Key information generation and live content processing method, apparatus and device

Abstract

The embodiment of the specification discloses a key information generation method, device and equipment, which is used for generating key information for original text including multiple sentences. The scheme includes: a word encoding step, which encodes each word in each sentence and determines the encoding result of each sentence; a sentence encoding step, which encodes the encoding result of each sentence and determines the overall encoding result of the original text; a decoding step, which decodes the encoding result of the original text and determines the target word; a key information generation step, which generates the key information of the original text by using multiple target words generated by multiple executions of the word encoding step, the sentence encoding step and the decoding step; wherein, when the word encoding step is executed each time, the key word in each sentence is determined according to the decoding state, and the encoding result of the sentence is determined based on the key word in each sentence.

Description

Technical Field

[0001] This specification relates to the field of machine learning technology, and in particular to methods, devices, and equipment for generating key information and processing live content. Background Art

[0002] With the development of Internet technology, the amount of information people face in their lives has exploded.

[0003] At present, in order to facilitate quick browsing of information, information providers usually extract key information from the information provided, and summarize the information content with a large amount of data in the form of a small amount of key information such as title and introduction, so that users can quickly understand various information and save browsing time.

[0004] Commonly used key information generation schemes use an attention mechanism to assign different weights to different parts of the text, and achieve end-to-end generation of long text sequences into short title sequences based on the same encoding results.

[0005] Based on this, a more accurate key information generation solution is needed. Summary of the Invention

[0006] One or more embodiments of this specification provide key information generation and live content processing methods, devices, equipment, and storage media to solve the following technical problems: a more accurate key information generation solution is needed.

[0007] To solve the above technical problems, one or more embodiments of this specification are implemented as follows:

[0008] One or more embodiments of this specification provide a key information generation method for generating key information for an original text including a plurality of sentences, the method comprising:

[0009] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0010] a sentence encoding step of encoding the encoding results of each of the sentences and determining the overall encoding result of the original text;

[0011] A decoding step of decoding the encoding result of the original text and determining the target word;

[0012] a key information generating step, which generates key information of the original text by using the plurality of target words generated by performing the word encoding step, the sentence encoding step, and the decoding step multiple times;

[0013] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0014] One or more embodiments of this specification provide a method for processing live content, including:

[0015] Get the live video of the anchor on the live broadcast platform;

[0016] Performing speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences;

[0017] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0018] a sentence encoding step, encoding the encoding results of each of the sentences and determining the overall encoding result of the live broadcast segment text;

[0019] A decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word;

[0020] a title generation step, which generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step, and serves as a title for the live video segment corresponding to the live segment text, for publishing in conjunction with the corresponding live video segment so that users can select and replay the segment;

[0021] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0022] One or more embodiments of this specification provide a key information generation device for generating key information for an original text including a plurality of sentences, the device comprising:

[0023] a word encoding module, executing a word encoding step, encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0024] A sentence encoding module, which performs a sentence encoding step, encodes the encoding result of each sentence and determines the overall encoding result of the original text;

[0025] A decoding module performs a decoding step, which decodes the encoding result of the original text and determines the target word;

[0026] a key information generating module, which generates key information of the original text by using the plurality of target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step;

[0027] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0028] One or more embodiments of this specification provide a live content processing device, including:

[0029] Video acquisition module, which obtains the live video of the anchor on the live broadcast platform;

[0030] A slicing processing module performs speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences;

[0031] a word encoding module, executing a word encoding step, encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0032] A sentence encoding module executes a sentence encoding step, which encodes the encoding result of each sentence and determines the overall encoding result of the live broadcast segment text;

[0033] A decoding module executes a decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word;

[0034] a title generation module, which generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step, and serves as a segment title of the live video segment corresponding to the live segment text, for publishing in conjunction with the corresponding live video segment so that users can select a segment for playback;

[0035] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0036] One or more embodiments of this specification provide a key information generation device for generating key information for an original text including a plurality of sentences, the device comprising:

[0037] at least one processor; and,

[0038] a memory communicatively connected to the at least one processor; wherein,

[0039] The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0040] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0041] a sentence encoding step of encoding the encoding results of each of the sentences and determining the overall encoding result of the original text;

[0042] A decoding step of decoding the encoding result of the original text and determining the target word;

[0043] a key information generating step, which generates key information of the original text by using the plurality of target words generated by performing the word encoding step, the sentence encoding step, and the decoding step multiple times;

[0044] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0045] One or more embodiments of this specification provide a live content processing device, including:

[0046] at least one processor; and,

[0047] a memory communicatively connected to the at least one processor; wherein,

[0048] The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0049] Get the live video of the anchor on the live broadcast platform;

[0050] Performing speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences;

[0051] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0052] a sentence encoding step, encoding the encoding results of each of the sentences and determining the overall encoding result of the live broadcast segment text;

[0053] A decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word;

[0054] a title generation step, which generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step, and serves as a title for the live video segment corresponding to the live segment text, for publishing in conjunction with the corresponding live video segment so that users can select and replay the segment;

[0055] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0056] One or more embodiments of this specification provide a non-volatile computer storage medium storing computer-executable instructions for generating key information for an original text comprising a plurality of sentences. The computer-executable instructions are configured to:

[0057] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0058] a sentence encoding step of encoding the encoding results of each of the sentences and determining the overall encoding result of the original text;

[0059] A decoding step of decoding the encoding result of the original text and determining the target word;

[0060] a key information generating step, which generates key information of the original text by using the plurality of target words generated by performing the word encoding step, the sentence encoding step, and the decoding step multiple times;

[0061] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0062] One or more embodiments of this specification provide a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to:

[0063] Get the live video of the anchor on the live broadcast platform;

[0064] Performing speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences;

[0065] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0066] a sentence encoding step, encoding the encoding results of each of the sentences and determining the overall encoding result of the live broadcast segment text;

[0067] A decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word;

[0068] a title generation step, which generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step, and serves as a title for the live video segment corresponding to the live segment text, for publishing in conjunction with the corresponding live video segment so that users can select and replay the segment;

[0069] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0070] At least one of the above-mentioned technical solutions employed in one or more embodiments of this specification can achieve the following beneficial effects: Each time a target word is generated (key information such as a title consists of multiple target words, with one target word generated each time), it is necessary to dynamically determine keywords in the original text, then represent sentences based on these keywords, and then use these represented sentences to represent the original text as a whole. This solution allows for the dynamic process of generating key information word by word to promptly adjust attention to different parts and levels of content in the original text, improving the accuracy of the identified target words, reducing the number of missed target words, and thus improving the accuracy of the generated key information. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0072] Figure 1 A flowchart of a key information generation method provided in one or more embodiments of this specification;

[0073] Figure 2 In an application scenario provided by one or more embodiments of this specification, Figure 1 A detailed framework and flowchart of the method;

[0074] Figure 3 A schematic diagram of the structure of a key information generating device provided in one or more embodiments of this specification;

[0075] Figure 4 A schematic diagram of the structure of a key information generating device provided in one or more embodiments of this specification;

[0076] Figure 5 A flowchart of a live content processing method provided in one or more embodiments of this specification;

[0077] Figure 6 A schematic diagram of the structure of a live content processing device provided in one or more embodiments of this specification;

[0078] Figure 7 A schematic diagram of the structure of a live content processing device provided in one or more embodiments of this specification. DETAILED DESCRIPTION

[0079] The embodiments of this specification provide methods, devices, equipment, and storage media for generating key information and processing live content.

[0080] In order to help those skilled in the art better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0081] For ease of description, the following embodiments illustrate this solution using key information such as title as an example.

[0082] In the process of title generation, since the content of the original text is usually long, we can first extract some content from the original text as key sentences, and then generate titles based on the extracted key sentences. In this scheme, the accuracy of key sentence extraction has a greater impact on the accuracy of title generation.

[0083] In one or more embodiments of the present specification, each time a target word constituting a title is generated, the current keywords in the original text are dynamically determined (the keywords determined for each target word generation may change; the currently determined keywords are referred to as the current keywords) and their encodings to update the encoding of the sentences in which they appear. Optionally, the current key sentences in the original text can also be dynamically determined (i.e., the currently determined key sentences may also change each time), to update the overall representation of the original text. Based on the grammatical relationships between the current key sentences and the sentences, the target words generated this time can be obtained through decoding to ultimately generate the title. This allows for timely adjustments to the keywords or keywords and key sentences of interest in the original text as the title is generated, improving the accuracy of target word generation and enhancing the effectiveness of the title. This solution can be applied in a variety of application scenarios, such as live broadcast publishing (e.g., generating titles for live broadcast content) and service integration (e.g., regenerating overall descriptions for integrated services). Taking the live broadcasting field as an example, this specification also uses the above solution to process live videos to determine the titles corresponding to live video segments, allowing users to more accurately and efficiently review the segments they are interested in without having to browse through the entire live video. The following is a detailed explanation based on this idea.

[0084] Figure 1 This is a flowchart of a key information generation method provided in one or more embodiments of this specification. This method is used to generate key information from a raw text consisting of multiple sentences and can be applied to various business areas, such as internet finance, e-commerce, instant messaging, gaming, and official business. The process can be executed by computing devices in the corresponding fields, and certain input parameters or intermediate results in the process can be manually adjusted to help improve accuracy.

[0085] In one or more embodiments of the present specification, the execution body of the process includes a word encoder, a sentence encoder, a decoder, etc. The word encoder and the sentence encoder are used to encode the original text hierarchically (from different granularity levels such as words, sentences, and original texts) and dynamically (dynamically selecting keywords each time to dynamically represent the original text as a whole). The decoder decodes the encoding results in sequence to obtain the target words in order.

[0086] Figure 1 The process in may include the following steps:

[0087] S102: a word encoding step, which encodes each word in each of the sentences and determines the encoding result of each of the sentences.

[0088] In one or more embodiments of this specification, each time the word encoding step is executed, the keywords in each sentence are determined based on the title generation state, and the encoding result of the sentence is determined based on the keywords in each sentence. It should be noted that the time mentioned in this solution is a summary of the time range of the execution process of one or more corresponding actions, and is not limited to a specific time point.

[0089] The title generation state includes, for example, a decoding state, where the decoding state may include a historical state at a previous decoding moment, such as a state at a previous moment, which is usually a hidden state.

[0090] The decoding moment indicates the moment when the target word in the title was generated. Each decoding moment corresponds to each generated target word. The target word may exist in the original text, meaning it is copied directly from the original text, or it may not exist in the original text but be summarized based on the key content of the original text.

[0091] The decoding state can reflect information such as the title composed of the target words that have been generated, or the overall representation of the previous original text, which will affect the generation of the next target word. In this step, it will affect the determination of keywords, and the keywords can be determined by paying different attention to different contents in the original text.

[0092] Raw text represents content that needs to be summarized and its central idea extracted to generate a title. This can be complete content, such as an article, an email, or a news report, or partial content, such as a section within an article. While raw text is in text form, this solution can also be applied to generating titles for non-textual content, such as video and audio. For non-textual content, it must first be converted into text through methods such as speech recognition to obtain the corresponding raw text before title generation.

[0093] Specifically, when processing the live video of the anchor on the live broadcast platform, the live video can be subjected to voice recognition and slicing processing to obtain multiple live segment texts, which are respectively used as the original texts corresponding to the corresponding live segments for title generation.

[0094] When encoding a word, you can encode it based solely on the sentence it appears in. The resulting code is considered a sentence-specific code for that word, used to represent the relationship between the keyword and other words within the sentence. Due to differences in sentence structure and other factors, the specific code corresponding to the same keyword in different sentences may be the same or different.

[0095] During each target word generation process, the current encoding (the encoding result of this sentence) of the original text corresponding to the current decoding moment (i.e., the current generation process) must be dynamically determined based on changes in the title generation state and keywords. The current encoding of the sentence containing the keyword can be determined based on the keyword's dedicated encoding. Sentence encodings may vary at different decoding moments.

[0096] It's important to note that the same keyword in different sentences can have different, specific encodings specific to that sentence. This allows for more granular and precise representation of the sentence, thus providing greater differentiation at the sentence level. Each sentence can encode its own words independently, and the encoding of the same keyword in different locations within the same sentence may also be different, especially when the keyword has different parts of speech.

[0097] According to the encoding results of each keyword, the weighted sum of the attention weights at the word granularity is calculated (the size of the attention weight itself can also indicate which words are the keywords of this time) to determine the encoding result of the sentence in which it is located.

[0098] S104: a sentence encoding step, which encodes the encoding results of each sentence and determines the overall encoding result of the original text.

[0099] During each target word generation process, the key sentence at the current decoding moment is determined based on the current encoding of the sentence and the current title generation state (e.g., as reflected by information such as sentence-level attention weights). The key sentences corresponding to different decoding moments may be different, and not all sentences containing the keywords described in S102 are necessarily key sentences.

[0100] S106: A decoding step, which decodes the encoding result of the original text and determines the target word.

[0101] In one or more embodiments of the present specification, decoding is performed to determine a target word based on the encoding result of the original text, the decoding state at the current moment, and one or more target words that have been generated previously.

[0102] In practical applications, there are two possible sources for target words: the first is to copy directly from words in the original text; the second is to generate words from a given preset vocabulary, which includes words not found in the original text. Decoding is used to determine which source is used to determine the target word. For example, the generation probability and vocabulary distribution of the preset vocabulary can be determined. Based on the keyword determination process, the copy probability of the words in the original text can be determined. The target word at the current moment is determined based on the generation probability, vocabulary distribution, and copy probability.

[0103] S108: A key information generating step, which generates key information of the original text by using the multiple target words generated by executing the word encoding step, the sentence encoding step and the decoding step multiple times; wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0104] By sequentially generating each target word, we can determine the title generated from the original text. Because each target word in the title requires attention to a different part of the original text, the selection of keywords must be dynamic. By determining the keywords in each sentence based on the decoding state during each word encoding step, we can ensure that each target word is generated accurately, achieving better generation results.

[0105] based on Figure 1 This specification also provides some specific implementation plans and extension plans of the method, which will be described below.

[0106] In one or more embodiments of this specification, a structure combining a graph convolution network (GCN) and a recurrent neural network (RNN) is employed to construct a word encoder and a sentence encoder, respectively. The RNN, for example, is composed of a gated recurrent unit (GRU), with the sentence encoder connected to the word encoder. The GCN encodes auxiliary information that depends on word or sentence granularity (e.g., independent syntactic relationships between words, grammatical relationships between sentences, etc.) to more accurately represent words and sentences. The RNN combines the basic content of words and sentences with this auxiliary information to perform holistic encoding influenced by historical state.

[0107] Syntactic relations reflect the structural and logical relationships between words and between words and sentences within a sentence. Based on these syntactic relations, words within a sentence can be more accurately encoded. Grammatical relations reflect the structural and logical relationships between sentences within a text. Based on these grammatical relations, sentences within the text can be more accurately encoded. Syntactic and grammatical relations can be expressed in forms such as matrices and structure diagrams.

[0108] In one or more embodiments of the present specification, when encoding each word in each sentence, a word encoder including a graph neural network layer and a recurrent neural network layer can be determined to extract the syntactic relationship between each word in each sentence (the auxiliary information described above), and obtain the embedding vector of each word (the basic content described above). The syntactic relationship is encoded by the graph neural network layer of the word encoder. Based on the encoding result of the embedding vector and the syntactic relationship, the word encoder is processed by the recurrent neural network layer to obtain the encoding result of each word. For example, each word in each of the sentences can be parsed to obtain the part of speech and the syntactic relationship between each word, and an embedding vector is generated for the corresponding word based on the part of speech, and an adjacency matrix for encoding that represents the syntactic relationship is generated. It should be noted that general word embedding vectors do not include part of speech. Here, in order to more accurately and dynamically select keywords, it is necessary to better understand the role of the word in the sentence, and part of speech is very helpful in understanding the role of the word. Therefore, this solution provides such a word embedding vector that additionally includes part of speech, which is more suitable for the concept of this solution.

[0109] Furthermore, when determining the keywords in each sentence, a similarity score can be determined based on the state of the word encoder and the state of the decoder at the previous moment. Based on the similarity score, the first attention weight of the word in each sentence at the current moment is determined, and based on the first attention weight, the keywords in each sentence are determined. Combined with the encoding results of each word output by the RNN, the encoding results of the keywords can also be obtained accordingly. The encoding results of the keywords can then be combined with their attention weights to determine the encoding results of the corresponding sentences.

[0110] In one or more embodiments of this specification, the sentence encoder process is similar to the word encoder process. The encoding results (e.g., the weighted sum of key words) generated by the word encoder are relatively loose, insufficiently influenced by other sentences, and fail to fully consider the status of each sentence within the entire original text. Therefore, further re-encoding is performed at the sentence level to more accurately represent the sentence and, by extension, the original text, by combining both the word and sentence levels.

[0111] Specifically, when encoding the encoding results of each sentence, a sentence encoder comprising a graph neural network layer and a recurrent neural network layer can be determined to extract the grammatical relationships between the sentences and obtain the encoding results of each sentence (based on the output of the word encoder). The grammatical relationships are then encoded using the graph neural network layer of the sentence encoder. Based on the encoding results of each sentence and the encoding results of the grammatical relationships, the sentences are processed using the recurrent neural network layer of the sentence encoder to obtain the re-encoded results of each sentence. For example, for grammatical relationships, the term frequency-inverse document frequency (TF-IDF) features of each sentence are extracted, the similarity of the TF-IDF features between each sentence is calculated, and a similarity matrix is ​​generated based on this to represent the grammatical relationships between the sentences.

[0112] Furthermore, when determining the overall encoding result of the original text, the second attention weight of each sentence in the original text at the current moment can be determined based on the decoding state (specifically, the state of the decoder at the previous moment) and the re-encoding result of each sentence. Based on the second attention weight, the key sentences in the original text can be determined, and based on the key sentences, the overall encoding result of the original text can be determined.

[0113] As can be seen from the above description, this solution employs two levels of attention: the first attention weight at the word level and the second attention weight at the sentence level, respectively, for each keyword selection and, furthermore, for each key sentence selection. This allows for more precise and condensed focus on the original text content when generating target words, while also helping to mitigate any noise from irrelevant content in the original text.

[0114] Furthermore, during the target word generation process, a copying mechanism can be used to directly use content from the original text as the target word. For each word in the original text, the likelihood of its being directly copied as the target word is expressed by a copy probability, which is positively correlated with this probability. If direct copying is not possible and a new word needs to be summarized and condensed, the target word to be generated can be determined from a preset word list. This preset word list contains words that are not present in the original text.

[0115] When determining the target word, the aforementioned first attention weight for the word and the second attention weight for the sentence are used to determine the probability of copying the original text (for copying words from the original text as generated words). Furthermore, based on this overall representation, the lexical distribution of the preset vocabulary is determined (for selecting words from the preset vocabulary as generated words, which may include words not included in the original text). The target word can then be predicted based on the copy probability and lexical distribution. For example, the target word can be determined by determining the larger or weighted difference between the copy probability and lexical distribution.

[0116] In one or more embodiments of the present specification, each time a target word is generated (key information such as a title is composed of multiple target words, and one target word is generated each time), it is necessary to dynamically determine the keywords in the original text, and then represent the sentence based on the keywords this time, and then dynamically determine the key sentence this time, and then represent the original text as a whole based on the key sentence represented this time. This solution can timely adjust the focus on different parts and different levels of content in the original text during the dynamic process of generating key information word by word, improve the accuracy of the determined target words, reduce the omission of target words, and thus improve the accuracy of the generated key information. In addition, through the combination of syntactic and grammatical relationships, the content of the original text is encoded differently at the two granularity levels of words and sentences, which can more accurately express the information of different granularities contained in the original text, which is conducive to enhancing the reliability of title generation.

[0117] In combination with the above description, one or more embodiments of this specification provide an application scenario in which: Figure 1 A detailed framework and process diagram of the method in the present invention is provided. The framework and process use some of the above optional solutions. The process is as follows: Figure 2 shown.

[0118] exist Figure 2 In , a hierarchical encoder consisting of a word encoder and a sentence encoder, and a corresponding decoder are used. Indicates the order of words in the original text. Represents the target words extracted sequentially from the original text.

[0119] Construct embedding vectors for the words in the original text in order to learn the specific encoding of the words. It should be noted that in addition to general embedding information, since words of different parts of speech may play different roles in a sentence, the part of speech of the word can also be used as embedding information to obtain an embedding vector that embeds it with more information, as shown in Formula 1.

[0120] ; (Formula 1)

[0121] in, Indicates time , Indicates the words, represents the embedding operation, Represents general embedded information, Indicates the part of speech of the corresponding word.

[0122] In addition, in order to obtain the intrinsic connection between different words in a sentence, GCN can be used to encode each word in each sentence separately and extract the syntactic relations (for example, represented as an adjacency matrix ) as the edge between any two words. For example, syntactic relations can be extracted by tools such as spacy. If there is a syntactic relation, the adjacency matrix The edge value between two words can be set to 1, otherwise it is set to 0. The calculation process is shown in Formula 2 and Formula 3.

[0123] ; (Formula 2)

[0124] ; (Formula 3)

[0125] in, for The hidden state of the output, for In terms of and are the same, but it should be noted that Output It will also be the next one Input, and so on, we finally get The encoding result of the output word can be regarded as the corresponding hidden state at the time of output .

[0126] For the sentence encoder, it continues to process after the word encoder. The structure is similar, but the input of the GRU is different. It is the weighted sum of the encoding results of the selected keywords in the corresponding sentence. (It can be regarded as the encoding of the sentence this time). GCN encodes grammatical relations rather than syntactic relations. For example, based on the tf-idf feature to represent the grammatical relationship, the cosine similarity between the tf-idf of any two sentences is calculated to construct a similarity matrix, which can be expressed as an adjacency matrix. , handed over to GCN for encoding, The value in can be set to 0 or 1 according to the preset similarity threshold.

[0127] Determined based on attention weights , that is, which words are selected as keywords to represent the sentence based on the different attention to the words. Specifically, at the decoding time , using a global attention layer for words to obtain sentences This encoding.

[0128] Use sentences The corresponding hidden states in the word encoder and the decoder's hidden state To calculate the similarity score , and then use the softmax function to calculate the attention weight Based on this, we can get the first attention weight mentioned above. The calculation process is shown in formulas 4, 5, and 6.

[0129] ; (Formula 4)

[0130] ; (Formula 5)

[0131] ; (Formula 6)

[0132] in, 、 、 、 are parameters that can be learned, such as scale parameters, weight parameters, bias parameters, etc.

[0133] The weighted sum of each sentence Through the above formula 2 and formula 3, GRU and GCN process the output vector Can be used as a sentence This encoding.

[0134] Similar to the above formula, another attention layer for the sentence is used to calculate the context vector (It should be noted that the context vector can generally also be expressed as characters To represent it, here we use the character The main purpose of expressing this is to show that the context vector is based on the state of the encoder, which can usually be expressed as Representation), thereby obtaining the overall representation of the original text. The specific calculation process and It is similar. You can refer to the above formula 2 to formula 6 and adjust the corresponding parameters. For example, you can Adjust to the encoding result of the sentence based on keywords, Adjust to the re-encoding result of the sentence, Adjust to ,Will (first attention weight) is adjusted to the second attention weight, and so on.

[0135] The context vector can be used to calculate the generation probability for the preset vocabulary using the following formulas 7 and 8: (used to decide whether to generate from a preset vocabulary or copy from the original text) and vocabulary distribution .

[0136] ; (Formula 7)

[0137] ; (Formula 8)

[0138] in, 、 、 、 are the parameters that can be learned, Indicates The target word generated at each moment.

[0139] In addition, based on the above-mentioned hierarchical processing and dynamic selection ideas, this solution proposes a hierarchical copying method, which specifically multiplies the attention weight of the word granularity in the same sentence and the attention weight of the sentence granularity to calculate the probability of copying the word in the original text, as shown in Formula 9. Formula 10 calculates the decoding time in the case of considering the existence of the preset word list. Time Words The final generation probability of .

[0140] ; (Formula 9)

[0141] . (Formula 10)

[0142] Based on the same idea, one or more embodiments of this specification also provide devices and apparatuses corresponding to the above methods, such as Figure 3 、 Figure 4 shown.

[0143] Figure 3 A schematic diagram of the structure of a key information generation device provided in one or more embodiments of this specification, the device comprising:

[0144] A word encoding module 302 performs a word encoding step, encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0145] The sentence encoding module 304 executes a sentence encoding step, which encodes the encoding results of each sentence and determines the overall encoding result of the original text;

[0146] The decoding module 306 performs a decoding step, which decodes the encoding result of the original text and determines the target word;

[0147] A key information generating module 308 generates key information of the original text by using the target words generated by performing the word encoding step, the sentence encoding step, and the decoding step multiple times;

[0148] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0149] Figure 4 A schematic diagram of a key information generation device provided in one or more embodiments of this specification, the device comprising:

[0150] at least one processor; and,

[0151] a memory communicatively connected to the at least one processor; wherein,

[0152] The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0153] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0154] a sentence encoding step of encoding the encoding results of each of the sentences and determining the overall encoding result of the original text;

[0155] A decoding step of decoding the encoding result of the original text and determining the target word;

[0156] a key information generating step, which generates key information of the original text by using the plurality of target words generated by performing the word encoding step, the sentence encoding step, and the decoding step multiple times;

[0157] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0158] Based on the same idea, one or more embodiments of this specification further provide a non-volatile computer storage medium corresponding to the above method, storing computer-executable instructions, wherein the computer-executable instructions are configured to:

[0159] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0160] a sentence encoding step of encoding the encoding results of each of the sentences and determining the overall encoding result of the original text;

[0161] A decoding step of decoding the encoding result of the original text and determining the target word;

[0162] a key information generating step, which generates key information of the original text by using the plurality of target words generated by performing the word encoding step, the sentence encoding step, and the decoding step multiple times;

[0163] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0164] Figure 5 This is a flowchart of a live content processing method provided in one or more embodiments of this specification. This method can be applied to various business areas, such as internet finance, e-commerce, instant messaging, gaming, and government affairs. The process can be executed by computing devices in the corresponding fields, and certain input parameters or intermediate results in the process can be manually adjusted to help improve accuracy.

[0165] Figure 5 The process in may include the following steps:

[0166] S502: Acquire the live video of the anchor on the live broadcast platform.

[0167] S504: Perform speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences.

[0168] For live videos on live platforms, in order to facilitate users to playback and view, they can be sliced ​​into several small video segments, and speech recognition is performed to obtain the live broadcast segment text (i.e. Figure 1 (original text in ).

[0169] S506: a word encoding step, which encodes each word in each of the sentences and determines the encoding result of each of the sentences.

[0170] S508: Sentence encoding step, which encodes the encoding results of each sentence and determines the overall encoding result of the live broadcast segment text.

[0171] S510: A decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word.

[0172] S512: A title generation step, which uses the multiple target words generated by multiple executions of the word encoding step, the sentence encoding step and the decoding step to generate a title as the segment title of the live video segment corresponding to the live segment text, and is used to be published in conjunction with the corresponding live video segment so that users can select segments for playback; wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0173] The segment titles of the live video segments are generated according to the obtained target words, so that users can directly understand the content of the corresponding live video segments through the segment titles, and can conveniently and quickly replay and view the live video segments they are interested in.

[0174] In one or more embodiments of this specification, based on changes in the state of title generation, the focus on different content within the live broadcast segment text is adjusted in a timely manner, improving the accuracy of identified key sentences and reducing the omission of key information, thereby improving the accuracy of the generated title. Furthermore, by combining syntactic and grammatical relationships, the content of the original text is encoded differently at both the word and sentence granularity levels, enabling more accurate representation of information at different granularities contained in the original text, thereby enhancing the reliability of title generation.

[0175] It should be noted that Figure 5 The scheme shown and Figure 1 The scheme shown is similar in nature, Figure 5 For parts not described in detail in the steps, please refer to Figure 1 The relevant description will not be repeated in this manual.

[0176] Figure 1 In addition to being applied to the above-mentioned live content processing, the solution shown can also be applied to information processing in various scenarios, and this specification does not limit this.

[0177] Based on the same idea, one or more embodiments of this specification also provide devices and apparatuses corresponding to the above methods, such as Figure 6 、 Figure 7 shown.

[0178] Figure 6 A schematic structural diagram of a live content processing device based on dynamic selection provided in one or more embodiments of this specification, the device comprising:

[0179] Video acquisition module 602, acquires the live video of the anchor on the live broadcast platform;

[0180] The slicing processing module 604 performs speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences;

[0181] A word encoding module 606 performs a word encoding step, encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0182] The sentence encoding module 608 executes a sentence encoding step, which encodes the encoding results of each sentence and determines the overall encoding result of the live broadcast segment text;

[0183] The decoding module 610 performs a decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word;

[0184] A title generation module 612 generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step. The title is used as a segment title of the live video segment corresponding to the live segment text, and is used to be published in conjunction with the corresponding live video segment so that users can select and playback the segment.

[0185] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0186] Figure 7 A schematic structural diagram of a live content processing device based on dynamic selection provided in one or more embodiments of this specification, the device comprising:

[0187] at least one processor; and,

[0188] a memory communicatively connected to the at least one processor; wherein,

[0189] The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0190] Get the live video of the anchor on the live broadcast platform;

[0191] Performing speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences;

[0192] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0193] a sentence encoding step, encoding the encoding results of each of the sentences and determining the overall encoding result of the live broadcast segment text;

[0194] A decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word;

[0195] a title generation step, which generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step, and serves as a title for the live video segment corresponding to the live segment text, for publishing in conjunction with the corresponding live video segment so that users can select and replay the segment;

[0196] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0197] Based on the same idea, one or more embodiments of this specification further provide a non-volatile computer storage medium corresponding to the above method, storing computer-executable instructions, wherein the computer-executable instructions are configured to:

[0198] Get the live video of the anchor on the live broadcast platform;

[0199] Performing speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences;

[0200] a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences;

[0201] a sentence encoding step, encoding the encoding results of each of the sentences and determining the overall encoding result of the live broadcast segment text;

[0202] A decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word;

[0203] a title generation step, which generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step, and serves as a title for the live video segment corresponding to the live segment text, for publishing in conjunction with the corresponding live video segment so that users can select and replay the segment;

[0204] Wherein, each time the word encoding step is executed, the keywords in each of the sentences are determined according to the decoding status, and the encoding result of the sentence is determined based on the keywords in each of the sentences.

[0205] In the 1990s, technological improvements could be clearly distinguished as either hardware improvements (for example, improvements to circuit structures like diodes, transistors, and switches) or software improvements (improvements to process flows). However, with the advancement of technology, many process flow improvements can now be considered direct improvements to hardware circuit structures. Designers almost always create the corresponding hardware circuit structure by programming the improved process flow into the hardware circuit. Therefore, it cannot be said that a process flow improvement cannot be implemented using physical hardware modules. For example, a programmable logic device (PLD), such as a field programmable gate array (FPGA), is an integrated circuit whose logical function is determined by user programming. Designers can "integrate" a digital system on a PLD by programming it themselves, without having to hire a chip manufacturer to design and manufacture a dedicated integrated circuit chip. Moreover, nowadays, instead of manually fabricating integrated circuit chips, this programming is mostly performed using software called a "logic compiler." This is similar to the software compilers used during program development. Before compilation, the original code must be written in a specific programming language, called a Hardware Description Language (HDL). There are many types of HDL, including ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art will also understand that simply by programming a method flow in one of these hardware description languages ​​and then programming it into an integrated circuit, a hardware circuit that implements the logic method flow can be easily obtained.

[0206] The controller can be implemented in any suitable manner. For example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the memory control logic. Those skilled in the art will also appreciate that, in addition to implementing the controller purely in computer-readable program code, the controller can also be implemented in the form of logic gates, switches, an application-specific integrated circuit, a programmable logic controller, an embedded microcontroller, etc. by logically programming the method steps. Therefore, such a controller can be considered a hardware component, and the means for implementing the various functions included therein can also be considered as structures within the hardware component. Alternatively, the means for implementing the various functions can be considered both a software module implementing the method and a structure within the hardware component.

[0207] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0208] For the convenience of description, the above devices are described as being divided into various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0209] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Thus, the embodiments of this specification may take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware. Furthermore, the embodiments of this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0210] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0211] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0212] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0213] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0214] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0215] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0216] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0217] This specification may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0218] The various embodiments in this specification are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from the other embodiments. In particular, the device, apparatus, and non-volatile computer storage medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simplified. For relevant details, refer to the descriptions of the method embodiments.

[0219] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0220] The foregoing description is merely one or more embodiments of this specification and is not intended to limit this specification. It will be apparent to those skilled in the art that various modifications and variations may be made to one or more embodiments of this specification. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of one or more embodiments of this specification are intended to be within the scope of the claims of this specification.

Claims

1. A key information generation method for generating key information from an original text comprising a plurality of sentences, the method comprising: a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences; a sentence encoding step of encoding the encoding results of each of the sentences and determining the overall encoding result of the original text; A decoding step of decoding the encoding result of the original text and determining the target word; a key information generating step, which generates key information of the original text by using the plurality of target words generated by performing the word encoding step, the sentence encoding step, and the decoding step multiple times; wherein, each time the word encoding step is performed, keywords in each of the sentences are determined according to the decoding state, and an encoding result of the sentence is determined based on the keywords in each of the sentences; The encoding of each word in each sentence specifically includes: Determine a word encoder including a graph neural network layer and a recurrent neural network layer; extract the syntactic relationship between each word in each sentence, and obtain the embedding vector of each word; encode the syntactic relationship through the graph neural network layer of the word encoder; and process the encoding results of the embedding vector and the syntactic relationship through the recurrent neural network layer of the word encoder to obtain the encoding results of each word.

2. The method according to claim 1, wherein extracting the syntactic relationship between the words in each sentence and obtaining the embedding vector of each word specifically comprises: Parsing each word in each of the sentences to obtain the part of speech and the syntactic relationship between the words; An embedding vector is generated for the corresponding word according to the part of speech, and an adjacency matrix for encoding is generated to represent the syntactic relationship.

3. The method of claim 1 , wherein the decoding step is performed by a decoder; Determining the keywords in each of the sentences according to the decoding status specifically includes: Determining a similarity score based on the state of the word encoder and the state of the decoder at a previous moment; Determining, based on the similarity score, a first attention weight of the words in each of the sentences at a current moment; Determine keywords in each sentence according to the first attention weight.

4. The method according to claim 1, wherein encoding the encoding results of each sentence comprises: Determine a sentence encoder that includes a graph neural network layer and a recurrent neural network layer; Extracting grammatical relations between the sentences and obtaining encoding results of the sentences; Encoding the grammatical relationship through a graph neural network layer of the sentence encoder; According to the encoding results of each of the sentences and the encoding results of the grammatical relations, the sentences are processed through the recurrent neural network layer of the sentence encoder to obtain the re-encoding results of each of the sentences.

5. The method according to claim 4, wherein extracting the grammatical relationship between the sentences comprises: Extracting tf-idf features of each of the sentences; The similarity of the tf-idf features between the sentences is calculated respectively, and a similarity matrix is ​​generated accordingly to represent the grammatical relationship between the sentences.

6. The method according to claim 4, wherein determining the overall encoding result of the original text specifically comprises: determining a second attention weight of each sentence in the original text at a current moment according to a decoding state and the re-encoding result of each sentence; Determining key sentences in the original text according to the second attention weight; An overall encoding result of the original text is determined based on the key sentence.

7. The method according to claim 1, wherein decoding the encoding result of the original text and determining the target word specifically comprises: Determining a generation probability and vocabulary distribution for a preset vocabulary list based on the encoding result of the original text, the decoding state at the current moment, and the target word generated at the previous moment; Determining the copy probability of the words in the original text according to the keyword determination process; A target word at a current moment is determined according to the generation probability, the vocabulary distribution, and the copy probability.

8. The method according to claim 7, wherein determining the probability of duplication of a word in the original text according to the process of determining the keyword specifically comprises: The copy probability of the words in the original text is determined based on the first attention weight for the words used to determine the keywords and the second attention weight for the sentences used to determine the overall encoding result of the original text.

9. A method for processing live content, comprising: Get the live video of the anchor on the live broadcast platform; Performing speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences; a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences; a sentence encoding step, encoding the encoding results of each of the sentences and determining the overall encoding result of the live broadcast segment text; A decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word; a title generation step, which generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step, and serves as a title for the live video segment corresponding to the live segment text, for publishing in conjunction with the corresponding live video segment so that users can select and replay the segment; wherein, each time the word encoding step is performed, keywords in each of the sentences are determined according to the decoding state, and an encoding result of the sentence is determined based on the keywords in each of the sentences; The encoding of each word in each sentence specifically includes: Determine a word encoder including a graph neural network layer and a recurrent neural network layer; extract the syntactic relationship between each word in each sentence, and obtain the embedding vector of each word; encode the syntactic relationship through the graph neural network layer of the word encoder; and process the encoding results of the embedding vector and the syntactic relationship through the recurrent neural network layer of the word encoder to obtain the encoding results of each word.

10. A key information generating device for generating key information for an original text comprising a plurality of sentences, the device comprising: a word encoding module, executing a word encoding step, encoding each word in each of the sentences and determining an encoding result for each of the sentences; A sentence encoding module, which performs a sentence encoding step, encodes the encoding result of each sentence and determines the overall encoding result of the original text; A decoding module performs a decoding step, which decodes the encoding result of the original text and determines the target word; a key information generating module, which generates key information of the original text by using the plurality of target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step; wherein, each time the word encoding step is performed, keywords in each of the sentences are determined according to the decoding state, and an encoding result of the sentence is determined based on the keywords in each of the sentences; The encoding of each word in each sentence specifically includes: Determine a word encoder including a graph neural network layer and a recurrent neural network layer; extract the syntactic relationship between each word in each sentence, and obtain the embedding vector of each word; encode the syntactic relationship through the graph neural network layer of the word encoder; and process the encoding results of the embedding vector and the syntactic relationship through the recurrent neural network layer of the word encoder to obtain the encoding results of each word.

11. A live content processing device, comprising: Video acquisition module, which obtains the live video of the anchor on the live broadcast platform; A slicing processing module performs speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences; a word encoding module, executing a word encoding step, encoding each word in each of the sentences and determining an encoding result for each of the sentences; A sentence encoding module executes a sentence encoding step, which encodes the encoding result of each sentence and determines the overall encoding result of the live broadcast segment text; A decoding module executes a decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word; a title generation module, which generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step, and serves as a segment title of the live video segment corresponding to the live segment text, for publishing in conjunction with the corresponding live video segment so that users can select a segment for playback; wherein, each time the word encoding step is performed, keywords in each of the sentences are determined according to the decoding state, and an encoding result of the sentence is determined based on the keywords in each of the sentences; The encoding of each word in each sentence specifically includes: Determine a word encoder including a graph neural network layer and a recurrent neural network layer; extract the syntactic relationship between each word in each sentence, and obtain the embedding vector of each word; encode the syntactic relationship through the graph neural network layer of the word encoder; and process the encoding results of the embedding vector and the syntactic relationship through the recurrent neural network layer of the word encoder to obtain the encoding results of each word.

12. A key information generating device for generating key information for an original text comprising a plurality of sentences, the device comprising: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences; a sentence encoding step of encoding the encoding results of each of the sentences and determining the overall encoding result of the original text; A decoding step of decoding the encoding result of the original text and determining the target word; a key information generating step, which generates key information of the original text by using the plurality of target words generated by performing the word encoding step, the sentence encoding step, and the decoding step multiple times; wherein, each time the word encoding step is performed, keywords in each of the sentences are determined according to the decoding state, and an encoding result of the sentence is determined based on the keywords in each of the sentences; The encoding of each word in each sentence specifically includes: Determine a word encoder including a graph neural network layer and a recurrent neural network layer; extract the syntactic relationship between each word in each sentence, and obtain the embedding vector of each word; encode the syntactic relationship through the graph neural network layer of the word encoder; and process the encoding results of the embedding vector and the syntactic relationship through the recurrent neural network layer of the word encoder to obtain the encoding results of each word.

13. A live content processing device, comprising: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: Get the live video of the anchor on the live broadcast platform; Performing speech recognition and slicing processing on the live video to obtain a live segment text including multiple sentences; a word encoding step of encoding each word in each of the sentences and determining an encoding result for each of the sentences; a sentence encoding step, encoding the encoding results of each of the sentences and determining the overall encoding result of the live broadcast segment text; A decoding step, which decodes the encoding result of the live broadcast segment text and determines the target word; a title generation step, which generates a title using the target words generated by repeatedly executing the word encoding step, the sentence encoding step, and the decoding step, and serves as a title for the live video segment corresponding to the live segment text, for publishing in conjunction with the corresponding live video segment so that users can select and replay the segment; wherein, each time the word encoding step is performed, keywords in each of the sentences are determined according to the decoding state, and an encoding result of the sentence is determined based on the keywords in each of the sentences; The encoding of each word in each sentence specifically includes: Determine a word encoder including a graph neural network layer and a recurrent neural network layer; extract the syntactic relationship between each word in each sentence, and obtain the embedding vector of each word; encode the syntactic relationship through the graph neural network layer of the word encoder; and process the encoding results of the embedding vector and the syntactic relationship through the recurrent neural network layer of the word encoder to obtain the encoding results of each word.

Images