CN116167382A

CN116167382A - Intention event extraction method and device, electronic equipment and storage medium

Info

Publication number: CN116167382A
Application number: CN202310014600.2A
Authority: CN
Inventors: 刘珮; 钱兵; 赵龙刚
Original assignee: China Telecom Corp Ltd
Current assignee: China Telecom Corp Ltd
Priority date: 2023-01-05
Filing date: 2023-01-05
Publication date: 2023-05-26

Abstract

The disclosure provides an intention event extraction method and device, electronic equipment and storage medium, and relates to the technical field of natural language processing. The intention event extraction method comprises the following steps: acquiring current statement data to be processed; extracting at least two semantic feature data of the current sentence data; correlating the at least two semantic feature data based on a preset dependency relationship to obtain a multi-form heterogeneous relationship graph; inputting the multi-form heterogeneous relation diagram into a pre-trained intention event extraction model to obtain a statement intention event corresponding to the current statement data. According to the technical scheme, through the constructed multi-form heterogeneous relation diagram, semantic information contained in sentence data is effectively mined, correlation between intra-sentence events and inter-sentence events is accurately expressed, and therefore accuracy of extracted sentence intention events is effectively improved.

Description

Intention event extraction method and device, electronic equipment and storage medium

Technical Field

The present disclosure relates to the field of natural language processing, and in particular, to an intention event extraction method, an intention event extraction device, an electronic apparatus, and a computer-readable storage medium.

Background

Natural language processing (Natural Language Processing, NLP) is an important direction in the fields of computer science and artificial intelligence, and has been mainly studied on various theories and methods for realizing effective communication between a person and a computer in natural language. At present, natural language processing is mainly applied to the aspects of machine translation, automatic summarization, text classification, question answering, text semantic comparison, voice recognition, optical character recognition (Optical Character Recognition, OCR), sentence intention event extraction and the like.

In the related scheme of sentence event extraction, ambiguity triggered by some types of events in sentences cannot be distinguished, or correlation between intra-sentence events and inter-sentence events cannot be understood, so that accuracy of sentence intention events extracted from sentence data is poor.

It should be noted that the information disclosed in the above background section is only for enhancing understanding of the background of the present disclosure and thus may include information that does not constitute prior art known to those of ordinary skill in the art.

Disclosure of Invention

An object of an embodiment of the present disclosure is to provide an intention event extraction method, an intention event extraction apparatus, an electronic device, and a computer-readable storage medium, thereby improving accuracy of a sentence intention event extracted from sentence data at least to some extent.

Other features and advantages of the present disclosure will be apparent from the following detailed description, or may be learned in part by the practice of the disclosure.

According to a first aspect of an embodiment of the present disclosure, there is provided an intention event extraction method, including:

acquiring current statement data to be processed;

extracting at least two semantic feature data of the current sentence data;

correlating the at least two semantic feature data based on a preset dependency relationship to obtain a multi-form heterogeneous relationship graph;

inputting the multi-form heterogeneous relation diagram into a pre-trained intention event extraction model to obtain a statement intention event corresponding to the current statement data.

In some example embodiments of the present disclosure, based on the foregoing, the at least two semantic feature data includes word feature data, entity feature data, context feature data, and pinyin feature data;

the extracting at least two semantic feature data of the current sentence data comprises:

word segmentation processing is carried out on the current sentence data to obtain sentence words, and word coding is carried out on the sentence words to obtain word characteristic data;

performing entity identification processing on the current sentence data to obtain sentence entities, and performing entity coding on the sentence entities to obtain entity characteristic data;

Acquiring adjacent sentence data corresponding to the current sentence data, and performing context coding on the adjacent sentence data to obtain context characteristic data;

and acquiring sentence pinyin corresponding to the current sentence data, and performing pinyin coding on the sentence pinyin to obtain the pinyin characteristic data.

In some example embodiments of the present disclosure, based on the foregoing solution, the word encoding the sentence word to obtain word feature data includes:

performing word embedding coding on the sentence words to obtain word embedding vectors;

performing part-of-speech tagging coding on the sentence words to obtain part-of-speech tag embedded vectors;

performing position coding on the sentence words to obtain position embedded vectors;

performing entity type labeling coding on the sentence words to obtain entity type label embedded vectors;

and splicing according to the word embedding vector, the part-of-speech tag embedding vector, the position embedding vector and the entity type tag embedding vector to generate word characteristic data.

In some example embodiments of the present disclosure, based on the foregoing solution, the performing context encoding on the adjacent sentence data to obtain context feature data includes:

Performing sentence coding on the adjacent sentence data to obtain adjacent sentence vectors;

and splicing the adjacent sentence vectors to obtain the context characteristic data corresponding to the current sentence data.

In some example embodiments of the present disclosure, based on the foregoing solutions, the preset dependency includes any one or more of a dependency between words, a dependency between words and trigger words, a dependency between words and entities, an association between words and entities, a dependency between contexts and entities, and a dependency between pinyin and trigger words.

In some example embodiments of the present disclosure, based on the foregoing solution, the inputting the multi-form heterogeneous relationship graph into a pre-trained intent event extraction model, to obtain a sentence intent event corresponding to the current sentence data includes:

inputting the multi-form heterogeneous relation diagram into a pre-trained intention event extraction model to obtain event trigger words and event content information;

and generating statement intention events corresponding to the current statement data based on the event trigger words and the event content information.

In some example embodiments of the present disclosure, based on the foregoing approach, the intent event extraction model includes a relationship graph convolution neural network R-GCN.

According to a second aspect of embodiments of the present disclosure, there is provided an intention event extraction apparatus including:

the sentence data acquisition module is used for acquiring current sentence data to be processed;

the semantic feature extraction module is used for extracting at least two semantic feature data of the current sentence data;

the semantic relation diagram construction module is used for associating the at least two semantic feature data based on a preset dependency relationship to obtain a multi-form heterogeneous relation diagram;

the statement intention event determining module is used for inputting the multi-form heterogeneous relation diagram into a pre-trained intention event extraction model to obtain a statement intention event corresponding to the current statement data.

According to a third aspect of embodiments of the present disclosure, there is provided an electronic device, comprising: a processor; and a memory having stored thereon computer readable instructions that when executed by the processor implement the method of intent event extraction of any of the above.

According to a fourth aspect of embodiments of the present disclosure, there is provided a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the intention event extraction method according to any one of the above.

The technical scheme provided by the embodiment of the disclosure can comprise the following beneficial effects:

according to the intention event extraction method in the example embodiment of the disclosure, at least two kinds of semantic feature data contained in current sentence data can be extracted, then the at least two kinds of semantic feature data are associated based on a preset dependency relationship to obtain a multi-form heterogeneous relationship graph, and the multi-form heterogeneous relationship graph can be input into a pre-trained intention event extraction model to obtain a sentence intention event corresponding to the current sentence data. Through the constructed multi-form heterogeneous relation diagram, semantic information contained in current sentence data can be effectively mined, semantic expression is enhanced, correlation between intra-sentence events and inter-sentence events is accurately represented, ambiguity between event trigger words is accurately distinguished, and therefore accuracy of extracted sentence intention events is effectively improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the disclosure and together with the description, serve to explain the principles of the disclosure. It will be apparent to those of ordinary skill in the art that the drawings in the following description are merely examples of the disclosure and that other drawings may be derived from them without undue effort. In the drawings:

FIG. 1 illustrates a schematic diagram of a system architecture of an exemplary application environment in which the intent event extraction method and apparatus of embodiments of the present disclosure may be applied;

FIG. 2 schematically illustrates a flow diagram of an intent event extraction method in accordance with some embodiments of the present disclosure;

FIG. 3 schematically illustrates a flow diagram for generating word feature data according to some embodiments of the present disclosure;

FIG. 4 schematically illustrates a flow diagram for implementing sentence intent event extraction in accordance with some embodiments of the present disclosure;

FIG. 5 schematically illustrates a schematic diagram of a multi-formal heterogeneous relationship graph, according to some embodiments of the present disclosure;

FIG. 6 schematically illustrates a schematic diagram of an intent event extraction device in accordance with some embodiments of the present disclosure;

FIG. 7 schematically illustrates a structural schematic diagram of a computer system of an electronic device, in accordance with some embodiments of the present disclosure;

fig. 8 schematically illustrates a schematic diagram of a computer-readable storage medium according to some embodiments of the present disclosure.

In the drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

Detailed Description

Example embodiments will now be described more fully with reference to the accompanying drawings. However, the exemplary embodiments may be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the example embodiments to those skilled in the art.

Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the disclosure. One skilled in the relevant art will recognize, however, that the disclosed aspects may be practiced without one or more of the specific details, or with other methods, components, devices, steps, etc. In other instances, well-known methods, devices, implementations, or operations are not shown or described in detail to avoid obscuring aspects of the disclosure.

Moreover, the drawings are only schematic illustrations and are not necessarily drawn to scale. The block diagrams depicted in the figures are merely functional entities and do not necessarily correspond to physically separate entities. That is, the functional entities may be implemented in software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor devices and/or microcontroller devices.

FIG. 1 illustrates a schematic diagram of a system architecture of an exemplary application environment in which the intent event extraction method and apparatus of embodiments of the present disclosure may be applied.

As shown in fig. 1, the system architecture 100 may include one or more of the

terminal devices

101, 102, 103, a network 104, and a server 105. The network 104 is used as a medium to provide communication links between the

terminal devices

101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, among others. The

terminal devices

101, 102, 103 may be a variety of electronic devices with artificial intelligence processing capabilities including, but not limited to, desktop computers, portable computers, smart phones, tablet computers, and the like. It should be understood that the number of terminal devices, networks and servers in fig. 1 is merely illustrative. There may be any number of terminal devices, networks, and servers, as desired for implementation. For example, the server 105 may be a server cluster formed by a plurality of servers.

The method for extracting an intention event provided by the embodiment of the present disclosure is generally performed in the server 105, and accordingly, the apparatus for extracting an intention event is generally disposed in the server 105. However, it is easily understood by those skilled in the art that the method for extracting an intention event provided in the embodiment of the present disclosure may be performed by the

terminal devices

101, 102, 103, and accordingly, the means for extracting an intention event may be provided in the

terminal devices

101, 102, 103, which is not particularly limited in the present exemplary embodiment.

For example, in an exemplary embodiment, the user may collect current sentence data through the

terminal devices

101, 102, 103, upload the current sentence data to the server 105, and after the server generates the sentence intention event through the intention event extraction method provided by the embodiment of the present disclosure, transmit the sentence intention event to the

terminal devices

101, 102, 103, and so on for processing.

In the related sentence intention event extraction scheme, only single-form semantic information of sentence data is considered, however, one word expresses completely different intention in different sentences, so that different events can be triggered, and for the situation, the ambiguity of triggering of certain types of events is often insufficient; or the relativity between the intra-sentence events and the inter-sentence events is difficult to model, because one sentence can express a plurality of related events at the same time, the modeling capability of the current scheme on the interdependence relationship between the related events is weaker, and the task is never extracted from the whole event, so that the accuracy of the finally extracted sentence intention is poor.

Based on the problems existing in the related art, in the present exemplary embodiment, there is provided first an intended event extraction method, and the intended event extraction method in the present embodiment will be specifically described below taking a server executing the method as an example.

Fig. 2 schematically illustrates a flow diagram of an intent event extraction method according to some embodiments of the present disclosure. Referring to fig. 2, the intention event extraction method may include the steps of:

step S210, acquiring current statement data to be processed;

step S220, extracting at least two semantic feature data of the current sentence data;

step S230, associating the at least two semantic feature data based on a preset dependency relationship to obtain a multi-form heterogeneous relationship graph;

step S240, inputting the multi-form heterogeneous relation diagram into a pre-trained intention event extraction model to obtain a statement intention event corresponding to the current statement data.

According to the method for extracting the intention event in the embodiment of the invention, semantic information contained in current sentence data can be effectively mined through the constructed multi-form heterogeneous relation diagram, semantic expression is enhanced, correlation between intra-sentence events and inter-sentence events is accurately represented, ambiguity between event trigger words is accurately distinguished, and thus the accuracy of the extracted sentence intention event is effectively improved.

Next, the intention event extraction method in the present exemplary embodiment will be further described.

In step S210, current sentence data to be processed is acquired.

In an example embodiment of the present disclosure, the current sentence data refers to sentence data selected for intent event extraction in a sentence paragraph, for example, for a sentence paragraph composed of 6 sentences, if intent event extraction is required for sentence 3 in a sentence paragraph at the current time, then sentence 3 may be the current sentence data.

The current sentence data to be processed may be sentence data input by the user in real time through a related interaction manner, or may be sentence data stored in the storage unit in advance, or may be sentence data obtained through recognition by a translation tool or an OCR tool, and the source of the current sentence data is not limited in this example embodiment.

In step S220, at least two kinds of semantic feature data of the current sentence data are extracted.

In an example embodiment of the present disclosure, the semantic feature data refers to data representing semantic information corresponding to current sentence data in any form, for example, the semantic feature data may be word features corresponding to the current sentence data, or may be entity features in the current sentence data, or may, of course, be context features, pinyin features, tone features, voice features, etc. corresponding to the current sentence data, and the type of the semantic feature data is not limited in any particular way in this example embodiment.

At least two kinds of semantic feature data contained in the current sentence data may be extracted, for example, the number of the semantic feature data may be 2, 3, 4, 5, etc., and the number of the semantic feature data is not particularly limited in this example embodiment, and in general, the more the semantic feature data may be extracted from the current sentence data, the better.

For example, word features, entity features, context features, pinyin features and the like of the current sentence data may be extracted, and it is assumed that the current sentence data has corresponding speech data, and tone features and intonation features (such as happy intonation, angry intonation and the like) corresponding to the speech data may also be used as semantic feature data of the current sentence data.

In step S230, the at least two semantic feature data are associated based on a preset dependency relationship, so as to obtain a multi-form heterogeneous relationship graph.

In an example embodiment of the present disclosure, the preset dependency relationship refers to preset data for mining the correlation between semantic feature data, for example, the preset dependency relationship may be any one or more of a dependency relationship between words, a dependency relationship between words and trigger words, a dependency relationship between words and entities, a dependency relationship between entities and entities, a dependency relationship between contexts and entities, and a dependency relationship between pinyin and trigger words, and of course, the preset dependency relationship may also be other types of data for mining the correlation between semantic feature data, such as a dependency relationship between contexts and word gases, a dependency relationship between words and trigger words, and the example embodiment is not limited thereto.

The multi-form heterogeneous relation graph is a relation graph obtained by taking a preset dependency relationship as an edge and semantic feature data as nodes and connecting the nodes in different forms through short paths by the edge, and the information of related events in the current sentence data can be aggregated through the multi-form heterogeneous relation graph, so that the interdependence relationship among the events is obtained, the extraction effect of the sentence intention events is effectively improved, and the accuracy of the extracted sentence intention events is improved.

In step S240, the multi-form heterogeneous relationship graph is input into a pre-trained intent event extraction model, so as to obtain a sentence intent event corresponding to the current sentence data.

In an example embodiment of the present disclosure, the intent event extraction model refers to a deep learning model for aggregating the relationship interactions between the node information and the inference perception node information in the multi-form heterogeneous relationship graph, for example, the intent event extraction model may be a deep learning model formed by a neural network R-GCN based on the relationship graph convolution, and of course, the intent event extraction model may also be other deep learning models capable of achieving the relationship interactions between the node information and the inference perception node information in the multi-form heterogeneous relationship graph, such as a recurrent neural network (Gate Recurrent Unit, GRU), which is not limited in this example embodiment.

Statement intent event refers to a trigger event extracted in the current statement data for characterizing statement intent, e.g. "at 2022, ming born in A city" for the current statement data, then the corresponding statement intent event may be "event trigger word: birth; event character: xiaoming; event location: a is true; event time: 2022 ", of course, is illustrative only and should not be taken as limiting in any way.

The multi-form heterogeneous relation diagram can be input into a pre-trained intention event extraction model in the form of feature vectors, sentence intention events corresponding to current sentence data can be generated, the generated sentence intention events can be used for chat reply scenes of intelligent customer service, scenes of intelligent robot text or voice form response questioners, intelligent text generation and the like, and the application scenes of the obtained sentence intention events are not limited in any way.

Through the constructed multi-form heterogeneous relation diagram, semantic information contained in current sentence data can be effectively mined, semantic expression is enhanced, correlation between intra-sentence events and inter-sentence events is accurately represented, ambiguity between event trigger words is accurately distinguished, and therefore accuracy of extracted sentence intention events is effectively improved; through the multi-form heterogeneous relation diagram and the pre-trained intention event extraction model, the extraction efficiency of statement intention events can be effectively improved, and the application range is improved.

Next, step S210 to step S240 will be described in detail.

In some example embodiments of the present disclosure, the at least two semantic feature data may include word feature data, entity feature data, context feature data, and pinyin feature data; optionally, if the current sentence data has corresponding voice data, tone feature data, mood feature data and the like corresponding to the current sentence data can be extracted according to the voice data, and the tone feature data and the mood feature data are also used as semantic feature data of the current sentence data; of course, other types of data characterizing the semantic features of the current sentence data are also possible, and the present exemplary embodiment is not limited thereto.

Specifically, word segmentation processing can be performed on current sentence data to obtain sentence words, and word encoding is performed on the sentence words to obtain word feature data. The word segmentation processing refers to a processing procedure of segmenting current sentence data to obtain words and characters, for example, the current sentence data can be segmented through a word dictionary to obtain sentence words, and the current sentence data can be segmented through a pre-trained word segmentation tool to obtain sentence words, which is not limited in this example embodiment.

Optionally, nonsensical word elimination can be performed on the data after word segmentation, for example, words without specific semantic information, such as "o", "y" and the like, in the current sentence data can be deleted, so that workload is reduced.

Alternatively, word coding of sentence words may be implemented by searching in an embedding matrix, specifically, word coding of sentence words may be implemented by steps in fig. 3, so as to obtain word feature data, which may specifically include, as shown in reference to fig. 3:

step S310, word embedding encoding is carried out on the sentence words to obtain word embedding vectors;

step S320, part-of-speech tagging encoding is carried out on the sentence words to obtain part-of-speech tag embedded vectors;

step S330, performing position coding on the sentence words to obtain position embedded vectors;

step S340, entity type labeling coding is carried out on the sentence words, and entity type label embedded vectors are obtained;

and step S350, splicing according to the word embedding vector, the part-of-speech tag embedding vector, the position embedding vector and the entity type tag embedding vector to generate word characteristic data.

The word embedding coding refers to a coding process of obtaining word embedding vectors by inquiring in a pre-trained word embedding matrix (such as GloVe) according to sentence words.

Part-Of-Speech (POS) -tag coding refers to a coding process Of generating Part-Of-Speech (POS) tag embedding vectors by looking up a POS tag embedding table initialized randomly, and Part-Of-Speech tag coding is actually used for classifying words, such as nouns, verbs, adjectives, adverbs, etc., and by studying POS tags, more relations between sentence words and sentence words can be learned, such as articles generally precede nouns, nouns may be followed by verbs, etc.

Position coding refers to a feature coding process for the position of a sentence word in current sentence data, e.g. assuming parameter w _i And w is equal to _c Can represent that the current sentence data has oneFrom the sentence words of the position relation, w can be obtained by searching a randomly initialized position embedded table _i To w _c The relative distance i-c of (c) is encoded into a real value vector to obtain a position embedding vector.

The entity type labeling coding refers to a coding process of labeling entity types for all sentence words in the current sentence data, for example, a process of labeling coding by referring to parts of speech, wherein the labeling of entity types for all sentence words in the current sentence data can be performed by adopting a BIO labeling mode, and an entity type label is converted into a real value vector through searching an embedding table, so as to obtain an entity type label embedding vector. BIO mode refers to labeling each sentence word with "B-X", "I-X", or "O", where "B-X" indicates that the sentence word belongs to an X type in the current sentence data (e.g., X type may include an entity type or not) and the beginning of the sentence word in the current sentence data, "I-X" indicates that the sentence word belongs to an X type in the current sentence data and the sentence word is in an intermediate position in the current sentence data, and "O" indicates that the sentence word does not belong to any type.

After the word embedding vector, the part-of-speech tag embedding vector, the position embedding vector and the entity type tag embedding vector are obtained, the word embedding vector, the part-of-speech tag embedding vector, the position embedding vector and the entity type tag embedding vector can be spliced to be used as word characteristic data of current sentence data.

By encoding the sentence words in the current sentence data from multiple aspects, deep semantic information in the current sentence data can be effectively mined, and encoding vectors in different forms are used as word feature data of the current sentence data, so that ambiguity of the sentence words in different sentences can be effectively distinguished, and the generated word feature data can accurately represent the semantic information of the current sentence data.

Specifically, entity recognition processing can be performed on the current sentence data to obtain sentence entities, and entity encoding is performed on the sentence entities to obtain entity characteristic data. The entity recognition refers to a process of performing entity matching on all sentence words in the current sentence data and determining words with entity meanings, for example, entity words in the current sentence data can be extracted through an entity dictionary, or entity words in the current sentence data can be extracted through a pre-trained entity recognition model, which is not limited in this example embodiment.

The word embedding vectors of all sentence words constituting the entity may be subjected to a mean pooling operation, and the entity embedding vectors, i.e., entity feature data, may be calculated. The entity encoding herein is not identical to the entity type annotation encoding mentioned in the previous word encoding: the entity type label embedding vector is used for marking entity types of all words in the whole sentence and then embedding; and the entity in the entity code refers to the entity in the sentence, such as a person name, a place name and the like, which is the entity, and some words in the sentence are not the entity.

Specifically, adjacent sentence data corresponding to the current sentence data can be obtained, and context coding is performed on the adjacent sentence data, so that context characteristic data is obtained. The adjacent sentence data refers to a sentence corresponding to a previous sentence and a sentence corresponding to a subsequent sentence of the current sentence data, for example, for the sentence paragraphs { sentence 1, sentence 2, sentence 3, sentence 4, sentence 5}, if the current sentence data is sentence 3, sentence 1, sentence 2, sentence 4, and sentence 5 all belong to adjacent sentence data of the current sentence data, wherein sentence 1 and sentence 2 belong to the previous adjacent sentence data of sentence 3, and sentence 4 and sentence 5 belong to the subsequent adjacent sentence data of sentence 3. Alternatively, the adjacent sentence data may be two sentences before and after the current sentence data, or one sentence before and after the current sentence data, or three sentences before and after the current sentence data, and the like, and the selection range of the adjacent sentence data is not limited in any way in this example embodiment.

Alternatively, adjacent sentence data can be subjected to sentence coding to obtain adjacent sentence vectors, and then the adjacent sentence vectors can be spliced to obtain context feature data corresponding to the current sentence data. For example, front and back sentence data corresponding to current sentence data are obtained and used as adjacent sentence data, word embedding vectors of all sentence words in the adjacent sentence data are determined for each adjacent sentence data, and the word embedding vectors are spliced to obtain adjacent sentence vectors corresponding to the adjacent sentence data; and finally, splicing the four adjacent sentence vectors to obtain the context characteristic data corresponding to the current sentence data.

Specifically, sentence pinyin corresponding to the current sentence data can be obtained, and pinyin coding is performed on the sentence pinyin to obtain pinyin characteristic data. The pinyin corresponding to the sentence words forming the current sentence data can be expressed by using a pre-trained pinyin embedded matrix to obtain pinyin characteristic data.

The sentence words in the current sentence data are encoded from a plurality of aspects such as words, entities, contexts, pinyin and the like, deep semantic information in the current sentence data can be effectively mined, and encoding vectors in different forms are used as word feature data of the current sentence data, so that ambiguity of the sentence words in different sentences can be effectively distinguished, and the generated word feature data can accurately represent the semantic information of the current sentence data.

In some example embodiments of the present disclosure, at least two kinds of semantic feature data may be associated based on a preset dependency relationship, resulting in a multi-form heterogeneous relationship graph.

The preset dependency relationship may include any one or more of a dependency relationship between words, a dependency relationship between words and trigger words, a dependency relationship between words and entities, an association relationship between words and entities, a dependency relationship between context and entities, and a dependency relationship between pinyin and trigger words.

Determining the dependency relationship between words through a syntactic dependency tree; the dependency relationship between the words and the trigger words can be determined through a trigger word lookup table; the subordinate relationship between the word and the entity can be determined through the belonging relationship between the word and the entity (such as that the word belongs to a certain entity); if an entity and a word appear in a certain event, the entity and the word can be used as the association relationship between the word and the entity; if the types of two entities are previously parameters involved in the same event, the two entities can be determined as dependency relationships between the entities; if an entity is present in the context, a dependency relationship between the context and the entity may be determined; and connecting the pinyin codes of the current sentence data with the trigger words to serve as the dependency relationship between the pinyin and the trigger words. Of course, other types of dependencies are possible, and the present exemplary embodiment is not particularly limited thereto.

The preset dependency relationship can be used as an edge of the multi-form heterogeneous relationship graph, semantic feature data (such as words, entities, contexts, pinyin and the like) can be used as nodes of the multi-form heterogeneous relationship graph, and the nodes in different forms are simultaneously connected through a short path by the edge to obtain the multi-form heterogeneous relationship graph corresponding to the current statement data.

In some example embodiments of the present disclosure, a multi-form heterogeneous relationship graph may be input into a pre-trained intent event extraction model to obtain event trigger words and event content information, and sentence intent events corresponding to current sentence data are generated based on the event trigger words and the event content information. For example, for the current sentence data "born in a city of a in 2022", the corresponding multi-form heterogeneous relation diagram is input into the pre-trained intent event extraction model, so that the corresponding event trigger word is "born", the corresponding event content information is that the event character is "small bright", the event place is "a city", the event time is "2022", and the current sentence data "born in 2022, the small bright in a city of a generates the" corresponding sentence intent event is "event trigger word": birth; event character: xiaoming; event location: market A; event time: 2022 ", of course, is illustrative only and should not be taken as limiting in any way.

In some example embodiments of the present disclosure, the intent event extraction model may be a relationship graph convolution neural network R-GCN.

The relationship graph convolutional neural network R-GCN can aggregate different forms of node information in the multi-form heterogeneous relationship graph and relationship interaction among the nodes to extract statement intention events. And the R-GCN is used for executing information propagation on the model, and can well process high-dimensional relation data when updating nodes, distinguish various different edge types and realize information propagation on graph nodes through aggregation. After the information is propagated for L times, the information of each node is propagated to the node with the distance L, and the node representation of the L-hop inference perception relationship is generated so as to update the embedded codes of the entity and the sentence words.

In this embodiment, the task of extracting the intended event is defined as a sequence marking task, each sentence word in the current sentence data is assigned a label for facilitating event annotation, specifically, a BIO annotation mode can be applied to assign a trigger label to each sentence word, and the label "O" indicates that the corresponding word is irrelevant to the target event. The other two labels "B-type" and "I-type" are made up of two parts. I.e. the position of the word in the trigger and any event type. After aggregating the word and entity node embedded representations from the R-GCN, inputting the word representation into a fully connected network, then calculating the distribution of all event types by using a Softmax function, and selecting the event label with the highest probability as a classification result. After obtaining a particular type of trigger candidate from the trigger tag, the role each entity plays in such events needs to be predicted, and the candidate vector and entity vector are triggered by aggregating word embedded representations along the average pooling of the sequence length dimension. Candidate vectors are composed of words that are combined into trigger words, which are concatenated together and input into a new fully concatenated network for predicting the parametric roles.

Fig. 4 schematically illustrates a flow diagram for implementing statement intent event extraction in accordance with some embodiments of the disclosure.

Referring to fig. 4, step S410 is performed to obtain current sentence data, and sentence coding (e.g. word coding, entity coding, pinyin coding) and context coding are performed on the current sentence data to obtain speech feature data such as word nodes, entity nodes, context nodes, pinyin nodes, etc.;

step S420, acquiring a preset dependency relationship, and modeling event correlation of word nodes, entity nodes, context nodes and pinyin nodes by taking the preset dependency relationship as an edge to obtain a multi-form heterogeneous relationship diagram;

step S430, the multi-form heterogeneous relation graph can be input into a pre-trained intention event extraction model R-GCN in the form of an embedded vector, aggregation node information is calculated through the R-GCN, and information propagation on graph nodes is realized through aggregation; after information is propagated for L times, information of each node is propagated to a node with a distance L, and node representation of an L-hop inference perception relationship is generated to update entity embedded codes and word embedded codes;

step S440, carrying out event detection and parameter extraction based on the R-GCN to obtain event trigger words and event content information, and determining statement intention events corresponding to the current statement data according to the event trigger words and the event content information.

For example, the sentence paragraph "small is old customer. Recently he felt that the web speed was not very fast. Thus, the Ming opens a 5G package at operator A. The 4G package of operator a has been used until now by the small. He has upgraded the package 8 years from the last time. "assuming that the current sentence data is" then, the Ming opens a 5G package at operator A ", the specific sentence intention event extraction procedure is as follows:

firstly, performing word segmentation processing on the current sentence data of' then, opening a 5G package at an operator A by using the Ming, removing stop words (namely nonsensical words), wherein the obtained sentence words can be { Ming, operator A, opening, 5G package }, and firstly performing word coding (comprising but not limited to word embedding coding, part-of-speech tagging coding, position coding and entity type tagging coding) on the sentence words to obtain word characteristic data; performing entity identification on the current sentence data to obtain an entity { Xiaoming, operators A,5G package }, and encoding the entity to obtain entity characteristic data; adjacent sentence data corresponding to the current sentence data comprise 'Xiaoming is old customer', 'recently he feels that the net speed is not very fast', '4G package of an operator A is always used before Xiaoming', '8 years before the last time he upgrades package', word embedding encoding is carried out on sentence words after word segmentation is carried out on the adjacent sentence data, the obtained word embedding vectors are spliced to obtain adjacent sentence vectors, and four adjacent sentence data are spliced to obtain context characteristic data corresponding to the current sentence data; sentence pinyin of the current sentence data can be extracted through a related algorithm, and the sentence pinyin is subjected to pinyin coding to obtain pinyin characteristic data.

Secondly, according to preset dependency relationships, such as a dependency relationship 1 between words, a dependency relationship 2 between words and trigger words, a dependency relationship 3 between words and entities, a dependency relationship 4 between words and entities, a dependency relationship 5 between entities, a dependency relationship 6 between contexts and entities and a dependency relationship 7 between pinyin and trigger words, the semantic feature data in different forms can be associated to obtain a multi-form heterogeneous relationship graph. Fig. 5 schematically illustrates a schematic diagram of a multi-formal heterogeneous relationship diagram according to some embodiments of the present disclosure, and referring to fig. 5, the multi-formal heterogeneous relationship diagram 500 may be represented by reference numerals of edges in the multi-formal heterogeneous relationship diagram 500, corresponding to the above 7 preset dependencies.

The multi-form heterogeneous relation diagram 500 is input into a pre-trained intention event extraction model in the form of an embedded vector to obtain current statement data, and then, statement intention events corresponding to the fact that the operator A opens a 5G package are event trigger words: opening; event character: zhang Jiang; event location: chinese telecom; the event content: 5G package. Of course, the statement intent event presentation may be otherwise, and should not be construed as limiting in any way to the exemplary embodiment, merely by way of illustration.

It should be noted that although the steps of the methods of the present disclosure are illustrated in the accompanying drawings in a particular order, this does not require or imply that the steps must be performed in that particular order or that all of the illustrated steps be performed in order to achieve desirable results. Additionally or alternatively, certain steps may be omitted, multiple steps combined into one step to perform, and/or one step decomposed into multiple steps to perform, etc.

Furthermore, in the present exemplary embodiment, an intention event extraction apparatus is also provided. Referring to fig. 6, the intention event extraction apparatus 600 includes:

the sentence data obtaining module 610 is configured to obtain current sentence data to be processed;

the semantic feature extraction module 620 is configured to extract at least two types of semantic feature data of the current sentence data;

the semantic relation diagram construction module 630 is configured to associate the at least two semantic feature data based on a preset dependency relationship to obtain a multi-form heterogeneous relation diagram;

the sentence intention event determining module 640 is configured to input the multi-form heterogeneous relationship graph into a pre-trained intention event extraction model, so as to obtain a sentence intention event corresponding to the current sentence data.

In some example embodiments of the present disclosure, based on the foregoing aspects, the at least two semantic feature data may include word feature data, entity feature data, context feature data, and pinyin feature data; the semantic feature extraction module 620 may be configured to:

In some example embodiments of the present disclosure, based on the foregoing approach, the semantic feature extraction module 620 may be to:

In some example embodiments of the present disclosure, based on the foregoing schemes, the preset dependency may include any one or more of a dependency between words, a dependency between words and trigger words, a dependency between words and entities, an association between words and entities, a dependency between contexts and entities, and a dependency between pinyin and trigger words.

In some example embodiments of the present disclosure, based on the foregoing approach, the statement intent event determination module 640 may be to:

In some example embodiments of the present disclosure, based on the foregoing approach, the intent event extraction model may include a relationship graph convolution neural network R-GCN.

The specific details of each module of the above intentional event extraction device have been described in detail in the corresponding intentional event extraction method, and thus are not described herein.

It should be noted that although several modules or units of the intention event extraction device are mentioned in the above detailed description, this division is not mandatory. Indeed, the features and functionality of two or more modules or units described above may be embodied in one module or unit in accordance with embodiments of the present disclosure. Conversely, the features and functions of one module or unit described above may be further divided into a plurality of modules or units to be embodied.

In addition, in an exemplary embodiment of the present disclosure, an electronic device capable of implementing the above-described intention event extraction method is also provided.

Those skilled in the art will appreciate that the various aspects of the present disclosure may be implemented as a system, method, or program product. Accordingly, various aspects of the disclosure may be embodied in the following forms, namely: an entirely hardware embodiment, an entirely software embodiment (including firmware, micro-code, etc.) or an embodiment combining hardware and software aspects may be referred to herein as a "circuit," module "or" system.

An electronic device 700 according to such an embodiment of the present disclosure is described below with reference to fig. 7. The electronic device 700 shown in fig. 7 is merely an example and should not be construed as limiting the functionality and scope of use of the disclosed embodiments.

As shown in fig. 7, the electronic device 700 is embodied in the form of a general purpose computing device. Components of electronic device 700 may include, but are not limited to: the at least one processing unit 710, the at least one storage unit 720, a bus 730 connecting the different system components (including the storage unit 720 and the processing unit 710), and a display unit 740.

Wherein the storage unit stores program code that is executable by the processing unit 710 such that the processing unit 710 performs steps according to various exemplary embodiments of the present disclosure described in the above-described "exemplary methods" section of the present specification. For example, the processing unit 710 may perform step S210 shown in fig. 1 to obtain current sentence data to be processed; step S220, extracting at least two semantic feature data of the current sentence data; step S230, associating the at least two semantic feature data based on a preset dependency relationship to obtain a multi-form heterogeneous relationship graph; step S240, inputting the multi-form heterogeneous relation diagram into a pre-trained intention event extraction model to obtain a statement intention event corresponding to the current statement data.

The memory unit 720 may include readable media in the form of volatile memory units, such as Random Access Memory (RAM) 721 and/or cache memory 722, and may further include Read Only Memory (ROM) 723.

The storage unit 720 may also include a program/utility 724 having a set (at least one) of program modules 725, such program modules 725 including, but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of which may include an implementation of a network environment.

Bus 730 may be a bus representing one or more of several types of bus structures including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

The electronic device 700 may also communicate with one or more external devices 770 (e.g., keyboard, pointing device, bluetooth device, etc.), one or more devices that enable a user to interact with the electronic device 700, and/or any device (e.g., router, modem, etc.) that enables the electronic device 700 to communicate with one or more other computing devices. Such communication may occur through an input/output (I/O) interface 750. Also, electronic device 700 may communicate with one or more networks such as a Local Area Network (LAN), a Wide Area Network (WAN) and/or a public network, such as the Internet, through network adapter 760. As shown, network adapter 760 communicates with other modules of electronic device 700 over bus 730. It should be appreciated that although not shown, other hardware and/or software modules may be used in connection with electronic device 700, including, but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, data backup storage systems, and the like.

From the above description of embodiments, those skilled in the art will readily appreciate that the example embodiments described herein may be implemented in software, or in combination with the necessary hardware. Thus, the technical solution according to the embodiments of the present disclosure may be embodied in the form of a software product, which may be stored in a non-volatile storage medium (may be a CD-ROM, a U-disk, a mobile hard disk, etc.) or on a network, and includes several instructions to cause a computing device (may be a personal computer, a server, a terminal device, or a network device, etc.) to perform the method according to the embodiments of the present disclosure.

In an exemplary embodiment of the present disclosure, a computer-readable storage medium having stored thereon a program product capable of implementing the method described above in the present specification is also provided. In some possible embodiments, the various aspects of the present disclosure may also be implemented in the form of a program product comprising program code for causing a terminal device to carry out the steps according to the various exemplary embodiments of the disclosure as described in the "exemplary methods" section of this specification, when the program product is run on the terminal device.

Referring to fig. 8, a program product 800 for implementing the above-described intent event extraction method according to an embodiment of the present disclosure is described, which may employ a portable compact disc read only memory (CD-ROM) and include program code, and may be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto, and in this document, a readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a combination of any of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium would include the following: an electrical connection having one or more wires, a portable disk, a hard disk, random Access Memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

The computer readable signal medium may include a data signal propagated in baseband or as part of a carrier wave with readable program code embodied therein. Such a propagated data signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination of the foregoing. A readable signal medium may also be any readable medium that is not a readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

Program code embodied on a readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

Program code for carrying out operations of the present disclosure may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device, partly on a remote computing device, or entirely on the remote computing device or server. In the case of remote computing devices, the remote computing device may be connected to the user computing device through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external computing device (e.g., connected via the Internet using an Internet service provider).

Furthermore, the above-described figures are only schematic illustrations of processes included in the method according to the exemplary embodiments of the present disclosure, and are not intended to be limiting. It will be readily appreciated that the processes shown in the above figures do not indicate or limit the temporal order of these processes. In addition, it is also readily understood that these processes may be performed synchronously or asynchronously, for example, among a plurality of modules.

From the above description of embodiments, those skilled in the art will readily appreciate that the example embodiments described herein may be implemented in software, or in combination with the necessary hardware. Thus, the technical solution according to the embodiments of the present disclosure may be embodied in the form of a software product, which may be stored in a non-volatile storage medium (may be a CD-ROM, a U-disk, a mobile hard disk, etc.) or on a network, and includes several instructions to cause a computing device (may be a personal computer, a server, a touch terminal, or a network device, etc.) to perform the method according to the embodiments of the present disclosure.

Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This application is intended to cover any adaptations, uses, or adaptations of the disclosure following, in general, the principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It is to be understood that the present disclosure is not limited to the precise arrangements and instrumentalities shown in the drawings, and that various modifications and changes may be effected without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. An intention event extraction method, characterized by comprising:

acquiring current statement data to be processed;

extracting at least two semantic feature data of the current sentence data;

2. The method of claim 1, wherein the at least two semantic feature data comprises word feature data, entity feature data, context feature data, and pinyin feature data;

3. The method for extracting an intended event according to claim 2, wherein said word encoding the sentence word to obtain word feature data comprises:

4. The method for extracting an intention event according to claim 2, wherein the performing context encoding on the adjacent sentence data to obtain context feature data includes:

5. The method for extracting an intention event according to claim 1, wherein the preset dependency relationship includes any one or more of a dependency relationship between words, a dependency relationship between words and trigger words, a dependency relationship between words and entities, an association relationship between words and entities, a dependency relationship between contexts and entities, and a dependency relationship between pinyin and trigger words.

6. The method for extracting an intention event according to claim 1, wherein the inputting the multi-form heterogeneous relationship graph into a pre-trained intention event extraction model to obtain a sentence intention event corresponding to the current sentence data comprises:

7. The method of claim 1 or 6, wherein the intent event extraction model comprises a relationship graph roll-up neural network R-GCN.

8. An intention event extraction device, characterized by comprising:

9. An electronic device, comprising:

a processor; and

a memory having stored thereon computer readable instructions that when executed by the processor implement the intended event extraction method of any of claims 1 to 7.

10. A computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the intended event extraction method of any of claims 1 to 7.