CN111370102B

CN111370102B - Department diagnosis guiding method, device and equipment

Info

Publication number: CN111370102B
Application number: CN202010081698.XA
Authority: CN
Inventors: 刘明录; 吴及; 刘喜恩
Original assignee: Tsinghua University; iFlytek Co Ltd
Current assignee: Tsinghua University; iFlytek Co Ltd
Priority date: 2020-02-06
Filing date: 2020-02-06
Publication date: 2024-07-05
Anticipated expiration: 2040-02-06
Also published as: CN111370102A

Abstract

The invention discloses a department diagnosis guiding method, a device and equipment. The design basis is 1) that the patient can conduct department diagnosis inquiry autonomously and self-help without configuring a great deal of medical manpower. 2) When the self-help query is carried out autonomously, the problem of two dimensions is overcome, and the trouble of professional medical information is removed from the aspect of patients; from the self-service system perspective, the recognition and analysis capability of fuzzy query information input by non-professional patients needs to be improved. Therefore, the main scheme provided by the invention is that the primary department identification and text validity judgment are firstly carried out on the input consultation text, the department recommendation or illegal consultation input advance treatment can be directly carried out, and further, the effective condition description which cannot be directly recommended to the department is deeply analyzed by adopting a double-strategy combination mode of text classification and key information retrieval, so that an accurate and reliable recommended consultation department is provided, a patient is guided to smoothly visit a doctor, and the pressure of hospital consultation guiding consultation is relieved.

Description

Department diagnosis guiding method, device and equipment

Technical Field

The invention relates to the technical field of intelligent application interaction, in particular to a department diagnosis guiding method, device and equipment combined with an artificial intelligent text processing technology.

Background

With the continuous exposure of contemporary medical problems, many problems remain to be solved. For example: because the number of counseling tables in the hospital is small and the number of patients is large, the workload of nurses is large, each patient cannot be helped, and the patients have a question about a registration department when in a hospital, but are difficult to obtain effective help in time, and further effective and rapid treatment cannot be obtained.

Even if many hospitals and enterprises release network inquiry/triage services, in terms of technical schemes, most of the existing triage systems are spot-selection triage schemes based on specific templates, and the department triage process only enables patients to select one of a plurality of disease option templates to meet the conditions of the patients or conduct multi-round questioning according to a certain option template.

Through practical application and feedback discovery, the disease coverage of the diagnosis guiding mode and the accuracy judgment of input information are poor, so that the situation that department diagnosis guiding fails or wrong department recommendation is given frequently occurs, and particularly the existing diagnosis guiding system too depends on a professional inquiry template, so that the user friendliness of common patients is poor, and the use difficulty is increased.

Disclosure of Invention

The invention provides a department diagnosis guiding method, a device and equipment aiming at the defects of the prior art, and correspondingly provides a computer readable storage medium and a computer program product.

The technical scheme adopted by the invention is as follows:

In a first aspect, the present invention provides a method for guiding a diagnosis in a department, including:

Performing preliminary department identification and text validity judgment on the input inquiry text;

based on the effective consultation text of the department which is not identified, obtaining a department classification result and a department retrieval result which are candidates by utilizing a text classification strategy and a key information retrieval strategy;

And determining a target recommended department by using the department classification result and the department retrieval result.

In one possible implementation, the text classification strategy includes outputting a number of department classification results by a number of classification models trained in advance.

In one possible implementation manner, the key information retrieval policy includes:

Determining key information from the effective inquiry text in combination with the processing of the classification model;

and searching the corresponding department in a preset department search library by utilizing the key information.

In one possible implementation, the classification model includes a convolutional neural network;

The processing in combination with the classification model, determining key information from the effective query text includes:

Determining characteristic parameters corresponding to each word in the effective inquiry text according to the output of the convolution layer of the convolution neural network;

And determining the distribution situation of the keywords in the effective inquiry text based on the optimal department classification result of the convolutional neural network and the characteristic parameters.

In one possible implementation manner, the determining key information from the valid inquiry text in combination with the processing of the classification model specifically includes:

acquiring a feature map of the output of the last convolution layer;

Reducing the dimension of the feature map to obtain a single-layer feature vector and feature parameters thereof for representing each character in the effective inquiry text;

Deriving the characteristic parameters by using the optimal department classification result to obtain class activation mapping corresponding to the optimal department classification result;

Determining the key degree of each word in the effective inquiry text according to the result of the class activation mapping;

and extracting a plurality of keywords according to the keyword degree.

In one possible implementation manner, the preliminary department identification and text validity judgment on the input inquiry text adopts various combinations of the following modes: keyword matching, named entity extraction and statement legitimacy classification;

The keyword matching is used for carrying out keyword matching according to the inquiry text and a preset inquiry word bank, and if the keyword matching is completed, a corresponding department is directly identified;

The named entity extraction is used for extracting corresponding entity information from the inquiry text according to a preset entity type label, and if the extraction is completed, a corresponding department is directly identified;

the statement validity classification is used for judging whether the inquiry text is of an effective inquiry statement type.

In one possible implementation manner, the performing preliminary department identification and text validity judgment on the input inquiry text includes:

Firstly, carrying out keyword matching on the inquiry text;

if the query text is not matched with the query text, extracting the named entity;

if not, classifying the statement legitimacy of the inquiry text;

If the classification result is an illegal sentence, requesting to input a new inquiry text;

And if the classification result is legal statement, obtaining the effective inquiry text.

In one possible implementation manner, the determining the target recommended department using the department classification result and the department search result includes:

Fusing all the department classification results;

The department search results are utilized to carry out weight adjustment on the merged department classification results;

and obtaining at least one target recommended department according to the weight adjustment result.

In a second aspect, the present invention provides a department diagnosis guiding device, including:

the preliminary diagnosis guiding module is used for carrying out preliminary department identification and text validity judgment on the input inquiry text;

The deep diagnosis guiding module is used for obtaining a department classification result and a department retrieval result serving as candidates by utilizing a text classification strategy and a key information retrieval strategy based on the effective consultation text of the unidentified department;

The target determining module is used for determining a target recommended department by using the department classification result and the department retrieval result.

In one possible implementation manner, the in-depth diagnosis guiding module includes: a text classification sub-module;

The text classification submodule is used for outputting a plurality of department classification results by a plurality of classification models trained in advance.

In one possible implementation manner, the in-depth diagnosis guiding module further includes: a key information retrieval sub-module;

the key information retrieval submodule specifically comprises:

the key information determining unit is used for determining key information from the effective inquiry text in combination with the processing of the classification model;

And the department retrieval unit is used for retrieving the corresponding department in a preset department retrieval library by utilizing the key information.

the key information determining unit is specifically configured to:

In one possible implementation manner, the key information determining unit specifically includes:

the characteristic diagram acquisition component is used for acquiring the characteristic diagram output by the last convolution layer;

The single-layer characteristic parameter determining component is used for reducing the dimension of the characteristic diagram to obtain single-layer characteristic vectors and characteristic parameters thereof for representing the characters in the effective inquiry text;

the class activation mapping operation component is used for deriving the characteristic parameters by using the optimal department classification result to obtain class activation mapping corresponding to the optimal department classification result;

The keyword determining component is used for determining the keyword degree of each word in the effective inquiry text according to the result of the class activation mapping;

And the keyword extraction component is used for extracting a plurality of keywords according to the keyword degree.

In one possible implementation manner, the preliminary diagnosis guiding module adopts the following various unit combinations: the system comprises a keyword matching unit, a named entity extraction unit and a statement legitimacy classification unit;

the keyword matching unit is used for matching keywords according to the inquiry text and a preset inquiry word bank, and if the keywords are matched, corresponding departments are directly identified;

The named entity extraction unit is used for extracting corresponding entity information from the inquiry text according to a preset entity type label, and if the extraction is completed, the corresponding department is directly identified;

the statement validity classification unit is used for judging whether the inquiry text is of an effective inquiry statement type.

In one possible implementation manner, the preliminary diagnosis guiding module is specifically configured to:

firstly, carrying out keyword matching on the inquiry text by utilizing the keyword matching unit;

if the query text is not matched with the query text, extracting the named entity by using the named entity extraction unit;

If not, classifying the statement legitimacy of the inquiry text by using the statement legitimacy classification unit;

In one possible implementation manner, the target determining module includes:

the classification result fusion unit is used for fusing all the department classification results;

The weight adjusting unit is used for adjusting the weight of the merged department classification result by utilizing the department retrieval result;

the target recommendation department determining unit is used for obtaining at least one target recommendation department according to the weight adjusting result.

In a third aspect, the present invention provides a department diagnosis guiding apparatus, comprising:

One or more processors, a memory, and one or more computer programs, the memory may employ a non-volatile storage medium, wherein the one or more computer programs are stored in the memory, the one or more computer programs comprising instructions which, when executed by the device, cause the device to perform the method as in the first aspect or any of the possible implementations of the first aspect.

It should be understood that the second to third aspects of the present invention are unified with the technical solution concept of the first aspect of the present invention, and the advantages achieved by the aspects and the corresponding possible embodiments are similar.

In a fourth aspect, the present invention provides a computer readable storage medium having stored therein a computer program which when run on a computer causes the computer to perform the method as in the first aspect or any of the possible implementations of the first aspect.

In a fifth aspect, the invention also provides a computer program product for performing the method of the first aspect or any of the possible implementations of the first aspect, when the computer program product is executed by a computer.

In a possible design of the fifth aspect, the relevant program related to the product may be stored in whole or in part on a memory packaged with the processor, or may be stored in part or in whole on a storage medium not packaged with the processor.

The core concept of the invention is mainly based on two preconditions, 1) the patients can conduct department diagnosis inquiry autonomously and self-help, and a large amount of medical manpower is not required to be configured. 2) The huge patient group belongs to non-medical professionals, and two dimensional problems are overcome together when autonomous self-help inquiry is carried out, namely, the trouble of professional medical information is required to be removed from the aspect of patients; from the self-service system perspective, the recognition and analysis capability of fuzzy query information input by non-professional patients needs to be improved.

Therefore, the main scheme idea provided by the invention is that firstly, preliminary department identification and text validity judgment are carried out on the input consultation text, and some consultation texts which can be directly recommended or illegally input for the consultation are processed in advance, and then, effective illness state descriptions which cannot be directly recommended to the departments are deeply analyzed by adopting a double-strategy combined mode of text classification and key information retrieval, so that accurate and reliable recommended consultation departments are provided, patients are guided to smoothly visit the doctor, and the pressure of hospital consultation guiding consultation is relieved.

Furthermore, the invention combines the application of expert knowledge system and artificial intelligence technology text processing, and utilizes the pre-modeling for processing text classification task to classify the intelligent departments, thereby being capable of better identifying the doctor-seeing departments corresponding to unusual illness or fuzzy illness description so as to help patients to register and seek doctor more efficiently.

Furthermore, the invention also provides a text-based class activation mapping processing thought on the basis of a specific model algorithm, so that classified diagnosis and treatment results can be reordered by means of the extracted core keywords and corresponding retrieval operations, the accuracy and recall rate of the output various department results can be further improved by the process, and the recommended results can be more stable while the accuracy is effectively improved.

In summary, compared with the prior art, the invention can better solve the problem that the prior art cannot solve or ignore in the real medical application environment and various scene challenges.

Drawings

For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described with reference to the accompanying drawings, in which:

FIG. 1 is a flowchart of an embodiment of a department consultation method provided by the present invention;

FIG. 2 is a flow chart of a method for determining an embodiment of a target recommendation department;

FIG. 3 is a flow chart of a method of an embodiment of key information acquisition provided by the present invention;

FIG. 4 is a process schematic diagram of an embodiment of a key information retrieval strategy provided by the present invention;

FIG. 5 is a schematic block diagram of an embodiment of a department guide device according to the present invention;

fig. 6 is a schematic structural diagram of an embodiment of a department diagnosis guiding apparatus provided by the present invention.

Detailed Description

Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to like or similar elements or elements having like or similar functions throughout. The embodiments described below by referring to the drawings are illustrative only and are not to be construed as limiting the invention.

In terms of the technical scheme, the existing point selection type diagnosis guiding scheme can realize the self-help inquiry function of the user, but the judgment essence is a decision problem. Such guided diagnostic systems rely on preset templates, simply to let the patient choose one of a plurality of disorder options that matches his own situation, or to follow up with a question according to a certain option. For example, the system may require the patient to be filled with basic information, such as: age, sex, etc.; the system then requires the patient to select the location where discomfort occurs, such as head, neck, chest, abdomen, back, etc.; further selecting uncomfortable symptoms and then listing possible diseases according to the information; finally, the patient is provided with detailed information of the doctor department, the etiology, the symptoms, the examination items, the treatment methods and the like of the diseases. Such a lead regimen has at least the following disadvantages: on the one hand, depending on the size of the template, only a limited range of options is usually given, so that it is not guaranteed to cover every disorder description for every patient, and it is not guaranteed to be accurate for the recommended department; on the other hand, the professional requirements on patients are high, and the patients are required to select according to the corresponding medical terms of subjective feeling matching, for example, lumbosacral pain and lumbocrural pain are distinguished, which is not different from the fact that the use difficulty and the applicability of the self-help diagnosis guiding system are additionally increased.

The invention, at the beginning of design, is analyzed to determine that the essence behind the defects is that the problems of using feelings of huge non-professional patient groups and diversity brought by group sizes are ignored in the conception of the prior art. Therefore, the invention aims at carrying out deep recognition, analysis and judgment on the input information only by organically combining the original input information with a text processing technology without making strict requirements on the self-help input illness state description of the patient and repeated professional questioning, thereby providing accurate and reliable department diagnosis guiding recommendation.

Specifically, the present invention provides an embodiment of a method for guiding a doctor in a department, as shown in fig. 1, which may include the following steps:

step S1, performing preliminary department identification and text validity judgment on an input inquiry text;

Step S2, based on the effective consultation text of the department which is not identified, obtaining a department classification result and a department retrieval result which are candidates by utilizing a text classification strategy and a key information retrieval strategy;

and step S3, determining a target recommended department by using the department classification result and the department retrieval result.

The concept of the present invention is to decompose the department guidance task into three stages of each department's task but related to each other: preliminary identification, deep identification and target screening. The above-mentioned frame is formed, and is closely related to the above-mentioned problems of the prior art analyzed by the foregoing, because the present invention does not make excessive demands on the input information of the patient user, the preliminary identification stage can filter out the inquiry information which can directly determine the department and the input information which is illegal and has weak correlation with the inquiry, and the rest can not directly and quickly locate the department and belongs to legal and effective inquiry information, and then enter the deep identification stage. The object processed in the deep recognition stage is still original and is subjected to inquiry information after preliminary recognition screening, so that no 'unfriendly' professional problem is required to be added to a user, no other inquiry information is required to be given by the user, and a computer carries out 'self' judgment according to a prefabricated algorithm related to two text processes to obtain candidate department recommendation results; here, the candidate inevitably contains tentative results with different credibility, so that the candidate goes to the final target screening stage, and the target recommendation department is selected by combining different results of different text processing strategies.

Specifically, the present invention is not limited to the original user input format, for example, the original user input format may be a voice input format, a keyboard input format or a handwriting input format, and of course, text-image input may be included in some scenes, but any input mode may be processed into text format, and the conversion from voice and image to text format belongs to the prior art, and the process is not in the scope of the present invention.

After the user entered inquiry text is taken, the preliminary recognition phase may be entered. The text information is used for identifying the effective content and judging whether the grammar rules are legal or not, and the like, so that various existing mature technologies in the fields of text processing and natural language understanding can be used for reference, and one or more of means such as keyword matching, named entity extraction, sentence legality classification and the like can be utilized but not limited to. It will be appreciated by those skilled in the art that the specific means described above is not of particular importance to the present invention as such is a text processing tool, and that the technical concept of the present invention is derived from the analysis of technical problems in the foregoing. The particular type of tool used and how it is used in the preliminary recognition stage can be adjusted as desired, but for ease of understanding and implementation of the present invention, the tool functions described above are schematically illustrated:

(1) Keyword matching (which is not related to the keyword matching in the deep recognition stage mentioned later) is mainly used for performing keyword matching according to the inquiry text and a preset inquiry word bank, and if the keyword matching is performed, the corresponding department can be directly recognized.

More specifically, a keyword matching method of self-weighted editing distance can be adopted, and related description is made in combination with the prior art: the edit distance (EDIT DISTANCE) refers to the minimum number of edit operations required to switch from one to the other between two strings, and if their distance is greater, it means that they are different. This particular embodiment uses self-weighted edit distances (WEIGHTED EDIT DISTANCE) for keyword matching. First, a standard inquiry word stock may be preset, and a disease name word stock, a symptom name word stock, a standard part name word stock, an inspection name word stock, and the like may be included.

The word stock is recorded as C, each element is C _k, k epsilon n, and the word stock comprises the following components: c= { C ₁,c₂,c₃,...,c_n }, let the word length of l _c ^k＝c_k. When calculating the edit distance, based on the traditional formula, the weight is considered in each operation, for example, the edit distance between the input word w and the c _k words in the inquiry word stock is calculated, so that the word length of l _w =w can be set up, and the matrix d with the dimension of l _c ^k×l_w can be established. The weight function for any position of d (the physical meaning of the weight function refers to determining the weight according to the position and the word stock, for example, the weight can be monotonically increased according to the position) is as follows:

α[i][j]＝f(w[i],c_k[j]),i∈(0,l_w)j∈(0,l_c ^k]

when calculating the editing distance, taking one step of calculation formula as an example:

d[i][j]＝min{d[i][j]+α[i][j],d[i-1][j]+α[i-1][j],d[i][j-1]+α[i][j-1]}

Finally, the edit distance between the input word w and the c _k word in the inquiry word library is recorded as D, then the edit distance can be compared with a preset threshold T, if D is more than T, the matching is successful, and if D is less than T, the keyword c _k is recorded.

For example, the input inquiry text is that the sleep quality is poor, and after long-term insomnia is matched with a preset inquiry word stock, words with higher specificity such as keywords 'insomnia' are obtained from the inquiry text, so that the inquiry text can be directly positioned to neurology or sleeping departments, namely, preliminary recognition results can be given through the keyword matching, and the patients can be rapidly and directly guided to visit corresponding departments.

(2) And extracting named entity, wherein the named entity mainly aims at extracting entity information corresponding to the named entity from the inquiry text according to a label of a preset entity type, and if the named entity is extracted to a related entity, the corresponding department is directly identified.

More specifically, the Named Entity Recognition (NER) may identify entities with specific meaning in the text, and in this embodiment, five tags in the following table are set together, so that entity Recognition and extraction can be performed on the patient input inquiry text:

likewise, when an entity corresponding to a preset tag (indicating a higher specificity) is acquired, the entity can be directly locked to the corresponding department.

(3) Statement legitimacy classification, which is mainly used for judging whether the inquiry text is a valid inquiry statement type. That is, whether the current inquiry text is related to the inquiry requirement is judged through semantic understanding and set rules, which mainly considers that the input of the patient user is not demanding for the purpose of facing the large-scale non-professional patient group, and thus illegal input or invalid input interfering with subsequent processing can be included.

Those skilled in the art will appreciate that the above-described tool is presented by way of example only and not by way of limitation. The means and combinations thereof may be changed, and for example, the above (1) and (3) may be used in combination, the (2) and (3) may be used in combination, or the (1), 2, and (3) may be used together. The sequence of the means can be adjusted according to the actual situation, for example, the means are executed in parallel, and the execution of other means is stopped when the result is obtained first by which means; or sequentially, e.g., by performing validity classification before keyword matching and/or entity extraction.

The invention provides at least one preferable preliminary recognition flow by combining the processing experience of the actual scene, namely, firstly carrying out the keyword matching on the inquiry text, wherein the keyword matching can be carried out firstly because the establishment of a word stock has the universality and the coverage limitation, and the keyword matching can be carried out firstly, and the keyword matching can be carried out quickly according to the words with obvious specificity to the corresponding departments through the introduction; if the keywords are not matched, the text information is indicated to exceed the coverage range of the word stock, the named entity extraction can be further carried out on the inquiry text, and the named entity extraction algorithm has wider coverage range compared with the manually constructed word stock, so that the keyword extraction algorithm can be used as a supplementary tool of a keyword matching tool; if the entity corresponding to the label is not extracted on the basis, the text information is beyond the range of an entity extraction algorithm, which is a common phenomenon in medical scenes, such as the analysis, the number of patient groups and the individual difference can lead to the difference of the described contents, so that the department result can not be primarily identified through the tool processing, at least two possibilities can be considered, one is legal inquiry information that the content input by a user is not effective and is eliminated; or the other is that the content input by the user is effective legal inquiry information, but because the expression specificity is low, the tool means in the primary identification cannot be identified, and then the candidate result can be determined by the joint strategy through the subsequent deep identification stage. Therefore, after the entity fails to extract, the statement legitimacy classification can be performed to separate out the valid inquiry text or the invalid text, and it should be noted that, when the entity finally determines that the text is invalid, in other embodiments of the present invention, prompt information about invalid inquiry can be given, and the user is requested to reenter the inquiry information by returning to the initial step, which is not limited in this embodiment.

In addition, it should be noted that, before the preliminary recognition, the obtained query text may be subjected to text preprocessing, for example, the high-dimensional space sentence input by the user may be embedded (embedding) into a low-dimensional continuous vector space according to the word2vec technology. And dividing the obtained data into long data and short data by taking the word number of each inquiry sentence after word segmentation as a standard according to the characteristics of the CNN and LSTM models. The long and short data classification standard may be that the number of segmented words is compared with a preset length (for example, 30), and data with a length greater than 30 is classified as long data, otherwise, classified as short data. For facilitating the subsequent operation, the lengths of the data can be unified, for example, the length of the long data is 150 by filling a0 value (padding), and if the length of the original data exceeds 150, the length of the original data is truncated at 150; or short data is made to have a length of 30 by filling a value of 0. Of course, text preprocessing methods conventional in the art, such as removal of stop words, may also be performed, and the present invention is merely exemplified above, and will not be described herein.

After preliminary processing, the input information which is not rapidly identified but belongs to the effective inquiry text can be obtained, and then the multi-strategy decision process combined by text classification and key information retrieval is carried out, namely, the legal inquiry with lower specificity is further identified by the deep identification stage, and the candidate department result is output, and the process is specifically described below. The target screening stage of determining the target department from the candidates is described first.

Text classification technology and key information retrieval technology have various optional tools in the field, as mentioned above, the core processing framework of the present invention does not focus on the isolated tools, but proposes a processing scheme integrating the interrelations and their own roles for the back cause of the deep-dug prior art problem. The text classification and retrieval tools available in the deep recognition stage may be multi-choice, and it is generally possible to obtain a score for each classification category and retrieval result after classifying and retrieving text in the art, where the score indicates the likelihood (e.g., probability value) of the classified category and the relevance (e.g., similarity score) of the retrieved category. In other words, the candidate department classification results and department search results may include a plurality of pending department names and corresponding processing scores obtained after classification and search, and in some embodiments may further be based on the scores, "cull" all classification results and "cull" all search results, e.g., department classification results may ultimately be ranked in score to obtain TOP5, and department search results may be ranked in score to obtain TOP3. Therefore, in the target screening stage, the names of the candidate departments may be used for screening, or all the candidate scores may be used for screening, for example, candidate results (for example, the candidate results with the repetition number of the names of the departments meeting the threshold value (for example, the candidate results are repeated in 8 candidates and the number of times is the greatest in the above example) may be extracted as target departments, or the candidate results may be directly sorted according to the scores of the candidate departments (for example, the scores of the candidate 8 candidates in the above example), and a plurality of candidate results with the scores meeting the threshold value may be selected as target departments. Of course, the name of the department and the score can be considered together to be taken as the screening basis, and the invention is not limited.

Here, the present invention further provides another preferred method for determining a target recommended department, as shown in fig. 2, which may include:

S31, fusing all the department classification results;

step S32, weighting adjustment is carried out on the merged department classification result by utilizing the department retrieval result;

And step S33, obtaining at least one target recommended department according to the weight adjustment result.

The key point of the processing thought is to update the candidate result of one strategy and take the candidate result of the other strategy as an increment weight, so that the updated strategy output result is more controllable, and particularly when a plurality of candidate results exist and a plurality of target departments which are finally recommended are expected, the candidate result with a relatively low score can be in a reasonable state of 'no-outlier'. Therefore, it can be understood that in some embodiments, the department search results may be fused first, and then the department classification results may be used to adjust the weight of the department search results, which is not described herein. It should be noted that, as illustrated in fig. 2, the fusion of all department classification results described herein includes at least two layers of meaning:

First, when a classification tool is adopted, the fusion can be understood that a plurality of department classification results output by the tool can be first fused internally, for example, but not limited to, all classification results of TOP5 can be subjected to score averaging operation, and the average value is used for further secondary filtering of department classification results (for example, TOP5 is selected as TOP2 according to the average value); or the average value is reserved, then when the weight adjustment operation is carried out by using the department search result, the search score and the average value (other fusion modes are not limited to the average value) can be weighted, summed, calculated and the like, and finally, each candidate result is reordered according to the final score, so that the target recommended department is obtained. It is to be understood that the above is only illustrative.

Secondly, because the invention does not limit the number of specific tools used by the text classification strategy and the key information retrieval strategy, when the classification tools adopt a plurality of types, the fusion can be understood as the fusion operation between a plurality of department classification results output by a plurality of classification tools, for example, but not limited to, selecting the repeated department names given by a plurality of classification tools, or accumulating the corresponding scores of the repeated departments given by a plurality of classification tools, reordering all department classification results after the scores are updated, and extracting TOPn with the top score. Similarly, the foregoing is merely illustrative.

Then, the weight adjustment is performed on the merged department classification results by using the department search results, which means that the merged current several department classification results are regarded as objects, and the importance degree of the related objects is improved by using the department detection results, namely, the department classification results are optimized. It should be noted that the weight adjustment is not limited to the adjustment of the score and the like, and is a weight adjustment as long as the importance of a certain object can be affected, for example, the object departments appearing in the merged department classification result and the department detection result are selected as the target recommendation side (of course, if the score level is considered, the score of the department detection result may be multiplied by the merged department classification result with the same name as described above, so that the score is "stand out" in the merged department classification result; if the result is not repeated, the department detection result can be used as a reference, the related departments can be searched in the merged department classification result, for example, the heart internal medicine appears in the search result, and the heart internal medicine appears in the merged classification result although the heart internal medicine does not exist, so that the importance degree of the heart internal medicine in the merged department classification result can be improved by the heart internal medicine of the search result, and of course, the related departments can be compared based on the literal similarity, and the semantic-level association can be performed by combining expert knowledge, for example, the sleep department and the neurology department have strong correlation, and the weight increment operation can be implemented between the two.

The weight adjustment is an optional measure without limitation, and is also applicable to the implementation mode of performing weight adjustment on the department search result by the department classification result, and finally, after the importance adjustment, a plurality of candidate departments which are out of the list can be used as target recommended departments.

Here, a multi-strategy decision process combined by text classification and key information retrieval will be specifically described, where the combination refers to parallel independent or interrelated collaboration of strategies with multiple different processing dimensions, and the purpose of the combination is to avoid limitation of a single processing strategy, so as to obtain more reliable or more widely covered candidate results.

The above-mentioned text classification method can be used as a reference, and thus, in some preferred embodiments, in combination with text processing technology in the field of artificial intelligence, it is proposed that in at least one possible implementation of the above-mentioned text classification strategy can refer to outputting a plurality of department classification results by a plurality of classification models trained in advance. The number, architecture and specific type of the classification model are not limited as such. Two preferred text processing models for which the concepts of the present invention may be applied are described herein for reference:

A Self-attention two-way long-short-term memory artificial neural network (Self-attention BiLSTM) for text classification processing.

RNN (Recurrent Neural Network) is a relatively common text classification method, and long-term memory models (Long Short Term Memory, LSTM) are proposed later to solve the problems of gradient explosion and gradient disappearance in RNN. The problems of RNN are solved by adding input gates, output gates and forget gates. The self-Attention model (self-Attention model) then utilizes the Attention mechanism to dynamically generate weights for different connections, processing variable length information sequences.

Assuming that the length of the inquiry text sentence is n and the dimension of each word embedding is k, self-Attention BiLSTM (hereinafter referred to simply as Self-Att) is input as an n×k matrix, and each time-step processes one word and completes a whole sentence through n time-steps. Taking the t moment as an example, the corresponding word at the t moment is x _t, the hidden state at the last moment is h _t-1,h_t-1 dimension is m, and the calculation method is as follows:

a) Forgetting the door: f _t＝σ(W_f[h_t-1,x_t]+b_f) outputting a value between 0 and 1, which determines how much history information to forget. 1 means complete retention, 0 means complete rejection.

B) An input door: i _t＝σ(W_i[h_t-1,x_t]+b_i)

C′_t＝tanh(W[h_t-1,x_t]+b_c)

And c _t＝f_t*c_t-1+i_t*C′_t, updating the current cell state, determining which historical information flows into the current cell (forgetting gate control), and determining which new information flows into the cell (input gate control).

C) Output door: o _t＝σ(W_o[h_t-1,x_t]+b_o)

h_t＝o_t*tanh(c_t)

Therefore, how much information in the output value flows into the hidden layer is determined, biLSTM is the combination of the forward LSTM and the backward LSTM, and only the forward and backward h _t is needed to be spliced to obtain a new h _t. Transforming h _t can obtain Y _t, and each moment can obtain corresponding Y _t, and finally, an output sequence y= { Y ₁,y₂,y₃,...,y_t,...y_n } containing n terms (the same as the sentence length) can be obtained.

Then, the Self-Attention mechanism is entered. Assume that: y _i ε Y, and i ε n, Y _i dimension is m. The Self-Attention calculation process is as follows:

v _i＝y_iW_v, wherein W _v has dimensions m x m, giving v _i a dimension m.

Beta _i＝q_i·y_i, the attention profile beta _i for each Y _i (each word) in Y, is a scalar.

The dimension is m for the final output of the vector from Self-Att.

Finally, the classification result and the score thereof can be obtained through a series of full connection layers and Softmax classification, and the department classification result TOPn under the tool can be obtained according to the above example of the present invention, and can be marked as a candidate department name R _LSTM and a score S _LSTM thereof.

And (II) convolutional neural network model (Text-Convolutional Neural Network, textCNN) for Text classification processing.

TextCNN is equivalent to a convolutional neural network in picture classification, which extracts features through convolutional operation by a plurality of fixed-length kernels, obtains simplified features through a pooling layer (pooling), and finally obtains classification results through a full-connection layer. When the convolution kernel length is n, the method is similar to the generation of n-gram features, and can capture local semantic information, so that a better effect can be obtained on local feature extraction.

Assuming that the ith word x _i∈R^k in the inquiry text sentence is a k-dimensional vector, a sentence of length n can be expressed asWherein the method comprises the steps ofRepresenting the concatenated symbols. X _i:i+j represents the juxtaposition vector of x _i,x_i:i+1 up to x _i:i+j. The filter w epsilon R ^hk is designed in each convolution layer, the width of the filter is consistent with k, and h words can be covered to generate a new feature map. For example, for feature c _i generated by overlay x _i:i+h-1, there is a calculation:

c_i＝f(wx_i:i+h-1+b)

Where b ε R is the bias and f is the activation function. When this filter is applied to the whole sentence: { x _1:h,x_2:h+1...x_n-h+1:n }, a feature vector can be obtained, and from the above illustration, a total of m '=p×m convolution kernels can be found, and there are m' feature vectors:

q_i＝[c₁,c₂,...,c_n-h+1],c∈R^n-h+1，i∈m′

The largest pooling is performed on each feature vector, and finally, the feature vectors are spliced into vectors with the dimension of m', and classification results and scores thereof can be obtained through the full connection layer and Softmax, and in the above example of the invention, department classification results TOPn under the tool can be obtained and can be recorded as candidate department names R _CNN and scores S _CNN.

In the above, it will be understood by those skilled in the art that the present invention does not limit the number and options of the modes used by the above text classification policies, nor does it limit the modes used by the key information retrieval policies, so that various existing schemes, such as recommendation ordering, semantic matching, association rules, etc. related to machine learning, can be used for the key information retrieval policies; it should be noted that the key information retrieval here is not the same processing logic as the keyword matching mentioned in the prior preliminary recognition, which is aimed at the desire to locate the corresponding department quickly and directly from the user input, so that the implementation thereof, for example the construction of the question word stock, can be relatively simple without the need for complex retrieval processing tools.

In connection with the foregoing, the present invention provides a key information retrieval preferred scheme associated and cooperated with a classification policy, that is, the specific implementation of the retrieval policy may combine with the processing procedure of the classification model to determine key information from the effective query text, and then use the key information to perform the retrieval of the corresponding department in the preset department retrieval library. The concept of the preferred scheme is that key information retrieval is not carried out on effective inquiry texts in an isolated mode, the contribution of the classification model is used as the basis or condition of a retrieval strategy, so that the retrieval result tends to be in a controllable range, and the method has operability and pertinence particularly for carrying out importance adjustment on classification strategy output by utilizing the output of the retrieval strategy.

The above-mentioned preferred key information retrieval method can be understood as being divided into a key information acquisition phase and an actual retrieval phase. For the key information obtaining stage, when different classification models are adopted, different corresponding schemes can be generated, and the text-based convolutional neural network embodiment is selected in combination with the classification models, so that the invention correspondingly provides a specific key information obtaining mode, and the implementation process can be summarized as a text-based class activation mapping mode: firstly, determining characteristic parameters corresponding to each word in the effective inquiry text according to the output of a convolution layer of the convolution neural network; and determining the distribution condition of the keywords in the effective inquiry text based on the optimal department classification result of the convolutional neural network and the characteristic parameters.

It can be stated that the ultimate goal of using this concept is to expect core key information that is more important to the outcome, and in particular, the present invention proposes Text-based class activation mapping (Text-based Classify Activation Mapping, text-CAM) to obtain core keywords in valid questionnaires Text related to the optimal department classification outcome, the principle of which is based on the class activation mapping (Class Activation Mapping, CAM) algorithm that replaces the fully connected layers in the traditional CNN network structure with GAP layers and takes them as features to derive the fully connected layers of the classification. Through the simple connection structure, important areas in the picture can be marked in a mode of mapping output layer weights back to the characteristics of the convolution layers. However, since most CNN networks currently do not use GAP (global averaging pooling) as feature extraction, the prior art requires modification of the network structure and retraining of the model in implementations using the CAM described above.

To this end, the present invention proposes a method of extracting key information without requiring GAP layer and without changing network structure, namely the above Text-based class activation map (Text-based Class Activation Mapping, text-CAM). The Text-CAM can form characteristic parameters corresponding to each word in the Text original sentence, and further judge the focus of the highest-score classification result output by the CNN relative to the Text original sentence, so that the distribution condition of key information is extracted from the Text original sentence.

More specifically, the implementation illustrated in fig. 3 may be employed in some embodiments of the present invention, and the critical information acquisition phase specifically includes:

S21, obtaining a feature map output by a last convolution layer;

S22, reducing the dimension of the feature map to obtain a single-layer feature vector and feature parameters thereof for representing each character in the effective inquiry text;

s23, deriving the characteristic parameters by using the optimal department classification result to obtain class activation mapping corresponding to the optimal department classification result;

Step S24, determining the key degree of each word in the effective inquiry text according to the result of the class activation mapping;

And S25, extracting a plurality of keywords according to the key degree.

For the embodiment of fig. 3, the operation level is illustrated again by way of example as follows: assuming that the vector dimension of the sentence in the effective inquiry text is n×k, the number of convolution kernels of the last layer of TextCNN is m ', the dimension of the feature map after convolution is n×m'. Wherein each column represents a feature vector. Any feature vector keeps the dimension consistent with the sentence length through padding, namely, the feature vector q _i, i epsilon m' and the dimension n. Based on this, the extraction of key information can be divided into two steps:

The first step, converting the multilayer feature into a single-layer feature, comprises the following steps:

Summing each q _i yields the parameter γ _i: gamma _i is a scalar, i e m'.

Vectorizing m' gamma to obtain gamma= [ gamma ₁,γ₂,γ₃,...,γ_i ].

And then, each row of the feature map is operated, so that each behavior vector p _j, j epsilon n, the dimension m', p _j respectively correspond to n words in the original sentence. Each p _j is multiplied by gamma to obtain a characteristic parameter f _j corresponding to each word: f _j is a scalar, j ε n.

Vectorizing n f _j, and finally converting the feature map with the dimension of n×m' into a single-layer feature parameter vector f with the dimension of n (same as the sentence length): f= [ f ₁,f₂,f₃,...,f_n ].

And secondly, deriving the single-layer characteristics. Assuming that the optimal class determined by TextCNN is c, which is scored as y _c, deriving for each f _j, j∈n:

And obtaining key feature distribution through normalization: Wherein the criticality of each word J ε n. In actual operation, a threshold lambda may be set, and if x _j > lambda, the information is determined to be a keyword.

The actual search stage may be understood as independent of the above-mentioned key information acquisition stage, that is, no matter what key information acquisition mode is adopted, in the actual search stage, search operations may be performed around the department corresponding to similar query data and/or the degree of correlation of the text corresponding to similar query data in a preset department search database (a department search database based on a large number of cases and query data may be specifically constructed by an expert knowledge system and a data acquisition technology).

In combination with the above preferred key information obtaining method, a complete key information retrieval strategy is shown in fig. 4, and a key word _Text-CAM of an effective query Text can be obtained through the Text-CAM, so that the key word _Text-CAM is used for retrieving in a department retrieval library containing disorder information and corresponding departments of medical treatment. It will be appreciated by those skilled in the art that the effective query text may be subjected to a pre-text process, such as word segmentation, word de-activation, etc. (note that the text pre-process may refer to a pre-process operation prior to the "preliminary recognition stage" or may refer to a re-pre-process operation after the effective query text is obtained via the "preliminary recognition stage"). The present invention may then, in some embodiments, further employ a more sophisticated retrieval system to accomplish the actual retrieval, such as, but not limited to, an elastic search (hereinafter ES) using the algorithm BM25, which returns a specified number of retrieval results for each valid query text and a similarity score for each retrieval result at the same time. In the manner of the previous example of the present invention, the department search result TOPn obtained by the tool may be noted as a candidate department name R _ES and its score S _ES.

In view of the above description of the in-depth recognition phase, the above-mentioned target screening phase may be described schematically as follows. In the schematic illustration, the text classification strategy adopts two tools, namely Self-Attention BiLSTM and TextCNN, and the key information retrieval strategy is to derive a class activation mapping extraction keyword for optimal classification according to a key feature vector generated by TextCNN, and retrieve candidate results in a preset department retrieval library.

Specifically, first, textCNN and a plurality of candidate departments obtained by Self-Attention BiLSTM can be subjected to fusion ranking. The actual operation mode can be that R _CNN and R _LSTM are compared, and the scores S _CNN and S _LSTM of repeated departments are added to obtain S _CNN-LSTM; and according to the updated score, TOP5 can be taken to construct an updated candidate department list and marked as R. Then comparing with R _ES obtained by Keyword _Text-CAM retrieval, if the department in R _ES appears in the updated candidate department list, multiplying S _ES by S _CNN-LSTM, and normalizing to obtain a final score; and finally, rearranging the list, and taking TOP2 with the highest score as a target recommended department to output. Score calculation for candidate department list j:

Finally, the deep recognition stage and the target screening stage are comprehensively described by an example: the effective inquiry text obtained through the preliminary identification stage is assumed to be' sudden naris bleeding when the old people eat meal at night and prepare for sleeping, measures such as high lifting, pressing and hemostasis by both hands are not stopped, dizziness, hypertension and hyperglycemia at ordinary times, heart diseases are caused, and the old people take a controlled medicine all the time after taking a bracket.

Firstly, the word segmentation can be carried out to obtain the 'old man/yesterday/night/eat meal/prepare/sleep/time/sudden/nostril/bleeding/high lift/both hands/pressing/hemostasis/waiting/measure/both/stopping/dizziness/normal times/hypertension/hyperglycemia/heart disease/done/stent/always/eat/control/medicine'. Then go through padding (long data fills 0 to 150) and perform word embedding according to word2vec to form word vectors with dimensions 150×300:

The results R _LSTM,S_LSTM are obtained by means of the Self-Attention BiLSTM model as follows:

Department name	Score value
		Cardiovascular internal medicine	0.6527
Dizziness department	0.0748
		Otorhinolaryngology department	0.0599
Neurosurgery department	0.0194
		Gastroenterology	0.0186

The results R _CNN,S_CNN were obtained by means of the TextCNN model as follows:

Department name	Score value
		Cardiovascular internal medicine	0.7851
Neurosurgery department	0.0512
		Dizziness department	0.0352
Endocrinology department	0.0121
		Department of cardiology	0.0102

The keyword list L is obtained by Text-CAM:

L= [ press, heart disease, yesterday evening, hyperglycemia, hypertension, stent, complete eating ]

Searching in a department search library according to L to obtain a search result R _ES (search score not shown):

high similarity text (after stopping words)	Department name
		Blood pressure, high pressure and low pressure of completed covered stent	Cardiovascular internal medicine
Hypertension heart disease hyperglycemia blood pressure noon hypertension	Cardiovascular internal medicine
		Blood pressure heart stent mounting	Cardiovascular internal medicine
Hypertension heart disease	Cardiovascular internal medicine
		Heart disease hypertension does nothing	Cardiovascular internal medicine

And calculating S _LSTMS_CNNS_ES to obtain a final recommended department list R and a corresponding score S. After reordering, top2 is taken as a recommended department, so that the final target recommended department R is:

R= { "cardiovascular department": 0.99, "vertigo family": 0.11}

It should be noted that the text, description of the condition, recommendation of the department, calculation of score, etc. in the search library referred to in the above examples are only schematic and are not referred to for accuracy or rationality.

In summary, the main scheme idea provided by the invention is that firstly, preliminary department identification and text validity judgment are carried out on the input consultation text, and some of the input consultation texts can be directly subjected to department recommendation or illegal consultation input advanced treatment, so that effective illness state descriptions which cannot be directly recommended to departments are deeply analyzed by adopting a double-strategy combination mode of text classification and key information retrieval, thereby providing accurate and reliable recommended consultation departments, guiding patients to smoothly visit the doctor, and further relieving the pressure of hospital consultation guiding consultation.

Therefore, compared with the prior art, the invention can better solve the problem that the prior art cannot solve or ignore in the real medical application environment and various scene challenges.

Corresponding to the above embodiments and preferred solutions, the present invention further provides an embodiment of a department diagnosis guiding device, as shown in fig. 5, which may specifically include the following components:

The preliminary diagnosis guiding module 1 is used for carrying out preliminary department identification and text validity judgment on an input inquiry text;

The deep diagnosis guiding module 2 is used for obtaining a department classification result and a department retrieval result serving as candidates by utilizing a text classification strategy and a key information retrieval strategy based on the effective consultation text of the unidentified department;

And the target determining module 3 is used for determining a target recommended department by using the department classification result and the department retrieval result.

the key information retrieval submodule specifically comprises:

the key information determining unit is specifically configured to:

In one possible implementation manner, the target determining module includes:

It should be understood that the division of the components of the department diagnosis guiding apparatus shown in fig. 5 is only a division of a logic function, and may be fully or partially integrated into a physical entity or may be physically separated. And these components may all be implemented in software in the form of a call through a processing element; or can be realized in hardware; it is also possible that part of the components are implemented in the form of software called by the processing element and part of the components are implemented in the form of hardware. For example, some of the above modules may be individually set up processing elements, or may be integrated in a chip of the electronic device. The implementation of the other components is similar. In addition, all or part of the components can be integrated together or can be independently realized. In implementation, each step of the above method or each component above may be implemented by an integrated logic circuit of hardware in a processor element or an instruction in the form of software.

For example, the above components may be one or more integrated circuits configured to implement the above methods, such as: one or more Application SPECIFIC INTEGRATED Circuits (ASIC), or one or more microprocessors (DIGITAL SINGNAL Processor (DSP), or one or more field programmable gate arrays (Field Programmable GATE ARRAY; FPGA), etc. For another example, these components may be integrated together and implemented in the form of a System-On-a-Chip (SOC).

In view of the foregoing examples and their preferred embodiments, those skilled in the art will appreciate that in practice the present invention is applicable to a variety of embodiments, and the present invention is schematically illustrated by the following carriers:

(1) A department guide apparatus may include:

One or more processors, memory, and one or more computer programs, wherein the one or more computer programs are stored in the memory, the one or more computer programs comprising instructions, which when executed by the device, cause the device to perform the steps/functions of the foregoing embodiments or equivalent implementations.

Fig. 6 is a schematic structural diagram of an embodiment of the office guiding device according to the present invention, where the device may be an electronic device or a circuit device built in the electronic device. The electronic equipment can be a cloud server, a mobile terminal (such as a mobile phone), an intelligent screen, self-service interaction equipment, a robot and the like. The specific form of the department diagnosis guiding apparatus is not limited in this embodiment.

As shown in fig. 6 in particular, the department guide device 900 includes a processor 910 and a memory 930. Wherein the processor 910 and the memory 930 may communicate with each other via an internal connection, and transfer control and/or data signals, the memory 930 is configured to store a computer program, and the processor 910 is configured to call and execute the computer program from the memory 930. The processor 910 and the memory 930 may be combined into a single processing device, more commonly referred to as separate components, and the processor 910 is configured to execute program code stored in the memory 930 to perform the functions described above. In particular implementations, the memory 930 may also be integrated within the processor 910 or separate from the processor 910.

In addition, in order to further improve the functionality of the department guide device 900, the device 900 may further include one or more of an input unit 960, a display unit 970, an audio circuit 980, a camera 990, a sensor 901, etc., which may further include a speaker 982, a microphone 984, etc. Wherein the display unit 970 may include a display screen.

Further, the department guide device 900 may further include a power supply 950 for providing power to various devices or circuits in the device 900.

It should be appreciated that the department guide apparatus 900 shown in fig. 6 can implement the respective procedures of the method provided in the foregoing embodiment. The operations and/or functions of the various components in the device 900 may be respectively for implementing the corresponding flows in the method embodiments described above. Reference is specifically made to the foregoing descriptions of embodiments of methods, apparatuses and so forth, and detailed descriptions thereof are appropriately omitted for the purpose of avoiding redundancy.

It should be understood that, the processor 910 in the department guide device 900 shown in fig. 6 may be a system on a chip SOC, where the processor 910 may include a central processing unit (Central Processing Unit; hereinafter referred to as a CPU), and may further include other types of processors, for example: an image processor (Graphics Processing Unit; hereinafter referred to as GPU) or the like, as will be described in detail below.

In general, portions of the processors or processing units within the processor 910 may cooperate to implement the preceding method flows, and corresponding software programs for the portions of the processors or processing units may be stored in the memory 930.

(2) A readable storage medium having stored thereon a computer program or the above-mentioned means, which when executed, causes a computer to perform the steps/functions of the foregoing embodiments or equivalent implementations.

In several embodiments provided by the present invention, any of the functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer-readable storage medium. Based on such understanding, certain aspects of the present invention may be embodied in the form of a software product as described below, in essence, or as a part of, contributing to the prior art.

(3) A computer program product (which may comprise the apparatus described above) which, when run on a terminal device, causes the terminal device to perform the department guide method of the previous embodiment or equivalent.

From the above description of embodiments, it will be apparent to those skilled in the art that all or part of the steps of the above described methods may be implemented in software plus necessary general purpose hardware platforms. Based on such understanding, the above-described computer program product may include, but is not limited to, an APP; in connection with the foregoing, the device/terminal may be a computer device (e.g., a mobile phone, a PC terminal, a cloud platform, a server cluster, or a network communication device such as a media gateway, etc.). Moreover, the hardware structure of the computer device may further specifically include: at least one processor, at least one communication interface, at least one memory and at least one communication bus; the processor, the communication interface and the memory can all communicate with each other through a communication bus. The processor may be a central Processing unit CPU, DSP, microcontroller or digital signal processor, and may further include a GPU, an embedded neural network processor (Neural-network Process Units; hereinafter referred to as NPU) and an image signal processor (IMAGE SIGNAL Processing; hereinafter referred to as ISP), where the processor may further include an ASIC (application specific integrated circuit) or one or more integrated circuits configured to implement embodiments of the present invention, and the processor may further have a function of operating one or more software programs, where the software programs may be stored in a storage medium such as a memory; and the aforementioned memory/storage medium may include: nonvolatile Memory (nonvolatile Memory), such as a non-removable magnetic disk, a USB flash disk, a removable hard disk, an optical disk, and so forth, and Read-Only Memory (ROM), random access Memory (Random Access Memory; RAM), and so forth.

In the embodiments of the present invention, "at least one" means one or more, and "a plurality" means two or more. "and/or", describes an association relation of association objects, and indicates that there may be three kinds of relations, for example, a and/or B, and may indicate that a alone exists, a and B together, and B alone exists. Wherein A, B may be singular or plural. The character "/" generally indicates that the context-dependent object is an "or" relationship. "at least one of the following" and the like means any combination of these items, including any combination of single or plural items. For example, at least one of a, b and c may represent: a, b, c, a and b, a and c, b and c or a and b and c, wherein a, b and c can be single or multiple.

Those of skill in the art will appreciate that the various modules, units, and method steps described in the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations of electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Those skilled in the art may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

And, each embodiment in the specification is described in a progressive manner, and the same and similar parts of each embodiment are mutually referred to. In particular, for embodiments of the apparatus, device, etc., as they are substantially similar to method embodiments, the relevance may be found in part in the description of method embodiments. The above-described embodiments of apparatus, devices, etc. are merely illustrative, in which modules, units, etc. illustrated as separate components may or may not be physically separate, i.e., may be located in one place, or may be distributed across multiple places, e.g., nodes of a system network. In particular, some or all modules and units in the system can be selected according to actual needs to achieve the purpose of the embodiment scheme. Those skilled in the art will understand and practice the invention without undue burden.

The construction, features and effects of the present invention are described in detail according to the embodiments shown in the drawings, but the above is only a preferred embodiment of the present invention, and it should be understood that the technical features of the above embodiment and the preferred mode thereof can be reasonably combined and matched into various equivalent schemes by those skilled in the art without departing from or changing the design concept and technical effects of the present invention; therefore, the invention is not limited to the embodiments shown in the drawings, but is intended to be within the scope of the invention as long as changes made in the concept of the invention or modifications to the equivalent embodiments do not depart from the spirit of the invention as covered by the specification and drawings.

Claims

1. A department consultation guiding method, which is characterized by comprising the following steps:

Based on the effective consultation text of the department which is not identified, obtaining a department classification result and a department retrieval result which are candidates by utilizing a text classification strategy and a key information retrieval strategy; the text classification strategy comprises outputting a plurality of department classification results by a plurality of classification models trained in advance; the key information retrieval strategy comprises the following steps: determining key information from the effective consultation text in combination with the processing of the classification model; searching corresponding departments in a preset department search library by utilizing the key information;

Wherein the classification model comprises a convolutional neural network; the determining key information comprises a text-based class activation mapping mode: determining characteristic parameters corresponding to each word in the effective inquiry text according to the output of the convolution layer of the convolution neural network; determining the distribution situation of keywords in the effective inquiry text based on the optimal department classification result of the convolutional neural network and the characteristic parameters;

2. The department guide method according to claim 1, wherein the determining key information from the effective inquiry text in combination with the processing of the classification model specifically includes:

acquiring a feature map of the output of the last convolution layer;

and extracting a plurality of keywords according to the keyword degree.

3. The department guide method according to claim 1, wherein the preliminary department identification and text validity judgment are performed on the inputted inquiry text by adopting various combinations of the following modes: keyword matching, named entity extraction and statement legitimacy classification;

4. The department guide method according to claim 3, wherein said performing preliminary department identification and text validity judgment on the inputted inquiry text comprises:

Firstly, carrying out keyword matching on the inquiry text;

if not, classifying the statement legitimacy of the inquiry text;

5. The department consultation guiding method according to any one of claims 1 to 4, characterized in that the determining a target recommended department using the department classification result and the department retrieval result includes:

Fusing all the department classification results;

6. A department diagnosis guiding device, comprising:

The deep diagnosis guiding module is used for obtaining a department classification result and a department retrieval result serving as candidates by utilizing a text classification strategy and a key information retrieval strategy based on the effective consultation text of the unidentified department; the text classification strategy comprises outputting a plurality of department classification results by a plurality of classification models trained in advance; the key information retrieval strategy comprises the following steps: determining key information from the effective consultation text in combination with the processing of the classification model; searching corresponding departments in a preset department search library by utilizing the key information;

7. The department consultation guide of claim 6, wherein the in-depth consultation guide module comprises: a text classification sub-module;

8. The department screening device of claim 7, wherein the in-depth screening module further comprises: a key information retrieval sub-module;

the key information retrieval submodule specifically comprises:

9. A department consultation guiding device, comprising:

one or more processors, memory, and one or more computer programs, wherein the one or more computer programs are stored in the memory, the one or more computer programs comprising instructions, which when executed by the device, cause the device to perform the department screening method of any one of claims 1-5.

10. A computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, which when run on a computer causes the computer to perform the department guide method as claimed in any one of claims 1 to 5.