Abstract generation system and method based on automated machine learning techniques

Abstract

This document describes a summary generation system and method based on automated machine learning technology. Specifically, it relates to a system and method for automatically generating text summaries using automated machine learning techniques. To address the problem that current summary generation methods cannot simultaneously ensure both effectiveness and timeliness, nor can they automatically select effective summarization methods based on various text features, resulting in inconsistent summary quality, the system comprises an acquisition module, a generation module, and an output module. The generation module includes sentence segmentation, text cleaning, word embedding, sentence embedding, and summary sentence evaluation modules. The method first acquires a text set and its corresponding summaries; then, it establishes an automated machine learning model, inputs the text set into the model, trains it using automated machine learning techniques, and outputs a summary for each text, resulting in a trained model; finally, it inputs the text to be summarized into the trained automated machine learning model, and outputs a summary of that text. This falls under the field of summary generation.

Description

Technical Field

[0001] This invention relates to a summary generation system and method, specifically to a system and method for automatically generating text summaries using automated machine learning technology, belonging to the field of summary generation. Background Technology

[0002] Automatic summary generation technology involves a system that, given a relevant text set or text, automatically generates a summary of the text set or text based on a certain algorithm; that is, the main content of the text set or text. Common automatic summary generation methods include rule-based automatic summary generation methods, graph model-based automatic summary generation methods, and redundancy avoidance-based automatic summary generation methods. Although there are many existing summarization methods, each with its own advantages and disadvantages, no single method can effectively balance the effectiveness and timeliness of the summary, nor can it automatically select the best summarization method based on the various features and attributes of the text set or text. This results in inconsistent quality of existing summaries, and in practical applications, there are no mobile apps or software applications that solve the summary generation problem. Summary of the Invention

[0003] To address the problem that summary generation methods cannot simultaneously ensure the effectiveness and timeliness of summaries, nor can they automatically select effective summarization methods based on various text features, resulting in inconsistent summary quality, this invention proposes a summary generation system and method based on automatic machine learning technology.

[0004] The technical solution adopted in this invention is:

[0005] It consists of an acquisition module, a generation module, and an output module, in that order.

[0006] The acquisition module is used to acquire text; it is also used to receive the output summary signal from the output module.

[0007] A generation module is used to generate a summary of the text;

[0008] The generation module includes, in sequence, a sentence segmentation module, a cleaning module, a word embedding module, a sentence embedding module, and a summary sentence evaluation module;

[0009] The sentence segmentation module is used to split the content of text into multiple sentences;

[0010] The cleaning module is used to clean each sentence;

[0011] The cleaning process includes removing empty sentences, converting languages ​​other than Chinese and English to spaces, converting all English letters to lowercase, and removing stop words.

[0012] The word embedding module is used to perform word embedding processing on the text after sentence cleaning;

[0013] The sentence embedding module is used to perform sentence embedding on the text after word embedding.

[0014] The summary sentence evaluation module is used to evaluate the text after sentence embedding processing, and to determine whether each sentence in the text can be used as a summary sentence; it is also used to integrate the sentences that the summary sentence evaluation module has determined to be summary sentences, and sort them according to their positions in the original text to form a summary.

[0015] The output module is used to output the summary obtained by the summary evaluation module; it is also used to send the signal of the output summary to the acquisition module.

[0016] The output module is connected to the summary evaluation module of the generation module.

[0017] Preferably, the acquisition module acquires one or more texts simultaneously;

[0018] When the acquisition module receives a text, it directly sends the text to the generation module.

[0019] When the acquisition module receives multiple texts, it first sends one text to the generation module, which then sends the summary result to the output module. The output module outputs the summary result and simultaneously sends a signal indicating that the summary result has been output to the acquisition module. After receiving the signal, the acquisition module sends the next text to the generation module, and repeats the above steps until the last text is output as a summary.

[0020] Preferably, the word embedding module includes a word frequency unit, a normalized word frequency unit, and a word vector unit;

[0021] The sentence embedding module includes a summation unit, a sentence embedding unit based on the KeyCluster algorithm, a sentence embedding unit based on the one-hot method, and an average smoothing processing unit;

[0022] The summary sentence evaluation module includes evaluation units based on proportion, evaluation units based on mean and standard deviation, evaluation units based on the MMR algorithm, and evaluation units based on the TextRank algorithm.

[0023] Preferably, the generation module further includes an adapter module, which is used to combine the units selected in the word embedding module, sentence embedding module, and summary sentence evaluation module, and the adapter module is connected to the generation module.

[0024] Preferably, when the generation module is applied to a text, the specific process is as follows:

[0025] ① Input the acquired text into the sentence segmentation module to obtain the processed text;

[0026] ② Input the sentence-segmented text into the cleaning module to obtain the cleaned text;

[0027] ③ Select one unit from each of the word embedding module, sentence embedding module, and summary sentence evaluation module in sequence. Combine the selected units in the order of selection to form a triple. Obtain multiple sets of triples in the above manner. Use automatic machine learning technology to select one or more sets of triples from the multiple sets of triples according to the selection conditions. Use the adapter module to connect the three units in each set of triples to obtain the connected triples. Use the connected triples to process the cleaned text to obtain the corresponding summary of the text.

[0028] ④ Input the summary corresponding to the obtained text into the output module and output the summary.

[0029] Preferably, it also includes a terminal application module and a client module, both of which are connected to the acquisition module and the output module;

[0030] The terminal application module is used to send the text input by the user to be summarized to the acquisition module, and also to receive the summary output by the output module and provide feedback to the user.

[0031] The client module is used to send the text input by the user and the text to be digested to the acquisition module, and also to receive the digest output by the output module and feed it back to the user;

[0032] The terminal application module includes mobile APP software and mini-programs.

[0033] The summary generation method based on automated machine learning technology includes the following steps:

[0034] S1. Obtain the text set and the summary corresponding to each text;

[0035] S2. Establish an automatic machine learning model. Input the text set and the summary corresponding to each text into the automatic machine learning model and train it using automatic machine learning technology. Output the summary of each text to obtain the trained model.

[0036] S3. Input the text to be summarized into the trained automatic machine learning model, and output the summary of the text.

[0037] Preferably, the text set in S1 includes one or more texts;

[0038] When the text set is a single text, the text is directly input into the automatic machine learning model, which outputs a summary of the text.

[0039] When the text set consists of multiple texts, a certain text is input into the model, and the corresponding summary of the text is output. Then, the next text is input into the model, and the corresponding summary is output. This process continues until the last text is output as its corresponding summary.

[0040] Preferably, it also includes an adapter.

[0041] Preferably, in step S2, an automatic machine learning model is established. The text set is input into the automatic machine learning model for training using automatic machine learning techniques, and a summary of each text is output to obtain the trained model. The specific process is as follows:

[0042] S21. Break down the content of a text into multiple sentences;

[0043] S22. Clean each sentence;

[0044] The cleaning process includes removing empty sentences, converting languages ​​other than Chinese and English to spaces, converting all English letters to lowercase, and removing stop words.

[0045] S23. Select one method from word embedding, sentence embedding, and summary sentence evaluation methods in sequence. Combine the three methods in the order of selection to form a triple. Obtain multiple triples in the above manner. Use automatic machine learning technology to select one or more triples from the multiple triples according to the selection criteria. Use an adapter to connect the three methods in each selected triple to obtain a connected triple. Use the connected triples to process the cleaned text to obtain the summary corresponding to the text.

[0046] The word embedding includes a word frequency representation method, a normalized word frequency representation method, and a word vector representation method;

[0047] The sentence embedding methods include a summation method, a sentence embedding method based on the KeyCluster algorithm, a sentence embedding method based on the one-hot algorithm, and an average smoothing method;

[0048] The summary evaluation methods include a proportion-based summary evaluation method, a mean and standard deviation-based summary evaluation method, an MMR algorithm-based evaluation method, and a TextRank algorithm-based evaluation method.

[0049] Beneficial effects:

[0050] This invention establishes a summary generation system based on automated machine learning technology. The system includes an acquisition module, a generation module, and an output module. First, the acquisition module acquires text and sends it to the generation module. The generation module sequentially performs sentence segmentation, cleaning, word embedding, sentence embedding, and summary sentence evaluation on the text to obtain a summary. The generated summary is then sent to the output module, completing the summary generation. The word embedding module includes a word frequency unit, a normalized word frequency unit, and a word vector unit. The sentence embedding module includes a summation unit, a sentence embedding unit based on the KeyCluster algorithm, a sentence embedding unit based on the one-hot encoding method, and an average smoothing unit. The summary sentence evaluation module includes evaluation units based on proportion, evaluation units based on mean and standard deviation, evaluation units based on the MMR algorithm, and evaluation units based on the TextRank algorithm.

[0051] In application, the acquired text is processed by the sentence segmentation and cleaning modules. Then, one unit is selected from each of the word embedding, sentence embedding, and summary evaluation modules to represent the corresponding module. Three selected units are grouped into a triplet. Multiple triplets are obtained in this way. Automatic machine learning technology is used to select one or more superior triplets from these multiple triplets according to selection criteria. An adapter is then used to combine each selected triplet, resulting in a combination algorithm for each triplet. This leads to multiple combination algorithms. The selection criteria include random sampling summarization algorithms, summarization algorithms recommended by machine learning models after extracting text features, and default summarization algorithms. Each combination algorithm after adapter processing generates a corresponding summary for each text, meaning one text corresponds to multiple summaries. When the acquisition module receives a text, it directly sends the text to the generation module to continue subsequent operations. When the acquisition module receives multiple texts, it iteratively generates summaries for each text until the last text outputs a summary.

[0052] This invention divides traditional summarizing algorithms into stages, with multiple replaceable components (methods or units) set up in each stage. One component (method or unit) is selected from each stage, and these selected components are combined to form triples. Multiple sets of triples are obtained in this way. Then, automatic machine learning technology is used to select one or more superior triples from these sets based on selection criteria. An adapter is then used to combine each selected triple, resulting in a combination algorithm for each triple. This yields multiple combination algorithms. Applying automatic machine learning technology to the problem of automatic summarizing allows for the identification of various feature attributes for different texts, enabling the selection of effective summarizing algorithms for different texts. By using automatic machine learning technology to select one or more effective combination algorithms from all triples and then combining them using an adapter, multiple summarizing algorithms can be obtained. For a single text, using these multiple summarizing algorithms will yield multiple summarizing results. These results not only consider the effectiveness and timeliness of the summarizing but also allow users to choose the desired summarizing. Attached Figure Description

[0053] Figure 1 This is a structural diagram of the abstract generation system;

[0054] Figure 2 This is a structural diagram of the generation module;

[0055] Figure 3 This is a flowchart of the automatic summary generation method; Detailed Implementation

[0056] Specific implementation method one: Combining Figures 1-2 This embodiment describes a summary generation system based on automatic machine learning technology, which includes, in sequence, an acquisition module, a generation module, and an output module.

[0057] The acquisition module is used to acquire text; it is also used to receive the output summary signal from the output module.

[0058] A generation module is used to generate a summary of the text;

[0059] The generation module includes, in sequence, a sentence segmentation module, a cleaning module, a word embedding module, a sentence embedding module, and a summary sentence evaluation module;

[0060] The sentence segmentation module is used to split the content of text into multiple sentences;

[0061] The cleaning module is used to clean each sentence;

[0062] The cleaning process includes removing empty sentences, converting languages ​​other than Chinese and English to spaces, converting all English letters to lowercase, and removing stop words.

[0063] An empty sentence is a sentence with a length of 0. First, the sentences are segmented into words. Then, based on the specific context of the sentences in the text, different stop word lists are used to remove stop words from each sentence. These stop word lists are sourced from publicly available sources such as GitHub and various labs. After removing stop words, sentences that were originally different may end up having the same meaning.

[0064] The word embedding module is used to perform word embedding processing on the text after sentence cleaning;

[0065] The sentence embedding module is used to perform sentence embedding on the text after word embedding.

[0066] The summary sentence evaluation module is used to evaluate the text after sentence embedding processing, and to determine whether each sentence in the text can be used as a summary sentence; it is also used to integrate the sentences that the summary sentence evaluation module has determined to be summary sentences, and sort them according to their positions in the original text to form a summary.

[0067] The output module is used to output the summary obtained by the summary evaluation module; it is also used to send the signal of the output summary to the acquisition module.

[0068] The output module is connected to the summary evaluation module of the generation module.

[0069] The resulting abstract takes into account both validity and timeliness.

[0070] Specific Implementation Method Two: Combining Figures 1-2 This embodiment describes a summary generation system based on automatic machine learning technology, wherein the acquisition module simultaneously acquires one or more texts.

[0071] When the acquisition module receives a text, it directly sends the text to the generation module.

[0072] When the acquisition module receives multiple texts (two or more), it first sends one text to the generation module, which then sends a summary result to the output module. The output module outputs the summary result and simultaneously sends a signal indicating that a summary result has been output to the acquisition module. Upon receiving this signal, the acquisition module sends the next text to the generation module, repeating the above steps until a summary is output for the last text. In other words, summarizing multiple texts is achieved through an iterative process.

[0073] Everything else is the same as in Specific Implementation Method 1.

[0074] Specific implementation method three: Combining Figures 1-2This embodiment describes a summary generation system based on automatic machine learning technology.

[0075] The word embedding module includes word frequency units, normalized word frequency units, word vector units, etc.

[0076] The term frequency unit is a set of words that are counted to count the frequency of each word in the cleaned text.

[0077] Normalized word frequency units are created by converting the frequency of each word in the cleaned text to a decimal between 0 and 1, forming a set of normalized word frequencies.

[0078] Word vector units are vectors corresponding to each word in the cleaned text, forming a set of word vectors. For English, a 100-dimensional Google GloVe model is used to transform each word into a 100-dimensional word vector. For Chinese, a 50-dimensional Chinese wiki encyclopedia model is used to transform each word into a 50-dimensional word vector.

[0079] The sentence embedding module includes a summation unit, a sentence embedding unit based on the KeyCluster algorithm, a sentence embedding unit based on the one-hot method, and an average smoothing processing unit.

[0080] Summation Unit: If the word embeddings of a sentence are 3, 2, and 1 respectively, then the sum of the word embeddings of the words in the sentence is 3 + 2 + 1 = 6, so the sentence embedding is 6.

[0081] Sentence embedding unit based on KeyCluster algorithm: The sentence embedding part of KeyCluster algorithm is used for calculation.

[0082] Sentence embedding unit based on one-hot encoding: The one-hot encoding method is used for computation.

[0083] Average smoothing unit: The word embeddings of each word in each sentence in the text are averaged and the denominator is smoothed (the smoothing factor is set to 0.001). For example, if the word embeddings of words in a sentence are 3, 2 and 1, then the average word embeddings of the words in the sentence are (3+2+1) / 3=2, and then a smoothing factor is added to the denominator for smoothing.

[0084] The summary evaluation module includes evaluation units based on proportion, evaluation units based on mean and standard deviation, evaluation units based on the MMR algorithm, and evaluation units based on the TextRank algorithm.

[0085] The evaluation unit based on proportion (top): The sentence embeddings of each sentence are sorted from high to low, and the top 20% of sentences are selected as summary sentences.

[0086] The evaluation unit based on mean (avg) and standard deviation (std) is as follows: The mean of the sentence embeddings of all sentences is calculated, and then added to half the standard deviation of the sentence embeddings of all sentences. The result is called the threshold. Sentences with embeddings greater than or equal to the threshold are selected as summary sentences. Sentence embedding transforms a sentence into a vector. If the vector has only one dimension, it can be considered as a number, i.e., score / rating.

[0087] Evaluation unit based on MMR algorithm: Evaluation is performed using the evaluation part of the MMR algorithm.

[0088] Evaluation unit based on TextRank algorithm: Evaluation is performed using the evaluation part of the TextRank algorithm.

[0089] The rest is the same as in specific implementation method one or two.

[0090] Specific implementation method four: Combination Figures 1-2 This embodiment describes a summary generation system based on automatic machine learning technology. The generation module further includes an adapter module, which is used to combine units selected from the word embedding module, sentence embedding module, and summary sentence evaluation module. The adapter module is connected to the generation module.

[0091] The different input constraints required for different units prevent arbitrary combination of units from the word embedding module, sentence embedding module, and summary sentence evaluation module. Therefore, this invention enables an adapter before combining units to satisfy the constraints between units that do not meet them, allowing mismatched units to be combined and ultimately achieving arbitrary combination of units. Other aspects are the same as in specific implementation methods one, two, or three.

[0092] Specific Implementation Method Five: Combining Figures 1-2 This embodiment describes a summary generation system based on automatic machine learning technology. When the generation module is applied to a text, the specific process is as follows:

[0093] ① Input the acquired text into the sentence segmentation module to obtain the processed text;

[0094] ② Input the sentence-segmented text into the cleaning module to obtain the cleaned text;

[0095] ③ Select one unit from each of the word embedding module, sentence embedding module, and summary sentence evaluation module in sequence. Combine the selected units in the order of selection to form a triple. Obtain multiple sets of triples in the above manner. Use automatic machine learning technology to select one or more sets of triples from the multiple sets of triples according to the selection conditions. Use the adapter module to connect the three units in each set of triples to obtain the connected triples. Use the connected triples to process the cleaned text to obtain the corresponding summary of the text.

[0096] The selection criteria include random sampling summarization algorithm, summarization algorithm recommended by machine learning model after extracting text features, and default summarization algorithm;

[0097] ④ Input the summary corresponding to the obtained text into the output module and output the summary.

[0098] The specific process is as follows: For a text, firstly, select one unit from each of the word embedding module, sentence embedding module, and summary sentence evaluation module. Combine the three selected units into a triplet, that is, combine one unit from the word embedding module, one unit from the sentence embedding module, and one unit from the summary sentence evaluation module into a triplet. Multiple triplets can be obtained in this way. Using automatic machine learning technology, the optimal selection is made from the multiple triplets based on one of the following algorithms: random sampling summarization algorithm, summarization algorithm recommended by machine learning model after extracting text features, and default summarization algorithm. Select one or more triplets with better performance as the summarization algorithm for the text. Input the text into each summarization algorithm in turn to obtain the corresponding summary, that is, one text corresponds to multiple summaries.

[0099] The rest is the same as in specific implementation methods one, two, three or four.

[0100] Specific Implementation Method Six: Combination Figures 1-2 This embodiment describes a summary generation system based on automatic machine learning technology, which further includes a terminal application module and a client module, both of which are connected to the acquisition module and the output module.

[0101] The terminal application module is used to send the text input by the user to be summarized to the acquisition module, and also to receive the summary output by the output module and provide feedback to the user.

[0102] The client module is used to send the text input by the user to be summarized to the acquisition module, and also to receive the summary output by the output module and provide feedback to the user.

[0103] The rest is the same as in specific implementation methods one, two, three, four or five.

[0104] Specific implementation method seven: Combining Figures 1-2 This embodiment describes a summary generation system based on automatic machine learning technology. The terminal application module includes a mobile app or mini-program, and the client module includes a webpage or computer application. Other aspects are the same as in specific embodiments one, two, three, four, five, or six.

[0105] Specific implementation method eight: Combination Figure 3 This embodiment describes a summarization method based on automated machine learning technology, which includes the following steps:

[0106] First, based on existing traditional automatic summarization algorithms, we analyze the definition, applicable scope, advantages and disadvantages of each algorithm for sentence importance, identify the similarities, differences, key points or features of each algorithm, and on this basis, explore the applicable scope of sentences and the key to each algorithm.

[0107] Therefore, this invention proposes a stage (each step below) division criterion to divide traditional summarization algorithms into stages. The difference in processing methods in each stage is the key to distinguishing different algorithms. Specifically, it adopts the pipeline idea, which divides the algorithm between input and output into multiple stages. Each stage often has some optional or replacement operations, which are called components. Replacing a component in a certain stage of the algorithm is equivalent to changing the entire algorithm. Therefore, the pipeline idea is to divide the intermediate process from input to output into multiple stages, select a component for each stage, and combine them into a new algorithm. The optimization goal of the pipeline is to minimize the error of the selected component on the text set, or maximize the evaluation index of the selected component on the test set. However, there are actually infinitely many possible components in each stage. Therefore, it is not feasible to try to find the optimal component combination by traversing all possible components for each text. At this point, how to select a suitable component becomes a problem. Therefore, this invention uses automatic machine learning (AutoML) technology to automatically select suitable components for each stage using certain algorithms, so that the text can achieve the best possible summarization result.

[0108] S1. Obtain the text set and the summary corresponding to each text.

[0109] First, the user inputs the text set for which a summary is to be generated and the summary corresponding to each text into the terminal application or client. The terminal application or client then retrieves the text set for which a summary needs to be generated, which includes one or more texts.

[0110] S2. Establish an automatic machine learning model. Input the text set and the summary corresponding to each text into the automatic machine learning model and train it using automatic machine learning technology. Output the summary of each text to obtain the trained model. The specific process is as follows:

[0111] An automated machine learning model is established. Given a text set (news, notifications, text, etc.), this model can quickly (within minutes) select the optimal overall algorithm for each text, accurately extracting the summary sentence for each text and ensuring good results in summary generation. It also achieves the ability to search for good algorithms within a vast summarization algorithm space, solving the problem that traversing the entire search space to find the optimal algorithm is impractical.

[0112] S21. Split the content of a text into multiple sentences.

[0113] The content of each text is automatically split into multiple sentences, abbreviated as SPLIT. Sentences belonging to the same text are grouped together for easier processing of each sentence later. Starting from this step, we will use a specific text as an example to explain the method for automatically generating summaries.

[0114] S22. Clean each sentence;

[0115] The cleaning process includes removing empty sentences, converting languages ​​other than Chinese and English to spaces, converting all English letters to lowercase, and removing stop words.

[0116] This step involves cleaning each sentence in the text, a process abbreviated as CLEAN. The specific cleaning process is as follows: remove sentences of length 0; convert all languages ​​except Chinese and English to spaces; convert all English letters to lowercase; segment sentences based on their specific characteristics; and then use different stop word lists to remove stop words from each sentence. These stop word lists are sourced from publicly available sources such as GitHub and various labs. After removing stop words, sentences that were originally different may become synonymous.

[0117] S23. Select one method from word embedding, sentence embedding, and summary sentence evaluation methods in sequence. Combine the three methods in the order of selection to form a triple. Obtain multiple triples in the above manner. Use automatic machine learning technology to select one or more triples from the multiple triples according to the selection criteria. Use an adapter to connect the three methods in each selected triple to obtain a connected triple. Use the connected triples to process the cleaned text to obtain the summary corresponding to the text.

[0118] The selection criteria include random sampling summarization algorithm, summarization algorithm recommended by machine learning model after extracting text features, and default summarization algorithm;

[0119] Hidden connections exist between components, often resulting in 1+1≠2: some components may be mutually exclusive, leading to 1+1<2 when used together for automatic summarization; others may be mutually compatible, resulting in 1+1>2 when used together for automatic summarization. In fact, these connections are unknown before the components are actually used. Therefore, this invention, based on a machine learning model, attempts to find these hidden connections between components at different stages.

[0120] Word embedding methods include word frequency representation methods, normalized word frequency representation methods, and word vector representation methods.

[0121] Word embedding, also known as word representation, is abbreviated as WORD. Automatic machine learning techniques automatically select word embedding methods to process all sentences after text cleaning, resulting in a set of sentences processed by the word embedding method. This invention includes three word embedding methods, but theoretically and in practice, the number of word embedding methods is unlimited. In application, the best method is selected, and not every word embedding method is used.

[0122] The sentence embedding methods include summation methods, sentence embedding methods based on the KeyCluster algorithm, sentence embedding methods based on the one-hot algorithm, and average smoothing methods.

[0123] Sentence embedding, also known as sentence representation or sentence scoring, is abbreviated as SENTENCE. This invention utilizes automated machine learning techniques to automatically select a sentence embedding method. Then, all sentences processed by the word embedding method obtained in S23 are processed using the sentence embedding method again, resulting in a complete set of sentences with the sentence embedding method. This invention includes four sentence embedding methods, but theoretically and in practice, the number of sentence embedding methods is unlimited. In application, the best method is selected, and not every sentence embedding method is used.

[0124] Summary evaluation methods include proportion-based summary evaluation methods, mean and standard deviation-based summary evaluation methods, MMR algorithm-based evaluation methods, and TextRank algorithm-based evaluation methods.

[0125] The summary sentence assessment is abbreviated as SUMMARIZE. An automated machine learning technique is used to select a summary sentence assessment method. Then, all sentences processed by the sentence embedding method are assessed using the summary sentence assessment method, ultimately yielding the summary of the text. This invention includes four summary sentence assessment methods, but theoretically and in practice, the number of methods is unlimited. In application, the best method is selected, and not every summary sentence assessment method is used.

[0126] This invention transforms the problem of selecting a summarizing algorithm into a search problem. It provides multiple processing methods for text in word embedding, sentence embedding, and summary sentence evaluation methods. Based on different texts, it uses automatic machine learning technology to automatically select the optimal processing method from these methods, forming one or more new summarizing algorithms. For a given text, in step S23, multiple sets of triples are first obtained. Each triple is represented by one method selected from word embedding, sentence embedding, and summary sentence evaluation methods, and these three methods are grouped together (i.e., triples). Multiple sets of triples are obtained using this method. Then, automatic machine learning technology is used to select one or more sets of triples from these sets. An adapter is then used to combine each set of triples, and each resulting set of triples becomes a combination algorithm. Thus, multiple combination algorithms for the text are obtained, resulting in multiple summarizing results for a single text. For example, if the triplet corresponding to a certain algorithm f in the automatic generation method is the word frequency method, the sentence embedding method based on the KeyCluster algorithm, and the evaluation method based on the TextRank algorithm, then it means that after processing by S21 and S22, algorithm f is composed of the word frequency method in the word embedding method, the sentence embedding method based on the KeyCluster algorithm in the sentence embedding method, and the evaluation method based on the TextRank algorithm in the summary sentence evaluation method.

[0127] S3. Input the text to be summarized into the trained automatic machine learning model, and output the summary of the text.

[0128] Specific Implementation Method Nine: Combining Figure 3 This embodiment describes a summarization method based on automated machine learning technology, wherein the text set in S1 includes one or more texts;

[0129] When the text set is a single text, the text is directly input into the automatic machine learning model, which outputs a summary of the text.

[0130] When the text set consists of multiple texts, a text is input into an automated machine learning model, which outputs a summary of that text. Then, the next text is input into the automated machine learning model, which outputs its corresponding summary, and so on, until the last text outputs its summary. The multiple texts are greater than or equal to 2.

[0131] Existing automated machine learning techniques require: labeled and clean datasets (each data point has a labeled category and each data point is cleaned), and sufficient domain knowledge (processing models or model components). Therefore, for each data point, it is necessary to include text features and the optimal combination of components corresponding to that text.

[0132] For example, in this invention, there are 3 components (units or methods) to choose from in the third stage, 4 components (units or methods) to choose from in the fourth stage, and 4 components (units or methods) to choose from in the fifth stage. This results in 1*1*3*4*4 = 48 possible component combinations, corresponding to 48 different summarization algorithms. Each text has 48 possible component combinations, so it's necessary to iterate through these combinations to find the optimal one as the labeled and clean dataset. Assuming only 500 texts are used, and each combination takes 5 seconds to complete, it would take approximately 500*48*5 / 60 / 60≈33 hours to obtain the labeled and clean dataset. This undoubtedly wastes a significant amount of time during the training phase. Therefore, this invention employs parallel processing to effectively save this time. Furthermore, although different traditional summarization algorithms can theoretically be decomposed into different components, they still require manual implementation; generating additional components requires additional manual operation, which is time-consuming and labor-intensive. Therefore, this invention employs automated machine learning technology to modify the hyperparameters within different components (units or methods) to generate new components, thereby obtaining a new algorithm. This algorithm produces more accurate summaries that reflect the text content, resulting in better summaries. Furthermore, when generating summaries from multiple texts simultaneously, an iterative approach is used to sequentially input multiple texts into the automated machine learning model for summary extraction and generation, outputting the summary for each text in turn. For different texts, various feature attributes are identified, allowing for the selection of a better summarization algorithm for different texts.

[0133] Everything else is the same as in Specific Implementation Method 8.

[0134] Specific Implementation Method Ten: Combining Figure 3 This embodiment describes a summary generation method based on automatic machine learning technology, which also includes an adapter.

[0135] To address the issue that different methods in word embedding, sentence embedding, and summary sentence evaluation require different input constraints, making it impossible to arbitrarily combine the methods from these three stages, this invention enables an adapter to satisfy the constraints between methods that do not meet them before combining the methods. This allows mismatched methods to be combined, ultimately achieving arbitrary combination of methods.

[0136] The rest is the same as in specific implementation method eight or nine.

[0137] Detailed Implementation Method Eleven: Combining Figures 1-2 This embodiment describes a method for using a summary generation system based on automated machine learning technology, which includes the following steps:

[0138] Step 1: The user inputs the text to be summarized into the terminal application module or client module. The terminal application module or client module sends the received text to the acquisition module, and the acquisition module sends the received text to the generation module.

[0139] When the acquisition module receives a text, it directly sends the text to the generation module.

[0140] When the acquisition module receives multiple texts, it first sends one text to the generation module, which then sends the summary result to the output module. The output module outputs the summary result and simultaneously sends a signal indicating that the summary result has been output to the acquisition module. After receiving the signal, the acquisition module sends the next text to the generation module, and repeats the above steps until the last text is output as a summary.

[0141] Assuming the number of texts input by the user into the terminal application module or client module is d, the terminal application module or client module will simultaneously send the received d texts to the acquisition module. The acquisition module will then send the received texts sequentially to the generation module according to the signal of the output summary result received from the output module.

[0142] Step 2: The generation module sends the received text sequentially to the sentence segmentation module, the cleaning module, the word embedding module, the sentence embedding module, and the summary sentence evaluation module to obtain the summary results, and then sends the summary results to the output module.

[0143] The resulting abstract takes into account both validity and timeliness.

[0144] Step 3: The output module sends the received summary results to the terminal application module or the client module.

[0145] Specific Implementation Method Twelve: Combining Figures 1-2 This embodiment describes the usage method of a summary generation system based on automatic machine learning technology. The specific process of using the terminal application module is as follows:

[0146] Step 1: The user enters one or more texts to be summarized into the terminal application module. The operation interface of the terminal application module will pop up a variable customization window. The user can set the variables in the variable customization window according to their own needs or skip the page and use the default value of the variables.

[0147] The variables include random sampling size, number of machine learning recommendations, and default summary algorithm;

[0148] The random sampling size is denoted as k. For each round, k triples need to be randomly selected for a text. These k triples correspond to k combination algorithms.

[0149] Let t be the number of machine learning recommendations. In each round, for a given text, its feature attributes are extracted and fed into the machine learning model to predict which triples will generate a better summary. This predicts t triples and t possible combinations of algorithms.

[0150] The default summarization algorithm is either the optimal algorithm for a particular text or a user-defined algorithm. It is not necessarily good for all text sets. Subsequent algorithm selection can generate additional combined algorithms (k+t), and these k+t+1 combined algorithms will generate summaries for the user to choose from.

[0151] The default values ​​for the variables are: random sampling size of 2, machine learning recommendation number of 1, and default summarization algorithm is the optimal combination algorithm corresponding to a randomly selected text input by the user.

[0152] Step 2: After the user completes or skips the variable customization window, a mode selection window will pop up, where the user can choose between traditional mode and automatic mode. If traditional mode is selected, the terminal application module will output the corresponding summary according to the selected traditional algorithm, and the step ends; if automatic mode is selected, the terminal application module will execute the summary generation method based on automatic machine learning technology of this invention.

[0153] Assuming the user inputs d texts, the summarization method of this invention executes d rounds iteratively. In the i-th (i = 1, ..., d) round, a summarization algorithm is selected for the i-th text. The specific process is as follows:

[0154] ① Clean the obtained text;

[0155] ② Randomly sample k components to form a triplet; this means randomly selecting one component from each stage to form a triplet, and randomly selecting k times to obtain k triplets, corresponding to k combination algorithms.

[0156] ③ Extract features from the i-th text, including but not limited to: total number of characters, number of sentences, number of words, number of keywords, number of sentences containing keywords in the uncleaned text, number of sentences containing keywords in the cleaned text, keyword ratio, average word length, average keyword length, average sentence length in the uncleaned text, average sentence length in the cleaned text, standard deviation of keyword position distribution, and standard mean of keyword position distribution. Then, use a machine learning model to recommend t optimal component combinations.

[0157] ④ Enable the adapter to make the selected components that do not meet the constraints meet the constraints, so that mismatched components can also be combined to achieve arbitrary combination of components;

[0158] ⑤ The k+t component combinations selected in ② and ③ above, together with the default component combination f0, form a total of k+t+1 combinations. These combinations serve as the summarization algorithm for the i-th text, outputting k+t+1 summaries for the i-th text for the user to choose from. The default component combination, i.e., the default summarization algorithm, is preset in Step 1—randomly selecting the optimal combination algorithm corresponding to a text input by the user, or specified by the user.

[0159] The terminal application module includes basic functions such as personalization settings, user login, privacy protection, and file reading.

[0160] Everything else is the same as in Specific Implementation Method Eleven.

Claims

1. A summarization system based on automated machine learning technology, characterized in that: It consists of an acquisition module, a generation module, and an output module, in that order. The acquisition module is used to acquire text; it is also used to receive the output summary signal from the output module. A generation module is used to generate a summary of the text; The generation module includes, in sequence, a sentence segmentation module, a cleaning module, a word embedding module, a sentence embedding module, and a summary sentence evaluation module; The sentence segmentation module is used to split the content of text into multiple sentences; The cleaning module is used to clean each sentence; The cleaning process includes removing empty sentences, converting languages ​​other than Chinese and English to spaces, converting all English letters to lowercase, and removing stop words. The word embedding module is used to perform word embedding processing on the text after sentence cleaning; The sentence embedding module is used to perform sentence embedding on the text after word embedding. The summary sentence evaluation module is used to evaluate the text after sentence embedding processing, and to determine whether each sentence in the text can be used as a summary sentence; it is also used to integrate the sentences that the summary sentence evaluation module has determined to be summary sentences, and sort them according to their positions in the original text to form a summary. The output module is used to output the summary obtained by the summary evaluation module; it is also used to send the signal of the output summary to the acquisition module. The output module is connected to the summary evaluation module of the generation module; The word embedding module includes a word frequency unit, a normalized word frequency unit, and a word vector unit; The sentence embedding module includes a summation unit, a sentence embedding unit based on the KeyCluster algorithm, a sentence embedding unit based on the one-hot method, and an average smoothing processing unit; The summary sentence evaluation module includes a ratio-based evaluation unit, a mean and standard deviation-based evaluation unit, an MMR algorithm-based evaluation unit, and a TextRank algorithm-based evaluation unit. The generation module also includes an adapter module, which is used to combine the units selected in the word embedding module, sentence embedding module, and summary sentence evaluation module. The adapter module is connected to the generation module. When the generation module is applied to a text, the specific process is as follows: ① Input the acquired text into the sentence segmentation module to obtain the processed text; ② Input the sentence-segmented text into the cleaning module to obtain the cleaned text; ③ Select one unit from each of the word embedding module, sentence embedding module, and summary sentence evaluation module in sequence. Combine the selected units in the order of selection to form a triple. Obtain multiple sets of triples in the above manner. Use automatic machine learning technology to select one or more sets of triples from the multiple sets of triples according to the selection conditions. Use the adapter module to connect the three units in each set of triples to obtain the connected triples. Use the connected triples to process the cleaned text to obtain the corresponding summary of the text. ④ Input the summary corresponding to the obtained text into the output module to output the summary; The specific process of selecting one or more sets of triples from multiple sets of triples based on selection criteria using automatic machine learning technology is as follows: A1: Let t be the number of recommendations made by the automatic machine learning model. A2: Extract the feature attributes of the input text. The feature attributes include: total number of characters, number of sentences, number of words, number of keywords, number of sentences containing keywords in the uncleaned text, number of sentences containing keywords after cleaning, keyword ratio, average word length, average keyword length, average sentence length in the uncleaned text, average sentence length after cleaning, standard deviation of keyword position distribution, and standard mean of keyword position distribution. A3: Input the feature attributes of the text into an automatic machine learning model for prediction processing to obtain t triples; Before connecting the three methods in each selected triplet, the adapter performs adaptation processing on the differences in input constraints between the three methods in each triplet, so that the methods that originally did not meet the input constraints meet the constraints.

2. The summary generation system based on automatic machine learning technology according to claim 1, characterized in that: The acquisition module acquires one or more texts simultaneously; When the acquisition module receives a text, it directly sends the text to the generation module. When the acquisition module receives multiple texts, it first sends one text to the generation module, which then sends the summary result to the output module. The output module outputs the summary result and simultaneously sends a signal indicating that the summary result has been output to the acquisition module. After receiving the signal, the acquisition module sends the next text to the generation module, and repeats the above steps until the last text is output as a summary.

3. The summary generation system based on automatic machine learning technology according to claim 2, characterized in that: It also includes a terminal application module and a client module, both of which are connected to the acquisition module and the output module; The terminal application module is used to send the text input by the user to be summarized to the acquisition module, and also to receive the summary output by the output module and provide feedback to the user. The client module is used to send the text input by the user and the text to be digested to the acquisition module, and also to receive the digest output by the output module and feed it back to the user; The terminal application module includes mobile APP software and mini-programs.

4. A summarization method based on automated machine learning technology, characterized in that: It includes the following steps: S1. Obtain the text set and the summary corresponding to each text; S2. Establish an automatic machine learning model. Input the text set and the summary corresponding to each text into the automatic machine learning model and train it using automatic machine learning technology. Output the summary of each text to obtain the trained model. S3. Input the text to be summarized into the trained automatic machine learning model, and output the summary of the text; The S2 automatic machine learning model also includes an adapter; In step S2, an automatic machine learning model is established. The text set is input into the automatic machine learning model for training using automatic machine learning techniques. A summary of each text is output to obtain the trained model. The specific process is as follows: S21. Break down the content of a text into multiple sentences; S22. Clean each sentence; The cleaning process includes removing empty sentences, converting languages ​​other than Chinese and English to spaces, converting all English letters to lowercase, and removing stop words. S23. Select one method from word embedding, sentence embedding, and summary sentence evaluation methods in sequence. Combine the three methods in the order of selection to form a triple. Obtain multiple triples in the above manner. Use automatic machine learning technology to select one or more triples from the multiple triples according to the selection criteria. Use an adapter to connect the three methods in each selected triple to obtain a connected triple. Use the connected triples to process the cleaned text to obtain the summary corresponding to the text. The specific process in S23 of selecting one or more triplet pairs from multiple triplet pairs based on selection criteria using automatic machine learning technology is as follows: A1: Let t be the number of recommendations made by the automatic machine learning model. A2: Extract the feature attributes of the input text. The feature attributes include: total number of characters, number of sentences, number of words, number of keywords, number of sentences containing keywords in the uncleaned text, number of sentences containing keywords after cleaning, keyword ratio, average word length, average keyword length, average sentence length in the uncleaned text, average sentence length after cleaning, standard deviation of keyword position distribution, and standard mean of keyword position distribution. A3: Input the feature attributes of the text into an automatic machine learning model for prediction processing to obtain t triples; Before connecting the three methods in each selected triplet, the adapter performs adaptation processing on the differences in input constraints between the three methods in each triplet, so that the methods that originally did not meet the input constraints meet the constraints.

5. The summarization method based on automatic machine learning technology according to claim 4, characterized in that: The text set in S1 includes one or more texts; When the text set is a single text, the text is directly input into the automatic machine learning model, which outputs a summary of the text. When the text set consists of multiple texts, a certain text is input into the automatic machine learning model, and the corresponding summary of the text is output. Then the next text is input into the automatic machine learning model, and the corresponding summary is output. This process continues until the last text outputs its corresponding summary.

6. The summarization method based on automatic machine learning technology according to claim 5, characterized in that: The word embedding includes a word frequency representation method, a normalized word frequency representation method, and a word vector representation method; The sentence embedding methods include a summation method, a sentence embedding method based on the KeyCluster algorithm, a sentence embedding method based on the one-hot algorithm, and an average smoothing method; The summary evaluation methods include a proportion-based summary evaluation method, a mean and standard deviation-based summary evaluation method, an MMR algorithm-based evaluation method, and a TextRank algorithm-based evaluation method.

Images