Information extraction method, electronic device, and storage medium

Abstract

The application discloses an information extraction method, an electronic device and a storage medium, and the method comprises the following steps: cutting dialogue text of a multi-round dialogue into at least one segment, each segment comprising a plurality of sentences; extracting semantic vectors of the plurality of sentences by using a pre-trained model network; determining the association relationship between sentences in the plurality of sentences in each segment based on the semantic vectors of the plurality of sentences; and fusing the association vector and the word vector of a sentence to determine whether the word is the beginning and end of an entity. According to the embodiment of the application, the dialogue text is cut into a segment comprising a plurality of sentences, the semantic vectors of the plurality of sentences are extracted, and the association relationship between sentences is determined according to the semantic vectors of the plurality of sentences, so that the problem that the calculation requirement of a model is very high due to the fact that the number of rounds of dialogue can be very long, dozens to thousands of sentences, and the length difference is too large for unified processing can be solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information extraction, and particularly relates to an information extraction method, an electronic device and a storage medium. BACKGROUND

[0002] Dialogue systems have a wide range of applications in social production and life, such as customer service systems, telephone sales, store clerk services, etc. In these scenarios, there is a wide range of automated information extraction needs, which can greatly facilitate the workload of copywriting recording and sorting. Information extraction is an important direction of natural language processing. In a limited application scenario, given an input text (query), the program outputs the corresponding "entity category" and "value". The main methods on the market at present are: 1. Entity category recognition and entity value extraction based on rule matching; 2. Entity category recognition and entity value extraction based on CRF algorithm; 3. Entity category recognition and entity value extraction based on deep learning.

[0003] Entity category recognition and entity value extraction based on rule matching: for each application scenario, an artificial person collates various scenarios of the required entities according to business needs, uses regular syntax, and designs corresponding rules to match various expressions. The weight of the rule needs to be adjusted by the artificial person, and the result of the rule with the highest score is output according to the weight. According to the rule matching result, the value of the corresponding slot is output as the entity value, and the entity category is determined by the matching rule.

[0004] Entity category recognition and entity value extraction based on CRF algorithm: Conditional Random Field (CRF) is an undirected graph model proposed by Lafferty et al. in 2001, which combines the characteristics of maximum entropy model and hidden Markov model. CRF is a sequence labeling algorithm that receives an input sequence such as [x1, x2,..., xn] and outputs a target sequence [y1, y2,..., yn]. It can also be regarded as a seq2seq model. According to the output sequence label value, the required information is extracted.

[0005] Intention recognition and slot extraction based on deep learning: there is no complex process of artificial feature extraction. The words are divided or directly used as input. The labeled sample training set is prepared, the deep learning model automatically learns the feature expression ability, and the problem is solved end-to-end. Common models include BERT (Bidirectional Encoder Representation from Transformers), LSTM (Long Short Term Memory Network), Transformer, etc.

[0006] The inventors find that the prior art methods are relatively general methods, which are theoretically applicable to various scenarios, but have a common problem: they cannot effectively utilize the characteristics of specific scenarios and grasp the structure of the specific scenarios; in the dialog system scenario, there is no corresponding optimization for the characteristics of the dialog; due to the long dialog text, these methods require high computing resources and performance, which is often difficult to meet. SUMMARY

[0007] Embodiments of the present application aim to at least solve one of the above technical problems.

[0008] In a first aspect, embodiments of the present application provide an information extraction method, comprising: dividing dialog text of a multi-turn dialog into at least one segment, each segment comprising a plurality of sentences; extracting semantic vectors of the plurality of sentences using a pre-trained model network; and determining an association relationship between sentences in the plurality of sentences in each segment based on the semantic vectors of the plurality of sentences.

[0009] In a second aspect, embodiments of the present application provide an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform any of the above information extraction methods.

[0010] In a third aspect, embodiments of the present application provide a storage medium, the storage medium storing one or more programs including execution instructions, the execution instructions being readable and executable by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to perform any of the above information extraction methods.

[0011] In a fourth aspect, embodiments of the present application further provide a computer program product, the computer program product comprising a computer program stored on a storage medium, the computer program comprising program instructions, when the program instructions are executed by a computer, causing the computer to perform any of the above information extraction methods.

[0012] Embodiments of the present application can solve the problem that the computing requirement of the model is very high due to the long number of turns of the dialog, tens to thousands of sentences, and the unified processing of the sentences with too large length difference by dividing the dialog text into segments containing a plurality of sentences, extracting semantic vectors of the plurality of sentences, and determining an association relationship between sentences based on the semantic vectors of the plurality of sentences. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0014] Figure 1 Flow chart of an embodiment of the information extraction method of the present application;

[0015] Figure 2 Flow chart of another embodiment of the information extraction method of the present application;

[0016] Figure 3 Flow chart of an information extraction process provided by an embodiment of the present application;

[0017] Figure 4 Structural schematic diagram of an embodiment of the electronic device of the present application. DETAILED DESCRIPTION

[0018] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without any creative effort under the premise of not making the present application fall into the public domain belong to the protection scope of the present application.

[0019] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0020] The present application can be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The present application can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are connected through a communication network. In a distributed computing environment, program modules can be located in both local and remote computer storage media including storage devices.

[0021] In the present application, "module", "device", "system" and the like refer to a related entity applied to a computer, such as hardware, a combination of hardware and software, software, or software in execution, and the like. In detail, for example, the element can be, but is not limited to, a process running on a processor, a processor, an object, an executable element, an execution thread, a program, and / or a computer. Also, an application program or a script program running on a server, a server can be an element. One or more elements can be in an execution process and / or thread, and the elements can be localized on one computer and / or distributed between two or more computers, and can be operated by various computer readable media. The elements can also communicate with each other according to a signal having one or more data packets, for example, a signal from a data packet from another element in a local system, a distributed system, and / or a network through a signal interacting with other systems over the Internet.

[0022] Finally, it should also be noted that in this document, relational terms such as first and second and the like can only be used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including", "containing", "consisting" and the like, not only include those elements recited, but also include other elements not explicitly listed or inherent to such process, method, article or apparatus. Without more limitations, the elements defined by the statement "comprising" do not exclude the presence of additional identical elements in the process, method, article or apparatus comprising the described elements.

[0023] The embodiment of the present application provides an information extraction method, which can be applied to an electronic device. The electronic device can be a computer, a server or other electronic products, and the present application is not limited thereto.

[0024] Please refer to Figure 1 , which shows an information extraction method provided by an embodiment of the present application.

[0025] As shown in Figure 1 , in step 101, the dialogue text of the multi-round dialogue is divided into at least one segment, and each segment includes a plurality of sentences;

[0026] In step 102, a pre-trained model network is used to extract semantic vectors of the plurality of sentences respectively.

[0027] In step 103, the association relationship between sentences in the plurality of sentences in each segment is determined based on the semantic vectors of the plurality of sentences.

[0028] In the embodiment, for step 101, the voice dialogue data input by the user is acquired, the voice dialogue data is converted into dialogue text, the converted dialogue text is segmented into at least one segment, and each segmented segment contains multiple sentences; then, for step 102, the semantic vectors of all sentences in the segment are extracted, and the semantic vectors of all sentences in the segment are extracted by using a pre-trained model network, for example, each sentence can obtain its semantic vector representing the sentence and the semantic vector of each word by using the same pre-trained model network, and the pre-trained model is used to extract the semantic vectors of the sentences and words.

[0029] Finally, for step 103, after the semantic vectors of all sentences in each segment are acquired, the association relationship between sentences in the multiple sentences in each segment is determined according to the semantic vectors of the multiple sentences in each segment.

[0030] The method of the embodiment can solve the problem that the calculation requirement of the model is very high due to the long number of rounds of dialogue, tens to thousands of sentences, and the unified processing of the sentences with too large length difference.

[0031] In some optional embodiments, after the association relationship between sentences is acquired, the semantic vector of the sentence and the word vector are fused, it is judged whether the word is the beginning and end of an entity, entity extraction is performed in the multiple sentences according to the association relationship between sentences, and the entity extraction method adopts multiple span model networks to perform entity extraction in the multiple sentences, each sentence is connected to a span model network, the span model network can judge the beginning and end of each entity in each sentence and identify the entity category of each entity, the entity extraction includes entity category identification and entity value extraction, for example, according to the relationship between sentences, each sentence is connected to a span model network, that is, N classification models, to judge whether it is the beginning (begin) and end (end) of an entity, and identify the entity category at the same time, and the sentence association relationship is used to guide the entity extraction in the sentence, and the span model is used for the entity extraction in the sentence.

[0032] The method of the embodiment can realize entity extraction and entity category identification of the sentence by using multiple span model networks after the association relationship between sentences is determined.

[0033] In some optional embodiments, the sentence semantic vectors of the plurality of sentences are connected by using an attention network, and the association between the sentences in the plurality of sentences is encoded by using the attention network, wherein the attention network can pass the semantic information at the sentence level, for example, an independent attention network is used to connect the sentence semantics of each sentence, and the attention network is used to encode the association between the sentences, and the attention network is also used to pass the semantic information at the sentence level.

[0034] The method provided in the embodiments of the present application can more accurately identify the semantic information at the sentence level by using the attention network to encode the association between the sentences in the plurality of sentences.

[0035] In some optional embodiments, the dialogue text of the multi-round dialogue is divided into at least one segment, wherein each segment includes a preset number of sentences, and the sentences correspond to the dialogue text of a role in each round of dialogue in the multi-round dialogue, for example, after the voice information input by the user is converted into dialogue text, the dialogue text is sliced, the dialogue text is converted into a dialogue format of one sentence per time, and the segments are divided according to the window, for example, a sliding window method can be used, and 20 sentences are sliced each time, and the sliding is performed every 10 sentences.

[0036] The method provided in the embodiments of the present application can solve the problem that the calculation requirement of the model is very high due to the unified processing of the dialogue with a long number of rounds, tens to thousands of sentences, and a large difference in length.

[0037] Please refer to Figure 2 which shows another information extraction method provided by an embodiment of the present application. The flowchart mainly further defines the flowchart Figure 1 of the steps, wherein the information extraction method is implemented by using an information extraction model, the information extraction model includes a pre-training model network, a span model network and an attention network, and the training method of the information extraction model includes the following steps.

[0038] As shown in Figure 2 , in step 201, the dialogue text of the multi-round dialogue is divided to obtain at least one segment, wherein each segment includes the annotations between the sentences and the annotations within the sentences;

[0039] In step 202, the at least one segment is input into the information extraction model for training.

[0040] In this embodiment, for step 201, the dialogue text obtained after multi-round dialogue conversion of a user is segmented to obtain at least one segment, each segment includes inter-sentence annotation and intra-sentence annotation, and the inter-sentence information and intra-sentence information of the sentence are applied simultaneously during annotation; for step 202, the plurality of segments obtained after segmentation of the dialogue text is input into an information extraction model for training, and after training is completed, each input segment is predicted, the prediction results are summarized, sorted and de-duplicated, and finally the output results are obtained, wherein the model complexity is close to the model complexity of a single sentence, and will not rapidly expand with the increase of the number of sentences.

[0041] The method of the embodiment of the application can solve the problem that the segments will not rapidly expand with the increase of the number of sentences by inputting the segments obtained after segmentation of the dialogue text into an information extraction model for training.

[0042] It should be noted that the pre-training model is used for each sentence in the application, rather than inputting the dialogue as a whole. If the dialogue is directly input into the model such as bert, the long input sentence will cause a large memory occupation, and the training speed is very slow. The general bert model limits the input length to 512 characters. The new network only uses bert in each sentence, and the association between sentences is extracted by another network, which greatly reduces the number of model parameters. The whole model complexity is close to the single sentence model complexity, and will not rapidly expand with the increase of the number of sentences.

[0043] The application also provides another information extraction method. On the basis of pre-training a large-scale language model, a classification model is directly used to convert entity categories and target values into fixed labels; the model is trained using the fixed labels to obtain a trained model, wherein the model includes a classification model; and the trained model is used to extract and classify entities from dialogue text of multi-round dialogue, for example, converting entity categories and values into fixed labels such as the format "age-18"; in this way, the classification model can be used to solve this problem. The training data is the corresponding classification model data, and the output data is the corresponding category and value. The method is relatively simple, common and easy to apply.

[0044] It should be noted that in the method provided by the application, the number of rounds of dialogue can be very long, from dozens to thousands of sentences, and the length difference is too large to be uniformly processed, which will result in high computational requirements of the model, and on the other hand, the performance of the prediction result is poor. According to the actual scene and data summary, we found that most of the information can be extracted according to the local dialogue, for example, the ten sentences before and after the information point are enough.

[0045] At the same time, a sliding window method is used, 20 sentences are segmented, and every 10 sentences are slid, and the global results are merged and de-duplicated. The window sliding method is as follows:

[0046] 1 from the first sentence to the 20th sentence is the first slice,

[0047] 2 from the 11th sentence to the 30th sentence is the second slice,

[0048] 3 from the 21st sentence to the 40th sentence is the third slice,

[0049] 3 and so on, each slice is 20 sentences, and there are 10 overlapping intersections between adjacent slices.

[0050] In this way, it is ensured that each information point has enough context information, avoiding the loss of information caused by the appearance of entities in the intersection of slices, and solving the problem of too long text. However, due to data intersection, the results may overlap, and finally need to be de-duplicated.

[0051] According to the special structural characteristics of the dialogue system, a deep learning algorithm with a multi-level network structure is designed. The method mainly has the following structure:

[0052] Use bert to do the encode layer, get the semantic vector of each sentence and the word semantic vector of each word through the bert model network; Here, the sentence semantics and word semantics are extracted using a pre-trained model, and the pre-trained model is a bert model. All sentences use the same bert model.

[0053] An independent attention network is connected to the semantic vector of each sentence to encode the relationship between sentences.

[0054] Each sentence is followed by a span model network, i.e. N classification models, to determine whether it is the beginning and end of an entity, as well as to identify the entity category.

[0055] This method solves the bottleneck of computing performance, mainly using the following two points:

[0056] 1. Divide the dialogue, reduce the document level to the paragraph level;

[0057] 2. The newly designed network model suitable for dialogue uses a pre-trained model for each sentence, rather than inputting the entire dialogue together.

[0058] If the dialogue is directly inputted into the model such as bert, the long input sentence will cause a large memory occupation, and the training speed will be very slow. Generally, the input length of the bert model is limited to 512 words. The new network only uses bert in each sentence, and the relationship between sentences is extracted through another network, which greatly reduces the number of model parameters. The complexity of the entire model is close to that of a single sentence model, and it will not quickly expand with the increase of the number of sentences.

[0059] In addition, the method is also an algorithm designed for the unique features of the dialogue system:

[0060] 1. The attention network between sentences mainly transmits sentence-level semantic information;

[0061] 2. Entity recognition and value extraction simultaneously utilize the information between sentences and the semantic information of the sentence itself.

[0062] Please refer to Figure 3 , which shows the implementation flowchart of the information extraction method of the present application. Although in Figure 3 , each sentence passes through a pre-trained model, in fact, the same pre-trained model is used, and the subsequent details are omitted.

[0063] As shown in Figure 3 , a) input the dialogue, and generate input data according to the format of each round of dialogue;

[0064] b) divide the input into a series of dialogue segments by sliding window segmentation;

[0065] c) input the model for each segment, and make a prediction after the training is completed;

[0066] d) summarize, sort, and remove duplicates of the prediction results to obtain the final output results, the predicted entity category and entity value information.

[0067] Specific process:

[0068] 1) Process the input, convert it into a one-sentence dialogue format according to the role, and divide it into a segment list according to the window;

[0069] 2) input the program for each segment, and input the pre-trained model for each sentence in the program, BERT is used in this scheme, other models can also be used, and the input is the text characters of each sentence;

[0070] 3) extract the semantic vector of the sentence and the semantic vector of each word (token). As shown in Figure 3 , the cls vector output by the BERT model represents the semantic vector of the entire sentence, and w1,…,wn represents the semantic vector of each word in the sentence;

[0071] 4) do attention on the semantics of all sentences to extract the context association relationship between sentences. Learn the association vector s between sentences by the inter-sentence attention network for the cls vector of each sentence, and each value in the s vector indicates the importance of the sentence in the existing task.

[0072] 5) Using sentence dependency to guide span extraction. The s vector is combined with the word vector of each sentence and input into the span model for classification. The classification mainly judges whether each word is the beginning and end of an entity, such as the 3rd word is the beginning of entity 1 and the 6th word is the end of entity 1, then the content of entity 1 is the 3rd-6th word.

[0073] 6) Finally, filter, deduplicate and sort all extracted entities to get the final result.

[0074] It should be noted that for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of actions, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, certain steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily necessary for the present application. In the above embodiments, the description of each embodiment is focused on, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0075] In some embodiments, the embodiments of the present application provide a non-volatile computer readable storage medium, the storage medium stores one or more programs including execution instructions, the execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to execute any information extraction method of the present application.

[0076] In some embodiments, the embodiments of the present application also provide a computer program product, the computer program product includes a computer program stored on a non-volatile computer readable storage medium, the computer program includes program instructions, when the program instructions are executed by a computer, the computer executes any information extraction method described above.

[0077] In some embodiments, the embodiments of the present application also provide an electronic device, which includes at least one processor, and a memory connected with the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute an information extraction method.

[0078] Figure 4 is a hardware structure schematic diagram of an electronic device for executing an information extraction method provided by another embodiment of the present application, as Figure 4 shown, the device includes:

[0079] one or more processors 410 and a memory 420, Figure 4The processor 410 is taken as an example.

[0080] The device for performing the information extraction method can further include an input device 430 and an output device 440.

[0081] The processor 410, the memory 420, the input device 430, and the output device 440 can be connected through a bus or other means, Figure 4 The connection through the bus is taken as an example.

[0082] The memory 420, as a non-volatile computer readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs, and modules, such as program instructions / modules corresponding to the information extraction method in the embodiments of the present application. The processor 410 performs various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in the memory 420, that is, implements the information extraction method in the above method embodiments.

[0083] The memory 420 can include a program storage area and a data storage area, wherein the program storage area can store an operating system and application programs required by at least one function; the data storage area can store data created according to the use of the information extraction device, etc. In addition, the memory 420 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 420 can optionally include a memory remotely arranged with respect to the processor 410, which can be connected to the information extraction device through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0084] The input device 430 can receive input digital or character information, and generate signals related to the user settings and function control of the information extraction device. The output device 440 can include a display device such as a display screen.

[0085] The one or more modules are stored in the memory 420, and when executed by the one or more processors 410, perform the information extraction method in any of the above method embodiments.

[0086] The above product can perform the method provided in the embodiments of the present application, has the corresponding function modules and beneficial effects of performing the method. Technical details not described in detail in the embodiments can be referred to the method provided in the embodiments of the present application.

[0087] The electronic device of the embodiments of the present application exists in various forms, including but not limited to:

[0088] (1) Mobile communication device: The feature of this kind of device is to have mobile communication function, and to provide voice and data communication as the main target. This kind of terminal includes: smart phone, multimedia phone, functional phone, and low-end phone, etc.

[0089] (2) Ultra-mobile personal computer device: This kind of device belongs to the category of personal computer, has computing and processing function, and generally has mobile internet feature. This kind of terminal includes: PDA, MID and UMPC device, etc.

[0090] (3) Portable entertainment device: This kind of device can display and play multimedia content. This kind of device includes: audio and video player, palm game machine, electronic book, and smart toy and portable car navigation device.

[0091] (4) Other on-board electronic devices with data interaction function, such as car machine device installed on vehicle.

[0092] The device embodiments described above are only illustrative, wherein the units illustrated as separate components can or can not be physically separated, and the components illustrated as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs.

[0093] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be realized by means of software plus a general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions or the part that contributes to the related art can be embodied in the form of software product, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, server, or network device, etc.) execute the method described in each embodiment or some part of the embodiment.

[0094] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An information extraction method, comprising: segmenting dialogue text of a multi-turn dialogue into at least one segment, each segment comprising a plurality of sentences, wherein the segmentation adopts a sliding window manner, each segment contains a preset number of sentences, and adjacent segments have an overlapping part, and the sentences correspond to dialogue text of a role in each turn of the multi-turn dialogue; extracting semantic vectors of the plurality of sentences respectively by using a pre-trained model network; determining a correlation between sentences in the plurality of sentences in each segment based on the semantic vectors of the plurality of sentences; performing entity extraction in the plurality of sentences based on the correlation, wherein the entity extraction comprises entity category identification and entity value extraction, and the entity extraction result is de-duplicated after the entity extraction. The information extraction method is implemented by using an information extraction model, the information extraction model comprises a pre-trained model network, a span model network and an attention network, and a training method of the information extraction model comprises: segmenting dialogue text of a multi-turn dialogue to obtain at least one segment, wherein each segment comprises inter-sentence labeling and intra-sentence labeling; and inputting the at least one segment into the information extraction model for training.

2. The method of claim 1, wherein, The entity extraction in the plurality of sentences based on the correlation comprises: performing entity extraction in the plurality of sentences by using a plurality of span model networks, wherein each sentence is connected with a span model network, and the span model network can determine the start and end of each entity in each sentence and identify the entity category of the entity.

3. The method of claim 1, wherein the extracting semantic vectors of the plurality of sentences respectively by using a pre-trained model network comprises: extracting semantic vectors of the plurality of sentences and semantic vectors of each word in each sentence respectively by using a pre-trained model network as an encoding layer.

4. The method of claim 1, wherein, The determining a correlation between sentences in the plurality of sentences based on the semantic vectors comprises: connecting sentence semantic vectors of the plurality of sentences by using an attention network, and encoding the correlation between sentences in the plurality of sentences by using the attention network, wherein the attention network can transmit sentence-level semantic information.

5. An information extraction method, comprising: converting an entity category and a target value into a fixed label; training a model by using the fixed label to obtain a trained model, wherein the model comprises a classification model; performing entity extraction and classification on dialogue text of a multi-turn dialogue according to the method of any one of claims 1 to 4 by using the trained model.

6. An electronic device comprising: At least one processor and a memory connected with the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the steps of the method of any one of claims 1 to 5.

7. A storage medium having stored thereon a computer program, characterized in that The program is executed by the processor to implement the steps of the method of any one of claims 1 to 5.

Images