A medical term extraction method and system based on multilingual parallel corpus

Abstract

The present invention relates to a method and system for extracting medical terms based on a multilingual parallel corpus, comprising: calculating semantic similarity of all sentences in medical knowledge data to obtain parallel corpus; translating the missing language of each parallel corpus and constructing sentences of all types of languages ​​into a multilingual parallel corpus; segmenting sentences in each parallel corpus, extracting medical terms and encoding and marking all sentences in combination with corresponding term types to obtain the encoding sequence of each sentence and training a neural network model in combination with corresponding sentences, identifying a medical text to be identified according to the trained neural network to obtain the encoding sequence of the medical text to be identified; determining the medical terms and corresponding term types in the medical text to be identified according to the encoding sequence of the medical text to be identified. The neural network model is trained by a multilingual parallel corpus and doctor terminology in the corresponding language, thereby accurately extracting medical terms of more language types.

Description

Technical Field

[0001] The present invention relates to the field of medical information processing, and in particular to a method and system for extracting medical terms based on a multilingual parallel corpus. Background Art

[0002] With the development of information technology, countries around the world have communicated spatially and in depth, and medical issues have become global issues. Since the outbreak of the new crown epidemic, the epidemic prevention situations in various countries have varied greatly. The reasons for this are complex. In addition to the different production and medical levels of various countries, the most important thing is the communication of epidemic prevention information. At present, research on term extraction at home and abroad is generally for term extraction in a single language or bilingual term extraction research. The term extraction work requires the annotation of a large amount of data, which is a huge workload. In order to reduce the workload, a method combining translation and term extraction is proposed to reduce the workload of annotation of term types. Based on the existing medical term extraction method, it is not suitable for multilingual types, and it is impossible to achieve accurate and efficient extraction of medical terms in more languages. In this regard, the present invention proposes a medical term extraction method and system based on a multilingual parallel corpus. Summary of the invention

[0003] The purpose of the present invention is to provide a medical terminology extraction method and system based on a multilingual parallel corpus. By constructing a multilingual parallel corpus and training a neural network terminology extraction model with doctor terminology in the corresponding language, the trained neural network terminology extraction model can be used to accurately extract medical terminology in more language types, which is conducive to the communication of medical issues around the world.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A medical term extraction method based on a multilingual parallel corpus, comprising:

[0006] Acquire medical knowledge data and calculate semantic similarity of all sentences in the medical knowledge data;

[0007] Sentences with semantic similarity values ​​greater than a first preset value are grouped into parallel corpora to obtain parallel corpora with multiple semantics; each parallel corpus includes sentences of at least one language type;

[0008] The missing language in each of the parallel corpora is translated to obtain sentences of the missing language; the parallel corpora include a preset number of language types; the preset number of language types include the types of the missing languages ​​and the types of the existing languages ​​in the parallel corpora.

[0009] Construct a multilingual parallel corpus of sentences from all types of languages;

[0010] Segmenting the sentences in each of the parallel corpora in the multilingual parallel corpus, and removing non-medical terms in each segmentation, to obtain medical terms for each of the parallel corpora;

[0011] Encoding and annotating all the sentences in the multilingual corpus according to the medical terms in the parallel corpus and the corresponding term types to obtain a coding sequence for each of the sentences;

[0012] A neural network term extraction model is trained using each of the sentences and the corresponding coding sequence to obtain a trained neural network term extraction model; a medical text to be identified is identified according to the trained neural network term extraction model to obtain a coding sequence of the medical text to be identified; and medical terms and corresponding term types in the medical text to be identified are determined according to the coding sequence of the medical text to be identified.

[0013] A medical term extraction system based on multilingual parallel corpus, comprising:

[0014] A sentence semantic similarity calculation module for medical knowledge data, used to obtain medical knowledge data and perform semantic similarity calculation on all sentences in the medical knowledge data;

[0015] A parallel corpus acquisition module, used to form parallel corpora from sentences whose semantic similarity values ​​are greater than a first preset value, to obtain parallel corpora of multiple semantics; each of the parallel corpora includes sentences of at least one language type;

[0016] A translation module, used for translating the missing language in each of the parallel corpora to obtain sentences of the missing language;

[0017] A multilingual parallel corpus construction module, which is used to construct sentences of all types of languages ​​into a multilingual parallel corpus;

[0018] A word segmentation module, used for performing word segmentation on the sentences in each parallel corpus in the multilingual parallel corpus, and removing non-medical terms in each word segmentation to obtain medical terms of each parallel corpus;

[0019] A sentence encoding module, used for encoding and marking all the sentences in the multilingual corpus according to the medical terms in the parallel corpus and the corresponding term types, to obtain an encoding sequence for each sentence;

[0020] The medical term extraction module is used to train a neural network term extraction model using each of the sentences and the corresponding coding sequence to obtain a trained neural network term extraction model; identify a medical text to be identified according to the trained neural network term extraction model to obtain a coding sequence of the medical text to be identified; and determine the medical terms and corresponding term types in the medical text to be identified according to the coding sequence of the medical text to be identified.

[0021] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0022] The present invention provides a medical term extraction method and system based on a multilingual parallel corpus, comprising: calculating semantic similarity of all sentences in medical knowledge data to obtain parallel corpora with multiple semantics; translating the missing language in each parallel corpus to obtain sentences of the missing language; constructing sentences of all types of languages ​​into a multilingual parallel corpus; segmenting the sentences in each parallel corpus to extract medical terms of each parallel corpus; encoding and marking all sentences according to medical terms and corresponding term types to obtain a coding sequence of each sentence; training a neural network term extraction model using each sentence and the corresponding coding sequence, identifying a medical text to be identified according to the trained neural network term extraction model to obtain a coding sequence of the medical text to be identified; and determining the medical terms and corresponding term types in the medical text to be identified according to the coding sequence of the medical text to be identified. By comparing the similarity of the sentences in the acquired medical knowledge data to obtain parallel corpora, the language types not included in each parallel corpus are translated, so that various language expressions of the sentences in the parallel corpus can be realized, so that the sentences of the constructed multilingual parallel corpus have more language expressions, and when subsequent term extraction is performed based on the multi-language parallel corpus, more doctor terms in multiple languages ​​can be extracted. The application of multilingual parallel corpora and neural network term extraction models to extract medical terms can ensure the accuracy of medical term extraction and improve the efficiency of medical term extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 A flow chart of a method for extracting medical terms based on a multilingual parallel corpus provided in Example 1 of the present invention;

[0025] Figure 2A schematic diagram of a translation method based on direct translation and a translation method based on an intermediate language provided in Embodiment 1 of the present invention;

[0026] Figure 3 A flowchart of a neural network translation model trained using an application domain adaptive method according to Embodiment 1 of the present invention;

[0027] Figure 4 A diagram of the relationship between term types and multilingual terms provided in Example 1 of the present invention;

[0028] Figure 5 A structural diagram of a neural network term extraction model provided in Example 1 of the present invention;

[0029] Figure 6 A block diagram of a medical term extraction system based on a multilingual parallel corpus provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] The purpose of the present invention is to provide a medical terminology extraction method and system based on a multilingual parallel corpus. By constructing a multilingual parallel corpus and training a neural network terminology extraction model with doctor terminology in the corresponding language, the trained neural network terminology extraction model can be used to accurately extract medical terminology in more language types, which is conducive to the communication of medical issues around the world.

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] Example 1

[0034] like Figure 1 As shown, this embodiment provides a medical term extraction method based on a multilingual parallel corpus, comprising:

[0035] Step S1: Obtain medical knowledge data and calculate semantic similarity for all sentences in the medical knowledge data.

[0036] Due to the scarcity of medical corpus, medical corpus is obtained from different channels. The main methods of obtaining corpus include: crawling different language versions of medical knowledge from the Internet, obtaining medical-related bilingual versions from hospitals, existing medical-related data sets, and translating medical corpus in existing languages ​​into the target language. The corpus form mainly includes chapters and sentences. The chapters need to be converted into sentence form, and subsequent processing steps are based on the sentence form. Obtaining medical knowledge data from multiple channels can ensure that the data of the subsequent multilingual parallel corpus is more sufficient.

[0037] In order to reduce the amount of calculation and improve the efficiency of term extraction, some sentences that do not contain medical terms or have a small amount of medical terms can be omitted before calculating the semantic similarity of all sentences in the medical knowledge data, and semantic extraction can be performed on sentences containing more terms to obtain their key points, which are more concise and easy to understand.

[0038] Step S2: Sentences with semantic similarity values ​​greater than a first preset value are grouped into parallel corpora to obtain parallel corpora of multiple semantics; each of the parallel corpora includes sentences of at least one language type.

[0039] Parallel corpora are different language descriptions of the same topic (semantics). Although there are certain differences between the language types, there are certain similarities in content. By comparing the word vectors of sentences, it can be considered that sentences with high word vector similarity have basically the same semantics, and such sentences can be used as an expansion of parallel corpora.

[0040] Step S3: translating each missing language in the parallel corpus to obtain sentences of the missing language; the parallel corpus includes a preset number of language types; the preset number of language types include the types of the missing languages ​​and the types of the existing languages ​​in the parallel corpus.

[0041] Considering that the language expression types of each sentence in the parallel corpus cannot meet the preset requirements, there are some situations where the language types do not exist, so it is necessary to supplement the missing languages. As shown in Table 1, it is assumed that the preset number of language types includes Chinese, German, English, Kazakh, Mongolian and Uyghur. The above language types are provided only to clearly illustrate this solution and have no limiting effect. Among them, the language type of sentence 1 can meet the preset requirements, while sentence 2 lacks Mongolian, sentence 3 lacks Chinese and Uyghur (Uyghur), and sentence 4 lacks Kazakh. The language missing from each sentence needs to be translated using sentences in existing languages.

[0042] Table 1 Language types of sentences in the parallel corpus

[0043] Chinese German English Kazakh Mongolian Uyghur Sentence 1 1 1 1 1 1 1 Sentence 2 1 1 1 1 0 1 Sentence 3 0 1 1 1 1 0 Sentence 4 1 1 1 0 1 1

[0044] 1 indicates that a sentence description of the language already exists, and 0 indicates that a sentence description of the language is missing.

[0045] In this embodiment, a translation method based on direct conversion and / or a translation method based on an intermediate language is applied. The translation method based on direct conversion can be regarded as a process of "independent analysis-related conversion-independent generation", which can be divided into six steps, each of which is completed by a corresponding translation rule, such as Figure 2 As shown in (a), the missing language sentence is directly translated using the existing language sentence.

[0046] Due to the different scopes and numbers of people using different languages ​​in the world, the amount of corpus for different languages ​​is also different. Some languages ​​are resource-rich languages, while others are low-resource languages. To build a multilingual parallel corpus, it is necessary to expand the low-resource corpus. The translation method based on the intermediate language adopts a knowledge representation structure called "intermediate language" and uses the intermediate language as a bridge between independent source language analysis and independent target language generation, thus truly realizing independent analysis and independent generation. Figure 2 As shown in (b), a missing language (language five) is selected as the intermediate language of another missing language (language one, two, three and four). For example, English and Chinese terminology-rich languages ​​can be used as the central language. French and German, which have relatively low corpus, have a high similarity with English, and parallel corpus are relatively easy to obtain. French can be converted into pseudo-corpus of English, and then the pseudo-corpus of English can be converted into pseudo-corpus of Chinese. Taking sentence 3 in Table 1 as an example, directly translating Uyghur into Uyghur based on existing languages ​​may have the problem of low accuracy. First, Chinese can be used as the intermediate language of Uyghur, and the Chinese expression of sentence 3 can be translated using the existing language, and then the Uyghur sentence can be translated from the Chinese sentence.

[0047] Whether it is a direct conversion translation method or an intermediate language-based translation method, when translating from one language to another, a neural network translation model with a translation relationship between the two languages ​​is used. For example, when translating English to Chinese, an English-to-Chinese neural network translation model is used. Considering that the present embodiment is to translate medical content, for languages ​​where medical corpus is difficult to obtain but have a large amount of general domain corpus, in order to improve the accuracy of model translation, a domain adaptive method can be used to complete the translation of medical corpus in an existing language to a target language. For example, pre-training is performed using Chinese general corpus or English general corpus that is easier to obtain data and Uyghur general corpus with less data. Then, the translation model is fine-tuned using the mixed general corpus and medical corpus, and finally the model is optimized using medical data. This method first uses general domain data plus some general domain monolingual data to train a neural network translation model for a general domain. Then gradually add monolingual data from the medical domain and the general domain to obtain pseudo corpus from the general domain and the medical domain. Then, we screened pseudo-corpus in the medical field and added it to the training of the translation model, gradually increasing the proportion of medical data in the training data of the translation model, and finally obtained a translation model in the medical field, such as Figure 3 shown.

[0048] From the perspective of the translation model structure:

[0049] The translation model consists of an encoder, a language sublayer, and a decoder.

[0050] ①Encoder

[0051] First, use the pre-trained language model to obtain the vector representation (h1, h2, ..., hn) of the general domain text. The data used in machine translation is mostly at the sentence level. By aligning the existing data, the sentence-level language pairs are obtained, and the parts with too large differences in sentence length in the language pairs are filtered out. The general ratio is set to 1 to 1.2. The encoder consists of a multi-head attention mechanism, a normalization layer, and a residual connection. The pre-training method can learn the language knowledge of the general domain, so that the encoder can better learn the grammatical content of the language. Then, the trained model is fine-tuned with the mixed general domain text and medical text. The vector representation (h1, h2, ..., hn) of the text is also obtained, and the sentence is extended to a fixed length. Finally, the trained model is fine-tuned with the medical domain text.

[0052] Finally, the encoder obtains a matrix representation containing the input language information, which is passed into the decoder as the representation of the source language.

[0053] ②Add sublayer

[0054] Add different language sublayers behind the encoder. Different sublayers will be used for translation when translating different languages. The parameter configuration of the encoder is shared here, and sublayers are added to achieve mutual translation between multiple languages.

[0055] ③Decoder

[0056] The composition of the decoder is similar to that of the encoder, except that it has an additional input. The input of the decoder contains the vector representation of the target language (t1, t2, ..., tn) and the output of the encoder, which contains the grammatical information of the source language and the grammatical information of the target language. The output is the predicted target language sequence.

[0057] The specific steps of the translation method based on direct conversion and the translation method based on intermediate language are described in detail below. Step S3 specifically includes:

[0058] Step S31: For each of the parallel corpora, a missing language is randomly selected.

[0059] Step S32: selecting a sentence in an existing language from the parallel corpus as an application language sentence according to the translation accuracy of each existing language in the parallel corpus when translated into the currently selected missing language; and translating a sentence in the currently selected missing language using the application language sentence;

[0060] The translation accuracy of each existing language when translated into the currently selected vacant language refers to the translation accuracy of the corresponding language translation model.

[0061] Step S32 specifically includes:

[0062] Step S321: determining the translation accuracy of each translation model applied when translating each existing language in the parallel corpus into the currently selected missing language;

[0063] Step S322: determining whether there is a translation model whose translation accuracy is greater than a first preset value;

[0064] If not, then selecting one of the remaining vacant languages ​​from the remaining vacant languages ​​according to the accuracy of translating the remaining vacant languages ​​into the selected vacant language;

[0065] Translating sentences in the selected remaining vacant language using the sentences in the existing language;

[0066] Translating the selected sentence of the missing language by using the selected sentence of the remaining missing language;

[0067] If so, the translation model with the translation accuracy greater than a preset value is used to translate the sentence of the missing language.

[0068] Considering that there are certain semantic differences in the translated sentences, it is necessary to judge the semantics of the sentences after translating the sentences of the missing language. That is, the translation of the sentences of the currently selected missing language by using the application language sentences further specifically includes:

[0069] Determine whether the similarity between the semantics corresponding to the currently selected sentence of the missing language and the semantics of the application language sentence is greater than a second preset value;

[0070] If yes, retain the translated sentence in the currently selected missing language;

[0071] If not, delete all the language type expression sentences under this semantics.

[0072] Step S33: randomly select a missing language from the remaining missing languages ​​and record it as the currently selected missing language, and return to step S32 until all the sentences of the missing languages ​​are translated.

[0073] Step S4: constructing all the sentences into a multilingual parallel corpus;

[0074] The sentences of each language type in each parallel corpus constitute a multilingual parallel corpus. Each parallel corpus includes sentences of the same language type. The breadth of the multilingual medical parallel corpus can lay the foundation for the breadth of world medical communication.

[0075] Step S5: segmenting the sentences in each parallel corpus in the multilingual parallel corpus, and removing non-medical terms in each segmentation to obtain medical terms for each parallel corpus;

[0076] Step S5 specifically includes:

[0077] For each of the parallel corpora, a sentence of a target language is selected according to the usage frequency of the language;

[0078] Performing lexical segmentation on the selected sentence to obtain various segmentation words;

[0079] Removing non-medical terms from each of the segmented words according to a term dictionary of the language corresponding to the segmented words to obtain medical terms in the target language;

[0080] Medical terms in other languages ​​are derived respectively according to the medical terms in the target language; and the medical terms in each of the parallel corpora include the medical terms in the target language and the medical terms in the other languages.

[0081] Segmenting sentences into words is to prepare for the work of medical terminology annotation. For Chinese, the Jieba word segmentation tool can be used to accurately segment sentences into words with specific meanings. For example, Wuhan Yangtze River Bridge will be segmented into Wuhan City\Yangtze River Bridge instead of Wuhan City\Jiang Bridge, retaining the minimum granularity of the original meaning of the words. The words obtained by word segmentation are compared with the term dictionary of the corresponding language type (Chinese term dictionary) for similarity, retaining medical terms and removing non-medical terms.

[0082] Step S6: encoding and annotating all the sentences in the multilingual corpus according to the medical terms in the parallel corpus and the corresponding term types to obtain a coding sequence for each sentence.

[0083] Step S6 specifically includes:

[0084] Step S61: For each of the sentences, the non-medical terms, medical terms and term types in the sentence are determined based on the medical terms of the corresponding language type; for example, a Chinese sentence is determined based on the Chinese medical terms for the non-medical term part and the medical term part in the Chinese sentence. Term types include disease names, drug names, pathogen names, symptoms, treatment methods, inspection methods, etc. Term categories include not only some entity terms, but also action terms and change terms. Entity terms generally include noun terms such as diabetes, action terms generally include operation terms such as craniotomy, and changes generally include terms such as carcinoma.

[0085] Step S62: Encode the medical terms in the sentence according to the term type.

[0086] For all sentences, the same coding is used for medical terms in various language expressions belonging to the same term type. No matter what language the sentence is in, the coding can accurately know that the same coding in each sentence corresponds to the same medical term, such as Figure 4 , showing a relationship diagram between term types and multilingual terms.

[0087] Step S63: Encode the non-medical terms in the sentence; in order to effectively distinguish between medical terms and non-medical terms and accurately extract medical terms, the encoding of the medical terms in the sentence is required to be different from the encoding of the non-medical terms.

[0088] Medical terms are coded according to the different term types. For example, the numbering starts from 1 according to the number of term types. One term type corresponds to a numerical or letter number. 0 can be defined as a non-medical term (in this case, 0 cannot be used as the code for medical terms), or other numerical values ​​or letters can be used.

[0089] Step S64: determining a coding sequence of the sentence according to the coding of the medical terms and the coding of the non-medical terms in the sentence.

[0090] For example, "He unfortunately suffered from diabetes." is a Chinese sentence example. Diabetes is a medical term and a disease name type. Each word in diabetes can be encoded as 1 (coded with a numerical value), and in order to distinguish the positions of the three words of diabetes, a position number is added after the code corresponding to each word. Therefore, the coding sequence of diabetes is 1-B 1-M 1-E, and the final coding sequence is: 0 0 0 0 0 0 1-B 1-M 1-E 0, and the last 0 numerical value represents the end of the sentence symbol. Each word in diabetes can be encoded as DISEASE with letters, and non-medical terms are encoded with O, then the coding sequence of diabetes is DISEASE-B DISEASE-M DISEASE-E, and the final coding sequence is: O OO OOO DISEASE-B DISEASE-M DISEASE-E O, and the last letter O represents the end of the sentence symbol. Take "Unfortunately, he suffered from diabetes." as an English sentence example, and the coding of the English sentence is O OO OO DISEASE-S O. Take the word as the smallest unit and encode each word.

[0091] Step S7: using each of the sentences and the corresponding coding sequence to train a neural network term extraction model to obtain a trained neural network term extraction model; identifying a medical text to be identified according to the trained neural network term extraction model to obtain a coding sequence of the medical text to be identified; determining the medical terms and corresponding term types in the medical text to be identified according to the coding sequence of the medical text to be identified.

[0092] The neural network term extraction model model includes neural network term extraction model models of multiple language types.

[0093] Considering that the language type of the medical text to be identified is uncertain, it may be a language type included in the multilingual parallel corpus or it may not be, so it is necessary to first determine the language type of the medical text to be identified, that is, the medical text to be identified according to the trained neural network term extraction model is identified to obtain the encoding sequence of the medical text to be identified, which specifically includes:

[0094] Determining whether the language type of the medical text to be identified is a language type in the multilingual parallel corpus;

[0095] If so, the trained neural network term extraction model is used to identify the medical text to be identified, and a coding sequence of the medical text to be identified is obtained; here, the trained neural network term extraction model has the same language type as the medical text to be identified.

[0096] If not, the medical text to be identified is translated into a sentence of any language type in the multilingual parallel corpus to obtain a translated medical sentence, and the trained neural network term extraction model is used to identify the translated medical sentence to obtain a coding sequence of the translated medical sentence.

[0097] The medical text to be identified is translated into sentences of any language type in the multilingual parallel corpus, and the language type after translation is determined according to demand. For example, if the medical text to be identified (Mongolian) needs to be translated into (Chinese), the Mongolian to Chinese translation model can be directly used for translation.

[0098] When translating the medical text to be recognized into the speech types included in the multilingual parallel corpus, the aforementioned translation method based on direct conversion and the translation method based on the intermediate language can still be used.

[0099] The applied neural network term extraction model uses the BERT-TRANSFORMER-CRF algorithm, which takes medical sentences and coding sequences as input and labels, and uses BERT language training to embed them in TRANSFORMER-CRF to achieve automatic extraction of language features. BERT learns the grammatical features of various languages, and TRANSFORMER-CRF encodes the word meaning vector input to the BERT layer and adds position information. The final result is output to the CRF layer.

[0100] Using the BERT model, a multilingual conditional random field model is constructed. The CRF algorithm inputs a character sequence or a word sequence, outputs the medical term type and sequence score, and normalizes the obtained score to obtain its probability distribution. The sequence-level likelihood function is used to solve the maximum a posteriori probability to improve the accuracy of medical term prediction. Finally, the highest score sequence is output as the prediction sequence.

[0101] Transformer is a deep learning model with excellent natural language processing effect. It mainly consists of a word embedding layer, a position embedding layer, an encoder and a decoder. Its encoder and decoder are composed of multiple Transformer_blocks, each of which consists of a multi-head attention mechanism, a normalization layer, a forward fully connected layer, and a residual link. The Q, K, and V of the encoder's multi-head attention mechanism are the same, which is the segmentation of the word vector of the input sentence. The Q and K of the decoder are the output of the encoder, and V represents the word vector of the encoded sequence. The output of the encoder will be passed to the CRF layer. The specific model is as follows Figure 5 As shown in the figure, the model parameters can be adjusted according to the number of parallel corpora. The default number of embedding layers is 512, the number of multi-head attention mechanisms is 4, the number of transformer_blocks is set to 3, and the number of forward fully connected layers is 512. The adam optimizer is used to adjust the learning rate. 4096 tokens are input at a time.

[0102] In order to enrich the data in the constructed multilingual parallel corpus, sentences not stored in the multilingual parallel corpus can be further stored in the term extraction process. Therefore, after determining the medical terms and corresponding term types in the medical text to be identified according to the coding sequence of the medical text to be identified, the following steps are also included:

[0103] Determining whether the medical text to be recognized is stored in the multilingual parallel corpus;

[0104] If not, the medical text to be recognized is stored in the multilingual parallel corpus, and the medical text to be recognized is translated into sentences of other language types in the multilingual parallel corpus and stored in the multilingual parallel corpus to enrich the data volume, increase the training data and model effect of the translation model and the recognition model, and whenever the newly saved data volume reaches 5% of the existing data volume, the model is trained on the original basis.

[0105] In this embodiment, medical term extraction can lock the key points of relevant medical knowledge in long articles or paragraphs, quickly extract medical expertise, speed up the dissemination and exchange of relevant medical information, have multilingual characteristics, use neural machine translation methods, enable multilingual translation in the medical field, and spread medical knowledge from medically developed areas to medically backward areas according to the language pairs selected by users. The more language categories, the wider the audience.

[0106] The establishment of a multilingual medical parallel corpus and a term extraction device based on the multilingual medical parallel corpus can accelerate the global exchange of medical knowledge, improve the health and medical literacy of people in all countries of the world, provide relevant knowledge and experience for the world to correctly understand and respond to medical problems, and provide solutions for the world to deal with sudden medical problems. The short and concise medical knowledge composed of medical terms can allow people to understand and remember the relevant content in the shortest time, and get the key points faster when users look up medical knowledge. Ultimately, the goal of improving the global medical and health level is achieved, and it also caters to my country's initiative to build a joint prevention and control network.

[0107] Example 2

[0108] like Figure 6 As shown, this embodiment provides a medical term extraction system based on a multilingual parallel corpus, including:

[0109] A sentence semantic similarity calculation module M1 for medical knowledge data, used to obtain medical knowledge data and perform semantic similarity calculation on all sentences in the medical knowledge data;

[0110] A parallel corpus acquisition module M2 is used to form parallel corpora from sentences whose semantic similarity values ​​are greater than a first preset value, so as to obtain parallel corpora of multiple semantics; each of the parallel corpora includes sentences of at least one language type;

[0111] A translation module M3, used for translating the missing language in each of the parallel corpora to obtain sentences of the missing language;

[0112] A multilingual parallel corpus construction module M4 is used to construct sentences of all types of languages ​​into a multilingual parallel corpus;

[0113] A word segmentation module M5, used for performing word segmentation on the sentences in each parallel corpus in the multilingual parallel corpus, and removing non-medical terms in each word segmentation to obtain medical terms of each parallel corpus;

[0114] A sentence encoding module M6 is used to encode and annotate all the sentences in the multilingual corpus according to the medical terms in the parallel corpus and the corresponding term types to obtain an encoding sequence for each sentence;

[0115] The medical term extraction module M7 is used to train a neural network term extraction model using each of the sentences and the corresponding coding sequence to obtain a trained neural network term extraction model; identify the medical text to be identified according to the trained neural network term extraction model to obtain the coding sequence of the medical text to be identified; and determine the medical terms and corresponding term types in the medical text to be identified according to the coding sequence of the medical text to be identified.

[0116] As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0117] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for extracting medical terms based on a multilingual parallel corpus, characterized in that: include: Acquire medical knowledge data and calculate semantic similarity of all sentences in the medical knowledge data; Sentences with semantic similarity values ​​greater than a first preset value are grouped into parallel corpora to obtain parallel corpora with multiple semantics; each parallel corpus includes sentences of at least one language type; Translating each missing language in the parallel corpus to obtain sentences of the missing language; the parallel corpus includes a preset number of language types; the preset number of language types include types of the missing languages ​​and types of existing languages ​​in the parallel corpus; Construct a multilingual parallel corpus of sentences from all types of languages; Segmenting the sentences in each of the parallel corpora in the multilingual parallel corpus, and removing non-medical terms in each segmentation, to obtain medical terms for each of the parallel corpora; Encoding and annotating all the sentences in the multilingual corpus according to the medical terms in the parallel corpus and the corresponding term types to obtain a coding sequence for each of the sentences; Using each of the sentences and the corresponding coding sequence to train a neural network term extraction model to obtain a trained neural network term extraction model; identifying a medical text to be identified according to the trained neural network term extraction model to obtain a coding sequence of the medical text to be identified; determining medical terms and corresponding term types in the medical text to be identified according to the coding sequence of the medical text to be identified; The step of translating the missing language in each of the parallel corpora to obtain sentences of the missing language specifically includes: For each of the parallel corpora, a missing language is randomly selected; Selecting a sentence in an existing language from the parallel corpus as an application language sentence according to the translation accuracy of each existing language in the parallel corpus when translated into the currently selected missing language; translating a sentence in the currently selected missing language using the application language sentence; A vacant language is randomly selected from the remaining vacant languages ​​and recorded as the currently selected vacant language, and the process returns to the step of "selecting a sentence of an existing language from the parallel corpus as an applied language sentence according to the translation accuracy of each existing language in the parallel corpus when translated into the currently selected vacant language", until all the sentences of the vacant language are translated.

2. The method according to claim 1, characterized in that The step of selecting a sentence in an existing language from the parallel corpus as an application language sentence according to the translation accuracy of each existing language in the parallel corpus when translated into the currently selected missing language specifically includes: Determining the translation accuracy of each translation model applied when translating each existing language in the parallel corpus into the currently selected missing language; Determining whether there is a translation model whose translation accuracy is greater than a first preset value; If not, then selecting one of the remaining vacant languages ​​from the remaining vacant languages ​​according to the accuracy of translating the remaining vacant languages ​​into the selected vacant language; Translating sentences in the selected remaining vacant language using the sentences in the existing language; Translating the selected sentence of the missing language by using the selected sentence of the remaining missing language; If so, the translation model with the translation accuracy greater than a preset value is used to translate the sentence of the missing language.

3. The method according to claim 1 or 2, characterized in that: The method of translating the currently selected sentence of the missing language by using the application language sentence further specifically includes: Determine whether the similarity between the semantics corresponding to the currently selected sentence of the missing language and the semantics of the application language sentence is greater than a second preset value; If yes, retain the translated sentence in the currently selected missing language; If not, then all sentences of the language expression type corresponding to the semantics of the application language sentence are deleted.

4. The method according to claim 1, characterized in that: The segmenting of sentences in each parallel corpus in the multilingual parallel corpus and removing non-medical terms in each segmented word to obtain medical terms in each parallel corpus specifically includes: For each of the parallel corpora, a sentence of a target language is selected according to the usage frequency of the language; Performing lexical segmentation on the selected sentence to obtain various segmentation words; Removing non-medical terms from each of the segmented words according to a term dictionary of the language corresponding to the segmented words to obtain medical terms in the target language; Medical terms in other languages ​​are derived respectively according to the medical terms in the target language; and the medical terms in each of the parallel corpora include the medical terms in the target language and the medical terms in the other languages.

5. The method according to claim 1 or 4, characterized in that: The step of encoding and marking all the sentences in the multilingual corpus according to the medical terms in the parallel corpus and the corresponding term types to obtain the encoding sequence of each sentence specifically includes: For each of the sentences, determining non-medical terms and medical terms and term types in the sentence based on medical terms of corresponding language types; encoding the medical terms in the sentences according to the term types; encoding the non-medical terms in the sentence; the encoding of the medical terms in the sentence is different from the encoding of the non-medical terms; A coding sequence of the sentence is determined according to the coding of the medical terms and the coding of the non-medical terms in the sentence.

6. The method according to claim 1, characterized in that The step of identifying the medical text to be identified according to the trained neural network term extraction model to obtain the encoding sequence of the medical text to be identified specifically includes: Determining whether the language type of the medical text to be identified is a language type in the multilingual parallel corpus; If yes, then using the trained neural network term extraction model to identify the medical text to be identified, and obtaining a coding sequence of the medical text to be identified; If not, the medical text to be identified is translated into a sentence of any language type in the multilingual parallel corpus to obtain a translated medical sentence, and the trained neural network term extraction model is used to identify the translated medical sentence to obtain a coding sequence of the translated medical sentence.

7. The method according to claim 1 or 6, characterized in that: After determining the medical terms and corresponding term types in the medical text to be identified according to the coding sequence of the medical text to be identified, the method further includes: Determining whether the medical text to be recognized is stored in the multilingual parallel corpus; If not, the medical text to be recognized is stored in the multilingual parallel corpus, and sentences of other language types in the multilingual parallel corpus are translated from the medical text to be recognized and stored in the multilingual parallel corpus.

8. A system based on the method according to any one of claims 1 to 7, characterized in that: include: A sentence semantic similarity calculation module for medical knowledge data, used to obtain medical knowledge data and perform semantic similarity calculation on all sentences in the medical knowledge data; A parallel corpus acquisition module, used to form parallel corpora from sentences whose semantic similarity values ​​are greater than a first preset value, to obtain parallel corpora of multiple semantics; each of the parallel corpora includes sentences of at least one language type; A translation module, used for translating the missing language in each of the parallel corpora to obtain sentences of the missing language; Translating the missing language in each of the parallel corpora to obtain sentences of the missing language specifically includes: For each of the parallel corpora, a missing language is randomly selected; Selecting a sentence in an existing language from the parallel corpus as an application language sentence according to the translation accuracy of each existing language in the parallel corpus when translated into the currently selected missing language; translating a sentence in the currently selected missing language using the application language sentence; Randomly select a vacant language from the remaining vacant languages ​​and record it as the currently selected vacant language, and return to the step of "selecting a sentence of an existing language from the parallel corpus as the application language sentence according to the translation accuracy of each existing language in the parallel corpus when translated into the currently selected vacant language", until all the sentences of the vacant language are translated; A multilingual parallel corpus construction module, which is used to construct sentences of all types of languages ​​into a multilingual parallel corpus; A word segmentation module, used for performing word segmentation on the sentences in each parallel corpus in the multilingual parallel corpus, and removing non-medical terms in each word segmentation to obtain medical terms of each parallel corpus; A sentence encoding module, used for encoding and marking all the sentences in the multilingual corpus according to the medical terms in the parallel corpus and the corresponding term types, to obtain an encoding sequence for each sentence; The medical term extraction module is used to train a neural network term extraction model using each of the sentences and the corresponding coding sequence to obtain a trained neural network term extraction model; identify a medical text to be identified according to the trained neural network term extraction model to obtain a coding sequence of the medical text to be identified; and determine the medical terms and corresponding term types in the medical text to be identified according to the coding sequence of the medical text to be identified.

Images