CN116795963A - Hotline allocation method based on knowledge graph question and answer - Google Patents
Hotline allocation method based on knowledge graph question and answer Download PDFInfo
- Publication number
- CN116795963A CN116795963A CN202310601972.5A CN202310601972A CN116795963A CN 116795963 A CN116795963 A CN 116795963A CN 202310601972 A CN202310601972 A CN 202310601972A CN 116795963 A CN116795963 A CN 116795963A
- Authority
- CN
- China
- Prior art keywords
- event
- vector
- graph
- text
- argument
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 32
- 239000013598 vector Substances 0.000 claims abstract description 164
- 238000000605 extraction Methods 0.000 claims abstract description 26
- 230000004927 fusion Effects 0.000 claims description 35
- 238000009826 distribution Methods 0.000 claims description 28
- 230000006870 function Effects 0.000 claims description 16
- 238000004364 calculation method Methods 0.000 claims description 12
- 238000010586 diagram Methods 0.000 claims description 11
- 238000004220 aggregation Methods 0.000 claims description 7
- 239000013604 expression vector Substances 0.000 claims description 7
- 230000011218 segmentation Effects 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 6
- 230000002776 aggregation Effects 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000013473 artificial intelligence Methods 0.000 abstract description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 101150104012 TOP2 gene Proteins 0.000 description 1
- 238000002679 ablation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000001149 cognitive effect Effects 0.000 description 1
- 230000001627 detrimental effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 238000004804 winding Methods 0.000 description 1
Landscapes
- Machine Translation (AREA)
Abstract
The invention relates to the technical field of artificial intelligence, in particular to a hotline allocation method based on knowledge graph question answering, which is characterized in that key information of an event is acquired by utilizing an event extraction model to construct a historical event knowledge graph, an event trigger word, an event type and an event role which are acquired according to event extraction are acquired, sub-graph retrieval is carried out, graph representation features are acquired by combining a fusion-GCN model, a text retrieval method is adopted to acquire a most relevant 'third' responsibility text of the event, a 'third' responsibility representation vector and an event text feature representation vector set are acquired by BERT, division score prediction is carried out by fusing event vector set elements, graph representation features and a 'third' responsibility representation vector, and an optimal allocation division is output according to sorting of a scoring result, so that accurate retrieval and final answer are realized, and the burden of hotline operators is facilitated to be reduced.
Description
Technical Field
The invention relates to the technical field of artificial intelligence, in particular to a hotline allocation method based on knowledge graph questions and answers.
Background
In recent years, hotline appeal relates to various aspects of urban governance, and various aspects of social governance are covered on incoming call topics, such as events of atmospheric pollution, medical insurance, noise disturbance, vehicle parking, property service, consumption disputes and the like. To efficiently implement hot line allocation, operators not only understand the reflected events fully, but also have a deep understanding of the roles of the departments. Without specialized training, operators have difficulty in choosing the right department from tens of treatment departments to complete hot line allocation, which results in low accuracy of manual allocation (i.e., incorrect dispatch) and a large number of secondary allocations. Some relatively complex complaints may require cross-department co-processing, and thus erroneous allocation is detrimental to the authority and executive power of the hotline.
Patent CN113919811B proposes a hot line event allocation method based on enhanced association. This method has the following problems: (1) For each hotline text, an event portrait needs to be constructed, and the event portrait formed by the history hotline text cannot be used for a new hotline text, namely, the event portrait cannot be precipitated as knowledge; (2) Although the 'three-dimensional' information plays a certain role in hotline allocation, uncertainty exists in strengthening association of the hotline allocation and the event portraits due to the existence of the problem (1), namely, the accuracy of hotline allocation is difficult to guarantee on the premise that the extraction effect of some event portraits is poor and even extraction errors exist.
Disclosure of Invention
Therefore, the invention aims to provide a hotline allocation method based on knowledge graph questions and answers, so as to solve the problem that the accuracy of hotline allocation is difficult to guarantee.
Based on the above purpose, the invention provides a hotline allocation method based on knowledge graph questions and answers, which is characterized by comprising the following steps:
s1, acquiring an event trigger word, an event type and an event argument role from an input event text by adopting an event extraction algorithm;
s2, using corresponding event number information and treatment department information in the historical event data, and establishing a relation pattern diagram by combining event trigger words, event types and event argument roles;
s3, integrating the information in the relation pattern diagram through knowledge fusion, and constructing corresponding triple information in each eventnZhang Zitu, fusing the repeated entities in the knowledge graph, and establishing the connection between the subgraphs to form a complete knowledge graph;
s4, obtaining a final sub-graph G by sub-graph retrieval based on a knowledge graph, and obtaining a text retrieval result by text retrieval based on 'three-dimensional' responsibilities;
s5, carrying out event association feature extraction on the final sub-graph G and the text retrieval result to respectively obtain a graph representation vector and a feature representation vector of a 'three-in' responsibility text, and combining event text features with a department entity set in the sub-graph to generate a fusion event vector set;
s6, carrying out department score prediction on all elements in the fusion event vector set through the graph representation vector and the characteristic representation vector of the 'three-dimensional' responsibility text, and outputting the optimal allocation department.
Preferably, in step S1, the step of obtaining the event trigger word includes:
word segmentation is carried out on the input event text by using a Chinese word segmentation tool to obtain a series of word sequences, wherein nWord sequence length obtained for word segmentation;
initializing and coding the word sequence by using word2vec coding model to obtain corresponding initialized semantic coding vectorFor each word, a corresponding position vector is constructed>And segment vector->；
Inputting the initialized coding vector, the position vector and the segment vector into a pre-training language model to obtain a corresponding semantic coding vectorThe semantically encoded vectors form a set +.>；
Constructing a classifier model, sequentially inputting vectors in word sequence semantic coding vectors into a classifier to obtain corresponding trigger probability distribution, wherein />，/>Is a parameter that can be learned;
after obtaining the classification probability distribution of the sequences, selecting word sequences corresponding to top-n in the probability distribution as final event trigger word prediction results;
optimizing predictions using a cross entropy loss function:
wherein Representing the actual trigger word classification result,/->Representing predicted trigger wordsAnd classifying the result.
Preferably, in step S1, the specific step of obtaining the event type includes:
for trigger word sequenceSemantic vector codes corresponding to the trigger words in the event type probability distribution are fused and input into a multi-classification network to obtain the corresponding event type probability distribution +.>And selecting the result with the highest probability as a predicted result of the event type:
wherein ,，/>for a learnable parameter->Representing maximum pooling operation, and respectively taking the maximum value corresponding to each dimension;
event classification optimization using cross entropy functions
。
Preferably, in step S1, the acquiring the event argument role specifically includes the following steps:
constructing a corresponding key information position vector according to the trigger word sequence generation result；
Set of semantically encoded vectorsCorresponding position vector->Key information position vector->Inputting the pre-training language model together to obtain corresponding coding vector +.>，
Encoding the encoding result and event typeSending the two information into an argument classifier together to generate corresponding argument character distribution probability
wherein ，/>For a learnable parameter, event type code +.>Constructing according to the generation result of the event type module to be a one-dimensional vector;
after probability distribution is obtained, selecting probability corresponding to probability topn for calculation, if the probability distribution of the first n distributions is similar, outputting n classification results at the same time, otherwise, outputting only the classification result with highest probability;
the argument character extraction is optimized by adopting a multi-classification cross entropy function:
wherein For the number of argument characters->As a sign function, if the sample->The true category equals->Then take 1, otherwise take 0, +.>Representation of sample->The predicted outcome category is->Is a probability of (2).
Preferably, step S3 specifically includes:
entity disambiguation: based on an entity pair formed by any two homonymous entities in the relation pattern diagram, after the two entities are associated with the surrounding event text, inputting a pre-training language model to obtain a corresponding vector code;
calculating semantic similarity between two entities, combining entity constitution and collection with similarity greater than a certain threshold, and distinguishing homonymous entities with different references
wherein Representing a vector encoding of an entity, +.>Representing the encoding length of the entity vector;
coreference resolution: and coding the argument and the sentence where the argument is positioned by using a pre-training language model, and fusing the two generated vectors to form fusion coding for a role of an argument. Clustering the argument character fusion codes, calculating fusion vector similarity in an argument fusion vector set aggregated in the same cluster, and combining argument pairs with similarity larger than a certain threshold value similarly to form unified entity references:
wherein Representing a fusion encoding of argument roles, +.>，/>Coding vectors of argument characters and sentences corresponding thereto, respectively>Is a weight that can be learned.
Preferably, in step S4, obtaining the final subgraph G using the subgraph search based on the knowledge graph includes:
the three types of entities of the event trigger word, the event type and the event argument roles are taken as seed entities and recorded as a set；
Pair aggregationThe elements in (a) identify other entity sets which are possibly related to the event by adopting a PPR algorithm +.>；
For elements in an entity set identified by a PPR algorithm, calculating word vectors of the elements and sentence vectors of the events in an inner product mode, taking a calculation result as edge weights of the elements and seed entities to obtain subgraphs of three different types of entities related to trigger words, event types and event roles, and combining the same entities in the subgraphs of the entities of the same type to form three subgraphs corresponding to the three types;
and merging the same entities in the three sub-graphs of the same type to obtain a final sub-graph G.
Preferably, in step S5, the extracting the event-related features from the final subgraph G includes:
forming G1, G2, G3 and G into pairs, i.e., (G1, G), (G2, G) and (G3, G), respectively;
for a key input node V, defining a graph convolution operation of two vertex domains to aggregate local features and global features of the node, wherein the graph convolution operation adopts GCN to extract features, and a calculation formula is as follows:
wherein ,a is the adjacency matrix of the input graph, I is the identity matrix,>，/> and />Are all learnable parameters, < >>；
The GCN output is fused in a multi-layer aggregation mode, and the calculation formula is as follows:
wherein ,l is the number of layers of the graph roll;
computing two vector pairs for each graph pair，/> and />Global and local feature representation vectors respectively representing nodes corresponding to single-type sub-graph { Gi } and fusion sub-graph G, and performing fusion based on the vectors to serve as final representation vector +.>，
wherein ,,/>d is the number of attention heads, which is a learnable parameter;
respectively calculating the expression vectors of three types of nodes of the event type, the trigger word and the event role, and splicingFusion resulting graph representation vectorAnd all department nodes in the subgraph represent a vector set。
Preferably, in step S4, text retrieval results are obtained using text retrieval based on "tricky" responsibilities, including:
vectorizing all sentences in a 'triangulated' responsibility retrieval corpus by using Glove word vectors to obtain vector representation sets of all sentences；
Vectorizing an input event text by adopting GloVe to obtain an event text representation vector Ve, and calculating Ve and VeThe inner products of the middle elements are sorted according to the inner product result, and top-n related sentences are obtained;
tracing the top-n related sentences, calculating the score of the 'triad' responsibility texts corresponding to the n sentences, and selecting one 'triad' responsibility text with the highest score as a text retrieval result.
Preferably, in step S5, performing event-related feature extraction on the text search result, and generating a fused event vector set by combining the event text feature with the department entity set in the subgraph includes:
for the problem text representation, the input event text and the sub-department entity vector set retrieved from the knowledge graph are combinedThe elements in (a) are fused to be expressed as a problem, and the method comprises the following steps:
vectorizing input event text using BERT to obtain event representation vectors,
Where q is the text of the input event,is a BERT pre-training model;
vectorAnd department entity vector set->Each element in the vector is fused to obtain a fusion event vector set +.>，
wherein , and />Are all learnable parameters;
for the searched "third-party" responsibility text, the BERT is adopted to directly extract the characteristics of the text so as to obtain the characteristic expression vector V of the "third-party" responsibility text sanding ，
wherein ,is BERT pre-training modelIs (1)>And (5) searching the result for the text.
Preferably, step S6 specifically includes:
representing vectors by graphFeature representation vector V with "triad" responsibility text sanding Splicing;
by matching problem vectorsAnd the probability prediction is carried out on the department entities obtained in the sub-graph retrieval with the spliced result vector,
wherein , and />Are all learnable parameters;
for fused problem vector setsAll elements in the department are predicted to obtain a prediction score set of the department；
Pair aggregation according to department output probabilityAnd sorting, and selecting the optimal department as a hotline allocation department.
The invention has the beneficial effects that: the method can effectively utilize the prior knowledge of the historical event and the prior knowledge of the 'three-setting' responsibility to improve the accuracy of the hot line allocation. The method comprises the steps of obtaining key information of an event by utilizing an event extraction model to construct a historical event knowledge graph, extracting obtained event trigger words, event types and event roles according to the event, carrying out sub-graph retrieval, obtaining graph representation features by combining a fusion-GCN model, obtaining a most relevant 'third-party' responsibility text of the event by adopting a text retrieval method, obtaining a 'third-party' responsibility representation vector and an event text feature representation vector set by BERT, carrying out department score prediction by fusing event vector set elements, graph representation features and 'third-party' responsibility representation vectors, and sequencing and outputting an optimal allocation department according to a score result. For complex incoming call consultation, a semantic analysis method can be used for analyzing consultation event text to obtain citizen intention, and then the citizen intention is converted into a structured logic query form by a template matching-based method, so that accurate retrieval is realized and a final answer is obtained. At the hot line intelligent distribution level: the more delicate distribution of the hotline is beneficial to reducing the burden of the hotline operators.
Drawings
In order to more clearly illustrate the invention or the technical solutions of the prior art, the drawings which are used in the description of the embodiments or the prior art will be briefly described, it being obvious that the drawings in the description below are only of the invention and that other drawings can be obtained from them without inventive effort for a person skilled in the art.
FIG. 1 is a schematic diagram of a hot wire distribution model according to an embodiment of the present invention;
FIG. 2 is a schematic diagram illustrating an event extraction process according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a relationship pattern according to an embodiment of the present invention;
fig. 4 is a schematic diagram of a sub-graph retrieval process according to an embodiment of the present invention.
Detailed Description
The present invention will be further described in detail with reference to specific embodiments in order to make the objects, technical solutions and advantages of the present invention more apparent.
It is to be noted that unless otherwise defined, technical or scientific terms used herein should be taken in a general sense as understood by one of ordinary skill in the art to which the present invention belongs. The terms "first," "second," and the like, as used herein, do not denote any order, quantity, or importance, but rather are used to distinguish one element from another. The word "comprising" or "comprises", and the like, means that elements or items preceding the word are included in the element or item listed after the word and equivalents thereof, but does not exclude other elements or items. The terms "connected" or "connected," and the like, are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "upper", "lower", "left", "right", etc. are used merely to indicate relative positional relationships, which may also be changed when the absolute position of the object to be described is changed.
As shown in fig. 1, for a given input event text, the invention firstly adopts an event extraction algorithm to acquire event trigger words, event types and event roles and performs sub-graph retrieval so as to acquire graph representation features in combination with a fusion-GCN model; secondly, the text retrieval method based on GloVe word vectors is adopted to obtain the most relevant 'third-order' responsibility text, and BERT is adopted to obtain the 'third-order' responsibility representation vector; then, extracting event text features by adopting a BERT model, and generating a fusion event vector set by combining a department entity set in the subgraph; finally, the invention fuses the event vector set elements, the graph representation features and the 'three-definition' responsibility representation vectors to conduct department score prediction, and sorts and outputs the optimal allocation departments according to the score result.
The method specifically comprises the following steps:
firstly, inputting an event description text into an event trigger word extraction module to generate a series of corresponding event trigger word lists, then inputting an event trigger word coding vector into an event classifier module to output a corresponding event type prediction result, and finally inputting an event classification result coding vector, the trigger word list coding vector and a word sequence coding vector into an event argument extraction module to generate a corresponding event argument role prediction result, as shown in figure 2.
Specifically, first, a Chinese word segmentation tool is used to segment an input text to obtain a series of word sequences（ wherein />Word2vec coding model is used for carrying out initialization coding on the word sequence to obtain a corresponding initialization semantic coding vector ++>. At the same time, for each word, a corresponding position vector is constructed>And segment vector->The two vectors are one-dimensional vectors, and the length is the word sequence length +.>Wherein the position vector->In the word corresponding position 1, the rest positions 0, and the segment vector is input as a text vector, and the full 0 vector is used in the initialization. After the construction is completed, the initialization coding vector, the position vector and the segment vector are input into a pre-training language model together to obtain a corresponding semantic coding vector:
semantic coding vector component set. In order to obtain the trigger words in the text, the invention converts the text trigger word extraction task into two trigger words for judging whether the words in the word sequence are the trigger words or notClassification problems. Constructing a classifier model, and sequentially inputting vectors in word sequence semantic coding vectors into a classifier to obtain corresponding trigger probability distribution:
wherein ，/>Is a parameter that can be learned. After the classification probability distribution of the sequences is obtained, selecting word sequences corresponding to top-n in the probability distribution as final trigger word prediction results. For this part of the module, the invention uses the cross entropy loss function to optimize:
wherein Representing the actual trigger word classification result,/->Representing the predicted trigger word classification result.
After trigger word extraction is completed, a trigger word sequence corresponding to the event text is obtained. Then fusing semantic vector codes corresponding to trigger words in the sequence, and inputting the semantic vector codes into a multi-classification network to obtain corresponding event type probability distribution +_>And selecting the result with the highest probability as a predicted result of the event type:
wherein ，/>For a learnable parameter->And representing maximum pooling operation, and respectively taking the maximum value corresponding to each dimension. The event classification module uses a second cross entropy function to optimize:
after event trigger word classification is completed, the method constructs the corresponding key information position vector according to the trigger word sequence generation resultThe key information position vector is a one-dimensional vector, the word position corresponding to the trigger word sequence is marked as 1, and the rest positions are 0, so that the position of the key information of the coding model is informed. Subsequently, the invention generates a vector corresponding to the descriptive text word sequence by using word2vec model>Corresponding position vector->Key information position vector->Inputting the pre-training language model together to obtain corresponding coding vector +.>：
After the coding is completed, the invention codes the coding result and the event typeSending the two components into an argument classifier together to generate corresponding argument role distribution probability:
wherein ，/>For a learnable parameter, event type code +.>And constructing a one-dimensional vector according to the generation result of the event type module, wherein the length is the size of the event type set, and the rest positions are 0 at the event type position 1 corresponding to the prediction result. After probability distribution is obtained, the probability corresponding to probability top2 is selected for calculation, if the probability distribution of the first two distributions is similar, two classification results are output at the same time, and otherwise, only the classification result with the highest probability is output. The argument character extraction module adopts multi-classification cross entropy functions for optimization:
wherein For the number of argument characters->Is of a sign ofNumber function, if sample->The true category equals->Then 1 is taken, otherwise 0 is taken. />Representation of sample->The predicted outcome category is->Is a probability of (2).
For the global model, the present invention uses a joint loss function consisting of three loss functions as the loss function of the entire event extraction model:
step 2: knowledge graph construction
Sub-step 1: information extraction
Besides the event trigger words, event types and event argument roles obtained by the event extraction module, the invention also uses corresponding event number information and disposal department information in each historical event data as other two types of entity information required for constructing a knowledge graph.
Based on the information, five types of knowledge entities are constructed, and a knowledge model frame corresponding to the five types of knowledge entities is constructed according to a cognitive framework of the prior knowledge on the basic element structure of the event, a constructed relation model diagram is shown in fig. 3, and a unified relation description mode is used between a plurality of meta-elements or a plurality of trigger words and the same event.
Sub-step 2: knowledge fusion
After the information extraction module finishes extraction and arrangement, outputting corresponding entity, relation and corresponding triplet information. The knowledge fusion mainly comprises two parts, namely entity links and knowledge merging.
The entity link is mainly divided into two parts of entity disambiguation and coreference resolution. The entity disambiguation part is used for associating two entities with surrounding event texts based on an entity pair formed by any two homonymous entities in the obtained event entity set, inputting a pre-training language model to obtain corresponding vector codes, calculating semantic similarity between the two entities, combining entity compositions and sets with similarity greater than a certain threshold value, and distinguishing homonymous entities with different references:
wherein Representing a vector encoding of an entity, +.>Representing the encoded length of the entity vector.
And performing coreference resolution on the entity set corresponding to the event argument role. Considering that co-finger ablation in an argument character is closely related to a context, after an argument character entity set is obtained, the method utilizes a pre-training language model to encode an argument and sentences in which the argument is positioned, and fuses two generated vectors to form fusion encoding of an argument character. Clustering the argument character fusion codes, calculating fusion vector similarity in an argument fusion vector set aggregated in the same cluster, and combining argument pairs with similarity larger than a certain threshold value similarly to form unified entity references:
wherein Representing a fusion encoding of argument roles, +.>，/>Coding vectors of argument characters and sentences corresponding thereto, respectively>Is a weight that can be learned.
After the entity disambiguation and coreference resolution are completed, the invention then constructs n Zhang Zitu for the corresponding triplet information in each event. And then fusing the repeated entities in the sub-graph to establish the connection between the sub-graphs, thereby forming a complete knowledge graph.
Step 3: event distribution
Sub-step 1: event association information retrieval
The invention uses two parallel pipes (two parallel pipelines) to obtain event correlation features: one is sub-graph retrieval based on the knowledge graph, namely, retrieval is carried out on the knowledge graph, and an entity set is returned; the other is text retrieval based on "tricky" responsibilities, i.e., text retrieval is performed on the "tricky" responsibilities text library D, returning to the document collection. The retrieved entity and document are then used as the associated information for the event to input the predictive model.
Firstly, extracting an event from an input event text, acquiring three types of entities including an event trigger word, an event type and an event role as seed entities, and recording the seed entities as a set. Then, the collection->The element in (a) adopts Personalized PageRank (PPR) algorithm to identify other entity sets which are possibly related to the event +.>. And for the elements in the entity set identified by the PPR algorithm, calculating the word vector of the element and the sentence vector of the event in an inner product mode, and taking the calculation result as the edge weight of the element and the seed entity. By the method, the subgraphs of three different types of entities related to the trigger word, the event type and the event role can be obtained, and the same entities in the subgraphs of the same type of entities (the trigger word, the event type and the event role) are combined to form three subgraphs corresponding to the three types.
And then merging the same entities in the three sub-graphs of the same type to obtain a final sub-graph G. For these consolidated entities, the present invention traces back to the root (i.e., seed entity) and back to the leaf in each subgraph. The invention only keeps entities and relations on the trace paths of all trees to form a subgraph.
Meanwhile, the text data with the function of 'third-order' responsibility is used as a text retrieval corpus, and sentence-level text retrieval is adopted, namely each sentence in the text retrieval corpus is retrieved. The text retrieval of the "triad" responsibilities is performed as follows:
firstly, using GloVe word vector to vector all sentences in the "trisection" responsibility retrieval corpus to obtain vector representation set of all sentences；
Secondly, carrying out vectorization on an input event text by adopting GloVe to obtain an event text representation vector Ve, and calculating Ve and VeThe inner products of the middle elements are sorted according to the inner product result, and top-5 related sentences are obtained;
and finally, tracing the top-5 related sentences, calculating the score of the 'triad' responsibility texts corresponding to the 5 sentences, and selecting one 'triad' responsibility text with the highest score as a text retrieval result.
Sub-step 2: event correlation feature extraction
For the constructed subgraphs G1, G2, G3 and G, the invention adopts a fusion graph convolution network to extract subgraph characteristics. Specifically, in order to extract graph representation features of each key node (trigger word, event type, event role, department), the present invention constructs G1, G2, G3 and G into pairs, i.e., (G1, G), (G2, G) and (G3, G), respectively. For a key input node V, the present invention defines a graph convolution operation of two vertex domains to aggregate the local and global features of that node. The graph convolution operation adopts GCN to extract the characteristics, and the calculation formula is as follows:
（1）
wherein ,a is the adjacency matrix of the input graph, I is the identity matrix,>，/> and />Are all learnable parameters, < >>. In order to further preserve the characteristic information of the graph, the GCN outputs are fused in a multi-layer aggregation mode, and the calculation formula is as follows:
（2）
wherein ,l is the number of windings, the invention l=3.
According to the formula, two vector pairs of each graph pair can be calculated，/>Andglobal and local feature representation vectors representing corresponding nodes of the single-type subgraph { Gi } and the fusion subgraph G, respectively. Vector Attention fusion based final representation vector as node>，
（3）
wherein ,,/>d is the number of attention heads, which is a learnable parameter.
According to the formula, respectively calculating the expression vectors of three types of nodes of the event type, the trigger word and the event role, and performing splicing and fusion to obtain the graph expression vectorAnd all department nodes in the subgraph represent vector set +.>。
At the same time, the question text representation and the "tricky" responsibility text are read. For the text representation of the questions, the invention searches the text and knowledge graph of the input eventSub-division entity vector set of (a)The elements in (a) are fused to represent a problem. The method comprises the following steps of vectorizing input event text by using BERT to obtain an event representation vector +.>,
（4）
Where q is the text of the input event,is a BERT pre-training model. The invention will vector->And department entity vector set->Fusing each element of the set to obtain a fused problem vector set +.>，
（5）
（6）
wherein , and />Are all learnable parameters.
For the retrieved "tricky" responsibility text, the invention uses BERT to directly extract its features,to obtain a feature representation vector V of the text of the "triad" responsibility sanding ，
（7）
wherein ,is a BERT pre-training model,/->And (5) searching the result for the text.
Sub-step 3: event dispatch department prediction
After event knowledge graph features, tri-responsibility features and input event vectorization are completed, the invention firstly represents vectors by the graphAnd "triad" responsibility represents vector V sanding Splicing is performed, and then the problem vector is matched>And the probability prediction is carried out on the department entities obtained in the sub-graph retrieval with the spliced result vector,
wherein , and />Are all learnable parameters. The invention adopts the two kinds of cross entropy as the training loss of the model, and takes the final output probability as the score of the department.
For fused problem vector setsAll elements in the system are predicted to obtain the prediction of departmentsGrouping into sets. Finally, the set is +_based on department output probability>And sorting, and selecting the optimal department as a hotline allocation department.
Those of ordinary skill in the art will appreciate that: the discussion of any of the embodiments above is merely exemplary and is not intended to suggest that the scope of the invention (including the claims) is limited to these examples; the technical features of the above embodiments or in the different embodiments may also be combined within the idea of the invention, the steps may be implemented in any order and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.
The present invention is intended to embrace all such alternatives, modifications and variances which fall within the broad scope of the appended claims. Therefore, any omission, modification, equivalent replacement, improvement, etc. of the present invention should be included in the scope of the present invention.
Claims (10)
1. A hotline allocation method based on knowledge graph questions and answers is characterized by comprising the following steps:
s1, acquiring an event trigger word, an event type and an event argument role from an input event text by adopting an event extraction algorithm;
s2, using corresponding event number information and treatment department information in the historical event data, and establishing a relation pattern diagram by combining event trigger words, event types and event argument roles;
s3, integrating the information in the relation pattern diagram through knowledge fusion, and constructing corresponding triple information in each eventnZhang Zitu, fusing the repeated entities in the knowledge graph, and establishing the connection between the subgraphs to form a complete knowledge graph;
s4, obtaining a final sub-graph G by sub-graph retrieval based on a knowledge graph, and obtaining a text retrieval result by text retrieval based on 'three-dimensional' responsibilities;
s5, carrying out event association feature extraction on the final sub-graph G and the text retrieval result to respectively obtain a graph representation vector and a feature representation vector of a 'three-in' responsibility text, and combining event text features with a department entity set in the sub-graph to generate a fusion event vector set;
s6, carrying out department score prediction on all elements in the fusion event vector set through the graph representation vector and the characteristic representation vector of the 'three-dimensional' responsibility text, and outputting the optimal allocation department.
2. The method for assigning hotline based on knowledge-graph questions and answers as claimed in claim 1, wherein in step S1, the step of obtaining event trigger words comprises:
word segmentation is carried out on the input event text by using a Chinese word segmentation tool to obtain a series of word sequences, wherein nWord sequence length obtained for word segmentation;
initializing and coding the word sequence by using word2vec coding model to obtain corresponding initialized semantic coding vectorFor each word, a corresponding position vector is constructed>And segment vector->；
Inputting the initialized coding vector, the position vector and the segment vector into a pre-training language model to obtain a corresponding semantic coding vectorThe semantically encoded vectors form a set +.>；
Constructing a classifier model, sequentially inputting vectors in word sequence semantic coding vectors into a classifier to obtain corresponding trigger probability distribution, wherein />，/>Is a parameter that can be learned;
after obtaining the classification probability distribution of the sequences, selecting word sequences corresponding to top-n in the probability distribution as final event trigger word prediction results;
optimizing predictions using a cross entropy loss function:
wherein Representing the actual trigger word classification result,/->Representing the predicted trigger word classification result.
3. The method for assigning hotline based on knowledge-graph questions and answers as claimed in claim 2, wherein in step S1, the specific step of obtaining event type comprises:
for trigger word sequenceSemantic vector codes corresponding to the trigger words in the code word are fused and input into a multi-classification network to obtain corresponding mattersPart type probability distribution->And selecting the result with the highest probability as a predicted result of the event type:
wherein ,，/>for a learnable parameter->Representing maximum pooling operation, and respectively taking the maximum value corresponding to each dimension;
event classification optimization using cross entropy functions
。
4. The method for assigning hotline based on knowledge-graph questions and answers as claimed in claim 2, wherein in step S1, the step of obtaining event argument roles specifically comprises the steps of:
constructing a corresponding key information position vector according to the trigger word sequence generation result；
Set of semantically encoded vectorsCorresponding position vector->Key information position vector->Inputting the pre-training language model together to obtain corresponding coding vector +.>，
Encoding the encoding result and event typeSending the two information into an argument classifier together to generate corresponding argument character distribution probability
wherein ，/>For a learnable parameter, event type code +.>Constructing according to the generation result of the event type module to be a one-dimensional vector;
after probability distribution is obtained, selecting probability corresponding to probability topn for calculation, if the probability distribution of the first n distributions is similar, outputting n classification results at the same time, otherwise, outputting only the classification result with highest probability;
the argument character extraction is optimized by adopting a multi-classification cross entropy function:
wherein For the number of argument characters->As a sign function, if the sample->The true category equals->Then 1 is taken, otherwise 0 is taken,representation of sample->The predicted outcome category is->Is a probability of (2).
5. The hotline allocation method based on knowledge-graph questions and answers as claimed in claim 1, wherein the step S3 specifically comprises:
entity disambiguation: based on an entity pair formed by any two homonymous entities in the relation pattern diagram, after the two entities are associated with the surrounding event text, inputting a pre-training language model to obtain a corresponding vector code;
calculating semantic similarity between two entities, combining entity constitution and collection with similarity greater than a certain threshold, and distinguishing homonymous entities with different references
wherein Representing a vector encoding of an entity, +.>Representing the encoding length of the entity vector;
coreference resolution: and coding the argument and the sentence where the argument is positioned by using a pre-training language model, and fusing the two generated vectors to form fusion coding for a role of an argument. Clustering the argument character fusion codes, calculating fusion vector similarity in an argument fusion vector set aggregated in the same cluster, and combining argument pairs with similarity larger than a certain threshold value similarly to form unified entity references:
wherein Representing a fusion encoding of argument roles, +.>，/>Coding vectors of argument characters and sentences corresponding thereto, respectively>Is a weight that can be learned.
6. The method for assigning hotline based on knowledge-graph questions and answers as claimed in claim 1, wherein in step S4, the retrieving the final sub-graph G using the sub-graph based on knowledge-graph comprises:
the three types of entities of the event trigger word, the event type and the event argument roles are taken as seed entities and recorded as a set；
Pair aggregationThe elements in (a) identify other entity sets which are possibly related to the event by adopting a PPR algorithm +.>；
For elements in an entity set identified by a PPR algorithm, calculating word vectors of the elements and sentence vectors of the events in an inner product mode, taking a calculation result as edge weights of the elements and seed entities to obtain subgraphs of three different types of entities related to trigger words, event types and event roles, and combining the same entities in the subgraphs of the entities of the same type to form three subgraphs corresponding to the three types;
and merging the same entities in the three sub-graphs of the same type to obtain a final sub-graph G.
7. The method for assigning hotline based on knowledge-graph questions and answers as claimed in claim 6, wherein in step S5, the extracting of event-related features from the final subgraph G comprises:
forming G1, G2, G3 and G into pairs, i.e., (G1, G), (G2, G) and (G3, G), respectively;
for a key input node V, defining a graph convolution operation of two vertex domains to aggregate local features and global features of the node, wherein the graph convolution operation adopts GCN to extract features, and a calculation formula is as follows:
wherein ,a is the adjacency matrix of the input graph, I is the identity matrix,>，/> and />Are all learnable parameters, < >>；
The GCN output is fused in a multi-layer aggregation mode, and the calculation formula is as follows:
wherein ,l is the number of layers of the graph roll;
computing two vector pairs for each graph pair，/> and />Global and local representing corresponding nodes of single-type sub-graph { Gi } and fusion sub-graph G respectivelyPartial feature representing vector, final representing vector based on attribute fusion to be node is carried out on vector +.>，
wherein ,, />d is the number of attention heads, which is a learnable parameter;
respectively calculating the expression vectors of three types of nodes of the event type, the trigger word and the event role, and performing splicing and fusion to obtain a graph expression vectorAnd all department nodes in the subgraph represent a vector set。
8. The method for assigning hotline based on knowledge-graph questions and answers as claimed in claim 1, wherein in step S4, the text search result is obtained by using the text search based on "triad" responsibilities, comprising:
vectorizing all sentences in a 'triangulated' responsibility retrieval corpus by using Glove word vectors to obtain vector representation sets of all sentences ；
Vectorizing an input event text by adopting GloVe to obtain an event text representation vector Ve, and calculating Ve and VeThe inner products of the middle elements are sorted according to the inner product result, and top-n related sentences are obtained;
tracing the top-n related sentences, calculating the score of the 'triad' responsibility texts corresponding to the n sentences, and selecting one 'triad' responsibility text with the highest score as a text retrieval result.
9. The method for hotline allocation based on knowledge-graph question-answering according to claim 8, wherein in step S5, performing event-related feature extraction on the text search result, and combining the event text features with the department entity set in the subgraph to generate a fused event vector set includes:
for the problem text representation, the input event text and the sub-department entity vector set retrieved from the knowledge graph are combinedThe elements in (a) are fused to be expressed as a problem, and the method comprises the following steps:
vectorizing input event text using BERT to obtain event representation vectors,
Where q is the text of the input event,is a BERT pre-training model;
vectorAnd department entity vector set->Each element in the file is fused to obtain the fusion event directionQuantity set，
wherein , and />Are all learnable parameters;
for the searched "third-party" responsibility text, the BERT is adopted to directly extract the characteristics of the text so as to obtain the characteristic expression vector V of the "third-party" responsibility text sanding ，
wherein ,is a BERT pre-training model,/->And (5) searching the result for the text.
10. The hotline allocation method based on knowledge-graph questions and answers as claimed in claim 1, wherein the step S6 specifically comprises:
representing vectors by graphFeature representation vector V with "triad" responsibility text sanding Splicing;
by matching problem vectorsAnd the probability prediction is carried out on the department entities obtained in the sub-graph retrieval with the spliced result vector,
wherein , and />Are all learnable parameters;
for fused problem vector setsAll elements in the department are predicted to obtain a prediction score set of the department；
Pair aggregation according to department output probabilityAnd sorting, and selecting the optimal department as a local and long hot line allocation department.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310601972.5A CN116795963A (en) | 2023-05-25 | 2023-05-25 | Hotline allocation method based on knowledge graph question and answer |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310601972.5A CN116795963A (en) | 2023-05-25 | 2023-05-25 | Hotline allocation method based on knowledge graph question and answer |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116795963A true CN116795963A (en) | 2023-09-22 |
Family
ID=88039049
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310601972.5A Pending CN116795963A (en) | 2023-05-25 | 2023-05-25 | Hotline allocation method based on knowledge graph question and answer |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116795963A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113111184A (en) * | 2021-04-27 | 2021-07-13 | 清华大学深圳国际研究生院 | Event detection method based on explicit event structure knowledge enhancement and terminal equipment |
| CN113919811A (en) * | 2021-10-15 | 2022-01-11 | 长三角信息智能创新研究院 | Hot line event distribution method based on strengthened correlation |
| CN113946684A (en) * | 2021-09-16 | 2022-01-18 | 国网四川省电力公司 | Electric power capital construction knowledge graph construction method |
| CN115982379A (en) * | 2022-12-20 | 2023-04-18 | 福州外语外贸学院 | User portrait construction method and system based on knowledge graph |
| CN116108169A (en) * | 2022-12-12 | 2023-05-12 | 长三角信息智能创新研究院 | Hot wire work order intelligent dispatching method based on knowledge graph |
-
2023
- 2023-05-25 CN CN202310601972.5A patent/CN116795963A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113111184A (en) * | 2021-04-27 | 2021-07-13 | 清华大学深圳国际研究生院 | Event detection method based on explicit event structure knowledge enhancement and terminal equipment |
| CN113946684A (en) * | 2021-09-16 | 2022-01-18 | 国网四川省电力公司 | Electric power capital construction knowledge graph construction method |
| CN113919811A (en) * | 2021-10-15 | 2022-01-11 | 长三角信息智能创新研究院 | Hot line event distribution method based on strengthened correlation |
| CN116108169A (en) * | 2022-12-12 | 2023-05-12 | 长三角信息智能创新研究院 | Hot wire work order intelligent dispatching method based on knowledge graph |
| CN115982379A (en) * | 2022-12-20 | 2023-04-18 | 福州外语外贸学院 | User portrait construction method and system based on knowledge graph |
Non-Patent Citations (9)
| Title |
|---|
| 中国人工智能学会: "《人工智能学科路线图》", 31 May 2022, 中国科学技术出版社, pages: 62 - 64 * |
| 刘宇: "《智能搜索和推荐系统 原理、算法与应用》", 31 January 2021, 机械工业出版社, pages: 50 - 52 * |
| 张文学: "《药品安全舆情的知识图谱获取方法与应用研究》", 31 December 2021, 华中科技大学出版社, pages: 107 - 110 * |
| 张朝阳: "《深入浅出 工业机器学习算法详解与实战》", 31 January 2020, 华中科技大学出版社, pages: 107 - 110 * |
| 潘峰;怀丽波;崔荣一;: "基于分布式图计算的学术论文推荐算法", 计算机应用研究, no. 06 * |
| 王然: "《人工智能路径下突发事件的模式识别与传媒预警》", 31 December 2022, 华中科技大学出版社, pages: 117 - 119 * |
| 胡盼盼: "《自然语言处理从入门到实战》", 30 April 2020, 中国铁道出版社, pages: 117 - 119 * |
| 邓佳佶: "《数智化转型 人工智能的金融实践》", 30 April 2022, 中国科学技术出版社, pages: 69 - 72 * |
| 陈钢等: "Joint Learning With BERT-GCN and Multi-Attention for EventText Classification and Event Assignment", 《IEEE ACCESS》, vol. 10, pages 27031 - 27040 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108573411B (en) | Mixed recommendation method based on deep emotion analysis and multi-source recommendation view fusion of user comments | |
| CN111190997B (en) | Question-answering system implementation method using neural network and machine learning ordering algorithm | |
| CN111291188B (en) | Intelligent information extraction method and system | |
| CN105393265A (en) | Active featuring in computer-human interactive learning | |
| CN107862561A (en) | A kind of method and apparatus that user-interest library is established based on picture attribute extraction | |
| CN112966091B (en) | Knowledge map recommendation system fusing entity information and heat | |
| CN110427480B (en) | Intelligent personalized text recommendation method and device and computer readable storage medium | |
| CN111767325B (en) | Multi-source data deep fusion method based on deep learning | |
| CN112597316A (en) | Interpretable reasoning question-answering method and device | |
| CN114332519A (en) | Image description generation method based on external triple and abstract relation | |
| CN114997288A (en) | Design resource association method | |
| CN117171333A (en) | Electric power file question-answering type intelligent retrieval method and system | |
| CN116561264A (en) | Knowledge graph-based intelligent question-answering system construction method | |
| CN114117000A (en) | Response method, device, equipment and storage medium | |
| Jin et al. | Textual content prediction via fuzzy attention neural network model without predefined knowledge | |
| CN117648984A (en) | Intelligent question-answering method and system based on domain knowledge graph | |
| CN117010373A (en) | Recommendation method for category and group to which asset management data of power equipment belong | |
| CN116628173A (en) | Intelligent customer service information generation system and method based on keyword extraction | |
| CN111259106A (en) | Relation extraction method combining neural network and feature calculation | |
| CN116795963A (en) | Hotline allocation method based on knowledge graph question and answer | |
| CN114022233A (en) | Novel commodity recommendation method | |
| CN113763084A (en) | Product recommendation processing method, device, equipment and storage medium | |
| Le-Minh et al. | Aspect-based sentiment analysis using mini-window locating attention for vietnamese e-commerce reviews | |
| CN117171413B (en) | Data processing system and method for digital collection management | |
| CN116702784B (en) | Entity linking method, entity linking device, computer equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |