A method for determining vocabulary relevance based on subject information and sub-domain prior knowledge

Abstract

The present invention discloses a method for determining vocabulary relevance based on subject information and sub-domain prior knowledge, the method comprising: obtaining a material text, a material text title, and a sub-domain material text, and processing the material text, the material text title, and the sub-domain material text; determining preliminary word representation data of scientific vocabulary in the material text and the material text title based on the processed material text and the material text title; determining representation data of keywords in the sub-domain material text based on the processed sub-domain material text; determining cosine similarity between scientific vocabulary based on the representation data of keywords and the preliminary word representation data, and determining the relevance between scientific vocabulary based on the cosine similarity. The present invention introduces text information rich in sub-domain prior knowledge into the word representation layer, enhances the ability of word vectors to represent scientific vocabulary in specific fields, screens scientific vocabulary based on similarity, and accurately obtains the relevance between scientific vocabulary.

Description

Technical Field

[0001] The present invention relates to the field of computer application technology, and in particular to a method for determining vocabulary relevance based on subject information and sub-domain prior knowledge. Background Art

[0002] Over time, the scientific achievements in the form of text in materials science literature have become increasingly rich, containing a wealth of knowledge worthy of exploration and research. Extracting and analyzing effective knowledge information from the massive unstructured materials science text corpus, and performing automated association and reasoning to discover new materials and predict material properties have become key challenges in materials informatics. For text mining in the field of materials science, the complexity and diversity of professional vocabulary (such as various crystal structures, interface structures, and other terms) pose a significant challenge to supervised natural language processing methods. This is because supervised learning requires a large number of well-annotated data sets, which inevitably leads to high manual annotation costs. Therefore, unsupervised techniques are more cost-effective.

[0003] To date, unsupervised word embedding technology has been applied to information extraction tasks from large-scale materials science literature. Its core concept is based on the distributional assumption: if two words have similar contexts, then they have similar semantics. In the field of materials science, if two materials have similar contexts in a text, given that contexts generally include descriptive terms about the material's structure, properties, and applications, semantic similarity implies that the two materials are also similar. This similarity can be categorized as the same material or have similar applications. Based on this concept, the word2vec model (an implementation of word embedding) has been applied to extract information from materials science text corpora. Word2vec includes the continuous bag-of-words (CBOW) model and the skip-gram model. Both models are three-layer neural networks with an input layer, a hidden layer, and an output layer. The difference between them is that CBOW takes the context of a word as input to predict the word itself, while skip-gram takes a word as input to predict its surrounding context.

[0004] However, the existing technology has the following drawbacks: 1) The distributed representation of materials science vocabulary is insufficiently rich. This is reflected in the fact that the corpus used for word embedding training is single, for example, word embedding is performed only on abstracts or body paragraphs, which results in a lack of main information about the words and an inability to enhance the context; 2) There is a lack of characteristic information about materials subfields. The field of materials science includes many subfields with distinct characteristics, such as lithium-ion battery cathode materials and catalytic materials. Simply embedding words in materials science texts will lead to a sparse prior knowledge of the subfields, further making it impossible to accurately associate and reason about scientific vocabulary in materials texts.

[0005] Therefore, the existing technology still needs to be improved and enhanced. Summary of the Invention

[0006] The technical problem to be solved by the present invention is that, in response to the above-mentioned defects of the existing technology, a method for determining vocabulary relevance based on subject information and sub-field prior knowledge is provided, aiming to solve the problem in the existing technology that, when extracting scientific vocabulary, the scientific vocabulary in the materials science text cannot be accurately associated due to the insufficient richness of the distributed representation of the scientific vocabulary and the lack of sub-field feature information.

[0007] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0008] In a first aspect, the present invention provides a method for determining vocabulary relevance based on subject information and sub-domain prior knowledge, wherein the method comprises:

[0009] Acquire a material text, a material text title, and a sub-field material text, and process the material text, the material text title, and the sub-field material text;

[0010] Determining preliminary representation data of scientific terms in the material text and the material text title based on the processed material text and the material text title;

[0011] Determining, based on the processed sub-field material text, representation data of keywords in the sub-field material text;

[0012] Based on the keyword representation data and the word preliminary representation data, the cosine similarity between the scientific terms is determined, and the relevance between the scientific terms is determined based on the cosine similarity.

[0013] In one implementation, the processing of the material text, the material text title, and the sub-field material text includes:

[0014] Deleting punctuation marks in the material text, the material text title, and the sub-field material text;

[0015] Standardizing the material text, the material text title, and the chemical formulas in the sub-field material text;

[0016] Deleting accent marks from the material text, the material text title, and the sub-field material text;

[0017] generating material science phrases based on the material text, the material text title, and the sub-field material text;

[0018] The first letters of the words in the material text, the material text title and the sub-field material text, excluding the chemical formula or element matching words, are lowercase.

[0019] In one implementation, the standardization of the material text, the material text title, and the chemical formula in the sub-field material text includes:

[0020] Screening the material text, the material text title, and the chemical formulas in the sub-field material text;

[0021] The elements in the screened chemical formula are arranged in alphabetical order and the atomic number ratios reflected in the screened chemical formula are normalized to the smallest integer ratio.

[0022] In one implementation, determining preliminary representation data of scientific terms in the material text and the material text title based on the processed material text and the material text title includes:

[0023] Performing word embedding on the processed material text and the material text title respectively to obtain corresponding word vectors;

[0024] Concatenate the word vectors to obtain a combined word vector;

[0025] The combined word vector is input into BiLSTM for encoding to obtain the preliminary representation data of the word.

[0026] In one implementation, determining the representation data of keywords in the sub-field material text based on the processed sub-field material text includes:

[0027] Calculate the word frequency-inverse document frequency value of each word in the sub-field material text in the material text and the material text title;

[0028] sorting the words based on the word frequency-inverse document frequency value, and using the top ranked words as keywords in the processed sub-field material text;

[0029] The keywords are grouped into keyword groups, and the keyword groups are passed through a word embedding matrix and a fully connected neural network to obtain representation data of the keywords.

[0030] In one implementation, determining the cosine similarity between the scientific terms based on the keyword representation data and the preliminary word representation data, and determining the relevance between the scientific terms based on the cosine similarity, includes:

[0031] Performing a keyword attention operation on the keyword representation data and the preliminary word representation data to obtain new representation data of each scientific word based on the keyword attention distribution;

[0032] Calculating the cosine similarity between the scientific terms based on the new representation data;

[0033] Based on the cosine similarity, the correlation between the scientific terms is determined.

[0034] In one implementation, the method further includes:

[0035] determining selected words from said scientific vocabulary,

[0036] sorting the remaining words according to the cosine similarity between the selected words in the scientific vocabulary and the remaining words to obtain a similar word list;

[0037] Based on the similar word list, it is determined whether new material exists.

[0038] In a second aspect, an embodiment of the present invention further provides a vocabulary relevance determination system based on subject information and sub-domain prior knowledge, wherein the system includes:

[0039] A text processing module, configured to obtain material text, material text title, and sub-field material text, and process the material text, the material text title, and the sub-field material text;

[0040] a preliminary representation determination module, configured to determine preliminary representation data of scientific terms in the material text and the material text title based on the processed material text and the material text title;

[0041] A keyword representation determination module, configured to determine representation data of keywords in the sub-field material text based on the processed sub-field material text;

[0042] The relevance determination module is configured to determine the cosine similarity between the scientific terms based on the keyword representation data and the word preliminary representation data, and determine the relevance between the scientific terms based on the cosine similarity.

[0043] In a third aspect, an embodiment of the present invention further provides a terminal device, comprising a memory, a processor, and a vocabulary relevance determination program based on subject information and sub-domain prior knowledge stored in the memory and runnable on the processor. When the processor executes the vocabulary relevance determination program based on subject information and sub-domain prior knowledge, the steps of the vocabulary relevance determination method based on subject information and sub-domain prior knowledge as described in any one of the above-mentioned schemes are implemented.

[0044] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, on which is stored a vocabulary relevance determination program based on subject information and sub-domain prior knowledge. When the vocabulary relevance determination program based on subject information and sub-domain prior knowledge is executed by a processor, the steps of the vocabulary relevance determination method based on subject information and sub-domain prior knowledge as described in any one of the above schemes are implemented.

[0045] Beneficial effects: Compared with the prior art, the present invention provides a method for determining vocabulary relevance based on subject information and sub-domain prior knowledge, the method comprising: obtaining a material text, a material text title and a sub-domain material text, and processing the material text, the material text title and the sub-domain material text; determining preliminary word representation data of scientific vocabulary in the material text and the material text title based on the processed material text and the material text title; determining representation data of keywords in the sub-domain material text based on the processed sub-domain material text; determining cosine similarity between the scientific vocabulary based on the representation data of the keywords and the preliminary word representation data, and determining the relevance between the scientific vocabulary based on the cosine similarity. Based on sub-domain prior knowledge, the present invention introduces text information rich in sub-domain prior knowledge into the representation layer of words, improves the density of sub-domain prior information, enhances the ability of word vectors to represent scientific vocabulary of materials in a specific field, and screens according to the similarity of scientific vocabulary, so as to accurately derive the relevance between scientific vocabulary. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flowchart of a specific implementation of a method for determining vocabulary relevance based on subject information and sub-domain prior knowledge provided by an embodiment of the present invention.

[0047] Figure 2This is a module design diagram of a method for determining vocabulary relevance based on subject information and sub-domain prior knowledge provided by an embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of a system for determining vocabulary relevance based on subject information and sub-domain prior knowledge, provided by an embodiment of the present invention.

[0049] Figure 4 A schematic diagram of a terminal device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0051] This embodiment provides a method for determining vocabulary relevance based on subject information and sub-domain prior knowledge. The method of this embodiment can accurately determine the relevance between scientific terms in a material text. In specific implementation, this embodiment first obtains a material text, a material text title, and a sub-domain material text, and processes the material text, the material text title, and the sub-domain material text. Then, based on the processed material text and the material text title, preliminary word representation data for scientific terms in the material text and the material text title is determined. Next, based on the processed sub-domain material text, representation data for keywords in the sub-domain material text is determined. Finally, based on the representation data for the keywords and the preliminary word representation data, the cosine similarity between the scientific terms is determined, and the relevance between the scientific terms is determined based on the cosine similarity. Thus, this embodiment, based on sub-domain prior knowledge, introduces text information rich in sub-domain prior knowledge into the word representation layer, improves the density of sub-domain prior information, enhances the ability of word vectors to represent scientific terms in a specific domain, and filters based on the similarity of scientific terms, accurately deriving the relevance between scientific terms.

[0052] Exemplary Methods

[0053] The vocabulary relevance determination method based on subject information and sub-domain prior knowledge of this embodiment can be applied to terminal devices, which can be intelligent terminal products such as computers and mobile phones. Specifically, Figure 1 As described in , the method for determining vocabulary relevance based on subject information and sub-domain prior knowledge of this embodiment includes the following steps:

[0054] Step S100: Acquire material text, material text title, and sub-field material text, and process the material text, the material text title, and the sub-field material text.

[0055] When acquiring text, this embodiment not only acquires the material text and the material text title, but also acquires the sub-domain material text to incorporate prior knowledge of the sub-domain material text. Furthermore, in this embodiment, after acquiring the material text, the material text title, and the sub-domain material text, regular expressions and rule-based methods are first used to process the material text, the material text title, and the sub-domain material text.

[0056] In one implementation, this embodiment includes the following steps when processing the material text, the material text title, and the sub-field material text:

[0057] Step S101: deleting punctuation marks in the material text, the material text title, and the sub-field material text;

[0058] Step S102: standardizing the material text, the material text title, and the chemical formula in the sub-field material text;

[0059] Step S102: deleting accent marks in the material text, the material text title, and the sub-field material text;

[0060] Step S13: generating material science phrases based on the material text, the material text title, and the sub-field material text;

[0061] Step S104: lowercase the first letters of the words in the material text, the material text title, and the sub-field material text, excluding the words that correspond to chemical formulas or elements.

[0062] During specific implementation, the present embodiment deletes all punctuation marks in the material text, the material text title and the sub-domain material text by software, to eliminate the useless information in the material text, the material text title and the sub-domain material text.Then the chemical formula in the material text, the material text title and the sub-domain material text is screened, retaining simple chemical formula, and the elements of the simple chemical formula retained are arranged in alphabetical order and the atomic number ratio reflected by the simple chemical formula is standardized to the minimum integer ratio, with standardized simple chemical formula. Further, the present embodiment deletes the accent mark in the material text, the material text title and the sub-domain material text, and considers the unit that often occurs in the materials science text. Therefore, software can ignore the character string of length 1, in order to avoid it being deleted by mistake. Because there are many phrases in the materials science text, for example " lithium battery ", this type of phrase only has practical significance when appearing in the combination vocabulary form, therefore, the present embodiment also generates materials science phrases based on the material text, the material text title and the sub-domain material text, thereby generates meaningful combination vocabulary. Finally, this embodiment also lowercases the first letter of any word in the material text, the material text title, and the sub-domain material text, excluding chemical formulas or element symbols. For example, if a scientific word is not a chemical formula or element symbol and only the first letter is capitalized, the software will lowercase that letter. This process lowercases words and proper nouns at the beginning of a sentence.

[0063] Step S200: Based on the processed material text and the material text title, preliminary representation data of scientific vocabulary in the material text and the material text title are determined.

[0064] In one implementation, this embodiment includes the following steps when determining preliminary word representation data:

[0065] Step S201: performing word embedding on the processed material text and the material text title respectively to obtain corresponding word vectors;

[0066] Step S202: concatenate the word vectors to obtain a combined word vector;

[0067] Step S203: Input the combined word vector into BiLSTM for encoding to obtain the preliminary representation data of the word.

[0068] Specifically, after processing the material text, the material text title, and the sub-field material text, this embodiment determines preliminary representation data of scientific vocabulary from the material text and the material text title. Since both the material text and the material text title contain subject information, this embodiment can perform word embedding on both the material text and the material text title to obtain their respective corresponding word vectors. The word embedding refers to converting a word into a vector representation. Figure 2 As shown in the combined word embedding module in , this embodiment concatenates the two obtained separate word vectors to obtain the combined word vector x n The combined word vector is shown in formula (1), t n The word vector representing the material text title, m n The word vector representing the material text, x n Represents the concatenated combined word vector.

[0069] x n =[t n ;m n ] (1)

[0070] Combined word vector x n Next, it will be input into BiLSTM (Bi-directional Long Short-Term Memory) for encoding to obtain the initial representation of the word h n The preliminary representation data of words is shown in formula (2), where P is the parameter of BiLSTM.

[0071] h n =BiLSTM(x n ;P) (2)

[0072] This embodiment takes into account that the material text title contains obvious subject information, so word embedding is performed on the material science text and the material text title respectively, and the two obtained word vectors are spliced. This operation incorporates explicit subject information into the word vector, which can strengthen the context and reflect the differences between the same word in different texts, thereby increasing the amount of information in the distributed representation of the word.

[0073] In addition, in order to enhance the main information and enrich the distributed representation of words, this embodiment uses the material text title as the carrier of the main information and combines its corresponding word vector with the word vector of the materials science text. In this process, the choice of the main information carrier is replaceable. In addition to the material text title, text keywords or other texts containing obvious main information can also be used. In addition, the method of combining the two word vectors is also flexible. This embodiment selects the direct splicing method. In addition, other reasonable methods such as multiplying by a weight coefficient and then combining can also be adopted.

[0074] Step S300: Determine the representation data of the keywords in the sub-field material text based on the processed sub-field material text.

[0075] In one implementation, determining the keyword representation data in this embodiment includes the following steps:

[0076] Step S301: Calculate the word frequency-inverse document frequency value of each word in the sub-field material text in the material text and the material text title;

[0077] Step S202: sorting the words based on the word frequency-inverse document frequency value, and using the top ranked words as keywords in the processed sub-field material text;

[0078] Step S203: The keywords are grouped into keyword groups, and the keyword groups are passed through a word embedding matrix and a fully connected neural network to obtain representation data of the keywords.

[0079] In specific implementation, this embodiment calculates the term frequency-inverse document frequency (TF-IDF, Term Frequency) of each word in the sub-field material text in the material text and the title of the material text.

[0080] Frequency–Inverse Document Frequency) is calculated, and the term frequency (Term

[0081] Frequency (TF) refers to the number of times a given word appears in the material text and the title of the material text. The main idea of ​​Inverse Document Frequency (IDF) is: if the word t i The fewer documents there are, the larger the IDF is, which means that the word has a good ability to distinguish categories. Specifically, in a certain category of documents d j The word t in i The TF-IDF value calculation formula is as follows:

[0082]

[0083] tfidf i,j =tf i,j ×idf i (5)

[0084] Among them, n i,j is the word in document d j The number of times it appears in k n k.j It is document d j The sum of the number of occurrences of all words in the corpus, |D| is the total number of documents in the corpus, |{j:t i ∈d j}| is a string containing the word t i The number of documents.

[0085] After calculating the word frequency-inverse document frequency value for each word in the sub-domain material text, the words are sorted based on the word frequency-inverse document frequency value, and the top ranked words are used as keywords. Next, the keywords are extracted and selected to form a keyword set. After the set is passed through the word embedding matrix and a fully connected neural network, the keyword representation data y is obtained. n .

[0086] This embodiment takes into account that materials science contains a large number of sub-fields. If only word embedding is performed on materials science text without prior knowledge of the sub-fields, the obtained word vector cannot be enriched with information about materials in specific sub-fields, which will directly affect the association and reasoning effect between materials science vocabulary. Therefore, this embodiment introduces prior knowledge of sub-fields by selecting keywords from material texts in specific sub-fields. Keyword-based attention operations can further enrich the representation of words and enhance the ability of word vectors to represent materials science vocabulary in specific sub-fields.

[0087] Furthermore, to introduce prior knowledge in sub-domains, this embodiment uses the TF-IDF algorithm to extract keywords from sub-domain material texts and uses keyword-based attention operations to enhance the quality of word vectors. In this process, the method of introducing prior knowledge is flexibly replaceable. For example, a Transformer can be used to extract feature information from sub-domain texts and then incorporate this feature information into word vectors, or a knowledge graph can be constructed for the materials sub-domain, providing sub-domain prior knowledge through the knowledge graph. In general, any solution that adopts the same design concept and improves the key vocabulary and reasoning effects of materials science by enhancing the main information and introducing prior knowledge in the materials science sub-domain falls within the scope of protection of this invention.

[0088] Step S400: Determine the cosine similarity between the scientific terms based on the keyword representation data and the preliminary word representation data, and determine the relevance between the scientific terms based on the cosine similarity.

[0089] In this embodiment, determining the relevance between the scientific terms includes the following steps:

[0090] Step S401: Perform keyword attention operation on the keyword representation data and the preliminary word representation data to obtain new representation data of each scientific word based on keyword attention distribution;

[0091] Step S402: Calculate the cosine similarity between the scientific terms based on the new representation data;

[0092] Step S403: Determine the relevance between the scientific terms based on the cosine similarity.

[0093] When implementing it specifically, Figure 2 As shown in the sub-domain keyword attention module in this embodiment, the keyword representation data y n The initial representation data of the word h n (i.e., the output of BiLSTM), perform keyword attention operation to obtain the new representation data x for each scientific word based on the keyword attention distribution ′ n The relevant calculation formula is as follows:

[0094]

[0095] x ′ n =∑ n α n y n (7)

[0096] Among them, α n is the attention weight, and l is the length of the sequence composed of the keyword set.

[0097] This embodiment can determine the relevance between scientific terms based on the calculated cosine similarity between scientific terms, that is, determine whether they belong to the same type of material.

[0098] In one implementation, after obtaining the cosine similarity between various scientific terms, this embodiment determines selected terms within the scientific terms. The remaining terms are then sorted based on the cosine similarity between the selected terms and the remaining terms within the scientific terms to obtain a list of similar terms. For example, when performing associative reasoning on a novel lithium-ion battery cathode material, the selected term is "cathode." Terms similar to "cathode" (i.e., the remaining terms from all scientific terms after removing the selected terms) are sorted to obtain a sorted list of similar terms, as shown in Table 1. In Table 1, for illustrative purposes, the sorted similar terms are represented by term 1, term 2, and so on.

[0099]

[0100] Table 1

[0101] Next, manually check the list of similar words to see if there are any new materials that have not been reported. Taking the new lithium-ion battery positive electrode material as an example, the sorted similar word list can be combined with literature and patent searches to check whether there are any positive electrode materials that have not been discovered.

[0102] In summary, the present embodiment first obtains the material text, the material text title and the sub-field material text, and processes the material text, the material text title and the sub-field material text. Then, based on the processed material text and the material text title, the preliminary representation data of the scientific vocabulary in the material text and the material text title is determined. Next, based on the processed sub-field material text, the representation data of the keywords in the sub-field material text is determined. Finally, based on the representation data of the keywords and the preliminary representation data of the words, the cosine similarity between the scientific vocabulary is determined, and the correlation between the scientific vocabulary is determined based on the cosine similarity. It can be seen that, in this embodiment, based on the sub-field prior knowledge, text information rich in sub-field prior knowledge is introduced into the representation layer of words, the density of the sub-field prior information is improved, the ability of the word vector to represent the scientific vocabulary of the material in a specific field is enhanced, and screening is performed according to the similarity of the scientific vocabulary, so that the correlation between the scientific vocabulary can be accurately obtained.

[0103] From a qualitative perspective, the vocabulary correlation determination method based on subject information and sub-domain prior knowledge in this embodiment enriches the distributed representation of materials science vocabulary, significantly improves the ability of word vectors to represent materials science vocabulary, and is beneficial to improving the correlation and reasoning effect between materials science vocabulary; from a quantitative perspective, the design of combined word embedding and sub-domain keyword attention mechanism can incorporate more text statistical information of materials science vocabulary into word vectors, so that the cosine similarity between word vectors can better represent the correlation relationship and degree between materials science vocabulary, and can better associate scientific vocabulary through text semantic information, facilitating reasoning.

[0104] Exemplary devices

[0105] Based on the above embodiments, the present invention also provides a vocabulary relevance determination system based on subject information and sub-domain prior knowledge, such as Figure 3 As shown in , the system includes: a text processing module 10, a preliminary representation determination module 20, a keyword representation determination module 30 and a relevance determination module 40. Specifically, the text processing module 10 is used to obtain material text, material text title and sub-field material text, and process the material text, the material text title and the sub-field material text. The preliminary representation determination module 20 is used to determine the preliminary representation data of scientific vocabulary in the material text and the material text title based on the processed material text and the material text title. The keyword representation determination module 30 is used to determine the representation data of keywords in the sub-field material text based on the processed sub-field material text. The relevance determination module 40 is used to determine the cosine similarity between the scientific vocabulary based on the representation data of the keywords and the preliminary representation data of the words, and determine the relevance between the scientific vocabulary based on the cosine similarity.

[0106] In one implementation, the text processing module 10 includes:

[0107] a punctuation mark processing unit, configured to delete punctuation marks in the material text, the material text title, and the sub-field material text;

[0108] A chemical formula standardization unit, configured to perform standardization processing on the chemical formulas in the material text, the material text title, and the sub-field material text;

[0109] an accent processing unit, configured to delete accents in the material text, the material text title, and the sub-field material text;

[0110] a phrase generating unit, configured to generate material science phrases based on the material text, the material text title, and the sub-field material text;

[0111] The letter lowercase processing unit is used to lowercase the first letters of the words in the material text, the material text title and the sub-field material text except the chemical formula or element matching words.

[0112] In one implementation, the chemical formula standardization unit includes:

[0113] A chemical formula screening unit, configured to screen the chemical formulas in the material text, the material text title, and the sub-field material text;

[0114] The normalization unit is used to arrange the elements in the screened chemical formula in alphabetical order and normalize the atomic number ratio reflected by the screened chemical formula to a minimum integer ratio.

[0115] In one implementation, the preliminary representation determination module 20 includes:

[0116] A word embedding unit, configured to perform word embedding on the processed material text and the material text title respectively to obtain word vectors corresponding to the scientific vocabulary;

[0117] A concatenation unit, configured to concatenate the word vectors to obtain a combined word vector;

[0118] The encoding unit is used to input the combined word vector into the BiLSTM for encoding to obtain the preliminary representation data of the word.

[0119] In one implementation, the keyword representation determination module 30 includes:

[0120] A calculation unit, configured to calculate a word frequency-inverse document frequency value of each word in the sub-field material text in the material text and the material text title;

[0121] a ranking unit, configured to rank the words based on the word frequency-inverse document frequency value, and use the top ranked words as keywords in the processed sub-field material text;

[0122] The data generating unit is used to group the keywords into keyword groups, and obtain the representation data of the keywords after the keyword groups are passed through a word embedding matrix and a fully connected neural network.

[0123] In one implementation, the relevance determination module 40 includes:

[0124] an attention operation unit, configured to perform a keyword attention operation on the keyword representation data and the preliminary word representation data to obtain new representation data of each scientific word based on the keyword attention distribution;

[0125] a cosine similarity calculation unit, configured to calculate the cosine similarity between the scientific terms based on the new representation data;

[0126] The relevance determination unit is configured to determine the relevance between the scientific terms based on the cosine similarity.

[0127] In one implementation, the system further includes:

[0128] a vocabulary determination module for determining selected vocabulary among the scientific vocabulary,

[0129] a sorting module, configured to sort the remaining words according to the cosine similarity between the selected words in the scientific vocabulary and the remaining words, to obtain a similar word list;

[0130] The new material determination module is used to determine whether a new material exists based on the similar word list.

[0131] The working principles of each module in this embodiment are the same as the principles of each step in the above method embodiment, and will not be repeated here.

[0132] Based on the above embodiment, the present invention further provides a terminal device, whose principle block diagram can be shown as follows: Figure 4 As shown. The terminal device includes a processor and a memory connected via a system bus. The processor of the terminal device is used to provide computing and control capabilities. The memory of the terminal device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the terminal device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for determining vocabulary relevance based on subject information and sub-domain prior knowledge is implemented.

[0133] Those skilled in the art will understand that Figure 4 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the solution of the present invention, and does not constitute a limitation on the terminal device to which the solution of the present invention is applied. The specific terminal device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0134] In one embodiment, a terminal device is provided. The terminal device includes a memory, a processor, and a method program for determining vocabulary relevance based on subject information and sub-domain prior knowledge, stored in the memory and executable on the processor. When the processor executes the method program for determining vocabulary relevance based on subject information and sub-domain prior knowledge, the following operating instructions are implemented:

[0135] Acquire a material text, a material text title, and a sub-field material text, and process the material text, the material text title, and the sub-field material text;

[0136] Determining preliminary representation data of scientific terms in the material text and the material text title based on the processed material text and the material text title;

[0137] Determining, based on the processed sub-field material text, representation data of keywords in the sub-field material text;

[0138] Based on the keyword representation data and the word preliminary representation data, the cosine similarity between the scientific terms is determined, and the relevance between the scientific terms is determined based on the cosine similarity.

[0139] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, operating database or other media used in the embodiments provided by the present invention may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0140] In summary, the present invention discloses a method for determining vocabulary relevance based on subject information and sub-domain prior knowledge, the method comprising: obtaining a material text, a material text title, and a sub-domain material text, and processing the material text, the material text title, and the sub-domain material text; determining preliminary word representation data of scientific vocabulary in the material text and the material text title based on the processed material text and the material text title; determining representation data of keywords in the sub-domain material text based on the processed sub-domain material text; determining cosine similarity between scientific vocabulary based on the representation data of keywords and the preliminary word representation data, and determining the relevance between scientific vocabulary based on the cosine similarity. The present invention introduces text information rich in sub-domain prior knowledge into the word representation layer, enhances the ability of word vectors to represent scientific vocabulary in specific fields, screens according to the similarity of scientific vocabulary, and accurately obtains the relevance between scientific vocabulary.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for determining vocabulary relevance based on subject information and sub-domain prior knowledge, characterized in that: The method comprises: Acquire a material text, a material text title, and a sub-field material text, and process the material text, the material text title, and the sub-field material text; Determining preliminary representation data of scientific terms in the material text and the material text title based on the processed material text and the material text title; Determining, based on the processed sub-domain material text, representation data of keywords in the sub-domain material text; determining cosine similarities between the scientific terms based on the keyword representation data and the preliminary word representation data, and determining associations between the scientific terms based on the cosine similarities; The determining, based on the processed material text and the material text title, preliminary representation data of scientific vocabulary in the material text and the material text title includes: Performing word embedding on the processed material text and the material text title respectively to obtain corresponding word vectors; Concatenate the word vectors to obtain a combined word vector; Input the combined word vector into BiLSTM for encoding to obtain preliminary representation data of the word; The preliminary data of the term are: h n =BiLSTM(x n ;P) Among them, P is the parameter of BiLSTM, x n Represents the combined word vector after splicing, x n =[t n ;m n ], t n The word vector representing the material text title, m n Word vectors representing material text; The step of determining the representation data of keywords in the sub-field material text based on the processed sub-field material text includes: Calculate the word frequency-inverse document frequency value of each word in the sub-field material text in the material text and the material text title; sorting the words based on the word frequency-inverse document frequency value, and using the top ranked words as keywords in the processed sub-field material text; The keywords are grouped into keyword groups, and the keyword groups are passed through a word embedding matrix and a fully connected neural network to obtain representation data of the keywords; The determining of the cosine similarity between the scientific terms based on the keyword representation data and the preliminary word representation data, and determining the relevance between the scientific terms based on the cosine similarity, includes: Performing a keyword attention operation on the keyword representation data and the preliminary word representation data to obtain new representation data of each scientific word based on keyword attention distribution; Calculating the cosine similarity between the scientific terms based on the new representation data; Determining the relevance between the scientific terms based on the cosine similarity; The new representation data of each scientific word based on keyword attention distribution is as follows: Among them, y n is the keyword representation data, α n is the attention weight, and l is the length of the sequence composed of the keyword set.

2. The method for determining vocabulary relevance based on subject information and sub-domain prior knowledge according to claim 1, characterized in that: The processing of the material text, the material text title and the sub-field material text includes: Deleting punctuation marks in the material text, the material text title, and the sub-field material text; Standardizing the material text, the material text title, and the chemical formulas in the sub-field material text; Deleting accent marks from the material text, the material text title, and the sub-field material text; generating material science phrases based on the material text, the material text title, and the sub-field material text; The first letters of the words in the material text, the material text title and the sub-field material text, excluding the chemical formula or element matching words, are lowercase.

3. The method for determining vocabulary relevance based on subject information and sub-domain prior knowledge according to claim 2, characterized in that: The standardization of the material text, the material text title, and the chemical formula in the sub-field material text includes: Screening the material text, the material text title, and the chemical formulas in the sub-field material text; The elements in the screened chemical formula are arranged in alphabetical order and the atomic number ratios reflected in the screened chemical formula are normalized to the smallest integer ratio.

4. The method for determining vocabulary relevance based on subject information and sub-domain prior knowledge according to claim 1, characterized in that: The method further comprises: determining selected words from said scientific vocabulary, sorting the remaining words according to the cosine similarity between the selected words in the scientific vocabulary and the remaining words to obtain a similar word list; Based on the similar word list, it is determined whether new material exists.

5. A vocabulary relevance determination system based on subject information and sub-domain prior knowledge, characterized in that: The system is used to implement the steps of the method for determining vocabulary relevance based on subject information and sub-domain prior knowledge according to any one of claims 1 to 4, and the system includes: A text processing module, configured to obtain material text, material text title, and sub-field material text, and process the material text, the material text title, and the sub-field material text; a preliminary representation determination module, configured to determine preliminary representation data of scientific terms in the material text and the material text title based on the processed material text and the material text title; A keyword representation determination module, configured to determine representation data of keywords in the sub-field material text based on the processed sub-field material text; The relevance determination module is configured to determine the cosine similarity between the scientific terms based on the keyword representation data and the word preliminary representation data, and determine the relevance between the scientific terms based on the cosine similarity.

6. A terminal device, characterized in that: The terminal device includes a memory, a processor, and a vocabulary relevance determination program based on subject information and sub-field prior knowledge stored in the memory and runnable on the processor. When the processor executes the vocabulary relevance determination program based on subject information and sub-field prior knowledge, the steps of the vocabulary relevance determination method based on subject information and sub-field prior knowledge as described in any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a vocabulary relevance determination program based on subject information and sub-domain prior knowledge. When the vocabulary relevance determination program based on subject information and sub-domain prior knowledge is executed by the processor, the steps of the vocabulary relevance determination method based on subject information and sub-domain prior knowledge as described in any one of claims 1 to 4 are implemented.

Images