparatus for generating dialogue model |
-
2021
- 2021-07-28 CN CN202110855573.2A patent/CN113591462A/en active Pending
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106101756A (en) * | 2016-06-07 | 2016-11-09 | 腾讯科技（北京）有限公司 | Barrage display packing, barrage adding method, Apparatus and system |
| CN107229610A (en) * | 2017-03-17 | 2017-10-03 | 咪咕数字传媒有限公司 | The analysis method and device of a kind of affection data |
| CN110895656A (en) * | 2018-09-13 | 2020-03-20 | 武汉斗鱼网络科技有限公司 | Text similarity calculation method and device, electronic equipment and storage medium |
| CN109408641A (en) * | 2018-11-22 | 2019-03-01 | 山东工商学院 | It is a kind of based on have supervision topic model file classification method and system |
| US20200294507A1 (en) * | 2019-03-11 | 2020-09-17 | Adobe Inc. | Pose-invariant Visual Speech Recognition Using A Single View Input |
| CN110035325A (en) * | 2019-04-19 | 2019-07-19 | 广州虎牙信息科技有限公司 | Barrage answering method, barrage return mechanism and live streaming equipment |
| CN110850983A (en) * | 2019-11-13 | 2020-02-28 | 腾讯科技（深圳）有限公司 | Virtual object control method and device in video live broadcast and storage medium |
| US20210200957A1 (en) * | 2019-12-27 | 2021-07-01 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and apparatus for generating dialogue model |
| CN111241295A (en) * | 2020-01-03 | 2020-06-05 | 浙江大学 | Knowledge map relation data extraction method based on semantic syntax interactive network |
| CN111783474A (en) * | 2020-07-16 | 2020-10-16 | 厦门市美亚柏科信息股份有限公司 | Comment text viewpoint information processing method and device and storage medium |
Non-Patent Citations (1)
| Title |
|---|
| 李卫疆;漆芳;: "基于多通道双向长短期记忆网络的情感分析", 《中文信息学报》, no. 12, 15 December 2019 (2019-12-15) * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2023108981A1 (en) * | 2021-12-15 | 2023-06-22 | 平安科技（深圳）有限公司 | Method and apparatus for training text generation model, and storage medium and computer device |
| CN114548115A (en) * | 2022-02-23 | 2022-05-27 | 北京三快在线科技有限公司 | Method and device for explaining compound nouns and electronic equipment |
| CN114548115B (en) * | 2022-02-23 | 2023-01-06 | 北京三快在线科技有限公司 | Method and device for explaining compound nouns and electronic equipment |
| CN114942986A (en) * | 2022-06-21 | 2022-08-26 | 平安科技（深圳）有限公司 | Text generation method and device, computer equipment and computer readable storage medium |
| CN114942986B (en) * | 2022-06-21 | 2024-03-19 | 平安科技（深圳）有限公司 | Text generation method, text generation device, computer equipment and computer readable storage medium |
| CN117521680A (en) * | 2024-01-04 | 2024-02-06 | 深圳和成视讯科技有限公司 | Online translation method of recorder and recorder |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP6818941B2 (en) | How to Train Multilingual Speech Recognition Networks, Speech Recognition Systems and Multilingual Speech Recognition Systems | |
| US20210390271A1 (en) | Neural machine translation systems | |
| US20240161732A1 (en) | Multi-dialect and multilingual speech recognition | |
| Tilk et al. | Bidirectional Recurrent Neural Network with Attention Mechanism for Punctuation Restoration. | |
| JP2020505650A (en) | Voice recognition system and voice recognition method | |
| CN113591462A (en) | Bullet screen reply generation method and device and electronic equipment | |
| Liu et al. | Reinforcement learning for emotional text-to-speech synthesis with improved emotion discriminability | |
| Khare et al. | Low Resource ASR: The Surprising Effectiveness of High Resource Transliteration. | |
| Nagaraj et al. | Kannada to English Machine Translation Using Deep Neural Network. | |
| KR20180062003A (en) | Method of correcting speech recognition errors | |
| CN112905772B (en) | Semantic correlation analysis method and device and related products | |
| Das et al. | Best of both worlds: Robust accented speech recognition with adversarial transfer learning | |
| CN112348073A (en) | Polyphone recognition method and device, electronic equipment and storage medium | |
| Granell et al. | Multimodal crowdsourcing for transcribing handwritten documents | |
| JP6810580B2 (en) | Language model learning device and its program | |
| CN116050425A (en) | Method for establishing pre-training language model, text prediction method and device | |
| Srinivasan et al. | Multimodal speech recognition with unstructured audio masking | |
| CN111508497B (en) | Speech recognition method, device, electronic equipment and storage medium | |
| Xie et al. | Dialogue breakdown detection using hierarchical bi-directional LSTMs | |
| CN111916052B (en) | Voice synthesis method and device | |
| KR102358485B1 (en) | Dialogue system by automatic domain classfication | |
| CN112989794A (en) | Model training method and device, intelligent robot and storage medium | |
| CN112686060B (en) | Text translation method, device, electronic equipment and storage medium | |
| KR20230156425A (en) | Streaming ASR model delay reduction through self-alignment | |
| Djeffal et al. | Automatic speech recognition with BERT and CTC transformers: A review |
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 |