CN112527977B - Concept extraction method, concept extraction device, electronic equipment and storage medium - Google Patents

Abstract

The embodiment of the invention provides a concept extraction method, a device, electronic equipment and a storage medium, wherein the method comprises the following steps: extracting terms from the text to be extracted according to a preset word list, obtaining a first candidate concept list, and carrying out entity linking on the text to be extracted according to a preset knowledge graph, so as to obtain a second candidate concept list; reordering each candidate concept in the first candidate concept list and the second candidate concept list, and acquiring a concept extraction result of the text to be extracted according to the reordered result; the text to be extracted is unstructured text. According to the concept extraction method, the device, the electronic equipment and the storage medium, the candidate concepts obtained by carrying out term extraction and entity link on the text to be extracted are reordered, and the concept extraction result is obtained according to the reordered result, so that the concepts can be extracted from unstructured texts more efficiently and accurately under the condition that the labeled data are less or even no labeled data exist.

Description

Concept extraction method, device, electronic device and storage medium

Technical Field

The present invention relates to the field of computer technology, and in particular to a concept extraction method, device, electronic equipment and storage medium.

Background technique

A concept, also known as a scientific concept, is a term or phrase used in scientific corpus to represent specific technologies and important knowledge points. For example, "binary tree" is an important concept in the computer field.

Traditional concept extraction methods mainly include three categories: key phrase and term extraction, entity linking, and concept/set expansion. Key phrase and term extraction generally obtains candidate phrases through methods such as word segmentation, and then sorts the candidate phrases by confidence, and selects the candidate phrases with higher scores as the extraction results. Entity linking is to find different mentions of entities in the background knowledge base from the text. Concept expansion tasks are to find concepts that belong to the same set as a small number of given seed concepts by analyzing large-scale resources such as corpus and external knowledge bases.

However, the above three methods rely on expert annotation. Under the premise of more annotated data, they can obtain relatively accurate concept extraction results. However, when there is less or no annotated data (such as unstructured text), the accuracy of the concept extraction results of the above three methods is low.

Summary of the invention

Embodiments of the present invention provide a concept extraction method, device, electronic device and storage medium to solve the defect in the prior art that the accuracy of concept extraction results is low when there is less labeled data, and to achieve more accurate concept extraction when there is less labeled data or even no labeled data.

An embodiment of the present invention provides a concept extraction method, including:

Extracting terms from the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and performing entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list;

Reordering the candidate concepts in the first candidate concept list and the second candidate concept list, and obtaining a concept extraction result of the text to be extracted according to the reordering result;

Wherein, the text to be extracted is unstructured text.

According to a concept extraction method of an embodiment of the present invention, the specific step of obtaining the concept extraction result of the text to be extracted according to the reordering result includes:

According to the re-ranking result and a preset scoring threshold, or according to the re-ranking result and a preset number, a plurality of the candidate concepts are selected as the concept extraction result of the text to be extracted.

According to a concept extraction method of an embodiment of the present invention, the specific steps of extracting terms from the text to be extracted according to a preset vocabulary and obtaining a first candidate concept list include:

Filtering the text to be extracted according to the preset vocabulary to obtain the intersection of the preset vocabulary and the text to be extracted;

The intersection is segmented and POS-tagged, and nouns in the intersection are obtained as candidate concepts to form the first candidate concept list.

According to a concept extraction method according to an embodiment of the present invention, the specific step of reordering the candidate concepts in the first candidate concept list and the second candidate concept list includes:

reordering the candidate concepts in the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted;

And/or, clustering is performed according to the word vectors of each candidate concept in the first candidate concept list and the second candidate concept list, and reordering is performed according to the clustering results and the word vectors of preset seed concepts.

According to a concept extraction method according to an embodiment of the present invention, the specific steps of reordering the candidate concepts in the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted include:

According to the TF-IDF method, obtaining the confidence of each candidate concept in the first candidate concept list;

A re-ranking result is obtained according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, and the second candidate concept list.

According to the concept extraction method of an embodiment of the present invention, the specific steps of clustering the word vectors of each candidate concept in the first candidate concept list and the second candidate concept list, and reordering the word vectors according to the clustering results and the preset seed concept include:

Obtaining the word vector of each candidate concept in the first candidate concept list and the second candidate concept list, and performing clustering according to the word vector of each candidate concept in the first candidate concept list and the second candidate concept list;

According to the similarity between each cluster center obtained by clustering and each pre-set sub-concept cluster center, the clusters to which each cluster center belongs are sorted;

For each of the clusters, the candidate concepts belonging to the cluster are ranked according to the similarity between the seed concept cluster center having the greatest similarity to the cluster center and the word vector of each candidate concept belonging to the cluster.

According to a concept extraction method of an embodiment of the present invention, the specific steps of obtaining the re-ranking result according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, and the second candidate concept list include:

For each candidate concept in the intersection of the first candidate concept list and the second candidate concept list, obtaining a score of each candidate concept in the intersection according to the confidence of each candidate concept in the intersection and the frequency of occurrence in the text to be extracted, and determining the score of each candidate concept that does not belong to the intersection as zero;

The re-ranking result is obtained according to the score of each candidate concept in the first candidate concept list and the second candidate concept list.

The embodiment of the present invention further provides a concept extraction device, comprising:

An extraction module is used to extract terms from the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and to perform entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list;

A screening module, used for reordering the candidate concepts in the first candidate concept list and the second candidate concept list, and obtaining a concept extraction result of the text to be extracted according to the reordering result;

Wherein, the text to be extracted is unstructured text.

An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the above-mentioned concept extraction methods are implemented.

An embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned concept extraction methods are implemented.

The concept extraction method, device, electronic device and storage medium provided in the embodiments of the present invention perform term extraction on the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and perform entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list, reorder the candidate concepts in the first candidate concept list and the second candidate concept list, and obtain the concept extraction result of the text to be extracted according to the reordering result. In the case of less or no annotated data, concepts can be extracted from unstructured text more efficiently and accurately, which can reduce manual work and improve the degree of automation.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are 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.

FIG1 is a schematic diagram of a flow chart of a concept extraction method provided by an embodiment of the present invention;

FIG2 is a schematic diagram of the structure of a concept extraction device provided by an embodiment of the present invention;

FIG. 3 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.

In the description of the embodiments of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the embodiments of the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the embodiments of the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

In the description of the embodiments of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the embodiments of the present invention can be understood according to specific circumstances.

In order to overcome the above-mentioned problems of the prior art, an embodiment of the present invention provides a concept extraction method, device, electronic device and storage medium. The inventive concept is to extract candidate concepts from unstructured text by different methods, sort the extracted candidate concepts according to the possibility of being concepts, and filter out concept extraction results according to the sorting, thereby filtering out the concepts that best fit the technical field.

FIG1 is a flow chart of a concept extraction method provided by an embodiment of the present invention. The concept extraction method of an embodiment of the present invention is described below in conjunction with FIG1. As shown in FIG1, the method includes: step S101, extracting terms from the text to be extracted according to a preset vocabulary, obtaining a first candidate concept list, and performing entity linking on the text to be extracted according to a preset knowledge graph, obtaining a second candidate concept list.

Among them, the text to be extracted is unstructured text.

Specifically, the text to be extracted may be unstructured text such as web page parsing results or subtitles in teaching videos. The concept extraction method provided by the embodiment of the present invention is particularly suitable for extracting concepts from a course in a certain technical field.

The text to be extracted can be used for preliminary concept extraction through term extraction and entity linking to obtain multiple candidate concepts.

The preliminary concept extraction of the extracted text by term extraction method can include:

According to the preset vocabulary, nouns in the text to be extracted are obtained as candidate concepts.

The preset vocabulary may be a large vocabulary, including a large number of words used in the language used by the text to be extracted.

The candidate concepts obtained by the term extraction method constitute the first candidate concept list.

The preliminary concept extraction of the extracted text through entity linking method can include:

Use the linking language to perform entity linking on the extracted text, and obtain each candidate concept from the preset knowledge graph based on the linking results.

The preset knowledge graph is a knowledge graph that includes knowledge in the technical field to which the text to be extracted belongs.

For example, Xlink can be used to link unstructured text and obtain candidate concepts from XLore.

XLink is an entity linking system based on the cross-language knowledge base XLORE. XLink can identify entities in text documents (such as news, blogs, etc.) input by users and link them to the corresponding entities (concepts, instances) in XLORE. XLink can bridge text information and knowledge graphs, providing external knowledge for text understanding. In addition, XLink can help readers understand ambiguous and uncommon entities and improve text comprehension.

XLore is the first large-scale knowledge graph that balances Chinese and English knowledge. XLORE extracts structured information from heterogeneous cross-language online encyclopedias, integrates Chinese and English Wikipedias, French Wikipedias and Baidu Encyclopedia, and structures and cross-language links on encyclopedia knowledge. It is a large-scale multilingual knowledge graph with a relatively balanced scale of Chinese and English knowledge.

The candidate concepts obtained by the entity linking method constitute the second candidate concept list.

Step S102: reorder the candidate concepts in the first candidate concept list and the second candidate concept list, and obtain a concept extraction result of the text to be extracted according to the reordering result.

Specifically, each candidate concept in the first candidate concept list and the second candidate concept list refers to each candidate concept in the union of the first candidate concept list and the second candidate concept list.

For each candidate concept in the union of the first candidate concept list and the second candidate concept list, the scores of each candidate concept are compared and sorted according to the comparison result to obtain a re-sorting result.

The score can reflect the probability that the candidate concept is a concept. The higher the score, the greater the probability that the candidate concept is a concept; the lower the score, the smaller the probability that the candidate concept is a concept.

The scores of each candidate concept in the union of the first candidate concept list and the second candidate concept list as a concept may be obtained, and the candidate concepts in the union may be sorted according to the probabilities that the candidate concepts are concepts.

It is also possible that instead of directly obtaining the probability that each candidate concept in the union of the first candidate concept list and the second candidate concept list is a concept, the probability that each candidate concept in the union is a concept is indirectly compared, and the candidate concepts are sorted according to the comparison result.

According to the re-ranking result, a certain number of candidate concepts can be screened out from the union of the first candidate concept list and the second candidate concept list as the concept extraction result.

It should be noted that the traditional key phrase and term extraction method extracts key phrases from the text. Generally, candidate phrases are obtained through word segmentation and other methods, and then they are ranked by confidence, and the parts with higher scores are selected as the extraction results. Generally, they are divided into supervised methods and unsupervised methods: supervised methods judge whether a word is a target phrase (i.e., concept) by training classification models, while unsupervised methods, such as TextRank, construct semantic graphs through context co-occurrence and other methods, and then propagate confidence to obtain results. However, for most actual concept acquisition scenarios, since the labeled data is very scarce, supervised methods are often difficult to implement, and the extraction accuracy is extremely low; for unsupervised methods, once the text length is too short (such as some web page parsing results, course subtitles in teaching, etc.), traditional context-based methods are difficult to build maps, so it is difficult to obtain reliable confidence results, and the extraction accuracy is low. In addition, since a large number of concepts actually appear at low frequencies in texts, it is difficult to achieve good results using only statistical methods such as TF-IDF.

Compared with traditional key phrase and term extraction methods, the embodiments of the present invention extract terms from the text to be extracted through a large vocabulary, extract noun phrases from the unstructured text as candidate concepts, obtain a first candidate concept list, and reorder the second candidate concept list obtained by entity linking according to a preset knowledge graph. This can overcome the shortcomings of traditional key phrase and term extraction methods to a certain extent, and can obtain concept extraction results with higher accuracy when there is less or even no annotated data.

The traditional entity linking method is to find different mentions of entities in the background knowledge base from the text. Its implementation requires a large-scale list of entities, and then matches and filters the target text to obtain the link results. The main defect of concept extraction through entity linking method is that it is difficult to evaluate the relevance and importance of the obtained results to the text: for entity linking methods, their main goal is to match the entities in the text with the existing knowledge base. However, in actual scenarios, not all entities that appear in the text should be considered as concepts. This makes it very easy to introduce noise to concept acquisition using only entity linking methods, resulting in low extraction accuracy.

Compared with the traditional entity linking method, the embodiment of the present invention, on the one hand, performs entity linking based on the knowledge graph, which can improve the extraction accuracy to a certain extent compared with the traditional entity linking method based on the entity list; on the other hand, the second candidate concept list obtained by entity linking is combined with the first candidate concept list obtained by term extraction of the text to be extracted through a large vocabulary, and re-sorted, which can further overcome the shortcomings of the traditional entity linking method and obtain more accurate concept extraction results when there is less annotated data or even no annotated data.

Traditional concept expansion/set expansion methods analyze large-scale resources such as corpus and external knowledge bases to find concepts that belong to the same set as a small number of given seed concepts. This method focuses on the number and accuracy of the results obtained. It is a method for concept acquisition under the setting of few annotated resources. It generally achieves the goal by analyzing semantic templates and ensemble learning methods. Concept expansion/set expansion methods are more in line with practical applications: there are a small number of high-quality annotation results, and qualified concepts are found from large-scale corpora. However, due to the lack of supervision signals, concepts are difficult to control during multiple rounds of expansion, resulting in serious semantic drift problems (as the set expands, its meaning gradually deviates from the initial meaning). In addition, since the concept expansion task depends on the evaluation of semantic templates, the discovery of new words and other processes, its performance is often constrained by the results of the pre-performed segmentation and sorting of candidate concepts. This makes it necessary to consume a lot of resources for evaluation and annotation of the expanded results in practical applications.

Compared with traditional concept expansion/set expansion methods, the embodiment of the present invention re-sorts the first candidate concept list obtained by term extraction of the extracted text based on a large vocabulary and the second candidate concept list obtained by entity linking based on a preset knowledge graph. In this way, more accurate concept extraction results can be obtained when there is less or no annotated data, without consuming a lot of resources for evaluation and annotation.

The embodiment of the present invention performs term extraction on the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and performs entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list, reorders the candidate concepts in the first candidate concept list and the second candidate concept list, and obtains a concept extraction result of the text to be extracted according to the reordering result. It can extract concepts from unstructured text more efficiently and accurately when there is less or no annotated data, and can reduce manpower and improve the degree of automation.

Based on the contents of the above embodiments, the specific steps of obtaining the concept extraction result of the text to be extracted according to the reordering results include: selecting multiple candidate concepts as the concept extraction result of the text to be extracted according to the reordering results and a preset scoring threshold, or according to the reordering results and a preset number.

Specifically, according to the re-ranking result, candidate concepts meeting preset conditions may be selected as concept extraction results of the text to be extracted.

If the re-ranking result is obtained by ranking the candidate concepts in the union of the first candidate concept list and the second candidate concept list according to the size of the concept scores, the preset condition may be met if the concept score is greater than a preset score threshold.

If the result of reordering is obtained by indirectly comparing the probability of each candidate concept in the union of the first candidate concept list and the second candidate concept list being a concept, the first N of the reordering results (arranged in descending order) that meet the preset condition may be a positive integer representing a preset number.

By presetting conditions, concepts that are more suitable for the technical field can be screened out from the candidate concepts.

The embodiment of the present invention screens candidate concepts according to the re-ranking result and preset conditions to obtain the concept extraction result of the text to be extracted, which can improve the accuracy of concept extraction.

Based on the contents of the above embodiments, the specific steps of extracting terms from the text to be extracted according to the preset vocabulary and obtaining the first candidate concept list include: filtering the text to be extracted according to the preset vocabulary and obtaining the intersection of the preset vocabulary and the text to be extracted.

Specifically, firstly, the text to be extracted is filtered by a vocabulary, and the text to be extracted is filtered by using a preset vocabulary to obtain the intersection of the preset vocabulary and the text to be extracted, that is, each word in the vocabulary in the text to be extracted is obtained to form the intersection.

The intersection is segmented and POS tagged, and the nouns in the intersection are obtained as candidate concepts to form a first candidate concept list.

Specifically, after the vocabulary filtering is performed, part-of-speech filtering may be performed.

The intersection can be segmented and POS tagged by Natural Language Processing (NLP) methods.

For example, the intersection can be segmented and POS tagged using open source segmentation tools such as Jieba, Ansj or Pangu segmentation.

After the intersection is segmented and POS tagged, the intersection can be screened according to the tagged POS to obtain nouns (including phrases) in the intersection as candidate concepts to form a first candidate concept list.

The embodiment of the present invention performs vocabulary filtering on the text to be extracted through a preset vocabulary, performs part-of-speech filtering on the result of the vocabulary filtering, and selects nouns in the text to be extracted as candidate concepts, which can improve the accuracy of concept extraction.

Based on the contents of the above-mentioned embodiments, the specific steps of reordering the candidate concepts in the first candidate concept list and the second candidate concept list include: reordering the candidate concepts in the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted; and/or, clustering the word vectors of each candidate concept in the first candidate concept list and the second candidate concept list, and reordering according to the clustering results and the word vectors of preset seed concepts.

Specifically, the candidate concepts in the first candidate concept list and the second candidate concept list are reordered, and formula reordering or word vector clustering reordering may be used alone, or formula reordering and word vector clustering reordering may be combined.

Formula reordering refers to reordering the candidate concepts in the first candidate concept list and the second candidate concept list according to a preset formula.

The preset formula is used to score the candidate concepts according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted. The score is used to describe the probability that the candidate concept is a concept.

The word vector clustering and re-ranking refers to clustering the word vectors of each candidate concept in the first candidate concept list and the second candidate concept list, and re-ranking according to the clustering results and the word vectors of the preset seed concepts.

Seed concepts refer to pre-determined concepts in the technical field to which the text to be extracted belongs.

The number of the preset seed concepts is multiple. Preferably, the number of the preset seed concepts may be less than 20.

Combining formula reordering and word vector clustering reordering means that after performing formula reordering and word vector clustering reordering respectively, the ranking results obtained by formula reordering and the ranking results obtained by word vector clustering reordering are merged, and the two ranking results are combined to obtain the final reordering result.

The sorting results obtained by formula reordering and the sorting results obtained by word vector clustering reordering are merged. The more accurate sorting result can be selected as the final reordering result, or the final reordering result can be obtained by weighted summing of the two results.

The embodiment of the present invention can obtain more accurate concept extraction results according to the reordering results by performing formula reordering and/or word vector clustering reordering on each candidate concept in the first candidate concept list and the second candidate concept list.

Based on the contents of the above embodiments, the specific steps of reordering the candidate concepts in the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted include: obtaining the confidence of each candidate concept in the first candidate concept list according to the TF-IDF method.

Specifically, for each candidate concept in the first candidate concept list, the confidence of the candidate concept can be obtained by using the TF-IDF method.

TF-IDF (Term Frequency Inverse Document Frequency) is a commonly used weighting technique for information retrieval and data mining.

TF-IDF is a statistical method used to assess the importance of a word to a document in a document set or a corpus. The importance of a word increases proportionally with the number of times it appears in a document, but decreases inversely with the frequency of its occurrence in the corpus.

The re-ranking result is obtained according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, and the second candidate concept list.

Specifically, the more important or basic the concept is, the higher the frequency of its appearance in the text to be extracted. Therefore, the re-ranking is not simply based on the confidence, but also takes into account the frequency of appearance in the text to be extracted. The confidence of the candidate concept is combined with the frequency of appearance in the text to be extracted to obtain the score of each candidate concept in the union of the first candidate concept list and the second candidate concept list.

According to the score of each candidate concept in the union, the candidate concepts in the union may be sorted in descending order according to the score to obtain a re-sorted result.

The embodiment of the present invention obtains the re-ranking result based on the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, as well as the second candidate concept list, which can more accurately reflect the probability of each candidate concept being a concept, thereby obtaining a more accurate concept extraction result based on the re-ranking result.

Based on the contents of the above embodiments, clustering is performed according to the word vectors of each candidate concept in the first candidate concept list and the second candidate concept list, and the specific steps of reordering according to the clustering results and the word vectors of the preset seed concepts include: obtaining the word vectors of each candidate concept in the first candidate concept list and the second candidate concept list, and clustering is performed according to the word vectors of each candidate concept in the first candidate concept list and the second candidate concept list.

Specifically, the Word2vec model may be used to obtain the word vector of each candidate concept in the union of the first candidate concept list and the second candidate concept list.

After obtaining the word vector of each candidate concept in the union, a preset clustering method can be used to perform clustering to obtain multiple clusters.

For example, a K-means clustering algorithm may be used to cluster the candidate concepts in the union according to the word vectors of the candidate concepts in the union.

For each cluster, the average vector of the word vectors of the candidate concepts belonging to the cluster can be obtained as the center of the cluster. The center of the cluster is also a word vector.

According to the similarity between each cluster center obtained by clustering and each pre-set sub-concept cluster center, the clusters to which each cluster center belongs are sorted.

Specifically, for any cluster center and any seed concept cluster center, the similarity between the cluster center and the word vector of the seed concept can be obtained.

Similarity can be cosine similarity, Euclidean distance, or Mahalanobis distance, etc.

It is understandable that clustering can be performed in advance based on the word vectors of various preset sub-concepts to obtain multiple seed concept clusters. The number of seed concept clusters can be the same as or different from the number of clusters of each candidate concept. The word vector of the seed concept is also obtained through the Word2vec model. The center of the seed concept cluster can be the average vector of the word vectors of various sub-concepts belonging to the seed concept cluster.

The clusters to which the various cluster centers belong can be sorted in descending order according to the maximum similarity between the cluster center and the various sub-concept cluster centers.

For each cluster, the candidate concepts belonging to the cluster are ranked according to the distance between the seed concept cluster center with the greatest similarity to the cluster center and the word vector of each candidate concept belonging to the cluster.

Specifically, for each cluster obtained by clustering, the distance between the seed concept cluster center with the greatest similarity to the cluster center and the word vector of each candidate concept belonging to the cluster can be obtained.

Distance can be cosine distance, Euclidean distance, or Mahalanobis distance, etc.

The candidate concepts belonging to the cluster can be sorted in ascending order according to the distance between the seed concept cluster center having the greatest similarity to the cluster center and the word vector of each candidate concept belonging to the cluster, thereby obtaining a re-sorting result.

It can be understood that, for the candidate concepts belonging to this cluster, the smaller the distance between the center of the seed concept cluster with the greatest similarity to the center of the cluster and the word vector of the candidate concept, the greater the probability that the candidate concept is the concept; the greater the distance between the center of the seed concept cluster with the greatest similarity to the center of the cluster and the word vector of the candidate concept, the smaller the probability that the candidate concept is the concept.

For any two candidate concepts with the same ranking within the clusters, the possibility that the candidate concept belonging to the cluster with the earlier ranking is the concept is higher than the possibility that the candidate concept belonging to the cluster with the later ranking is the concept.

The embodiment of the present invention obtains a reordering result by clustering the word vectors of each candidate concept in the first candidate concept list and the second candidate concept list, which can more accurately reflect the relative size of the probability that each candidate concept is a concept, thereby obtaining a more accurate concept extraction result based on the reordering result.

Based on the contents of the above embodiments, according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, and the second candidate concept list, the specific steps of obtaining the reordering results include: for each candidate concept in the intersection of the first candidate concept list and the second candidate concept list, according to the confidence of each candidate concept in the intersection and the frequency of occurrence in the text to be extracted, obtain the score of each candidate concept in the intersection, and determine the score of each candidate concept that does not belong to the intersection as zero.

Specifically, the confidence of the candidate concept may be combined with the frequency of occurrence in the text to be extracted to obtain the score of each candidate concept in the union of the first candidate concept list and the second candidate concept list.

First, the intersection of the first candidate concept list and the second candidate concept list is obtained.

For any candidate concept, if the candidate concept belongs to the intersection, the probability that the candidate concept is a concept is higher than the probability that the candidate concept not belonging to the intersection is a concept, and the score of each candidate concept not belonging to the intersection can be determined to be zero.

For each candidate concept belonging to the intersection, the score of the candidate concept can be obtained according to the following formula:

R＝ln(freq)*max(conf-c,0)

Among them, R represents the score of the candidate concept; freq represents the frequency of occurrence (i.e., the number of occurrences) of the candidate concept in the text to be extracted; conf represents the confidence of the candidate concept; and c represents the confidence threshold.

The re-ranking result is obtained according to the score of each candidate concept in the first candidate concept list and the second candidate concept list.

Specifically, according to the score of each candidate concept in the union of the first candidate concept list and the second candidate concept list, the candidate concepts in the union may be sorted in descending order according to the score to obtain a re-sorted result.

The embodiment of the present invention obtains the re-ranking result based on the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, as well as the second candidate concept list, which can more accurately reflect the probability of each candidate concept being a concept, thereby obtaining a more accurate concept extraction result based on the re-ranking result.

By manually selecting 18 preset seed concepts, taking the subtitles of the "Data Structure" course in MOOCCube as the text to be extracted, and using the methods provided in the above embodiments of the present invention to perform concept extraction, 138 concepts can be obtained from the top 200 candidate concepts; further, the above 138 concepts are combined with the above 18 concepts as seed concepts, and the "Data Structure" related encyclopedia is taken as the text to be extracted, and the methods provided in the above embodiments of the present invention are used to perform concept extraction, 174 concepts can be obtained from the top 200 candidate concepts.

It can be seen that the methods provided by the above embodiments of the invention can extract concepts from unstructured text more efficiently and accurately when there is less or no annotated data.

The concept extraction device provided by the embodiment of the present invention is described below. The concept extraction device described below and the concept extraction method described above can be referenced to each other.

Fig. 2 is a schematic diagram of the structure of a concept extraction device according to an embodiment of the present invention. Based on the contents of the above embodiments, as shown in Fig. 2, the device includes an extraction module 201 and a screening module 202, wherein:

The extraction module 201 is used to extract terms from the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and to perform entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list;

A screening module 202 is used to reorder the candidate concepts in the first candidate concept list and the second candidate concept list, and obtain a concept extraction result of the text to be extracted according to the reordering result;

Among them, the text to be extracted is unstructured text.

Specifically, the extraction module 201 and the screening module 202 are electrically connected.

The extraction module 201 may include a first extraction submodule and a second extraction submodule.

The first extraction submodule is used to extract terms from the text to be extracted according to a preset vocabulary, obtain multiple candidate concepts, and form a first candidate concept list.

The second extraction submodule is used to perform entity linking on the text to be extracted according to a preset knowledge graph, obtain multiple candidate concepts, and form a second candidate concept list.

The screening module 202 compares the scores of the candidate concepts in the union of the first candidate concept list and the second candidate concept list, and sorts the candidate concepts according to the comparison results to obtain a re-sorting result.

The score can reflect the probability that the candidate concept is a concept. The higher the score, the greater the probability that the candidate concept is a concept; the lower the score, the smaller the probability that the candidate concept is a concept.

The screening module 202 may include a reordering submodule and a screening submodule.

The screening submodule is used to select candidate concepts that meet the preset conditions according to the re-ranking results as the concept extraction results of the text to be extracted.

The first extraction submodule may include: a vocabulary filtering unit and a part-of-speech filtering unit.

A vocabulary filtering unit, used to filter the text to be extracted according to a preset vocabulary, and obtain the intersection of the preset vocabulary and the text to be extracted;

The part-of-speech filtering unit is used to perform word segmentation and part-of-speech tagging on the intersection, obtain the nouns in the intersection as candidate concepts, and form a first candidate concept list.

The reordering submodule may include a formula reordering unit and/or a word vector clustering reordering unit.

A formula reordering unit, used to reorder each candidate concept in the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted;

The word vector clustering and reordering unit is used to cluster the word vectors of each candidate concept in the first candidate concept list and the second candidate concept list, and reorder them according to the clustering results and the word vectors of the preset seed concepts.

The formula re-ranking unit may include a confidence acquisition sub-unit and a scoring sub-unit.

The confidence acquisition subunit is used to acquire the confidence of each candidate concept in the first candidate concept list according to the TF-IDF method.

The scoring subunit is used to obtain the re-ranking result according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, as well as the second candidate concept list.

The word vector clustering re-ranking unit may include a clustering sub-unit, an inter-class ranking sub-unit and an intra-class ranking sub-unit.

The clustering subunit is used to obtain the word vector of each candidate concept in the first candidate concept list and the second candidate concept list, and to perform clustering according to the word vector of each candidate concept in the first candidate concept list and the second candidate concept list.

The inter-class sorting subunit is used to sort the clusters to which each class cluster center belongs according to the similarity between each class cluster center obtained by clustering and each preset sub-concept class cluster center.

The intra-class sorting subunit is used to sort the candidate concepts belonging to the cluster for each cluster according to the similarity between the seed concept cluster center with the greatest similarity to the cluster center and the word vector of each candidate concept belonging to the cluster.

The scoring subunit is specifically used to obtain the score of each candidate concept in the intersection of the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the intersection and the frequency of occurrence in the text to be extracted, and determine the score of each candidate concept that does not belong to the intersection as zero; obtain the re-ranking result according to the score of each candidate concept in the first candidate concept list and the second candidate concept list.

The concept extraction device provided in an embodiment of the present invention is used to execute the concept extraction method provided in the above-mentioned embodiments of the present invention. The specific methods and processes for each module included in the concept extraction device to implement the corresponding functions are detailed in the embodiments of the above-mentioned concept extraction method and will not be repeated here.

The concept extraction device is used in the concept extraction method of the above embodiments. Therefore, the description and definition of the concept extraction method in the above embodiments can be used for understanding each execution module in the embodiments of the present invention.

The embodiment of the present invention performs term extraction on the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and performs entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list, reorders the candidate concepts in the first candidate concept list and the second candidate concept list, and obtains a concept extraction result of the text to be extracted according to the reordering result. It can extract concepts from unstructured text more efficiently and accurately when there is less or no annotated data, which can reduce manpower and improve the degree of automation.

Figure 3 illustrates a schematic diagram of the physical structure of an electronic device. As shown in Figure 3, the electronic device may include: a processor 301, a memory 302 and a bus 303; wherein the processor 301 and the memory 302 communicate with each other through the bus 303; the processor 301 is used to call computer program instructions stored in the memory 302 and executable on the processor 301 to execute the concept extraction method provided by the above-mentioned method embodiments, the method comprising: performing term extraction on the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and performing entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list; reordering each candidate concept in the first candidate concept list and the second candidate concept list, and obtaining a concept extraction result of the text to be extracted according to the reordering result; wherein the text to be extracted is unstructured text.

In addition, the logic instructions in the above-mentioned memory 302 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

On the other hand, an embodiment of the present invention further provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the concept extraction method provided by the above-mentioned method embodiments, the method including: extracting terms on the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and performing entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list; reordering each candidate concept in the first candidate concept list and the second candidate concept list, and obtaining a concept extraction result of the text to be extracted according to the reordering result; wherein the text to be extracted is unstructured text.

On the other hand, an embodiment of the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it is implemented to execute the concept extraction method provided by the above-mentioned embodiments, the method comprising: extracting terms from the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and performing entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list; reordering each candidate concept in the first candidate concept list and the second candidate concept list, and obtaining a concept extraction result of the text to be extracted according to the reordering result; wherein the text to be extracted is unstructured text.

The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM/RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

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 embodiments of the present invention.

Claims (7)

1. A concept extraction method, characterized by comprising: Extracting terms from the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and performing entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list; Reordering the candidate concepts in the first candidate concept list and the second candidate concept list, and obtaining a concept extraction result of the text to be extracted according to the reordering result; Wherein, the text to be extracted is unstructured text; The specific steps of reordering the candidate concepts in the first candidate concept list and the second candidate concept list include: reordering the candidate concepts in the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted; and/or, clustering the word vectors of the candidate concepts in the first candidate concept list and the second candidate concept list, and reordering the word vectors according to the clustering results and the word vectors of the preset seed concepts; The specific step of reordering the candidate concepts in the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted includes: According to the TF-IDF method, obtaining the confidence of each candidate concept in the first candidate concept list; Obtaining a re-ranking result according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, and the second candidate concept list; The specific steps of clustering the word vectors of the candidate concepts in the first candidate concept list and the second candidate concept list and reordering the word vectors according to the clustering results and the word vectors of the preset seed concepts include: Obtaining the word vector of each candidate concept in the first candidate concept list and the second candidate concept list, and performing clustering according to the word vector of each candidate concept in the first candidate concept list and the second candidate concept list; According to the similarity between each cluster center obtained by clustering and each pre-set sub-concept cluster center, the clusters to which each cluster center belongs are sorted; For each of the clusters, the candidate concepts belonging to the cluster are ranked according to the similarity between the seed concept cluster center having the greatest similarity to the cluster center and the word vector of each candidate concept belonging to the cluster. 2. The concept extraction method according to claim 1 is characterized in that the specific step of obtaining the concept extraction result of the text to be extracted according to the reordering result comprises: According to the re-ranking result and a preset scoring threshold, or according to the re-ranking result and a preset number, a plurality of the candidate concepts are selected as the concept extraction result of the text to be extracted. 3. The concept extraction method according to claim 1 is characterized in that the specific steps of extracting terms from the text to be extracted according to the preset vocabulary and obtaining the first candidate concept list include: Filtering the text to be extracted according to the preset vocabulary to obtain the intersection of the preset vocabulary and the text to be extracted; The intersection is segmented and POS-tagged, and nouns in the intersection are obtained as candidate concepts to form the first candidate concept list. 4. The concept extraction method according to claim 1 is characterized in that the specific step of obtaining the re-ranking result according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, and the second candidate concept list comprises: For each candidate concept in the intersection of the first candidate concept list and the second candidate concept list, obtaining a score of each candidate concept in the intersection according to the confidence of each candidate concept in the intersection and the frequency of occurrence in the text to be extracted, and determining the score of each candidate concept that does not belong to the intersection as zero; The re-ranking result is obtained according to the score of each candidate concept in the first candidate concept list and the second candidate concept list. 5. A concept extraction device, characterized by comprising: An extraction module is used to extract terms from the text to be extracted according to a preset vocabulary to obtain a first candidate concept list, and to perform entity linking on the text to be extracted according to a preset knowledge graph to obtain a second candidate concept list; A screening module, used for reordering the candidate concepts in the first candidate concept list and the second candidate concept list, and obtaining a concept extraction result of the text to be extracted according to the reordering result; Wherein, the text to be extracted is unstructured text; The specific steps of reordering the candidate concepts in the first candidate concept list and the second candidate concept list include: reordering the candidate concepts in the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted; and/or, clustering the word vectors of the candidate concepts in the first candidate concept list and the second candidate concept list, and reordering the word vectors according to the clustering results and the word vectors of the preset seed concepts; The specific step of reordering the candidate concepts in the first candidate concept list and the second candidate concept list according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted includes: According to the TF-IDF method, obtaining the confidence of each candidate concept in the first candidate concept list; Obtaining a re-ranking result according to the confidence of each candidate concept in the first candidate concept list and the frequency of occurrence in the text to be extracted, and the second candidate concept list; The specific steps of clustering the word vectors of the candidate concepts in the first candidate concept list and the second candidate concept list and reordering the word vectors according to the clustering results and the word vectors of the preset seed concepts include: Obtaining the word vector of each candidate concept in the first candidate concept list and the second candidate concept list, and performing clustering according to the word vector of each candidate concept in the first candidate concept list and the second candidate concept list; According to the similarity between each cluster center obtained by clustering and each pre-set sub-concept cluster center, the clusters to which each cluster center belongs are sorted; For each of the clusters, the candidate concepts belonging to the cluster are ranked according to the similarity between the seed concept cluster center having the greatest similarity to the cluster center and the word vector of each candidate concept belonging to the cluster. 6. An electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the concept extraction method as claimed in any one of claims 1 to 4 are implemented. 7. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the concept extraction method as described in any one of claims 1 to 4 are implemented.