Method for extracting text events with overlapping and nesting features

Abstract

The invention discloses an extraction method for text events with overlapping and nesting characteristics, which comprises the following steps of: performing attention interaction on original text word embedding and event type word embedding through a symmetric bidirectional attention mechanism to obtain corresponding word embedding after semantic enhancement; original text word embedding and semantic enhanced original text word embedding are carried out, important texts are extracted by extracting weights, and the important texts are integrated to obtain optimized text word embedding; splicing the original text word embedding, the optimized text word embedding and the corresponding event type word embedding after semantic enhancement to obtain a spliced text word embedding, and obtaining a trigger according to the spliced text word embedding; and obtaining the start and end probabilities of each word embedded in the spliced text word in the question and answer task, and obtaining the event element of the current question role through a position decoding algorithm to complete the event extraction process. According to the method, overlapped and nested events can be better understood, and event type information in a complex structure can be represented.

Description

Technical Field

[0001] The invention relates to the technical field of text event extraction, and in particular to a method for extracting text events with overlapping and nested features. Background Art

[0002] Event extraction is a key link in the information extraction task. When a sentence input is given, the task of event extraction is to identify and extract its related parameters: triggers, event types, and parameters. However, the parameters in actual events are often complicated, and overlaps and nestings often occur. This is a problem that event extraction cannot avoid. The overlapping and nesting problem refers to the fact that the same sentence may contain multiple intertwined or nested events, in which some words (tokens) may belong to multiple event types at the same time and have multiple labels. Dealing with these problems requires the model to be able to identify and distinguish these complex relationships to ensure accurate extraction of key information for each event. These challenges are particularly prominent when dealing with tokens with multiple meanings, requiring more complex mechanisms to understand and parse the context.

[0003] Most of the existing event extraction methods target flat events, assuming that events do not overlap or nest, and ignore unconventional event extraction scenarios. Existing event extraction methods use a unified decoder to extract triggers and parameters through traditional joint methods, usually only labeling once. However, the nature of overlapping and nested events leads to label conflicts, and only one label is assigned during the prediction stage. There are also methods that handle overlapping events through cascade decoding. Although this strategy can effectively handle the overlap problem, it is still insufficient for nested events. In addition, although the traditional pipeline method is committed to solving the overlap problem, the information between the two subtasks cannot be effectively utilized, and the information of sentences and labels is not fully utilized. Recent studies have improved the effect of event extraction through reading comprehension and formulated event extraction as question-answering questions to improve the effect, but these methods are relatively fixed in design and lack contextual connections. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method for extracting text events with overlapping and nested features, so as to solve the technical problems in the prior art of inaccurate extraction of overlapping and nested events, inability to adapt to tokens with multiple semantics, and inability to understand the positional relationship of tokens in the context.

[0005] The present invention provides a method for extracting text events with overlapping and nested features, comprising the following steps:

[0006] Step 1: Obtain the original text word embedding through the pre-trained encoder, and preset the event type word embedding;

[0007] Step 2: The original input vector is subjected to attention interaction between the original text word embedding and the event type word embedding through a symmetric bidirectional attention mechanism to obtain the semantically enhanced original text word embedding and the semantically enhanced event type word embedding;

[0008] Step 3: Set the extraction weights, extract the important texts of each word embedding from the original text word embedding and the original text word embedding after semantic enhancement through the extraction weights, and integrate the extracted important texts to obtain the optimized text word embedding;

[0009] Step 4: Concatenate the original text word embedding, the optimized text word embedding, and the corresponding semantically enhanced event type word embedding to obtain the concatenated text word embedding, and obtain the trigger corresponding to the event type according to the concatenated text word embedding;

[0010] Step 5: Construct the corresponding question-answering task according to the preset question role, obtain the start and end probabilities of each word in the concatenated text word embedding in the question-answering task, and obtain the event element of the current question role through the position decoding algorithm, where the position decoding algorithm is specifically as follows:

[0011] When the start and end probabilities of the current word are greater than the start and end probabilities of all separators, and the current word is not in the question sentence of the question-answering task, the probability sum of the start and end probabilities of the current word is calculated, and the word with the largest probability sum among all words is used as the event element of the current question role;

[0012] Step 6: After obtaining the event elements of all problem roles, construct the event extraction results to complete the event extraction process.

[0013] Furthermore, the symmetric bidirectional attention mechanism in step 2 is expressed as:

[0014]

[0015] Where, d h is the scaling factor; Q, K, V are the query, key, and value tensors respectively.

[0016] Furthermore, in step 3, the specific formula for extracting the important text embedded in each word by extracting the weight is as follows:

[0017]

[0018] In the formula, g is the extraction weight; m and n represent two input quantities respectively.

[0019] Furthermore, the setting formula of the extraction weight is:

[0020] g=σ(W g [m;n]+b g )

[0021] Where σ(·) represents the sigmoid activation function; Indicates element-by-element multiplication; [;] indicates a cascade operation; W g and b g These are all training parameters within the model.

[0022] Furthermore, the specific process of step 3 is as follows:

[0023] Step 31: The original text word embedding and the original text word embedding after semantic enhancement are used to obtain extraction weights through a gated equation, and important texts of the original text word embedding and the original text word embedding after semantic enhancement are extracted respectively according to the extraction weights, and the extracted important texts are integrated to obtain a transitional text word embedding;

[0024] Step 32: The transitional text word embedding and the semantically enhanced event type word embedding are embedded, and the extraction weights are obtained through the gating equation. The important texts of the transitional text word embedding and the corresponding semantically enhanced event type word embedding are extracted according to the extraction weights, and the extracted important texts are integrated to obtain the optimized text word embedding.

[0025] Furthermore, in step 4, the trigger is determined and obtained by embedding the concatenated text words into binary labels.

[0026] Furthermore, in step 5, the specific method for obtaining the start and end probabilities of each word in the concatenated text word embedding in the question-answering task is:

[0027] First, each input word is converted into a corresponding embedding vector through the RoBERTa model, and then the obtained embedding vector is normalized through the softmax function to obtain the start and end probabilities of each word.

[0028] Furthermore, the embedding vector includes: a word embedding vector, a position embedding vector and a separator embedding vector.

[0029] Beneficial effects of the present invention:

[0030] The present invention uses a symmetric bidirectional attention mechanism to embed two types of words for mutual interaction to strengthen the semantic interaction between event types and texts. The model can effectively capture the multiple label meanings of complex tokens, especially when dealing with overlapping and nested events, and can better understand and represent the event type information in these complex structures.

[0031] The present invention can effectively filter out non-important information in the text by using extraction weights, reduce the attention paid to non-important information, highlight key information, and perform well in the face of overlapping and nested events with multiple tags, thereby improving the accuracy of trigger recognition. The important text extraction method provided by the present invention can improve the recognition and extraction capabilities of such events while ensuring the extraction accuracy of independent events, in the case of overlapping events in the text and the case where one event completely contains another event.

[0032] The present invention adopts a position decoding algorithm to obtain the relationship between the starting position and the ending position of a word, and considers the starting and ending position relationship between each token more finely according to the algorithm. It not only enhances the model's ability to judge the starting and ending positions of arguments, but also better handles the uniqueness of arguments, making the extraction of arguments more accurate in complex overlapping and nested event relationships. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:

[0034] Figure 1 is a flow chart of a specific embodiment of the present invention;

[0035] Figure 2 is a specific flow chart of event element acquisition according to a specific embodiment of the present invention;

[0036] Figure 3 It is a code representation of the position decoding algorithm of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0038] The present invention is further illustrated below in conjunction with specific embodiments. Those skilled in the art should understand that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention, and modifications to various equivalent forms of the present invention fall within the scope defined by the appended claims of this application.

[0039] like Figure 1 As shown, a specific embodiment of the present invention provides a method for extracting text events with overlapping and nested features.

[0040] The following text about finance with overlapping events is taken as an example to explain the steps in detail: "In the context of global economic slowdown, central banks around the world are implementing interest rate cuts. Meanwhile, banks in some countries are increasing their support for small and medium-sized enterprises by lowering financing costs in an attempt to stimulate economic recovery."

[0041] Step 1: Pass the original text through the pre-trained BERT encoder to obtain the original text word embedding and preset the event type word embedding;

[0042] BERT is pre-trained with a bidirectional Transformer encoder to learn rich semantic representations in a large amount of text. The transformers library in Python is used to load and use the pre-trained BERT model and tokenizer. The tokenizer first decomposes the sentence into word fragments (tokens). The decomposition result of the case is:

[0043] ["In","the","context","of","global","economic","slowdown",",","central","banks","around","the","world","are","implementing","in terest","rate","cuts",".","Meanwhile",",","banks","in","some","countries","are","increasing","their","support","for","small","a nd","medium-sized","enterprises","by","lowering","financing","costs","in","an","attempt","to","stimulate","economic","recovery","."],

[0044] Each word fragment will be mapped to a word vector and combined with the positional encoding and segment encoding to form the input of the model. BERT analyzes and processes each word in the sentence through its bidirectional Transformer layer to capture the relationship between it and other words. For example, BERT will understand the causal relationship between "global economic growth slowdown" and "major central banks have adopted interest rate cuts", as well as the target relationship between "banks in some countries have increased loan support for small and medium-sized enterprises" and "stimulating economic recovery by reducing financing costs". At the same time, the event type words are also initially randomized.

[0045] After step 1, BERT will generate a high-dimensional vector representation (768-dimensional vector) and obtain the original text word embedding preprocessing result:

[0046] "global"->[0.22,-0.55,...,0.19]

[0047] "economic"->[0.48,-0.12,...,0.67]

[0048] "slowdown"->[-0.34,0.77,...,-0.23]

[0049] "background"->[0.11,-0.24,...,0.88]

[0050] "central bank"->[0.65,0.12,...,-0.45]

[0051] Preprocessing results of the preset event type word embedding (768-dimensional vector):

[0052] "Investment"->[-0.52,0.17,...,0.44]

[0053] "Policy"->[0.69,-0.24,...,0.66]

[0054] Step 2: The original input vector is subjected to attention interaction between the original text word embedding and the event type word embedding through a symmetric bidirectional attention mechanism to obtain the semantically enhanced original text word embedding and the semantically enhanced event type word embedding;

[0055] In order to better understand complex events and distinguish the different effects of different event types on the same token, the attention weights between the input query (Q), key (K), and value (V) of the original text word embedding and the semantically enhanced event type word embedding are calculated, and the weighted sum is output. Specifically, the dot product QK^T is calculated through torch.matmul, and then normalized through softmax, and finally multiplied by the value V to generate the final attention representation. This process helps the model focus on the most relevant event types and text information when processing complex semantics. The symmetric bidirectional attention can be expressed as:

[0056]

[0057] Where, d h is the scaling factor; Q, K, V are the query, key, and value tensors respectively.

[0058] The representation of event type obtained through the pre-trained BERT model is: E = {e 1 ,e 2 ,e 3 ,…,e M}, where M represents the number of event types, and the symmetric bidirectional attention between event type and text is expressed as:

[0059] E g =Attention(W qe H,W ke E,W ve E),

[0060] H g =Attention(W qh E,W kh H,W vh H),

[0061] Among them, E g and H g represents the output of the attention layer, which is equivalent to the semantically enhanced word embedding; W qe , W ke , W ve , W qh , W kh , W vh These are all built-in training parameters within the model.

[0062] Step 3: Set the extraction weights, extract the important texts of each word embedding from the original text word embedding and the original text word embedding after semantic enhancement through the extraction weights, and integrate the extracted important texts to obtain the optimized text word embedding. The specific process is as follows:

[0063] Step 31: The original text word embedding and the original text word embedding after semantic enhancement are used to obtain extraction weights through a gated equation, and important texts of the original text word embedding and the original text word embedding after semantic enhancement are extracted respectively according to the extraction weights, and the extracted important texts are integrated to obtain a transitional text word embedding;

[0064] Step 32: The transitional text word embedding and the corresponding semantically enhanced event type word embedding are used to obtain extraction weights through a gated equation, and important texts of the transitional text word embedding and the corresponding semantically enhanced event type word embedding are extracted respectively according to the extraction weights, and the respectively extracted important texts are integrated to obtain optimized text word embedding;

[0065] The extraction weight can be expressed as:

[0066] g=σ(W g [m;n]+b g )

[0067] Where g represents the extraction weight, m ​​and n represent the input vector; σ(·) represents the sigmoid activation function; Indicates element-by-element multiplication, implemented by np.multiply(); [;] indicates a cascade operation, implemented by np.concatenate(); W g and b g These are training parameters.

[0068] The extraction weights are generated through a fully connected layer with a sigmoid function to adaptively control the input information.

[0069] The extraction weight fuses information from two different sources (the original text word embedding and the original text word embedding after semantic enhancement, the transitional text word embedding and the corresponding semantically enhanced event type word embedding) through the following formula, and controls the flow of information according to the extracted weight g:

[0070]

[0071] Among them, g represents the extraction weight, m ​​and n represent the input vector; Represents element-by-element multiplication, implemented by np.multiply().

[0072] By adaptively deciding which information to keep and which to ignore, the model can improve its ability to handle multiple label tokens. The vector g is generated by a fully connected layer with a sigmoid function, which adaptively controls the input information. The formula of the filter node (Node) is shown in formulas (3) and (4), which generates the final output through the two input vectors m and n.

[0073] In the example of the specific embodiment of the present invention, the event type information (such as "interest rate cut policy") and the text information (the context of the sentence) are integrated through extraction weights to generate a more expressive representation. The original text is represented as H g , the text after symmetric bidirectional attention is represented as E g , the target event type corresponding to the text is represented by e t , then the extraction process is expressed as:

[0074] S g =Node(H g ,E g ),

[0075] V t =Node(S g ,e t ),

[0076] Among them, e t ∈E represents the target event type embedding, V t ={v1,v2,v3,…,v L} represents the word representation of the last perceived event type.

[0077] Step 4: Concatenate the original text word embedding, the optimized text word embedding, and the semantically enhanced event type word embedding to obtain the concatenated text word embedding, and judge and obtain the trigger by classifying the concatenated text word embedding into binary labels;

[0078] For the concatenated representation, combined_context = torch.cat([sentence_context, event_context.rep eat(sentence_context.size(1),1).unsqueeze(0)], dim = -1) is used to concatenate the original text word embedding, the optimized text word embedding, and the semantically enhanced event type word embedding information. The classifier is used to obtain the probability value of each word, which indicates the possibility of each token being a trigger. Finally, trigger_labels = (trigger_probabili ties>0.5).int() is used to generate binary labels based on the output probability of the model. Words with probability values ​​greater than 0.5 are marked as triggers (label 1), otherwise they are marked as non-triggers (label 0).

[0079] For the case provided, it can be determined that one of the triggers is: interest rate cuts, and the corresponding role is Impact on investment.<Interrogative words> is the <role>in <trigger>The specific design of the specific question is: What is the impact of interest rate cuts on investment? The question can be interpreted as: interest rate cuts are the trigger, and we are asking about its impact on investment.

[0080] To distinguish it from the text, this article adds [CLS] and [SEP] as separators. The standard input of the model will be expressed as: [CLS]What is the impact of interest rate cuts on investment? [SEP]In the context of global economic slowdown, central banks around the world are implementing interest rate cuts. Meanwhile, banks in some countries are increasing their support for small and medium-sized enterprises by lowering financing costs in an attempt to stimulate economic recovery. [SE P].

[0081] like Figure 2 As shown in Figure 5, step 5: construct the corresponding question-answering task according to the preset question role. First, convert each input word into a corresponding embedding vector through the RoBERTa model, and then normalize the obtained embedding vector through the softmax function to obtain the start and end probabilities of each word, where

[0082] The RoBERTa model converts each input token into a corresponding embedding vector, including three types of embedding: word embedding: simply mapping each token to a vector representation; position embedding: introducing position information for each token so that the model knows the position of the token in the sequence; separator embedding: RoBERTa distinguishes different parts through the [SEP] boundary. The embedded text will be passed to RoBERTa's Transformer encoder. RoBERTa uses a multi-layer self-attention mechanism, and each layer will update the correlation representation based on each token and other tokens in the input sequence. Then the softmax function is used to obtain the start and end probabilities of each token.

[0083] Use the Transformers library to load the pre-trained RoBERTa model and tokenizer. Pass the normalized input to the RoBERTa model to get the token embedding, specifically represented as: c1,c2,c3,…,c m =RoBERTa(a1,a2,a3,…,a m ). Introduce two new parameter matrices W start and W end , for c i W start and c i W end Apply softmax normalization to obtain probabilities and calculate the start and end probabilities of each token, i.e., by softmax(c i W start ) and softmax(c ui W end ) operation to obtain the start and end probabilities.

[0084] The event element of the current problem role is obtained through the position decoding algorithm, where the position decoding algorithm is specifically as follows:

[0085] When the start and end probabilities of the current word are greater than the start and end probabilities of all separators, and the current word is not in the question sentence of the question-answering task, the probability sum of the start and end probabilities of the current word is calculated, and the word with the largest probability sum among all words is used as the event element of the current question role;

[0086] like Figure 3 The code shown in the figure takes the starting position probability vector and the ending position probability vector as input for position decoding, and outputs a trigger word candidate group, with [i, j, score] representing the start, end, and score respectively (the start and end positions are closed intervals). start (1) (this position corresponds to the [SEP] special separator in the original token) is the lowest value of the starting score, and the first value I of the probability vector of the ending position is end (1) As the minimum value of the termination score, after screening, the argument candidate group is obtained. The member with the highest score in the obtained candidate group is selected as the answer to the argument extraction.

[0087] For this case, the candidates result is:

[0088] [[2,10,0.85],

[0089] [3,12,0.87],

[0090] [5,15,0.90],

[0091] [7,20,0.88],

[0092] [10,22,0.84]].

[0093] Step 6: After obtaining the event elements of all problem roles, construct the event extraction results to complete the event extraction process.

[0094] Finally, taking the case as an example, the result of event extraction is:

[0095] Event 1

[0096] Event Type: Investment

[0097] Event Trigger: "increasing"

[0098] Event Arguments:

[0099] Banks - Role: Agent

[0100] Small and Medium-Sized Enterprises-Role: Beneficiary

[0101] Loan Support-Role: Resource

[0102] Argument Role:

[0103] Agent: "Bank"

[0104] Beneficiary: "Small and medium-sized enterprises"

[0105] Funds(Resource):"Loan support"

[0106] Event 2

[0107] Event Type: Policy

[0108] Event Trigger: "interest rate cut"

[0109] Event Arguments:

[0110] Central Banks - Role: Agent

[0111] Policy-Role: Policy Measure

[0112] Argument Role:

[0113] Agent: "Central Bank"

[0114] Policy Measure: "interest rate cut"

[0115] The following is a comparison of the effects of the present invention and the prior art under four indicators:

[0116]

[0117] Table 1.

[0118] As shown in Table 1, the present invention (Ours) has achieved experimental results of 89.0, 80.0, 76.4, and 73.5 in the four F1 indicators of TI (%), TC (%), AI (%), and AC (%). In terms of overall performance, the F1 scores of the present invention in TI and TC tasks are 89.0% and 80.0%, respectively, which are comparable to OneEE and CasEE, and better than most BERT series models, especially in trigger word recognition and classification tasks. In terms of fine-grained performance, in the AI ​​task, the F1 of the present invention is 76.4%, slightly higher than OneEE (76.2%) and much higher than PLMEE (70.6%), indicating that the argument recognition effect is strong. In the AC task, the F1 of the present invention is 73.5%, which is better than CasEE (71.4%), and performs well in argument classification. Compared with the classic model: Compared with the BERT series, the present invention has better recall and F1 scores in multiple tasks. Although PLMEE has a slight advantage in TI tasks, it performs worse than the present invention in TC and AI tasks.

[0119] The following is a comparison of the effects of the present invention and the prior art under the Argument ID+Classification indicator:

[0120]

[0121] Table 2

[0122] As shown in Table 2, the present invention (Ours) is compared with Random NE. The Random NE method basically performs argument recognition and classification randomly, and the accuracy is extremely low. The F1 score of the present invention method reaches 66.13%, which is 40.47% higher than Random NE, and the precision and recall rates are also greatly improved, indicating that the present invention can effectively identify and classify arguments in events, rather than random operations. Compared with BERT_QA_Arg w / Role name and BERT_QA_Arg w / Type+Role, the present invention is slightly improved. This present invention makes better use of input information, thereby improving the overall performance of argument recognition and classification, showing that the present invention is more accurate in identifying event arguments, especially in tasks that require accurate distinction between roles and types. It is more robust.

[0123] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.< / trigger> < / role>

Claims

1. A method for extracting text events with overlapping and nested features, characterized in that: The steps include: Step 1: Obtain the original text word embedding through the pre-trained encoder, and preset the event type word embedding; Step 2: The original input vector is subjected to attention interaction between the original text word embedding and the event type word embedding through a symmetric bidirectional attention mechanism to obtain the semantically enhanced original text word embedding and the semantically enhanced event type word embedding; Step 3: Set the extraction weights, extract the important texts of each word embedding from the original text word embedding and the original text word embedding after semantic enhancement through the extraction weights, and integrate the extracted important texts to obtain the optimized text word embedding; Step 4: Concatenate the original text word embedding, the optimized text word embedding, and the corresponding semantically enhanced event type word embedding to obtain the concatenated text word embedding, and obtain the trigger corresponding to the event type according to the concatenated text word embedding; Step 5: Construct the corresponding question-answering task according to the preset question role, obtain the start and end probabilities of each word in the concatenated text word embedding in the question-answering task, and obtain the event element of the current question role through the position decoding algorithm, where the position decoding algorithm is specifically as follows: When the start and end probabilities of the current word are greater than the start and end probabilities of all separators, and the current word is not in the question sentence of the question-answering task, the probability sum of the start and end probabilities of the current word is calculated, and the word with the largest probability sum among all words is used as the event element of the current question role; Step 6: After obtaining the event elements of all problem roles, construct the event extraction results to complete the event extraction process.

2. The method for extracting text events with overlapping and nested features as claimed in claim 1, characterized in that: The symmetric bidirectional attention mechanism in step 2 is expressed as: Where, d h is the scaling factor; Q, K, V are the query, key, and value tensors respectively.

3. The method for extracting text events with overlapping and nested features as claimed in claim 1, characterized in that: In step 3, the specific formula for extracting the important text embedded in each word by extracting the weight is as follows: In the formula, g is the extraction weight; m and n represent two input quantities respectively.

4. The method for extracting text events with overlapping and nested features as claimed in claim 3, characterized in that: The setting formula of the extraction weight is: g=σ(W g [m;n]+b g ) Where σ(·) represents the sigmoid activation function; Indicates element-by-element multiplication; [;] indicates a cascade operation; W g and b g These are all training parameters within the model.

5. The method for extracting text events with overlapping and nested features as claimed in claim 1, characterized in that: The specific process of step 3 is as follows: Step 31: The original text word embedding and the original text word embedding after semantic enhancement are used to obtain extraction weights through a gated equation, and important texts of the original text word embedding and the original text word embedding after semantic enhancement are extracted respectively according to the extraction weights, and the extracted important texts are integrated to obtain a transitional text word embedding; Step 32: The transitional text word embedding and the semantically enhanced event type word embedding are embedded, and the extraction weights are obtained through the gating equation. The important texts of the transitional text word embedding and the corresponding semantically enhanced event type word embedding are extracted according to the extraction weights, and the extracted important texts are integrated to obtain the optimized text word embedding.

6. The method for extracting text events with overlapping and nested features as claimed in claim 1, characterized in that: In step 4, the trigger is determined and obtained by embedding the concatenated text words into binary labels.

7. The method for extracting text events with overlapping and nested features as claimed in claim 1, characterized in that: In step 5, the specific method for obtaining the start and end probabilities of each word in the concatenated text word embedding in the question-answering task is: First, each input word is converted into a corresponding embedding vector through the RoBERTa model, and then the obtained embedding vector is normalized through the softmax function to obtain the start and end probabilities of each word.

8. The method for extracting text events with overlapping and nested features as claimed in claim 7, characterized in that: The embedding vector includes: a word embedding vector, a position embedding vector and a separator embedding vector.

Images