Methods and devices for interpreting articles

Abstract

The embodiments of this specification provide a method and device for interpreting an article, wherein the method for interpreting an article includes: extracting multiple corresponding entities from the article according to multiple entity types used to describe the article; performing relationship extraction using the extracted multiple entities and the text of the article to obtain the association relationship between the entities; constructing a multi-tuple forest for describing the article based on the association relationship between the entities, wherein the multi-tuple forest includes one or more tree structures with entities as nodes and association relationships between entities as edges; and inputting the multi-tuple forest into an interpretation rule engine to obtain an interpretation result. In view of the complexity of the article text, this method structures the article text into a multi-tuple forest, and the multi-tuple forest accurately and completely describes the content of the article. The multi-tuple forest is then input into an interpretation rule engine constructed based on the multi-tuple forest structure for interpretation, and can accurately obtain the interpretation result, thereby achieving accurate and efficient article interpretation.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of computer technology, and more particularly to a method for interpreting an article. One or more embodiments of this specification also relate to a device for interpreting an article, a computing device, and a computer-readable storage medium. Background Art

[0002] In different scenarios, there are various types of articles that reflect work situations, product notifications, contract notifications, health notifications, etc. Since the content of the articles often contains a lot of professional knowledge, it is sometimes difficult for readers to understand the true meaning of the articles, which leads to misunderstandings, making decisions that do not conform to the articles, buying wrong products, and other problems. Therefore, it is very necessary to help people interpret the articles.

[0003] At present, traditional article interpretation is generally done offline with the help of experts. With the development of the Internet industry, some companies have also begun to adopt online article interpretation, but in essence, it is still done manually by experts online. When the expert experience is limited or the amount of articles is large, it is impossible to provide accurate and efficient interpretation services. Summary of the invention

[0004] In view of this, the embodiments of this specification provide a method for interpreting an article. One or more embodiments of this specification also relate to an apparatus for interpreting an article, a computing device, and a computer-readable storage medium to solve the technical defects existing in the prior art.

[0005] According to the first aspect of the embodiments of this specification, a method for interpreting an article is provided, comprising: extracting a plurality of corresponding entities from the article according to a plurality of entity types used to describe the article; performing relationship extraction using the extracted plurality of entities and the text of the article to obtain association relationships between entities; constructing a multi-tuple forest for describing the article based on the association relationships between the entities, the multi-tuple forest comprising one or more tree structures with entities as nodes and association relationships between entities as edges; and inputting the multi-tuple forest into an interpretation rule engine to obtain an interpretation result.

[0006] Optionally, the method uses the extracted multiple entities and the text of the article to perform relationship extraction to obtain the association relationship between entities, including: using the extracted multiple entities and the text of the article to construct a text and entity-based graph network; using a graph convolution algorithm to perform feature extraction on entity nodes in the graph network to obtain vectors of entity nodes; and obtaining the association relationship between entities by calculating the vector similarity between entity nodes in the graph network.

[0007] Optionally, the method utilizes the extracted multiple entities and the text of the article to construct a text- and entity-based graph network, including: encoding each sentence and each entity in the text of the article as an independent node to obtain sentence nodes and entity nodes of the graph network; and completing the construction of the graph network by establishing an edge between each pair of sentence nodes, establishing an edge between each pair of entity nodes in the same sentence, and establishing an edge between each entity node and the sentence node to which it belongs.

[0008] Optionally, the association relationship between entities is obtained by calculating the vector similarity between entity nodes in the graph network, including: calculating the vector similarity between every two entity nodes; and determining entity nodes with an association relationship by comparing the vector similarity between every two entity nodes with a preset similarity threshold.

[0009] Optionally, before extracting entities corresponding to the plurality of entity types, the method further includes: filtering irrelevant information from the initial text of the article to obtain the filtered text as the text of the article.

[0010] Optionally, the method of filtering irrelevant information from the initial text of the article to obtain the filtered text as the text of the article includes: converting the initial text of the article into a sentence sequence; encoding each sentence through a sentence-level language model to obtain an initial semantic vector for each sentence; inputting the initial semantic vector of each sentence into a chapter-level language model to classify relevant information and irrelevant information to obtain an updated semantic vector for each sentence and a sentence classification probability sequence, wherein the updated semantic vector incorporates the semantic information of the entire text of the sentence sequence; passing the sentence classification probability sequence through a fully-linked network to obtain an optimized classification result; filtering out irrelevant information based on the classification result to obtain the filtered text.

[0011] Optionally, the article is a medical report, and the multiple entity types include: location, organ, tissue, indicator, and indicator value.

[0012] Optionally, extracting corresponding multiple entities from the article based on multiple entity types used to describe the article includes: inputting the text of the article into a language model for text semantic extraction to obtain vectors corresponding to each input character that integrate the semantic information of the entire text; passing the vectors of each character through a fully linked network to obtain labels corresponding to each character; extracting multiple references corresponding to the multiple entity types based on the labels corresponding to each character; and extracting corresponding multiple entities by linking the multiple references to corresponding entities in an existing knowledge base.

[0013] Optionally, before linking the multiple references to corresponding entities in the existing knowledge base, it also includes: judging whether each extracted reference has a corresponding negative meaning description in the text of the article; if so, adding the negative meaning description to the corresponding reference.

[0014] Optionally, the judging whether each extracted reference has a corresponding negative meaning description in the text of the article comprises: performing binary classification of the position of each reference in the text of the article, the reference text, and the text of the article into a language model as to whether it has a negative meaning; and determining whether each reference has a corresponding negative meaning description based on the result of the binary classification.

[0015] According to the second aspect of the embodiments of this specification, there is provided a device for interpreting an article, comprising: an entity extraction module, configured to extract corresponding multiple entities from the article according to multiple entity types used to describe the article. A relationship extraction module, configured to perform relationship extraction using the extracted multiple entities and the text of the article to obtain the association relationship between the entities. A forest construction module, configured to construct a multi-tuple forest for describing the article based on the association relationship between the entities, the multi-tuple forest comprising one or more tree structures with entities as nodes and association relationships between entities as edges. An engine interpretation module, configured to input the multi-tuple forest into an interpretation rule engine to obtain an interpretation result.

[0016] According to a third aspect of an embodiment of the present specification, a computing device is provided, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions: extracting corresponding multiple entities from an article according to multiple entity types used to describe the article; performing relationship extraction using the extracted multiple entities and the text of the article to obtain associations between entities; constructing a multi-tuple forest for describing the article based on the associations between the entities, the multi-tuple forest comprising one or more tree structures with entities as nodes and associations between entities as edges; inputting the multi-tuple forest into an interpretation rule engine to obtain an interpretation result.

[0017] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the method for interpreting an article described in any embodiment of this specification are implemented.

[0018] An embodiment of the present specification provides a method for interpreting an article. The method extracts corresponding multiple entities from the article according to multiple entity types used to describe the article, obtains the association relationship between the entities using the extracted multiple entities and the text of the article, and constructs a multi-tuple forest for describing the article based on the association relationship between the entities. The multi-tuple forest includes one or more tree structures with entities as nodes and the association relationship between entities as edges. The multi-tuple forest is input into an interpretation rule engine to obtain an interpretation result. It can be seen that the method proposes a multi-tuple forest structure in response to the complexity of the article text, so that the article text can be structured into a multi-tuple forest, and the multi-tuple forest accurately and completely describes the article content. The multi-tuple forest is then input into an interpretation rule engine constructed based on the multi-tuple forest structure for interpretation, which can accurately obtain the interpretation result and achieve accurate and efficient article interpretation. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of a method for interpreting an article provided by an embodiment of this specification;

[0020] Figure 2 This is a schematic diagram of an interface interaction provided by an embodiment of this specification;

[0021] Figure 3 It is a schematic diagram of a processing process of irrelevant information filtering provided by an embodiment of this specification;

[0022] Figure 4 is a schematic diagram of a processing process of named entity recognition provided by an embodiment of this specification;

[0023] Figure 5 is a schematic diagram of a negative detection processing process provided by an embodiment of this specification;

[0024] Figure 6 It is a network diagram provided by an embodiment of this specification;

[0025] Figure 7 It is a process flow chart of a method for interpreting an article provided by an embodiment of this specification;

[0026] Figure 8 is a call link diagram provided by an embodiment of this specification;

[0027] Fig. 9 is a schematic diagram of a quintuple forest provided by an embodiment of this specification;

[0028] Fig.10 It is a structural schematic diagram of a device for interpreting articles provided by an embodiment of this specification;

[0029] Fig.11 is a structural schematic diagram of a device for interpreting an article provided by another embodiment of the present specification;

[0030] Fig.12 It is a structural block diagram of a computing device provided by an embodiment of this specification. DETAILED DESCRIPTION

[0031] Many specific details are described in the following description to facilitate a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited to the specific implementation disclosed below.

[0032] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0033] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0034] First, the terms involved in one or more embodiments of this specification are explained.

[0035] References are words actually used in the text of the article that refer to the entity type in this specification. An entity can have one or more references with the same meaning. For example, the references "tumor", "lump", and "nodule" all correspond to the entity "lump".

[0036] Entity type is used to describe the content and structure of an article.

[0037] Entity is an entity in an existing knowledge base that is used to describe the content and structure of an article. It is a standard name in an existing knowledge base.

[0038] Medical reports: single examination and test reports, such as ultrasound reports, blood routine reports, comprehensive reports, such as complete physical examination reports, etc.

[0039] OCR: Optical Character Recognition, optical character recognition technology, can convert images of paper documents into text.

[0040] NER: Named Entity Recognition, named entity recognition technology, can identify entity nouns (directions, organs, organizations, indicators, indicator values, etc.) in a text.

[0041] Entity linking refers to: Entity Linking, which associates the entities identified by NER with the corresponding entities in the existing knowledge base.

[0042] Relation extraction: It can identify the relationship between entities in a text (indicator-description-organization, organization-location-organ).

[0043] Negation detection: Identify whether certain entities that do appear in a text are semantically positive or negative. For example, in "I ate an apple this morning", the entity "apple" is semantically positive; in "I ate pears this morning, not apples", the entity "apple" is semantically negative.

[0044] Quintuple: refers to five entities: location, organ, tissue, indicator, and indicator value.

[0045] In this specification, a method for interpreting an article is provided. This specification also relates to an apparatus for interpreting an article, a computing device, and a computer-readable storage medium, which are described in detail one by one in the following embodiments.

[0046] Figure 1 A flowchart of a method for interpreting an article provided according to an embodiment of the present specification is shown, including steps 102 to 108.

[0047] Step 102: Extract corresponding entities from the article according to the multiple entity types used to describe the article.

[0048] Step 104: Relationship extraction is performed using the extracted multiple entities and the text of the article to obtain association relationships between entities.

[0049] Step 106: Based on the association relationships between the entities, a tuple forest for describing the article is constructed, wherein the tuple forest includes one or more tree structures with entities as nodes and association relationships between entities as edges.

[0050] It is understandable that, according to the association relationship between entities, the hierarchical relationship between entities can be derived accordingly. For example, an entity node at the upper level is the parent node of the entity node with the association relationship at the lower level. Therefore, according to the association relationship between entities, one or more tree structures can be constructed.

[0051] Step 108: Input the tuple forest into the interpretation rule engine to obtain the interpretation result.

[0052] For example, the interpretation rule engine may be a program designed according to a plurality of pre-set interpretation rules to identify the structure of a tuple forest. For example, each tree structure in a tuple forest may include one or more paths from a root node to a leaf node. Similarly, one interpretation rule corresponds to one path from a root node to a leaf node, and one path may derive a corresponding interpretation result. By matching each path from a root node to a leaf node in a tuple forest with a plurality of pre-set interpretation rules, a matching interpretation rule may be found, and then the interpretation result corresponding to the matching interpretation rule may be determined.

[0053] Since the method extracts corresponding multiple entities from the article according to multiple entity types used to describe the article, obtains the association relationship between the entities using the extracted multiple entities and the text of the article, and constructs a multi-tuple forest for describing the article based on the association relationship between the entities, the multi-tuple forest includes one or more tree structures with entities as nodes and the association relationship between entities as edges, and the multi-tuple forest is input into the interpretation rule engine to obtain the interpretation result, it can be seen that the method proposes a multi-tuple forest structure in response to the complexity of the article text, so that a plain text article can be structured into a multi-tuple forest, and the multi-tuple forest accurately describes the article content, and then the multi-tuple forest is input into the interpretation rule engine constructed based on the multi-tuple forest structure for interpretation, which can accurately obtain the interpretation result, thereby providing accurate and efficient article interpretation results.

[0054] For example, the method provided in the embodiments of this specification can be applied to the interpretation of medical reports. After interpreting the medical report, it can further support scenarios such as intelligent underwriting, intelligent claims settlement, and medical service recommendations. For example, the corresponding underwriting results such as risk level prompts are provided to users. For example, the basic insurance capabilities of the health notice center, pricing center, and underwriting center are abstractions of health notices during the insurance process, and their core models are health notices and health notice contracts. The health notice is a health notice of an insurance product, which defines the content that users need to actively inform when insuring. The health notice contract is a contract signed by the user when insuring, which carries the user's notification situation and serves as an important reference content for the user's claims settlement. Since the content in the health notice often contains a lot of professional medical knowledge, it is sometimes difficult for users to understand the true meaning of the health notice. In order to avoid users from mistakenly insuring products that do not meet the health notice, the method provided in the embodiments of this specification can be used to intelligently interpret the user's health notice report and provide the user with corresponding interpretation results. For example, the interpretation results may include: risk warnings and other information, prompting users to focus on the health notice clauses or automatically judging whether the user meets the health notice, and can also provide users with corresponding medical explanations, so that users without a medical background can also fully understand the health notice content. by Figure 2 Taking the interface interaction diagram shown in the figure as an example, the left picture is the upload page, where users can take a photo of their physical examination report and upload it. The right picture is the interpretation result display page, which displays the medical explanations and suggestions based on the report content.

[0055] From the execution environment, the method provided in the embodiments of this specification can be applied to the server side or to the local user side. For example, when applied to the server side, the server side can receive the article image or text uploaded by the user side, interpret the article, and feed back the interpretation result to the user side, so as to provide the user with online interpretation of the article.

[0056] In order to make the interpretation of the article more accurate and efficient, in one or more embodiments of this specification, before extracting the entities corresponding to the multiple entity types, it also includes: filtering the initial text of the article for irrelevant information, and obtaining the filtered text as the text of the article. Through this embodiment, after filtering the irrelevant information, the article content containing only the relevant information is interpreted, making the interpretation of the article more accurate and efficient.

[0057] Among them, the specific processing means for filtering irrelevant information can adopt any possible implementation method. For example, irrelevant information can be filtered out by detecting formats, detecting keywords, etc. In order to make the filtering of irrelevant information more accurate, in one or more embodiments of this specification, the processing of filtering irrelevant information on the initial text of the article to obtain the filtered text as the text of the article includes: converting the initial text of the article into a sentence sequence; encoding each sentence through a sentence-level language model to obtain an initial semantic vector for each sentence; inputting the initial semantic vector of each sentence into a chapter-level language model to classify relevant information and irrelevant information, obtaining an updated semantic vector for each sentence and a sentence classification probability sequence, and the updated semantic vector integrates the semantic information of the entire text of the sentence sequence; passing the sentence classification probability sequence through a fully linked network to obtain an optimized classification result; filtering out irrelevant information according to the classification result to obtain the filtered text. In this embodiment, the classification technology based on the language model of deep learning achieves more accurate filtering of irrelevant information.

[0058] Specifically, for example, Figure 3 As shown in the schematic diagram of the processing process of irrelevant information filtering shown in FIG, the sentence-level language model and the article-level language model in this embodiment can adopt the BERT model. For an input sentence sequence, each sentence can be first encoded with a sentence-level BERT model to obtain a semantic vector for each sentence (at this time, each vector only contains the semantic information of the sentence itself). Then the semantic vector of the sentence is input into an article-level BERT model to obtain a new semantic vector for each sentence. At this time, the new semantic vector of the sentence contains the semantic information of the entire sentence sequence to a certain extent. Finally, the obtained sentence classification probability sequence is passed through the CRF layer to optimize the classification result from the perspective of the rationality of the label sequence.

[0059] The method provided in the embodiments of this specification is not limited to the specific implementation methods of extracting entities. For example, the method extracts corresponding multiple entities from the article according to the multiple entity types used to describe the article, including: inputting the text of the article into a language model for text semantic extraction, and obtaining vectors corresponding to each input word that integrate the semantic information of the whole text; passing the vectors of each word through a fully linked network to obtain labels corresponding to each word; extracting multiple references corresponding to the multiple entity types according to the labels corresponding to each word; extracting the corresponding multiple entities by linking the multiple references to the entities corresponding to the existing knowledge base. In this embodiment, the NER (Named Entity Recognition, referred to as NER) technology based on deep learning is used to extract the required information through semantic features. Compared with the extraction technology that relies on the vocabulary, the generalization is greatly enhanced and will not be strictly restricted by the vocabulary or annotation data formulated by experts.

[0060] The method provided in the embodiments of this specification does not limit which NLP (natural language processing) model is used as the language model. For example, the language model in this embodiment can be a BERT model. The training of the BERT model is mainly divided into two stages: a pre-training stage and a fine-tuning stage. The pre-training stage is trained on a large data set based on some pre-training tasks. The fine-tuning stage is used for fine-tuning in some downstream tasks, such as text classification, part-of-speech tagging, question-answering systems, etc. BERT can be fine-tuned on different tasks without adjusting the structure. For example, Figure 4 As shown in the schematic diagram of the processing process of named entity recognition, in this embodiment, the input of the BERT model is the words of the text, and the output of the BERT model can be the vectors corresponding to each word that integrate the semantic information of the whole text. After the vectors of each word are passed through the fully linked network CRF, the NER label sequence, i.e., the labels corresponding to each word, can be obtained. Furthermore, multiple references corresponding to multiple entity types can be extracted based on the labels corresponding to each word.

[0061] Among them, the entity chaining process of linking multiple references to entities corresponding to the existing knowledge base can adopt any possible implementation method. For example, when the number of entities involved in the graph is small and there are basically no cases where different entities have the same name, the entity chaining can be matched by ha3 search and literal similarity calculation. For another example, when the graph content expands and there are many ambiguous entities, deep learning can be used to perform entity chaining.

[0062] In order to make the extracted reference more accurate, in one or more embodiments of the specification, before linking the multiple references to the corresponding entities in the existing knowledge base, it also includes: determining whether each extracted reference has a corresponding negative meaning description in the text of the article; if so, adding the negative meaning description to the corresponding reference. In this embodiment, by performing a negative detection on the previously extracted reference, the negative meaning description corresponding to the reference is added to the corresponding reference, making the extracted reference more accurate.

[0063] Among them, the specific implementation method of negation detection is not limited, for example, it can be detected by sensitive word detection. For another example, in order to make the negation detection more accurate, in one or more embodiments of the present specification, a deep learning method is used for negation detection, so that the extracted reference is more accurate. For example, the judgment of whether each extracted reference has a corresponding negative meaning description in the text of the article includes: the position of each reference in the text of the article, the reference text, and the text of the article are input into the language model for binary classification as to whether it has a negative meaning; based on the result of the binary classification, determine whether each reference has a corresponding negative meaning description. For example, if Figure 5 As shown in the schematic diagram of the processing process of negation detection, the language model for binary classification of negation meaning can adopt the BERT model. Figure 5 As shown in the figure, the NER result and the original text of the article are input into the BERT model, and the output of the BERT model passes through the fully linked network to obtain the probability of the reference being denied.

[0064] In the methods provided in the embodiments of this specification, the specific implementation methods of relationship extraction are not limited. For example, in one or more embodiments of this specification, a relationship extraction technology based on deep learning technology is used, and the generalization is greatly enhanced, avoiding the problem of insufficient generalization of relationship extraction capabilities. Specifically, for example, the use of the extracted multiple entities and the text of the article to extract relationships to obtain the association relationship between entities includes: using the extracted multiple entities and the text of the article to construct a graph network based on text and entities; using a graph convolution algorithm to extract features of entity nodes in the graph network to obtain vectors of entity nodes; and calculating the vector similarity between entity nodes in the graph network to obtain the association relationship between entities. In this embodiment, the graph convolution algorithm is used to extract features of entity nodes in the graph network, so that the features of the entity nodes are transmitted and updated through the structure of the graph network, and a more accurate node vector is obtained, thereby accurately determining the association relationship between entities based on vector similarity.

[0065] The method provided in the embodiments of this specification does not limit the specific implementation method of constructing the graph network. For example, the method of constructing a graph network based on text and entities using the extracted multiple entities and the text of the article includes: encoding each sentence and each entity in the text of the article as an independent node to obtain sentence nodes and entity nodes of the graph network; and completing the construction of the graph network by establishing an edge between each pair of sentence nodes, establishing an edge between each pair of entity nodes in the same sentence, and establishing an edge between each entity node and the sentence node to which it belongs.

[0066] Due to the complexity of the article, related entities are often located in different sentences. Therefore, we can first build a graph network based on sentences and entities, and then use the graph convolution algorithm to extract features from the entity nodes in the graph network to obtain the vectors of the entity nodes. By calculating the vector similarity between the entity nodes in the graph network, we can get the association relationship between the entities. Here, each sentence and entity is regarded as a node, and each sentence and each entity can be encoded by, for example, the BERT model as a vector representation of the node. An edge is connected between any two sentence nodes. If an entity belongs to a sentence, an edge is connected between the entity and the sentence. If two entities are in the same sentence, an edge is also constructed. For example, based on the content of the article "There is a nodule above the left breast, the size is 3*3mm, the border is smooth, and no cyst is found. There is a low echo area below with unclear boundaries", we can construct the following Figure 6 The graph network shown.

[0067] In the method provided in the embodiments of this specification, there is no restriction on how to use vector similarity to determine the association relationship between entities. For example, in order to accurately and efficiently obtain the association relationship between entities, in one or more embodiments of this specification, a preset similarity threshold is used to determine the entity nodes with an association relationship by comparing with the preset similarity threshold. For example, the association relationship between entities is obtained by calculating the vector similarity between the entity nodes in the graph network, including: calculating the vector similarity between every two entity nodes; determining the entity nodes with an association relationship by comparing the vector similarity between every two entity nodes with the preset similarity threshold. For example, if the vector similarity is higher than the preset similarity threshold (the specific threshold value can be set according to manual experience), it is considered that there is an association relationship between the two entities.

[0068] The following combination Figure 7 , taking the application of the method for interpreting articles provided in this specification in medical reports as an example, the method for interpreting articles is further described. In this embodiment, the multiple entity types may include: position, organ, tissue, indicator, and indicator value. Among them, Figure 7 A flowchart of a processing process of a method for interpreting an article provided by an embodiment of the present specification is shown, and the specific steps include step 702 to step 720.

[0069] Step 702: Receive a photo of a medical report uploaded by a user.

[0070] Step 704: Process the photo of the medical report through OCR (Optical Character Recognition) to obtain the initial text.

[0071] For example, photos are converted into sentence texts through OCR systems.

[0072] Step 706: Convert the initial text into a sentence sequence.

[0073] Step 708: Filter irrelevant information from the sentence sequence.

[0074] In the step of filtering irrelevant information in the sentence sequence, information irrelevant to the person to whom the medical report belongs can be filtered out (for example, the description of the report grading system, "BIRADS level I means xxx, BIRADS level II means xxx, ...").

[0075] Step 710: Perform named entity recognition and negation detection on the filtered text to obtain multiple references corresponding to any one or more entity types of location, organ, tissue, indicator, and indicator value.

[0076] For example, in the steps of named entity recognition and negation detection of filtered sentences, corresponding references can be extracted: such as orientation class (reference, such as "left side"), such as organ class (reference, such as "breast"), such as tissue class (reference, such as "nodule"), such as indicator class (reference, such as "boundary"), such as indicator value class (reference, such as "unclear"), and it is determined whether each reference has a negative meaning in the sentence (such as "no nodule was found").

[0077] Step 712: extracting corresponding entities by linking the multiple references to corresponding entities in the existing knowledge base.

[0078] Step 714: perform relationship extraction using the extracted multiple entities and the text of the report to obtain association relationships between entities.

[0079] Step 716: Construct a quintuple forest based on the association relationships between entities.

[0080] Step 718: Input the quintuple forest into the interpretation rule engine to obtain the interpretation result.

[0081] For example, expert rules can be set according to the experience of medical experts to form an interpretation rule engine. Of course, corresponding prompts can also be added to the interpretation rules, and the interpretation rule engine directly outputs the corresponding prompts.

[0082] Step 720: Providing corresponding prompts such as medical explanations, risk levels, etc. to the user based on the interpretation results.

[0083] For example, information such as prompts can be output to the user side and displayed to the user.

[0084] based on Figure 7 The processing process shown in the figure, the method for interpreting an article provided in the embodiment of this specification can be formed as follows Figure 8 The overall call chain shown. Among them, Figure 8 The "irrelevant information filtering module", "NER+negation detection module", "entity linking module" and "relationship extraction module" shown can all be implemented based on the natural language processing method of deep learning technology. The specific implementation method can be found in the description of other embodiments and will not be described in detail here.

[0085] It can be seen from the above embodiments that after being processed by the method provided in the embodiments of this specification, the medical report can be structured in the form of a quintuple forest. The quintuple is the five entities of location, organ, tissue, indicator, and indicator value, and there is a hierarchical relationship between these five entities. With the five entity nodes and the hierarchical relationship as the edge, tree structures can be constructed. The collection of these tree structures (forest) can fully describe the content of a medical report. For example, the report content "There is a nodule above the left breast, the size is 3*3mm, the border is smooth, and no cyst is found. There is a low echo area below the right breast with unclear boundaries." After being processed by the method provided in the embodiments of this specification, a forest such as the following can be constructed. Fig. 9 Assume that the interpretation rule engine contains the interpretation rule "right -> breast -> hypoechoic area -> boundary -> unclear = breast nuclear preservation risk", then Fig. 9 The five-tuple forest shown is input into the interpretation rule engine, and an interpretation result such as "Breast nuclear preservation risk" can be obtained. It can be seen that for content with complex contextual logical relationships such as "A nodule was found above the left breast... The boundary is unclear", the method provided in the embodiment of this specification can be applied to accurately describe the content of the article based on the five-tuple forest after extracting the association relationship based on deep learning natural language processing technology. By accurately describing the content of the article, it is possible to accurately interpret the article.

[0086] Corresponding to the above method embodiment, this specification also provides an embodiment of a device for interpreting an article. Fig.10 FIG. 1 is a schematic diagram showing a structure of a device for interpreting an article provided by an embodiment of the present specification. Fig.10 As shown, the device includes: an entity extraction module 1002, a relationship extraction module 1004, a forest construction module 1006 and an engine interpretation module 1008.

[0087] The entity extraction module 1002 may be configured to extract corresponding entities from the article according to multiple entity types used to describe the article.

[0088] The relationship extraction module 1004 may be configured to perform relationship extraction using the extracted multiple entities and the text of the article to obtain association relationships between entities.

[0089] The forest construction module 1006 may be configured to construct a tuple forest for describing the article based on the association relationships between the entities, wherein the tuple forest includes one or more tree structures with entities as nodes and association relationships between entities as edges.

[0090] The engine interpretation module 1008 may be configured to input the multi-tuple forest into an interpretation rule engine to obtain an interpretation result.

[0091] Since the device extracts corresponding multiple entities from the article according to multiple entity types used to describe the article, obtains the association relationship between the entities using the extracted multiple entities and the text of the article, and constructs a multi-group forest for describing the article based on the association relationship between the entities, the multi-group forest includes one or more tree structures with entities as nodes and the association relationship between entities as edges, and the multi-group forest is input into the interpretation rule engine to obtain the interpretation result, it can be seen that the device proposes a multi-group forest structure in response to the complexity of the article text, so that a plain text article can be structured into a multi-group forest, and the multi-group forest accurately describes the article content, and then the multi-group forest is input into the interpretation rule engine constructed based on the multi-group forest structure for interpretation, and the interpretation result can be obtained intelligently, thereby providing accurate and efficient article interpretation results.

[0092] Fig.11 FIG. 2 shows a schematic diagram of a device for interpreting an article provided by another embodiment of the present specification. Fig.11 As shown, in order to make the interpretation of the article more accurate and efficient, the device may also include: an irrelevant filtering module 1010, which can be configured to filter irrelevant information from the initial text of the article before the entity extraction module 1002 extracts the entities corresponding to the multiple entity types respectively, and obtains the filtered text as the text of the article.

[0093] The specific processing means for filtering irrelevant information may adopt any possible implementation. In one or more embodiments of the specification, the irrelevant filtering module 1010 may include: a sentence sequence conversion submodule 1010a, a sentence-level model processing submodule 1010b, a paragraph-level model processing submodule 1010c, an irrelevant classification output submodule 1010d, and a filtering processing submodule 1010e.

[0094] The sentence sequence conversion submodule 1010a may be configured to convert the initial text of the article into a sentence sequence.

[0095] The sentence-level model processing submodule 1010b may be configured to encode each sentence using a sentence-level language model to obtain an initial semantic vector for each sentence.

[0096] The chapter-level model processing submodule 1010c can be configured to input the initial semantic vector of each sentence into the chapter-level language model to classify relevant information and irrelevant information, and obtain an updated semantic vector for each sentence and a sentence classification probability sequence, wherein the updated semantic vector incorporates the semantic information of the entire text of the sentence sequence.

[0097] The irrelevant classification output submodule 1010d may be configured to pass the sentence classification probability sequence through a fully linked network to obtain an optimized classification result.

[0098] The filtering processing submodule 1010e may be configured to filter out irrelevant information according to the classification result to obtain filtered text.

[0099] In this embodiment, the classification technology based on the language model of deep learning achieves more accurate filtering of irrelevant information.

[0100] The method provided in the embodiments of this specification is not limited to the specific implementation method of extracting entities. Fig.11 As shown, the entity extraction module 1002 may include: a word vector extraction submodule 1002a, a tag output submodule 1002b, a reference extraction submodule 1002c and an entity link submodule 1002d.

[0101] The word vector extraction submodule 1002a may be configured to input the text of the article into a language model for text semantic extraction, and obtain vectors corresponding to each input word that incorporate the semantic information of the entire text.

[0102] The label output submodule 1002b may be configured to pass the vectors of each word through a fully connected network to obtain labels corresponding to each word.

[0103] The reference extraction submodule 1002c may be configured to extract multiple references corresponding to the multiple entity types according to the labels corresponding to the characters.

[0104] The entity linking submodule 1002d may be configured to extract corresponding entities by linking the multiple references to corresponding entities in an existing knowledge base.

[0105] In this embodiment, the NER technology based on deep learning is used to extract the required information through semantic features. Compared with the extraction technology relying on the vocabulary, the generalization is greatly enhanced and is not strictly restricted by the vocabulary or annotation data formulated by experts.

[0106] In order to make the extracted references more accurate, in one or more embodiments of the present specification, the entity extraction module 1002 may also include: a negation detection submodule 1002e, which can be configured to determine whether each extracted reference has a corresponding negative meaning description in the text of the article before the entity linking submodule 1002d links the multiple references to the corresponding entities in the existing knowledge base; if so, add the negative meaning description to the corresponding reference.

[0107] Among them, the specific implementation method of negation detection is not limited, for example, it can be detected by sensitive word detection. For another example, in order to make negation detection more accurate, in one or more embodiments of this specification, a deep learning method is used for negation detection, so that the extracted reference is more accurate. For example, the negation detection submodule 1002e can be configured to classify the position of each reference in the text of the article, the reference text, and the text input language model of the article into two categories as to whether it has a negative meaning; according to the result of the two-category classification, determine whether each reference has a corresponding negative meaning description.

[0108] The specific implementation method of relationship extraction in the embodiments of this specification is not limited. Fig.11 As shown, the relationship extraction module 1004 may include: a graph network construction submodule 1004a, a graph convolution calculation submodule 1004b and an association relationship determination submodule 1004c.

[0109] The graph network construction submodule 1004a may be configured to construct a text- and entity-based graph network using the extracted multiple entities and the text of the article.

[0110] The graph convolution calculation submodule 1004b can be configured to use a graph convolution algorithm to perform feature extraction on entity nodes in the graph network to obtain vectors of entity nodes.

[0111] The association relationship determination submodule 1004c may be configured to obtain the association relationship between entities by calculating the vector similarity between entity nodes in the graph network.

[0112] In this embodiment, the features of the entity nodes in the graph network are extracted through a graph convolution algorithm, so that the features of the entity nodes can be transmitted and updated through the structure of the graph network to obtain more accurate node vectors, and then the association relationship between entities is accurately determined based on vector similarity.

[0113] The method provided in the embodiments of this specification does not limit the specific implementation method of constructing the graph network. For example, the graph network construction submodule 1004a can be configured to encode each sentence and each entity in the text of the article as an independent node to obtain sentence nodes and entity nodes of the graph network; by establishing an edge between each pair of sentence nodes, establishing an edge between each pair of entity nodes in the same sentence, and establishing an edge between each entity node and its sentence node, the construction of the graph network is completed.

[0114] In the method provided in the embodiments of this specification, there is no limitation on how to use vector similarity to determine the association relationship between entities. For example, the association relationship determination submodule 1004c can be configured to calculate the vector similarity between every two entity nodes; by comparing the vector similarity between every two entity nodes with a preset similarity threshold, the entity nodes with an association relationship are determined.

[0115] The above is a schematic scheme of a device for interpreting articles of this embodiment. It should be noted that the technical scheme of the device for interpreting articles and the technical scheme of the method for interpreting articles described above are of the same concept, and the details of the technical scheme of the device for interpreting articles that are not described in detail can be found in the description of the technical scheme of the method for interpreting articles described above.

[0116] Fig.12 The block diagram of a computing device 1200 according to one embodiment of the present specification is shown. The components of the computing device 1200 include but are not limited to a memory 1210 and a processor 1220. The processor 1220 is connected to the memory 1210 via a bus 1230, and the database 1250 is used to store data.

[0117] The computing device 1200 also includes an access device 1240 that enables the computing device 1200 to communicate via one or more networks 1260. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 1240 may include one or more of any type of network interface (e.g., a network interface card (NIC)) whether wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a World Wide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like.

[0118] In one embodiment of the present specification, the above components of the computing device 1200 and Fig.12Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Fig.12 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0119] The computing device 1200 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. The computing device 1200 may also be a mobile or stationary server.

[0120] The processor 1220 is used to execute the following computer executable instructions:

[0121] Extracting multiple entities corresponding to the article according to multiple entity types used to describe the article;

[0122] Relationship extraction is performed using the extracted multiple entities and the text of the article to obtain association relationships between the entities;

[0123] Based on the association relationship between the entities, a multi-tuple forest for describing the article is constructed, wherein the multi-tuple forest includes one or more tree structures with entities as nodes and association relationships between entities as edges;

[0124] The multi-tuple forest is input into the interpretation rule engine to obtain the interpretation result.

[0125] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above-mentioned method for interpreting articles belong to the same concept, and the details not described in detail in the technical scheme of the computing device can be referred to the description of the technical scheme of the above-mentioned method for interpreting articles.

[0126] An embodiment of the present specification further provides a computer-readable storage medium storing computer instructions, which are used when executed by a processor to:

[0127] Extracting multiple entities corresponding to the article according to multiple entity types used to describe the article;

[0128] Relationship extraction is performed using the extracted multiple entities and the text of the article to obtain association relationships between the entities;

[0129] Based on the association relationship between the entities, a multi-tuple forest for describing the article is constructed, wherein the multi-tuple forest includes one or more tree structures with entities as nodes and association relationships between entities as edges;

[0130] The multi-tuple forest is input into the interpretation rule engine to obtain the interpretation result.

[0131] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the method for interpreting the article above are of the same concept, and the details not described in detail in the technical scheme of the storage medium can be referred to the description of the technical scheme of the method for interpreting the article above.

[0132] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0133] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0134] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0135] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0136] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technicians in the relevant technical field can well understand and use this specification. This specification is only limited by the claims and their full scope and equivalents.

Claims

1. A method for interpreting an article, comprising: Extracting multiple entities corresponding to the article according to multiple entity types used to describe the article; Relationship extraction is performed using the extracted multiple entities and the text of the article to obtain association relationships between the entities; Based on the association relationship between the entities, a multi-tuple forest for describing the article is constructed, wherein the multi-tuple forest includes one or more tree structures with entities as nodes and association relationships between entities as edges; The tuple forest is input into the interpretation rule engine to obtain an interpretation result, wherein the tuple forest is used to match the path from the root node to the leaf node with the interpretation rules preset in the interpretation rule engine to determine the matching interpretation rules.

2. According to the method of claim 1, the step of extracting relationships using the extracted entities and the text of the article to obtain association relationships between the entities comprises: Using the extracted entities and the text of the article, a text- and entity-based graph network is constructed; Using a graph convolution algorithm to extract features of entity nodes in the graph network to obtain vectors of entity nodes; By calculating the vector similarity between entity nodes in the graph network, the association relationship between entities is obtained.

3. According to the method of claim 2, the step of constructing a graph network based on text and entities by using the extracted multiple entities and the text of the article comprises: Each sentence and each entity in the text of the article is regarded as an independent node and encoded to obtain sentence nodes and entity nodes of the graph network; The graph network is constructed by establishing an edge between each pair of sentence nodes, an edge between each pair of entity nodes in the same sentence, and an edge between each entity node and the sentence node to which it belongs.

4. The method according to claim 2, wherein the obtaining the association relationship between entities by calculating the vector similarity between entity nodes in the graph network comprises: Calculate the vector similarity between every two entity nodes; By comparing the vector similarity between every two entity nodes with a preset similarity threshold, entity nodes with an associated relationship are determined.

5. The method according to claim 1, before extracting the corresponding multiple entities, further comprises: The original text of the article is processed to filter out irrelevant information, and the filtered text is obtained as the text of the article.

6. The method according to claim 5, wherein filtering the original text of the article for irrelevant information to obtain the filtered text as the text of the article comprises: converting the initial text of the article into a sequence of sentences; Encode each sentence through a sentence-level language model to obtain the initial semantic vector of each sentence; Inputting the initial semantic vector of each sentence into the paragraph-level language model to classify relevant information and irrelevant information, obtaining an updated semantic vector of each sentence and a sentence classification probability sequence, wherein the updated semantic vector incorporates the semantic information of the entire sentence sequence; Passing the sentence classification probability sequence through a fully linked network to obtain an optimized classification result; Irrelevant information is filtered out according to the classification result to obtain filtered text.

7. The method according to claim 1, wherein the article is a medical report, and the multiple entity types used to describe the article include: Direction, organ, tissue, indicator, indicator value.

8. The method according to claim 1, wherein extracting corresponding entities from the article according to the multiple entity types used to describe the article comprises: Input the text of the article into a language model to extract text semantics, and obtain vectors corresponding to each word of the input that integrate the semantic information of the entire text; Passing the vector of each word through a fully linked network to obtain labels corresponding to each word, wherein the labels are set according to multiple entity types used to describe the article; Extracting multiple references corresponding to the multiple entity types according to the labels corresponding to the characters; By linking the multiple references to corresponding entities in an existing knowledge base respectively, a corresponding multiple entities are extracted.

9. The method according to claim 8, before linking the plurality of references to corresponding entities in the existing knowledge base, further comprising: Determine whether each extracted reference has a corresponding negative meaning description in the text of the article; If so, add the negative meaning description to the corresponding reference.

10. The method according to claim 9, wherein the step of determining whether each extracted reference has a corresponding negative meaning description in the text of the article comprises: Input the position of each reference in the text of the article, the reference text, and the text of the article into the language model for binary classification of whether it is a negative meaning; According to the result of the binary classification, it is determined whether each reference has a corresponding negative meaning description.

11. A device for interpreting an article, comprising: An entity extraction module is configured to extract corresponding entities from the article according to multiple entity types used to describe the article; A relationship extraction module is configured to perform relationship extraction using the extracted multiple entities and the text of the article to obtain association relationships between the entities; A forest construction module is configured to construct a multi-tuple forest for describing the article based on the association relationship between the entities, wherein the multi-tuple forest includes one or more tree structures with entities as nodes and association relationships between entities as edges; The engine interpretation module is configured to input the tuple forest into the interpretation rule engine to obtain an interpretation result, wherein the tuple forest is used to match the path from the root node to the leaf node with the interpretation rules pre-set in the interpretation rule engine to determine the matching interpretation rules.

12. A computing device comprising: Memory and processor; The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions: Extracting multiple entities corresponding to the article according to multiple entity types used to describe the article; Relationship extraction is performed using the extracted multiple entities and the text of the article to obtain association relationships between the entities; Based on the association relationship between the entities, a multi-tuple forest for describing the article is constructed, wherein the multi-tuple forest includes one or more tree structures with entities as nodes and association relationships between entities as edges; The tuple forest is input into the interpretation rule engine to obtain an interpretation result, wherein the tuple forest is used to match the path from the root node to the leaf node with the interpretation rules preset in the interpretation rule engine to determine the matching interpretation rules.

13. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the method for interpreting an article as claimed in any one of claims 1 to 10.

14. A computer program product, characterized in that The method comprises computer instructions, which, when executed by a processor, implement the steps of the method for interpreting an article as claimed in any one of claims 1 to 10.

Images