A method and system for identifying language steps in scientific paper abstracts based on saliency attribution

Abstract

The present application discloses a method and system for identifying speech steps in scientific paper abstracts based on saliency attribution, wherein the method includes: extracting the context information and corresponding position feature information of each sentence in the scientific paper abstract to be identified; encoding the context information and the position feature information of each sentence in the scientific paper abstract to be identified, and concatenating the feature vector of the encoded context information with the encoding vector of the corresponding position feature information to obtain the identification data of the corresponding sentence; inputting the identification data of each sentence in the scientific paper abstract to be identified into a speech step identification model obtained through saliency attribution training, and predicting the speech step label of each sentence in the scientific paper abstract to be identified. Through the present application, the influence of the sentences before and after the sentences to be identified on the label prediction of the sentences to be identified can be fully explored, thereby improving the accuracy of speech step label recognition.

Description

Technical Field

[0001] The present invention relates to data intelligent classification technology, and in particular to a method and system for identifying language steps in scientific paper abstracts based on significance attribution. Background Art

[0002] The recognition of speech steps in scientific paper abstracts is to use concise speech step words to summarize the abstract sentence in order to quickly locate the specific information of the paper. It is of great significance as the basis for text processing related research.

[0003] At present, the existing speech step recognition solutions are mainly based on traditional machine learning and deep learning methods. During the model training process, only the sentences in the training set are annotated with speech step labels. The speech step recognition model trained with such a training set only focuses on the overall structural characteristics of the identified sentences when performing speech step recognition. It does not fully explore and utilize the relationship and features between words and between words and labels in the identified sentences, resulting in a decrease in the accuracy of speech step recognition results. Summary of the invention

[0004] In order to fully explore the influence of the sentences before and after the sentence to be identified on the label prediction of the sentence to be identified and improve the accuracy of speech step label recognition, the present application provides a method and system for identifying speech steps in scientific paper abstracts based on significance attribution.

[0005] In a first aspect, the present application provides a method for identifying language steps in scientific paper abstracts based on saliency attribution, comprising the following steps:

[0006] Extract the context information and corresponding position feature information of each sentence in the abstract of the scientific paper to be identified;

[0007] Encoding the context information and the position feature information of each sentence of the scientific paper abstract to be identified, and concatenating the encoded feature vector of the context information with the encoding vector of the corresponding position feature information to obtain identification data of the corresponding sentence;

[0008] Input the recognition data of each sentence in the scientific paper abstract to be recognized into a speech step recognition model obtained through saliency attribution training, and predict the speech step label of each sentence in the scientific paper abstract to be recognized; wherein the training process of the speech step recognition model obtained through saliency attribution training includes:

[0009] Use the training data set for the first training;

[0010] Based on the saliency attribution and the speech step labels obtained by the first training prediction, the recognition data corresponding to the training data in the training data set is corrected, and the recognition data after the correction of the training data is used for the second training to obtain a speech step recognition model.

[0011] Preferably, the extracting of context information and corresponding position feature information of each sentence in the abstract of the scientific paper to be identified comprises the following steps:

[0012] Obtain the two sentences before and after the current sentence to be recognized, and concatenate them with the current sentence to be recognized as the context information of the sentence to be recognized;

[0013] Extracting the absolute position information of the sentence to be identified in the abstract of the scientific paper to be identified; the absolute position information refers to the number of the sentence to be identified in the abstract of the scientific paper;

[0014] Obtaining normalized position information of the current sentence to be identified according to the total number of sentences in the scientific paper abstract to be identified and the absolute position information of the current sentence to be identified; wherein the normalized position information = the absolute position information / the total number of sentences;

[0015] splicing the absolute position information and the normalized position information to obtain the position feature information of the sentence to be recognized;

[0016] Determine whether all sentences in the scientific paper abstract to be identified have been extracted. If not, loop the above steps until all sentences in the scientific paper abstract to be identified have completed the extraction of context information and corresponding position feature information.

[0017] Preferably, in the process of obtaining two sentences before and after the current sentence to be identified and concatenating them with the current sentence to be identified as the context information of the sentence to be identified, if there are no adjacent sentences before and after the sentence to be identified, a special identifier is used to replace the adjacent sentence before the sentence to be identified and / or the adjacent sentence after the sentence to be identified.

[0018] Preferably, encoding the context information and the position feature information of each sentence in the abstract of the scientific paper to be identified, and concatenating the encoded feature vector of the context information with the encoding vector of the corresponding position feature information to obtain the identification data of the corresponding sentence, comprises the following steps:

[0019] Use the BERT model as a text encoder to semantically encode the context information of the current sentence to be recognized;

[0020] Obtain the CLS vector of the encoded context information as the feature vector of the current sentence to be recognized;

[0021] The feature vector of the current sentence to be recognized is concatenated with the encoding vector of the position feature information to obtain recognition data of the current sentence to be recognized.

[0022] Preferably, the training process of the speech step recognition model obtained based on saliency attribution training includes:

[0023] Collect a large number of scientific paper abstracts and annotate each scientific paper abstract with a language step label to obtain a training corpus;

[0024] Extracting context information and position feature information of each sentence in each scientific paper abstract in the training corpus, and associating and storing the sentence, the corresponding context information and the position feature information in a training data set, wherein each training data in the training data set includes a sentence annotated with a label and the context information and position feature information corresponding to the sentence;

[0025] Inputting the training data of the training data set into the pre-training model, encoding the context information of the current sentence to be recognized, and concatenating the feature vector of the encoded context information with the encoding vector of the corresponding position feature information to obtain recognition data of the current sentence to be recognized;

[0026] Input all the recognition data of all the training data of the training data set into a multilayer perceptron, predict the speech step label of the current sentence to be recognized corresponding to the recognition data, and complete the first training;

[0027] Calculating the predicted significance value of each word pair in each sentence of the training data set according to the speech step labels predicted by the training data set;

[0028] Correcting the feature vector of the context information of the sentence by using the significance value predicted for each word pair of the sentence, and concatenating the corrected feature vector of the context information with the encoding vector of the corresponding position feature information to obtain new recognition data of the sentence;

[0029] All new recognition data of the training data set are input into the multilayer perceptron that has been trained for the first time, and the second training is completed to obtain a speech step recognition model.

[0030] Preferably, calculating the predicted significance value of each word pair in each sentence of the training data set according to the speech step labels predicted by the training data set comprises:

[0031] Perform word segmentation on the sentence to obtain multiple words that make up the sentence;

[0032] The ensemble gradient method is used to calculate the predicted significance value of each word pair of the sentence; where the loss function L is used as the expected model and evaluated accordingly, the word w of the sentence i The significance value of the prediction The calculation formula is as follows:

[0033]

[0034] Where S = [w1 ,w 2 ,…,w N ] is a sentence consisting of N words; G s is the sentence-level event label of sentence S; i represents the i-th word of sentence S; xi is the word w of the sentence i The vector representation in the BERT model, X S is the input embedding of sentence S in BERT; X' is a sequence of all-zero vectors (used as reference input), x' i represents the i-th element in X'; α is the independent variable of the definite integral, ranging from 0 to 1;

[0035]

[0036] Where N is the total number of words in sentence S; n is the subscript of the series summation formula; || || represents the regularization function L 2 Norm.

[0037] Preferably, the feature vector of the context information of the sentence is corrected by using the significance value predicted by each word pair of the sentence, and the feature vector of the corrected context information is spliced ​​with the encoding vector of the corresponding position feature information to obtain new recognition data of the sentence, including the following steps:

[0038] Compare the significance value predicted by each word pair of the sentence with a preset threshold value, if the significance value predicted by a certain word pair for a certain speech step label is greater than the preset threshold value, mark the significance feature predicted by the word pair as 1, otherwise, mark it as 0;

[0039] Setting a learnable 512-dimensional significant feature vector for each word whose significant feature is 1 or 0, adding the original encoding vector of the word to the corresponding significant feature vector to generate a new encoding vector for the word;

[0040] Update the encoding vector of the context information of each sentence in the scientific paper based on the encoding vector of each word in the sentence;

[0041] The feature vector is extracted from the updated encoding vector of each sentence context information, and is concatenated with the encoding vector of the position feature information corresponding to the sentence to obtain new recognition data.

[0042] In a second aspect, the present application provides a system for identifying language steps in scientific paper abstracts based on significance attribution, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, a method for identifying language steps in scientific paper abstracts based on significance attribution as described in any one of the first aspects is implemented.

[0043] The beneficial effects of the present invention are:

[0044] 1. Speech step recognition is performed through the speech step recognition model obtained through saliency attribution training. When the speech step recognition model extracts the feature vector of the sentence to be recognized, it extracts the feature vector including the context information of the sentence to be recognized and the adjacent sentences of the sentence to be recognized. Then, the recognition data spliced ​​by this feature vector and the position feature information encoding vector is used to predict the sentence's speech step label, fully exploring the influence of the sentences before and after the sentence to be recognized on the label prediction of the sentence to be recognized, and improving the accuracy of speech step label recognition;

[0045] 2. During the training, the speech step recognition model was adjusted based on the predicted significance value of each word pair in the sentence, so that the speech step recognition model can learn the ability to pay attention to and predict the most relevant text of the speech step label, further improving the accuracy of speech step recognition and effectively optimizing the effect of speech step recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of a method for identifying language steps in scientific paper abstracts based on saliency attribution provided in an embodiment of the present application;

[0047] Figure 2 This is a training flow chart of a speech step recognition model obtained based on saliency attribution training in an embodiment of the present application;

[0048] Figure 3 A schematic diagram of the structure of a scientific paper abstract speech step recognition system based on saliency attribution provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0050] In order to solve the problems in the prior art and focus on the contribution of each word in the identified sentence to the label, the present application embodiment provides a method for identifying the language steps of scientific paper abstracts based on saliency attribution, such as Figure 1 As shown, the following steps are included:

[0051] Step S100: extracting context information and corresponding position feature information of each sentence in a scientific paper abstract (scientific paper abstract to be identified).

[0052] Step S200, encode the context information and position feature information of each sentence in the scientific paper abstract, and concatenate the feature vector of the encoded context information with the encoding vector of the corresponding position feature information to obtain the recognition data of the corresponding sentence.

[0053] Step S300: input recognition data of each sentence in the scientific paper abstract into a speech step recognition model obtained through saliency attribution training to predict the speech step label of each sentence in the scientific paper abstract.

[0054] The embodiment of the present application performs speech step recognition through a speech step recognition model obtained through saliency attribution training. When the speech step recognition model extracts feature vectors for sentences, it extracts feature vectors including context information of the sentence to be recognized and adjacent sentences of the sentence to be recognized, and then predicts the speech step label of the sentence through recognition data spliced ​​by this feature vector and the position feature information encoding vector, fully exploits the influence of the sentences before and after the sentence to be recognized on the label prediction of the sentence to be recognized, and improves the accuracy of speech step label recognition. More importantly, during training, the speech step recognition model adjusts the training process based on the predicted saliency value of each word pair in the sentence, so that the speech step recognition model learns the ability to pay attention to the text most relevant to predicting the speech step label, further improves the accuracy of speech step recognition, and effectively optimizes the effect of speech step recognition.

[0055] In some embodiments of the present application, step S100 includes the following steps:

[0056] Step S101: Obtain the two sentences before and after the current sentence to be identified, and concatenate them with the current sentence to be identified as the context information of the sentence to be identified. In this embodiment, if there are no adjacent sentences before and after the sentence to be identified, a special identifier is used to replace the adjacent sentence before the sentence to be identified or / and the adjacent sentence after the sentence to be identified; wherein the special identifier is <none>, if there is no adjacent sentence before the sentence A to be recognized, then the context information of the sentence A to be recognized is: <none>+ Sentence A to be recognized + Sentence B following A.

[0057] Step S102, extracting the absolute position information of the sentence to be recognized in the scientific paper abstract; the absolute position information refers to the sentence number of the sentence to be recognized in the scientific paper abstract.

[0058] Step S103, obtaining normalized position information of the sentence to be recognized according to the total number of sentences in the scientific paper abstract and the absolute position information of the sentence to be recognized; wherein, normalized position information=absolute position information / total number of sentences.

[0059] Step S104: concatenate the absolute position information and the normalized position information to obtain the position feature information of the sentence to be recognized.

[0060] Step S105, determine whether all the sentences in the scientific paper abstract have been extracted (which can be determined by the absolute position information of the current sentence to be identified and the total number of sentences in the scientific paper abstract). If not, obtain the next sentence and execute step S101; otherwise, complete the extraction and end the program.

[0061] In the embodiment of the present application, step S200 includes the following steps:

[0062] Step S201: Using the BERT model as a text encoder, semantically encode the context information of the current sentence to be recognized;

[0063] Step S202, obtaining the CLS vector of the encoded context information as the feature vector of the current sentence to be recognized;

[0064] Step S203: concatenate the feature vector of the current sentence to be recognized and the encoding vector of the position feature information to obtain recognition data of the current sentence to be recognized.

[0065] In the embodiment of the present application, the training process of the speech step recognition model obtained by the saliency attribution training includes performing a first training using a training data set; then using the speech step labels obtained based on the saliency attribution and the first training prediction to correct the recognition data corresponding to the training data in the training data set, and performing a second training using the recognition data corrected by the training data to obtain a speech step recognition model, such as Figure 2 As shown, the specific steps include:

[0066] Step S301, collect a large number of scientific paper abstracts, and annotate each scientific paper abstract with a speech step label to obtain a training corpus, that is, the training corpus includes a large number of scientific papers, and each sentence in each scientific paper is annotated with a label.

[0067] In this embodiment, the collection of scientific paper abstracts is not limited to channels; scientific paper abstracts are not limited to the scope of fields, including natural language processing, image visual analysis, machine learning and other disciplines, to ensure the diversity of the scope of the corpus in the training set, and then ensure that the speech step recognition model trained by the training set has stronger field generalization ability when performing speech step recognition. The speech step labels of this embodiment are based on linguists' research on linguistic theory and the writing characteristics of scientific paper abstracts, and abstractly define eight speech step labels, including "background", "difference", "method", "purpose", "conclusion", "result", "contribution" and "influence".

[0068] Step S302, extract the context information of each sentence in each scientific paper abstract in the training corpus and the position feature information of the sentence, and store the sentence, the corresponding context information and the position feature information in the training data set in association, that is, the training data set includes a large amount of training data, each training data includes a labeled sentence and the corresponding context information and position feature information of the sentence; the specific extraction process of extracting the context information of each sentence in each scientific paper abstract in the training corpus and the position feature information of the sentence in this embodiment is the same as step S100, which will not be repeated here.

[0069] Step S303, taking the pre-trained model as the benchmark model, inputting the training data of the training data set into the pre-trained model, encoding the context information of the current sentence to be recognized, and concatenating the feature vector of the encoded context information with the encoding vector of the corresponding position feature information to obtain the recognition data of the current sentence to be recognized; in this embodiment, the context information of the current sentence to be recognized is encoded, and the feature vector of the encoded context information is concatenated with the encoding vector of the corresponding position feature information to obtain the recognition data of the current sentence to be recognized. The specific steps are the same as step S200 and will not be repeated here.

[0070] Step S304: input the obtained recognition data into a multi-layer perceptron, train the multi-layer perceptron, and predict the speech step label of the current sentence to be recognized corresponding to the recognition data.

[0071] Step S305, determine whether all the recognition data corresponding to the training data of the training data set have been input, if so, execute step S306; otherwise, obtain the next training data and execute step S303; wherein the initial value of the number of training times is 0.

[0072] Step S306, add 1 to the number of training times, and determine whether the number of training times is 2. If not, execute step S307; otherwise, end the training and complete the training of the multilayer perceptron. The multilayer perceptron obtained by training is the speech step recognition model. In the embodiment of the present application, each training time can also include a preset number of iterative training rounds.

[0073] Step S307, calculate the significance value of each word pair in each sentence of the training data set according to the speech step labels predicted by the training data set; wherein the significance value can reflect the contribution of each word pair to the prediction of the speech step label of the sentence in which it is located.

[0074] In this embodiment, the significance value of each word pair prediction in each sentence of the training data set is calculated based on the speech step labels predicted by the training data set, that is, each word in the sentence is regarded as a feature, and its contribution to the predicted speech step label is calculated according to the feature attribution method, which specifically includes:

[0075] Perform word segmentation on the sentence to obtain multiple words that make up the sentence;

[0076] The ensemble gradient method is used to calculate the significance value of each word pair prediction in the sentence; the word w of the sentence i In BERT’s vector representation x i ∈X s , using the loss function L as the desired model and evaluating accordingly, the word w of the sentence i The significance value of the prediction The calculation formula is as follows:

[0077]

[0078] Where S = [w 1 ,w 2 ,…,w N ] is a sentence consisting of N words; G s is the sentence-level event label of sentence S; i represents the i-th word of sentence S; X S is the input embedding of sentence S in BERT; X' is a sequence of all-zero vectors (used as reference input), x' i represents the i-th element in X'; α is the independent variable of the definite integral, ranging from 0 to 1;

[0079]

[0080] Where N is the total number of words in sentence S; n is the subscript of the series summation formula; || || represents the regularization function L 2 Norm.

[0081] Step S308, correct the feature vector of the context information of the sentence by using the predicted significance value of each word pair of the sentence, and concatenate the corrected feature vector of the context information with the encoding vector of the corresponding position feature information to obtain new recognition data of the sentence, and then execute step S304.

[0082] The speech step recognition model training method provided in this embodiment uses the significance value as an additional feature to adjust the learning process of the speech step recognition model, and uses the significance value as a basis to force the trained speech step recognition model to focus on the text most relevant to the label, thereby correcting the feature vector of the training data, and then performing secondary training on the speech step recognition model with the corrected feature vector, thereby improving the recognition accuracy of the speech step recognition model and optimizing the speech step recognition effect of the speech step recognition model.

[0083] In the embodiment of the present application, step S308 includes the following steps:

[0084] Compare the predicted significance value of each word pair in the sentence with the preset threshold. If the significance value predicted by a word pair for a certain speech step label is greater than the preset threshold, mark the predicted significance feature of the word pair as 1, otherwise, mark it as 0.

[0085] For words with salient features of 1 and 0, a learnable 512-dimensional salient feature vector is set respectively, and the original encoding vector of the word is added to the corresponding salient feature vector to generate a new encoding vector for the word;

[0086] Update the encoding vector of the context information of each sentence in the scientific paper according to the encoding vector of each word in the sentence, that is, obtain a new encoding vector of the sentence according to the encoding vector of each word in the sentence, and obtain a new encoding vector of the context information of each sentence according to the new encoding vector of each sentence in the scientific paper;

[0087] The feature vector is extracted from the updated encoding vector of each sentence context information (the updated encoding vector of each sentence context information is input into the BERT model, and the feature vector is extracted using the BERT model), and is concatenated with the encoding vector of the position feature information corresponding to the sentence to obtain new recognition data, that is, the recognition data update of all training data in the training data set is completed through step S308.

[0088] The present application embodiment provides a system for identifying language steps of scientific paper abstracts based on saliency attribution. Specifically, the device includes: one or more processors and a memory, such as Figure 3 As shown, a processor 200 and a memory 100 are taken as an example. The processor 200 and the memory 100 may be connected via a bus or other methods, such as taking the connection via a bus as an example.

[0089] The memory 100 is a non-transitory computer-readable storage medium that can be used to store non-transitory software programs and non-transitory computer executable programs, such as a method for identifying language steps in scientific paper abstracts based on saliency attribution in an embodiment of the present application. The processor 200 implements the method for identifying language steps in scientific paper abstracts based on saliency attribution in the embodiment of the present application by running the non-transitory software programs and instructions stored in the memory 100.

[0090] The memory 100 may include a program storage area and a data storage area, wherein the program storage area may store an operating device, an application required for at least one function; the data storage area may store data required to execute a method for identifying language steps of scientific paper abstracts based on significance attribution in the above-mentioned embodiment, etc. In addition, the memory 100 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0091] The non-transient software program and instructions required to implement the above-mentioned method for identifying the language steps of scientific paper abstracts based on saliency attribution are stored in the memory. When executed by one or more processors, the above-mentioned method for identifying the language steps of scientific paper abstracts based on saliency attribution is executed, for example, the above-mentioned method is executed. Figure 1 Method steps S100 to S300, method steps S101 to S105, method steps S201 to S203, Figure 2 Method steps S301 to S30 in.

[0092] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0093] Although the embodiments of the present application have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present application, and that the scope of the present application is defined by the appended claims and their equivalents.< / none> < / none>

Claims

1. A method for identifying the language steps of scientific paper abstracts based on saliency attribution, characterized in that: The following steps are involved: Extract the context information and corresponding position feature information of each sentence in the abstract of the scientific paper to be identified; Encoding the context information and the position feature information of each sentence of the scientific paper abstract to be identified, and concatenating the encoded feature vector of the context information with the encoding vector of the corresponding position feature information to obtain identification data of the corresponding sentence; Input the recognition data of each sentence in the scientific paper abstract to be recognized into a speech step recognition model obtained through saliency attribution training, and predict the speech step label of each sentence in the scientific paper abstract to be recognized; wherein the training process of the speech step recognition model obtained through saliency attribution training includes: Use the training data set for the first training; Based on the saliency attribution and the speech step labels obtained by the first training prediction, the recognition data corresponding to the training data in the training data set is corrected, and the recognition data corrected by the training data is used for the second training to obtain a speech step recognition model; The training process of the speech step recognition model obtained based on saliency attribution training includes: Collect a large number of scientific paper abstracts and annotate each scientific paper abstract with a language step label to obtain a training corpus; Extracting context information and position feature information of each sentence in each scientific paper abstract in the training corpus, and associating and storing the sentence, the corresponding context information and the position feature information in a training data set, wherein each training data in the training data set includes a sentence annotated with a label and the context information and position feature information corresponding to the sentence; Inputting the training data of the training data set into the pre-training model, encoding the context information of the current sentence to be recognized, and concatenating the feature vector of the encoded context information with the encoding vector of the corresponding position feature information to obtain recognition data of the current sentence to be recognized; Input all the recognition data of all the training data of the training data set into a multilayer perceptron, predict the speech step label of the current sentence to be recognized corresponding to the recognition data, and complete the first training; Calculating the predicted significance value of each word pair in each sentence of the training data set according to the speech step labels predicted by the training data set; Correcting the feature vector of the context information of the sentence by using the significance value predicted for each word pair of the sentence, and concatenating the corrected feature vector of the context information with the encoding vector of the corresponding position feature information to obtain new recognition data of the sentence; Inputting all new recognition data of the training data set into the multilayer perceptron that has been trained for the first time, completing the second training, and obtaining a speech step recognition model through training; Calculating the predicted significance value of each word pair in each sentence of the training data set according to the speech step labels predicted by the training data set includes: Perform word segmentation on the sentence to obtain multiple words that make up the sentence; The ensemble gradient method is used to calculate the predicted significance value of each word pair of the sentence; where the loss function L is used as the expected model and evaluated accordingly, the word w of the sentence i The significance value of the prediction The calculation formula is as follows: Where S = [w1,w2,…,w N ] is a sentence consisting of N words; G s is the sentence-level event label of sentence S; i represents the i-th word in sentence S; xi is the word w of the sentence i The vector representation in the BERT model, X S is the input embedding of sentence S in BERT; X' is a sequence of all zero vectors, x' i represents the i-th element in X'; α is the independent variable of the definite integral, ranging from 0 to 1; Where N is the total number of words in sentence S; n is the subscript of the series summation formula; || || represents the L2 norm of the regularization function.

2. A method for identifying language steps in scientific paper abstracts based on saliency attribution according to claim 1, characterized in that: The step of extracting the context information and the corresponding position feature information of each sentence in the abstract of the scientific paper to be identified comprises the following steps: Obtain the two sentences before and after the current sentence to be recognized, and concatenate them with the current sentence to be recognized as the context information of the sentence to be recognized; Extracting the absolute position information of the sentence to be identified in the abstract of the scientific paper to be identified; the absolute position information refers to the number of the sentence to be identified in the abstract of the scientific paper; Obtaining normalized position information of the current sentence to be identified according to the total number of sentences in the scientific paper abstract to be identified and the absolute position information of the current sentence to be identified; wherein the normalized position information = the absolute position information / the total number of sentences; splicing the absolute position information and the normalized position information to obtain the position feature information of the sentence to be recognized; Determine whether all sentences in the scientific paper abstract to be identified have been extracted. If not, loop the above steps until all sentences in the scientific paper abstract to be identified have completed the extraction of context information and corresponding position feature information.

3. A method for identifying language steps in scientific paper abstracts based on saliency attribution according to claim 2, characterized in that: In the process of obtaining the two sentences before and after the current sentence to be recognized and concatenating them with the current sentence to be recognized as the context information of the sentence to be recognized, if there are no adjacent sentences before and after the sentence to be recognized, a special identifier is used to replace the adjacent sentence before the sentence to be recognized and / or the adjacent sentence after the sentence to be recognized.

4. The method for identifying the language steps of scientific paper abstracts based on saliency attribution according to claim 1, characterized in that: The encoding of the context information and the position feature information of each sentence of the scientific paper abstract to be identified, and concatenating the encoded feature vector of the context information with the encoding vector of the corresponding position feature information to obtain the identification data of the corresponding sentence, comprises the following steps: Use the BERT model as a text encoder to semantically encode the context information of the current sentence to be recognized; Obtain the CLS vector of the encoded context information as the feature vector of the current sentence to be recognized; The feature vector of the current sentence to be recognized is concatenated with the encoding vector of the position feature information to obtain recognition data of the current sentence to be recognized.

5. The method for identifying the language steps of scientific paper abstracts based on saliency attribution according to claim 1, characterized in that: The feature vector of the context information of the sentence is corrected by using the significance value predicted by each word pair of the sentence, and the feature vector of the corrected context information is spliced ​​with the encoding vector of the corresponding position feature information to obtain new recognition data of the sentence, including the following steps: Compare the significance value predicted by each word pair of the sentence with a preset threshold value, if the significance value predicted by a certain word pair for a certain speech step label is greater than the preset threshold value, mark the significance feature predicted by the word pair as 1, otherwise, mark it as 0; Setting a learnable 512-dimensional significant feature vector for each word whose significant feature is 1 or 0, adding the original encoding vector of the word to the corresponding significant feature vector to generate a new encoding vector for the word; Update the encoding vector of the context information of each sentence in the scientific paper based on the encoding vector of each word in the sentence; The feature vector is extracted from the encoding vector of the updated context information of each sentence, and is concatenated with the encoding vector of the position feature information corresponding to the sentence to obtain new recognition data.

6. A system for identifying language steps in scientific paper abstracts based on saliency attribution, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, a method for identifying language steps in scientific paper abstracts based on significance attribution as described in any one of claims 1 to 5 is implemented.

Images