CN107291687A - It is a kind of based on interdependent semantic Chinese unsupervised open entity relation extraction method - Google Patents

Abstract

The invention relates to a Chinese unsupervised open entity relationship extraction method based on dependency semantics. The method includes the following steps: preprocessing input text: performing Chinese word segmentation, part-of-speech tagging and dependency syntax analysis on the input text; and naming entities on the input text Recognition; randomly select two entities from the identified entities to form a candidate entity pair; find the dependency path between the two entities in the candidate entity pair; analyze whether the syntactic structure mapped by the dependency path is consistent with the paradigm of the dependency semantic paradigm set Match, if yes, then extract words or phrases from the rest of the input text according to the matched paradigm as relational words, and the extracted relational words and candidate entity pairs form a relation triplet, otherwise proceed to the next set of candidate entity pairs Pattern matching; output relation triples. Compared with the prior art, the present invention has the advantages of low computational complexity, high extraction efficiency, overcoming the limitation of distance and location, and being able to extract even a single sentence.

Description

A Chinese Unsupervised Open Entity Relationship Extraction Method Based on Dependency Semantics

technical field

The invention relates to information extraction research in the fields of artificial intelligence and natural language processing, in particular to a Chinese unsupervised open entity relationship extraction method based on dependency semantics.

Background technique

The wave of big data seems to be surging like the Qiantang River, and the data accumulated on the Internet is growing explosively. Faced with the massive amount of information on the web, it becomes very difficult for users to quickly find the information they care about. Traditional search engines can only return a large number of webpages related to the user's query content to the user, and the information that the user needs can only be obtained after browsing the webpage. This single search method of returning webpages can no longer meet the actual needs of users facing massive network data. The Internet provides people with an inexhaustible source of information, how to quickly and accurately mine valuable information automatically has become the focus of attention.

Information extraction technology came into being. The task of information extraction is to output the unstructured information contained in the text in a structured or semi-structured form, quickly obtain the content that users really care about, and provide intelligent and humanized information services. For example, from the news reports of the plane crash, information such as people, time, location, number of casualties, and cause of the accident are extracted, allowing users to quickly obtain the cause of the incident. Named entity relationship extraction is a core subtask of information extraction, also known as entity relationship extraction or relationship extraction, which extracts the semantic relationship between related named entities from unstructured natural language texts and organizes them into structured relational triples. Group (Entity1, RelationWords, Entity2), where Entity1 and Entity2 are entity pairs with relationships, and RelationWords are words or word sequences describing the semantic relationship between entities.

Entity relationship extraction has important research value and is widely used in many artificial intelligence fields such as knowledge graphs, intelligent search engines, automatic question answering systems, text mining, and machine translation.

Traditional information extraction uses trained extractors to identify target relation types, which requires pre-defined relation types and a large amount of labeled training corpus. Traditional Chinese relation extraction is based on supervised machine learning algorithms, mainly including feature-based methods and kernel-based methods. This type of method has several shortcomings: first, it is very difficult to define a comprehensive entity relationship type system; second, it relies heavily on large-scale labeled training corpus, manual labeling of corpus is time-consuming and laborious, and the quality of labeling is difficult to control. Finally, open network texts are massive and cannot be pre-defined, so traditional methods cannot meet the needs of open domain information extraction. Open entity relationship extraction technology overcomes the disadvantages of traditional relationship extraction, and can automatically discover any relationship type in network texts, which has important development prospects and research value. In the research of open relation extraction, clustering algorithm is mainly applied. By means of position restriction and distance restriction, the candidate entity pairs are extracted, and then clustered to generate clusters of similar entity pairs, and then the relationship class labels are marked for each type of cluster, and more representative words are selected as the relationship descriptors of the class. There are two problems with this method: the clustering algorithm requires a considerable number of related entity pairs, that is, effective results cannot be obtained for a single or a small number of entity pairs, and the effect of such methods will be seriously affected when the training corpus is insufficient; it is difficult to determine Whether the final core relational word can become an effective relational characteristic word, and the descriptor of the finally determined class family may not be suitable for every pair of entities in the cluster. In addition, some scholars have studied methods based on deep syntactic analysis or semantic role labeling, and achieved good results. Research in this area mainly focuses on English corpus.

The research on open relational extraction on English corpus has achieved remarkable results, but there are relatively few studies on Chinese corpus. Chinese corpus has its unique flexibility and complexity in terms of word formation, sentence construction and expression, and its research difficulty is much greater than that of English. Therefore, some existing English entity relationship extraction systems cannot adapt to Chinese corpus. It is necessary to carefully study Chinese lexical and syntax and introduce them into entity relationship extraction in order to obtain an entity relationship extraction system suitable for the Chinese field.

The study found that when extracting entity relationships, there is often a certain syntactic relationship between entity pairs that have relationships. For example, if two entities are the subject and object of the sentence respectively, then the relation feature word of the entity pair is most likely to be the predicate verb. If the syntactic relationship between entity pairs is known in advance, then the characteristic words of the relationship between entity pairs can be determined more accurately. Dependency syntactic analysis can reflect the semantic modification relationship between the components of a sentence. Since the named entity in a sentence must appear in the dependency structure as a noun phrase, the dependency path between entities must also reflect the relationship characteristics of the corresponding entity pair.

To sum up, in order to make the entity relationship extraction method more suitable for Chinese corpus, based on the unique syntax and semantic features of Chinese, it fully demonstrates the adaptability and effectiveness of unsupervised methods in the open field. The present invention proposes an unsupervised Chinese open relation extraction method - Dependency Semantic Normal Forms (DSNFs). Bring innovative results to the research field of Chinese open relational extraction.

Contents of the invention

The purpose of the present invention is to provide a Chinese unsupervised open entity relationship extraction method based on dependency semantics in order to overcome the above-mentioned defects in the prior art. The purpose of the present invention is to avoid the disadvantages of traditional extraction methods such as high requirements for training corpus, poor portability and scalability, and inability to adapt to open network texts, and the extraction method under the English corpus due to the complex and flexible characteristics of Chinese in terms of morphology and grammar. It cannot be transplanted to Chinese, and the present invention proposes an open unsupervised entity relationship extraction method for network texts based on the characteristics of Chinese language.

In order to solve the above technical problems, the present invention is based on the mapping between the entity relationship and the dependency analysis tree, deeply digs the dependency semantics contained in the shortest dependency path, and uses the characteristics of dependency relationship, part-of-speech information and position relationship as limitations to obtain the dependency semantics Paradigm, proposed and implemented a novel unsupervised Chinese open relation extraction method.

The purpose of the present invention can be achieved through the following technical solutions:

A Chinese unsupervised open entity relationship extraction method based on dependency semantics, the method includes the following steps:

S1. Preprocessing the input text: perform Chinese word segmentation, part-of-speech tagging and dependency syntax analysis on the input text;

S2. Perform named entity recognition on the input text;

S3. Randomly select two entities from the identified entities to form a candidate entity pair;

S4. Find a dependency path between two entities in the candidate entity pair;

S5. Analyze whether the syntactic structure mapped by the dependency path between the two entities in the candidate entity pair matches the paradigm of the dependent semantic paradigm set, and if so, extract words or words from the rest of the input text according to the matched paradigm. Phrases are used as relational words, and the extracted relational words and candidate entity pairs constitute a relational triplet, otherwise, the paradigm matching of the next group of candidate entity pairs is performed;

S6. Outputting a relation triplet.

The form of the relation triplet is: (Entity1, RelationWords, Entity2), where Entity1 and Entity2 are entity pairs with relation, and RelationWords are words or phrases describing the semantic relation between entities.

The dependency semantic paradigm includes the first type of pre-modified structure, the second type of parallel structure, the third type of verb-related type, the fourth type of templated type and other types.

The first type of pre-modification structure class includes a combined attributive structure and a structure connected by the structural particle "of" and the head, and the combined attributive structure corresponds to the dependent semantic paradigm "Entity1+AttWord1(+AttWord2)+Entity2", which is formed by The structure of the connection between the structural particle "的" and the head word corresponds to the semantic paradigm "Entity1+的+Noun+Entity2" or "Entity1+的+Entity2+Noun", in which Entity1 and Entity2 are entity pairs with a relationship, and AttWord1 and AttWord2 are different definitions. Words, Noun is a noun.

The second type of coordinating structures includes coordinating noun structures and coordinating verb structures.

The described parallel noun structure includes a parallel entity as a subject structure, a parallel entity as a predicate object structure, a parallel entity as a prepositional object structure and a mixed structure of the first three, and a parallel entity as a subject structure corresponding to the dependent semantic paradigm "Entity2+Conj+(Entity1++) +Pred+Entity3", the parallel entity as a predicate object structure corresponds to the dependent semantic paradigm "Entity2+Pred+Entity3+Conj+(Entity1++)", the parallel entity as a preposition object structure corresponds to the dependent semantic paradigm "Entity2+Prep+Entity3+Conj+(Entity1++) +Pred (+Dobj)", where Entity2 and Entity3 are entity pairs that have a relationship, (Entity1++) indicates that there are one or more parallel entities, Conj is a conjunction, Pred is a predicate, Prep is a preposition, and Dobj is a direct object.

The said coordinating verb structure includes a verb conjunction structure and a coordinating compound sentence structure.

The third category of verb-related classes includes a subject-verb-verb-object structure and a subject-predicate-interface-object structure, the subject-predicate-verb-object structure corresponds to the dependent semantic paradigm "Entity1+Pred+Entity2", and the subject-verb-verb-object structure corresponds to the dependent semantic paradigm "Entity1 +Prep+Entity2+Pred(+Dobj)", wherein, Entity1 and Entity2 are entity pairs with relationship, Pred is a predicate, Prep is a preposition, and Dobj is a direct object.

Compared with the prior art, the present invention has the following advantages:

1) The method proposed by the present invention has sufficient ability to deal with complex Chinese syntax. In the extraction process, there is no need to limit the relative position of entity pairs and relative words, and avoid the disadvantages caused by position restrictions in traditional methods;

2) The method proposed by the present invention can obtain richer results, and can extract relational phrases with verbs or nouns as the core. In contrast, other extractors with better effects can only extract verbs as relational words;

3) The method proposed by the present invention can better identify long-span dependencies, especially in the case of juxtaposed structures, can extract co-occurring relational triples, avoiding the disadvantages caused by distance restrictions in traditional methods;

4) The method proposed by the present invention does not require model training corpus, and a sentence can also be used for relationship extraction, with low computational complexity and high extraction efficiency, and can meet high real-time requirements.

Description of drawings

Fig. 1 is a schematic flow chart of the extraction method of the present invention;

Figure 2 is a graph model of the dependency semantic paradigm DSNF1;

Figure 3 is a graph model of the dependency semantic paradigm DSNF2;

Figure 4 is a graph model of the dependency semantic paradigm DSNF3;

Figure 5 is a graph model of the dependency semantic paradigm DSNF4;

Figure 6 is a graph model of the dependency semantic paradigm DSNF5;

Figure 7 is a graph model of the dependency semantic paradigm DSNF6;

Figure 8 is a graph model of the dependency semantic paradigm DSNF7;

Figure 9 is a graph model of the dependency semantic paradigm DSNF8;

Figure 10 is the graph model of the dependency semantic paradigm DSNF9.

detailed description

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to 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 them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

Example

A Chinese unsupervised open entity relationship extraction method based on dependency semantics proposed by the present invention is an entity relationship extraction method based on dependency semantic paradigm (DSNFs), which can realize automatic extraction without manual intervention, and the input is without any processing Natural language sentences, the output is entity-relationship triples. As shown in Figure 1, the whole process can be described as follows:

Step 1: Preprocess the input text. Each sentence will undergo a series of natural language processing operations such as word segmentation, part-of-speech tagging, and dependency syntax analysis to prepare for the next steps. The method proposed by the present invention uses the natural language processing technology provided by the "Language Technology Platform (LTP)" developed by the Social Computing and Information Retrieval Research Center of Harbin Institute of Technology to perform the above operations.

Step 2: Select candidate entity pairs. The entity recognition of the input text is carried out through the named entity recognition module, and then all the recognized candidate entities are paired in pairs. This method uses the named entity recognition technology provided by the language technology platform of Harbin Institute of Technology and the iterative heuristic method for named entity recognition. The latter is to obtain the maximum noun phrase by merging connected nouns, in which the part of speech of nouns can only be {ni, nh, ns, nz, j}, which respectively represent organization names, personal names, geographical names, other proper nouns and acronyms . The two methods complement each other and are carried out simultaneously.

Step 3: Match the dependency semantic paradigm. For the candidate entity pairs obtained in step 2, analyze whether the syntax structure mapped by the shortest path between entities can match a certain DSNF,

Step 4: Output relation triplets. After the execution of step 3, if it matches, then extract the relation word from it and output the relation triplet; if not match, then match the next set of candidate entity pairs.

The core of the method proposed by the present invention lies in the dependency semantic paradigm, and the relevant content will be introduced emphatically below:

Through the statistical analysis of a large number of relational examples, it is found that relational triples always appear in certain fixed syntactic structures, and the syntactic structures that represent entity relations include: subject-predicate relations, verb-object relations, prepositional objects, and parallel components and grooming relationships, etc. Mapping these structures into dependency trees yields Dependency Semantic Normal Forms (DSNFs). DSNFs are composed of word sequences, parts of speech, dependency paths and their associated dependency labels. This method divides the paradigm set into five categories: premodification, juxtaposition, verb correlation, templated, and others. In each category, one or more DSNFs can be obtained to provide a reasonable basis for relation extraction.

1. Pre-Modification Class (PreMod)

Pre-modification is a very important type of modification in Chinese phrases. From the perspective of Chinese linguistics, the relational expression of the PreMod syntactic class is a positive structure, which can form a positive phrase. However, the structure of partial phrases is composed of the attributive head and the modifier pair, and the attributive is the front additional component in the noun partial phrases. The composition of attributives has a wide range. Except for adverbs and "de" phrases, other content words (nouns, verbs and adjectives) and phrases can act as attributives. In addition, the complexity of an attributive also lies in its multi-layered nature, which can be defined and described from different sides and superimposed before a central term at the same time, so that a central term can have multiple attributives.

From the perspective of formal structure, attributives can be divided into the following two types:

1) Combined attributive, an attributive that is added directly before the head subject without "的" in the middle, that is, "attributive + head word". For example, in "<ORG>Grade 2 Class 3</ORG>Head Teacher <PER>Wang</PER>", "Grade 2 Class 3" is the attributive of "Senior Class 3", and "Grade 2 Class 3" is the attributive of "Wang". The attributive, "head teacher" also expresses the semantic relationship between the entity "Grade 2 Class 3" and "Wang", thus forming a relational triple (Grade 2 Class 3, head teacher, Wang). Due to the multi-level nature of attributives, multiple word combinations may be used as relationship feature words. For example, "<ORG>a company</ORG>CEO <PER>Zhao</PER>" can extract the relationship (a company , Chief Executive Officer, Zhao), wherein the combination of "Chief" and "Executive Officer" is a relational feature word. PER means the name of a person, and ORG means the name of an organization.

The combined attributive structure mapping is shown in the dependency analysis: the attributive is dependent on the central term, and the dependence relationship is a "centered relationship". The attributive word or directly depends on the head word, and the dependency relationship is also a "fixed-middle relationship". According to statistical research, in the actual relationship extraction, we mainly consider the structure of two-layer attributives and three-layer attributives, and obtained the relationship extraction paradigm DSNF1: "Entity1+AttWord1(+AttWord2)+Entity2", the dependency analysis is shown in Figure 2 Show. In addition, the limitation of part of speech should also be considered, only considering the situation that attributive words (AttWord1, AttWord2) are nouns, if "AttWord1" is an occupation-related noun (mainly including nouns related to institutions and work, such as chairman, general manager, county magistrate etc.); or "AttWord1" is a common noun (relative to occupation-related nouns) and "Entity2" is a character entity, and relationship extraction will only be performed when these two restrictions are met.

2) The attributive connected by the structural particle "的" and the head, that is, "attributive+de+head". For example, "<PER1>Zhang</PER1>'s wife <PER2>Sun</PER2>" can extract relational tuples (Zhang, wife, Sun). For another example, "teacher <PER>Pei</PER> from <ORG>university</ORG>" and "teacher <PER>Pei</PER> from <ORG>university</ORG>" , although the structure is different, but express the same meaning. Therefore, it can be expressed as two relational extraction paradigms DSNF2 and DSNF3: "+Noun+Entity2 of Entity1+" or "+Entity2+Noun of Entity1+". From these two structures a relational triple (Entity1, Noun, Entity1) can be extracted. It can be mapped to the form of dependency syntax analysis, as shown in Figure 3 and Figure 4.

It is also possible to encounter such a situation in relation extraction. The partial positive phrase contains only one entity noun, such as "Teacher Liu Moumou visited Shanghai", "Xiaoming's wife is Xiaohong", etc. This partial positive phrase is often contained in Among other relational syntax classes. At this time, the entity is used as an attributive modifier to modify the head term. In the dependency syntax analysis, the entity will no longer be directly used as the subject or object, but the modified head term becomes the main component of the syntactic structure. This situation is fully considered in the process of relationship extraction, and the head language is used as a "pseudo-entity (Pe)" to make corresponding conversions during dependency analysis. For example, "<PER>Liu Moumou</PER><Pe-PER>teacher</Pe-PER> visit <LOC>Shanghai</LOC>", extract the relationship between the pseudo-entity "teacher" and the entity "Shanghai" "Travel", then convert and output relational triples (Liu XX, Tour, Shanghai). This situation will not be repeated in the following analysis. Pe-PER represents pseudo-entities of personal names.

2. Verb related (Verbal Class, VERB)

In this class, one of the two related entities is often in the position of the subject, while the other is in the position of the object, which can be the object of the verb (verb-object structure), or the object of the preposition (preposition, Prep) (preposition, prep). structure), and the relationship between entities can be directly expressed by a predicate (predicate, Pred). According to the different objects, it can be further divided into "subject-verb-verb-object" structure and "subject-verb-interobject" structure.

1) For the "subject-verb-verb object" structure, for example, "<PER>Liu XX</PER>Visits <LOC>Shanghai</LOC>", in this example sentence "Liu XX" is the subject, "Shanghai" is the object, and "Tour" is the predicate verb associated with the two entities, and triples (Liu XX, Tour, Shanghai) can be extracted. Mapping the "subject-verb-verb-object" structure to the dependency analysis diagram, both entities depend on the core verb, and the dependency relationship is "subject-verb relationship" and "verb-object relationship". The available relation extraction paradigm DSNF4: "Entity1+Pred+Entity2", can extract relation triples (Entity1, Pred, Entity2). Dependency analysis is shown in Figure 5. LOC means geographical noun,

2) For the "subject-predicate-intermediate" structure, such as "<PER>Liu XX</PER> conducted an in-depth tour of <LOC>Shanghai</LOC>", the subject is the entity "Liu XX", and the verb "go on " is the predicate of the sentence, the subject entity depends on the predicate verb, and the dependency relationship is the "subject-predicate relationship". "Dui Shanghai" constitutes a prepositional phrase, and the entity "Shanghai" depends on the preposition "Dui", and the dependency relationship is "preposition-object relationship". The noun phrase "deep tour" is the direct object of the predicate, from which the relational tuple (Liu Moumou, conducting deep tour, Shanghai) can be extracted. It is worth noting that since entity 2 is in the position of the predicate phrase, it has an indirect dependency relationship with the predicate verb through the preposition. Therefore, in order to make the relationship extraction result have clearer semantics, this paper combines the predicate phrase and the direct object of the predicate (direct object , Dobj) together as relational feature words. The "subject-predicate-subject" structure can be mapped to the relational extraction paradigm DSNF5: "Entity1+Prep+Entity2+Pred(+Dobj)", which can extract relational triples (Entity1, Pred-Dobj, Entity2) dependency analysis as shown in Figure 6 Show.

In particular, for the "subject-predicate-predative" structure, if the preposition is "by, be", etc. to represent passive words, at this time, the positions of Entity1 and Entity2 are exchanged to form a relational triple (Entity2, Pred-Dobj, Entity1).

3. Coordination Class (COOR)

Parallel relations are also quite common in Chinese sentences. Parallel means that sentences or phrases are related to each other, either at the same time, or at the same time. The parallel relationship can be different things that are related to each other, or different aspects of the same thing, or different actions of the same subject. Coordinated phrases are also called parallel phrases, which are generally composed of two or more nouns, verbs, adjectives, pronouns or quantifiers, etc., and the parts of speech of the constituent words generally require the same. There is a parallel relationship between words, and commas or conjunctions (conjunction, Conj) such as "and, and, and, and, and" are often used in the middle. In relation extraction, two kinds of coordinating nouns and coordinating verbs are mainly considered.

For example, in "<PER1>Liu XX</PER1> and <PER2>Peng XX</PER2>visit <ORG>Shanghai</ORG>", "Liu XX" and "Peng XX" are two Nouns that have parallel relationships. When two entities have this noun phrase juxtaposition, they produce the same behavior and act on another common entity. In the example, the relationship triplet (Liu XX, tour, Shanghai) can be extracted. At the same time, the parallel component "Peng XX" of "Liu XX" also has a "tour" relationship with "Shanghai", and the relationship element can be extracted Group (Peng Moumou, tour, Shanghai). In fact, the COOR syntactic class needs to exist depending on other syntactic classes. For example, in the above example, the relational tuple (Liu XX, Tour, Shanghai) should belong to the VERB syntactic class. Because the entity "Peng XX" is dependent on the entity "Liu XX", the dependency relationship is a "parallel relationship", so the relationship that occurs on the entity "Liu XX" also applies to the entity "Peng XX". According to the positions of entities in the syntax, there are mainly three types: subject position, predicate-object position and preposition-object position, thus,

1) When the parallel noun is used as the subject, the relation extraction paradigm DSNF6 is extracted: "Entity2+Conj+(Entity1++)+Pred+Entity3", (where (Entity1++) means that there are one or more parallel entities, the same below). From the relationship triplet (Entity2, Pred, Entity3), the triplet (Entity1, Pred, Entity3) can be obtained, and the dependency relationship is shown in FIG. 7 .

2) When the parallel noun is used as the predicate object, the relation extraction paradigm DSNF7 is extracted: "Entity2+Pred+Entity3+Conj+(Entity1++)", from the relational triplet (Entity2, Pred, Entity3), the triplet (Entity2, Pred , Entity1), the dependency relationship is shown in Figure 8.

3) When a co-ordinated noun is used as a prepositional object, the relational extraction paradigm DSNF8 is extracted: "Entity2+Prep+Entity3+Conj+(Entity1++)+Pred(+Dobj)", which can be obtained from the relational triplet (Entity2, Pred-Dobj, Entity3) The triplet (Entity2, Pred-Dobj, Entity1) is obtained, and the dependency relationship is shown in Figure 9.

4) A mixture of the first three types. For example, "<PER1>Li Moumou</PER1> and <PER2>Zhang Moumou</PER2> together invited <PER3 in <ORG1>Shanghai</ORG1> and <ORG2>Hangzhou</ORG2> >Zhang XX</PER3> and <PER4>Gao XX</PER4>." It is a mixture of the first three types.

Coordinating verbs mainly describe two different actions performed by the same subject at the same time. There are two types of situations,

1) The first type of situation is the joint use of verbs. When composing sentences in Chinese, when a verb cannot fully describe the meaning of an action, two verbs are often used in conjunction, the first verb complements the second verb, and the second verb is a transitive verb, so the distance is generally extracted The second verb that is closer to the object serves as a relational character. For example, "<PER>Zhang Moumou</PER>Snow Tour <LOC>Lushan</LOC>", in which "Snowwalking" and "Tour" form a parallel relationship, and the relationship can be extracted (Zhang Moumou, tour, Lushan Mountain) .

2) The second type of situation is the parallel compound sentence, which means that several clauses in the compound sentence have an equal and parallel relationship in semantics. If there is a parallel listing relationship between two or more events, but there is no causal connection, then a parallel complex sentence can be formed. Clauses are often separated by commas and conjunctions such as "and, also, and". For example, "<ORG1>a company</ORG1> manager <PER>Gao</PER> visited <ORG2>factory</ORG2>, and issued production guidance suggestions in <ORG3>a workshop</ORG3>." Comma The compound sentence is divided into two clauses, expressing two events respectively, and the subject is the same as the entity "Gao Mou", so the two clauses are juxtaposed. The predicates "visit" and "publish" in the parallel clause constitute a parallel, and the dependent relationship is "parallel relationship". When mapped to dependency syntax, it can be described as: if entity 2 is dependent on predicate verb 2 as an object, and this verb 2 forms a parallel relationship with another verb 1 (the dependency relationship is "parallel relationship"), and entity 1 exists as a subject dependent on the verb 1, then it can be inferred that there is a relationship between entity 1 and entity 2, and the characteristic word of the relationship is verb 2. Therefore, the relationship extraction paradigm DSNF9 can be obtained: "Entity1+Pred1+Pred2+Entity2", and the dependency analysis is shown in Figure 10. Normal form DSNF9 can cover the above two types of situations.

It is worth noting that the parallel structure is nested in other syntactic classes. Normal forms DSNF6, DSNF7, DSNF8 and DSNF9 only express the performance of coordinating nouns depending on the "subject-verb-verb-object" structure in the VERB syntactic class. Other conditions will not be repeated. The actual extraction steps are similar. When Entity1 and Entity2 have a parallel relationship, if the triplet (Entity2, RelationWord, Entity3) is established, then the relational triplet (Entity1, RelationWord, Entity3) can be obtained; if the triplet (Entity3, RelationWord, Entity2) is established, then the relation triplet (Entity3, RelationWord, Entity1) can be obtained.

4. Modeled (Formulaic Class, FORM)

The types of FORM often appear in Chinese and cannot be classified into the previous relational syntax categories, but generally have a fixed expression format. For example, "Wang Mou, a professor of a certain university, published...", there is no corresponding conjunction between "Wang Mou" and "a professor of a certain university", and there is no direct modification relationship, so they do not meet the above types. But entity-relationship triples (Wang, professor, certain university) can be extracted from this sentence. Similar expressions in writing are very common, and it is a Chinese writing habit. For these special grammatical expression structures, you only need to extract templates for hard matching to achieve good results.

5. Other Class

This method generalizes all other relation types that cannot be distinguished so far into this category. Due to the uncertainty of this category, this article does not make an in-depth study of this category.

The invention discloses a Chinese unsupervised open entity relationship extraction method based on dependency semantics, which avoids the disadvantages of traditional methods such as manual labeling of large dependencies and unreasonable results, based on the unique and flexible syntax features of Chinese, and uses entity relationship and dependency Based on the mapping between analysis trees, the dependency semantics contained in the shortest dependency path are deeply excavated, and the dependency semantics normal forms (DSNFs) are obtained by using the characteristics of dependency relationship, part-of-speech information and position relationship as limitations. Using this paradigm set, we can learn from massive Fast and accurate automatic extraction of entity relationships from data. Fully automatic extraction can be realized without any manual work, no need to rely on model training corpus, low computational complexity, high extraction efficiency, and can meet high real-time requirements. The present invention can be widely used in artificial intelligence fields such as knowledge maps, intelligent search engines, automatic question answering systems, text mining, and machine translation.

The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can easily think of various equivalents within the technical scope disclosed in the present invention. Modifications or replacements shall all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (8)

1. it is a kind of based on interdependent semantic Chinese unsupervised open entity relation extraction method, it is characterised in that this method bag Include following steps：

S1, pretreatment input text：Chinese word segmentation, part-of-speech tagging and interdependent syntactic analysis are carried out to input text；

S2, to input text be named Entity recognition；

S3, any from the entity identified select two entities and constitute candidate's entities pair；

Interdependent path between S4, two entities of searching candidate's entity centering；

S5, analyze the syntactic structure that is mapped of interdependent path between two entities of candidate's entity centering whether with interdependent semanteme The normal form matching of normal form collection, if so, then extracting word or phrase from the remainder of input text according to the normal form being matched As relative, the relative and candidate's entity of extraction are to constituent relation triple, if otherwise carrying out next group of candidate's entity pair Normal form matching；

S6, output relation triple.

2. it is according to claim 1 a kind of based on interdependent semantic Chinese unsupervised open entity relation extraction method, Characterized in that, described relation triple form is：(Entity1, RelationWords, Entity2), wherein Entity1, Entity2 are the entities pair that there is relation, and RelationWords is the word or short of semantic relation between description entity Language.

3. it is according to claim 1 a kind of based on interdependent semantic Chinese unsupervised open entity relation extraction method, Moved characterized in that, described interdependent semantic normal form includes modification structure class, Equations of The Second Kind parallel construction class, the 3rd class before the first kind Word associated class, the 4th class template class and other classes.

4. it is according to claim 3 a kind of based on interdependent semantic Chinese unsupervised open entity relation extraction method, Characterized in that, before the described first kind modification structure class include combined type attribute structure and by structural auxiliary word " " and head The structure of connection, the interdependent semantic normal form of combined type attribute structure correspondence " Entity1+AttWord1 (+AttWord2)+ Entity2 ", by structural auxiliary word " " semanteme normal form " Entity1++Noun+ corresponding with the structure that head is connected Entity2 " or " Entity1++Entity2+Noun ", wherein Entity1, Entity2 are the entities pair that there is relation, AttWord1 and AttWord2 is different attribute word, and Noun is noun.

5. it is according to claim 3 a kind of based on interdependent semantic Chinese unsupervised open entity relation extraction method, Characterized in that, described Equations of The Second Kind parallel construction class includes coordinate noun structure and verb structure arranged side by side.

6. it is according to claim 5 a kind of based on interdependent semantic Chinese unsupervised open entity relation extraction method, Characterized in that, described coordinate noun structure includes entity arranged side by side as subject structure, entity arranged side by side is used as predicate object knot Structure, entity arranged side by side is as object of preposition structure and the mixed structure of first three, and entity arranged side by side is interdependent as subject structure correspondence Semantic normal form " Entity2+Conj+ (Entity1++)+Pred+Entity3 ", entity arranged side by side is used as predicate object structure correspondence Interdependent semantic normal form " Entity2+Pred+Entity3+Conj+ (Entity1++) ", entity arranged side by side is used as object of preposition structure The interdependent semantic normal form " Entity2+Prep+Entity3+Conj+ (Entity1++)+Pred (+Dobj) " of correspondence, wherein Entity2, Entity3 are the entity pair that there is relation, and (Entity1++) represents there are one or more entities arranged side by side, Conj For conjunction, Pred is predicate, and Prep is preposition, and Dobj is direct object.

7. it is according to claim 5 a kind of based on interdependent semantic Chinese unsupervised open entity relation extraction method, Characterized in that, structure and class complex sentence structure arranged side by side is used in conjunction including verb in described verb structure arranged side by side.

8. it is according to claim 3 a kind of based on interdependent semantic Chinese unsupervised open entity relation extraction method, Characterized in that, the 3rd described class verb associated class includes subject-predicate V-O construction and subject-predicate guest's Jie structure, subject-predicate V-O construction The interdependent semantic normal form " Entity1+Pred+Entity2 " of correspondence, the interdependent semantic normal form " Entity1+ of subject-predicate guest Jie structure correspondence Prep+Entity2+Pred (+Dobj) ", wherein, Entity1, Entity2 are the entities pair that there is relation, and Pred is predicate, Prep is preposition, and Dobj is direct object.