CN117056451A - New energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement - Google Patents
New energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement Download PDFInfo
- Publication number
- CN117056451A CN117056451A CN202311024888.8A CN202311024888A CN117056451A CN 117056451 A CN117056451 A CN 117056451A CN 202311024888 A CN202311024888 A CN 202311024888A CN 117056451 A CN117056451 A CN 117056451A
- Authority
- CN
- China
- Prior art keywords
- text
- representation
- context
- complaint
- viewpoint
- 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
- 238000000605 extraction Methods 0.000 title claims abstract description 25
- 239000013598 vector Substances 0.000 claims abstract description 61
- 238000000034 method Methods 0.000 claims abstract description 34
- 230000007246 mechanism Effects 0.000 claims abstract description 23
- 238000012216 screening Methods 0.000 claims abstract description 7
- 238000002372 labelling Methods 0.000 claims abstract description 6
- 230000008569 process Effects 0.000 claims description 20
- 238000011176 pooling Methods 0.000 claims description 18
- 238000012549 training Methods 0.000 claims description 14
- 238000004364 calculation method Methods 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 10
- 238000001514 detection method Methods 0.000 claims description 8
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 claims description 7
- 230000002457 bidirectional effect Effects 0.000 claims description 6
- 238000012546 transfer Methods 0.000 claims description 5
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 230000035945 sensitivity Effects 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 238000012935 Averaging Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 claims description 2
- 230000007774 longterm Effects 0.000 claims description 2
- 238000012512 characterization method Methods 0.000 claims 1
- 230000009193 crawling Effects 0.000 claims 1
- 230000008447 perception Effects 0.000 claims 1
- 230000003044 adaptive effect Effects 0.000 abstract description 2
- 238000003058 natural language processing Methods 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 3
- 239000002131 composite material Substances 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000012938 design process Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/31—Indexing; Data structures therefor; Storage structures
- G06F16/313—Selection or weighting of terms for indexing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/284—Lexical analysis, e.g. tokenisation or collocates
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
- G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/01—Customer relationship services
- G06Q30/015—Providing customer assistance, e.g. assisting a customer within a business location or via helpdesk
- G06Q30/016—After-sales
Abstract
The invention discloses a new energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement, which belongs to the field of natural language processing and comprises the following steps: s1: acquiring new energy automobile complaint data, and performing part-of-speech tagging and text screening treatment on the new energy automobile complaint data; s2: acquiring text feature vectors based on an entity extraction method of a conditional random field, and using CRF to carry out sequence labeling to obtain an entity set; s3: fusing text features and theme features to obtain enhanced text context representation; s4: based on the tri-affine mechanism, the context is expressed as auxiliary information, the capability of the model to perceive the relation between the aspect words and the viewpoint words is enhanced, and therefore the adaptive aspect-viewpoint pairs are screened out. The method acquires the enhanced context representation by fusing the theme features and the text features, and takes the enhanced context representation as auxiliary information to improve the capability of the model to perceive the relation between the aspect words and the viewpoint words, thereby providing reliable data reference for improving products of new energy automobile merchants.
Description
Technical Field
The invention relates to the field of natural language processing, in particular to a new energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement.
Background
With the rapid development of new energy products, more and more people select new energy automobiles as travel tools. The new product brings convenience to people, and simultaneously causes a plurality of new problems, thereby causing a great deal of complaints of customers. Customer complaint information often exposes defects existing in the product design process, so that comments of customers on specific aspects of products in complaint texts are mined, reliable data reference can be provided for adjustment and decision making of enterprise operation, and enterprises are helped to design products meeting public demands.
Currently, research on extraction is lacking in terms of new energy automobile complaint texts, namely views. Wherein the research on aspect-viewpoint extraction mainly relies on word distance information or syntax dependency for relation detection, namely target word information. However, since the complaint text of the new energy automobile has the characteristics of large entity density, long sentence pattern and the like, the ideas and the method only consider word-level information, neglect the influence of the context of the text on word matching, and easily cause misjudgment on matching of aspect words and viewpoint words. Because the subject information of the complaint text is highly summarized to the text content and has low probability of being interfered by redundant information, the invention considers the fusion of the subject characteristics and the text characteristics to acquire enhanced text context representation, and takes the context representation as auxiliary information for relation detection to enhance the capability of the model to sense the relation between the aspect words and the viewpoint words so as to improve the extraction accuracy.
Disclosure of Invention
The invention aims to provide a new energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement, which is characterized by integrating theme features and text features as context representations and using the theme features and the text features as auxiliary information enhancement models to identify the relation among entities. And in the entity extraction stage, capturing context information through an entity extraction model to obtain text characteristics, and inputting the text characteristics into the CRF to obtain an entity set. The relation detection stage obtains enhanced context representation by fusing the theme features and the text features, and screens out adapted aspect-perspective pairs with the context representation as an aid by using a tri-affine mechanism.
The invention adopts the following technical scheme for realizing the purpose:
a new energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement comprises the following steps:
(1) Acquiring new energy automobile complaint data and preprocessing a text;
and acquiring new energy automobile complaint data, and performing part-of-speech tagging and text screening processing on the new energy automobile complaint data.
(2) Acquiring an aspect word set and a viewpoint word set in a complaint text based on an entity extraction method of a conditional random field;
and taking the text word vector as the input of a bidirectional long-short-time memory network, updating the weight of the feature vector through a multi-head attention mechanism, and using CRF to carry out sequence labeling so as to obtain an entity set.
(3) Fusing text features and theme features to obtain enhanced text context representation;
and taking the theme of the complaint text as the input of the BERT pre-training model, acquiring a theme feature representation, and fusing the theme feature and the text feature to acquire an enhanced text context representation.
(4) Based on a tri-affine mechanism, the context is expressed as auxiliary information, and the capability of the model for sensing the relation between aspect words and viewpoint words is enhanced;
the context representation, the aspect word representation and the viewpoint word representation are used as inputs of a three affine mechanism, word pair matching scores are calculated, and the adaptive aspect-viewpoint pairs are screened out.
The specific operations of the new energy automobile complaint data acquisition and text screening pretreatment in the step (1) are as follows:
(1.1) acquiring complaint data of the vehicle mass network platform through a crawler technology.
(1.2) removing stop words. Nonsensical phrases that have only structural functions but do not provide useful information are removed. For example, "principal purchases … … in a month of a year", "principal spends a plurality of money … …", "let continue to use … …", etc.
(1.3) screening data. And removing complaint texts with the number of words less than 30 words and more than 200 words in the crawl data, so as to ensure the balance of the data set.
And (1.4) marking the parts of speech of the new energy automobile complaint text by using an NITK part of speech marking tool, and marking the data by using a BIO marking method, thereby constructing a new energy automobile complaint text data set containing 4680 pieces of data.
In the step (2), based on entity relation extraction, the specific steps of obtaining the aspect word set and the viewpoint word set in the complaint text are as follows:
and (2.1) acquiring word vectors of each word of the text by using the BERT pre-training model, and inputting the acquired vector representations into a bidirectional long-short-time memory network to obtain characteristic representations of the text.
And (2.2) obtaining richer text feature representation while reducing redundant information interference in the new energy automobile complaint text by increasing weights occupied by aspect words and viewpoint words by considering multiple groups of attention scores through a multi-head attention mechanism.
The multi-head attention mechanism calculation process is as follows:
wherein Q is a query vector, K is a key vector, V is a value vector matrix, d k The term vector dimension Attention (Q, K, V) represents a single self-Attention score, head j Representing the calculation result of the j-th attention head, W is a weight matrixMHA (Q, K, V) represents the final attention score.
And inputting the feature vector representation of the obtained complaint text into the CRF for sequence labeling, marking out aspect words and viewpoint words, and obtaining an aspect word set and a viewpoint word set.
The global optimal solution is sought by calculating the joint probability distribution of the whole sequence, so that the problem that most of terms in the field of new energy automobiles are in a nested and composite structure is effectively solved. Let the current observation sequence be x= { x 1 ,x 2 ,x 3 ,..,x n Tag sequence y= { y } 1 ,y 2 ,y 3 ,..,y n The calculation process is as follows:
wherein W is 1 、W 2 Are trainable weight vectors, and F (y, x) is expressed as a combination of a transfer characteristic function and a state characteristic function.
In the step (3), based on fusion of text features and theme features, the specific steps for obtaining the enhanced text context representation are as follows:
(3.1) obtaining sentence representations of the subject using the BERT pre-training model as an encoder. By adding special marks [ CLS ] to the head and tail of the subject sentence]And [ SEP ]]As input, the hidden state representation of each position is obtained after the coding layerThe acquisition process is as follows:
wherein the CLS vector is the hidden state representation of the first position in the input sequence, and represents the sentence characteristic of the complaint subject, and is marked as h CLS 。
(3.2) in order to promote the relevance between the entity identification and the relation detection two modules, a pooling operation is adopted to select the part which has an important role in the extraction task. The new feature representation is then passed through a max pooling operation. The calculation process is as follows:
(3.2.1) representing the text feature as h t And carrying out averaging pooling to obtain a vector with the same dimension as the theme feature vector, and splicing the new text feature vector and the theme feature vector to obtain a new feature vector, so that semantic information of an entity can be better expressed by considering each feature vector uniformly. The calculation process is as follows:
wherein V is A And the average information of the text features is represented, and V represents the context representation after the topic features and the text features are spliced.
And (3.2.2) carrying out maximum pooling on the new spliced feature vectors to obtain a final context feature representation. The purpose of the maximum pooling is to reduce the sensitivity of the model to noise, enhance the characterizability of the context information, and furthermore use of the maximum pooling reduces the training parameters and increases the training speed for the whole model.
h t =max pooling(V)
Wherein h is t Representation of enhanced contextual feature representations
The context is expressed as auxiliary information based on a affine mechanism in the step (4), so that the capability of the model to perceive the relation between the aspect words and the viewpoint words is enhanced, and the process is as follows:
and respectively projecting the query vector, the key vector and the value vector into different low-dimensional spaces by adopting a three-affine mechanism, and calculating the association relation among the three by introducing additional linear transformation.
Pairing aspect items and viewpoint items pairwise to obtain an aspect-viewpoint pair set, and for each target word pair (S a ,S o ) Distribution relation labelThereby transformingInto separate classes of tasks.
Obtaining aspect vector representation a by position coding t ={a i ,a i+1 ,...,a i+q Sum-view vector representation o t ={o i ,o i+1 ,...,o i+p Finally, aspect representations, perspective representations, and contextual feature representations are taken as TriAffine inputs. The specific process is as follows:
wherein W represents a parameter that can be learned,and->Respectively representing an aspect representation and a perspective representation from an entity identification module, h c Representing context feature vectors, Φ a,o,t Representing the probability of word pair matching, r a,o The classification result is represented.
The context-enhanced new energy automobile complaint text aspect-viewpoint pair extraction method provided by the invention has the following advantages:
(1) According to the invention, the text vector is acquired through the long-short-term memory network, more accurate and comprehensive text characteristic representation can be obtained, and then the noise interference caused by redundant information in the complaint text is reduced through the multi-head attention mechanism, so that technical assistance is provided for subsequent relation detection.
(2) The method utilizes the method of fusing the theme characteristics and the text characteristics, so that the context semantics of the obtained complaint text are more accurate; meanwhile, a tri-affine mechanism is used to represent the association relationship between entity words as auxiliary computation by context, and the target aspect-viewpoint pair is obtained through a Sigmoid function.
Drawings
FIG. 1 is a flow chart of a new energy automobile complaint text aspect-perspective pair extraction method based on context enhancement;
FIG. 2 is a schematic diagram of a process for obtaining entity word sets by an entity extraction method based on a conditional random field;
FIG. 3 is a diagram of a context representation process for merging text features with subject feature acquisition enhancements;
fig. 4 is a process diagram of a relationship detection method based on a affine mechanism.
Detailed Description
The invention is further illustrated by the following examples.
Embodiment one: the invention provides a new energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement, which is shown in figure 1. The method comprises the following specific steps:
s1, acquiring a new energy automobile complaint text and preprocessing the text;
s1.1, firstly, acquiring a new energy automobile complaint text. 5000 new energy automobile related complaint data with different styles are crawled from a vehicle quality network through a crawler technology, and meanwhile, a theme sentence corresponding to each complaint text is obtained and stored in an excel file mode.
S1.2, removing stop words. Nonsensical phrases that have only structural functions but do not provide useful information are removed. For example, "principal purchases … … in a month of a year", "principal spends a plurality of money … …", "let continue to use … …", etc.
S1.3, screening data. And removing complaint texts with the word numbers of less than 30 and more than 200 in the crawl data, so as to ensure the balance of the data set.
S1.4, using NITK part-of-speech tagging tool to tag part-of-speech of new energy automobile complaints text to obtain vectorized representation V= [ V ] 1 ,v 2 ,...,v n ]And then marking the data by adopting a BIO marking method, thereby constructing a new energy automobile complaint text data set containing 4680 pieces of data.
S2, acquiring an aspect word set and a viewpoint word set in the complaint text by using an entity extraction method based on a conditional random field. The following description is made in connection with fig. 2:
s2.1 through a bidirectional long and short time memory network BiLSTM obtains text forward hidden stateAnd the backward hidden state->And splicing to obtain a complete sequence to capture long-term dependency in the sequence.
S2.1.1 complaint text data P= { C 1 ,C 2 ,...,C n N is greater than or equal to 2, each sentence contains m words, C p Represents the p-th sentence, w p,j Representing the jth word in the p-th sentence.
S2.1.2 the text word vector with semantic representation is obtained by using the BERT pre-training model as an encoder to encode sentences, and is expressed as follows:
W S ={w 1 ,w 2 ,w 3 ,...,w n },w k a vector representation representing the kth word.
S2.1.3, capturing the bidirectional information through the bidirectional long-short-time memory network can transmit the context information containing the candidate aspect items to the corresponding aspect items, and simultaneously transmit the context information containing the candidate aspect items to the corresponding aspect items so as to improve the interactivity between target entities.
S2.2, through a multi-head attention mechanism, by considering a plurality of groups of attention scores, the weights occupied by aspect words and viewpoint words are increased so as to obtain richer text feature representation while reducing redundant information interference in new energy automobile complaint texts.
S2.3, searching a global optimal solution by calculating joint probability distribution of the whole sequence, so that the problem that most of terms in the field of new energy automobiles are in a nested and composite structure is effectively solved. Let the current observation sequence be x= { x 1 ,x 2 ,x 3 ,..,x n Tag sequence y= { y } 1 ,y 2 ,y 3 ,..,y n The calculation process is as follows:
wherein W is 1 、W 2 Are trainable weight vectors, and F (y, x) is expressed as a combination of a transfer characteristic function and a state characteristic function.
S3, fusing the text features and the theme features to obtain the enhanced text context representation. The description is given in connection with fig. 3:
s3.1, a BERT pre-training model is adopted as an encoder to acquire sentence representations of the subject. By adding special marks [ CLS ] to the head and tail of the subject sentence]And [ SEP ]]As input, the hidden state representation t= [ h ] of each position is obtained after passing through the coding layer t 1 ,h t 2 ,h t 3 ...,h t n ]. The calculation process is as follows:
wherein the CLS vector is the hidden state representation of the first position in the input sequence, and represents the sentence characteristic of the complaint subject, and is marked as h CLS
S3.2, after the theme features are acquired, the text features and the theme features are fused, and the enhanced context representation is obtained.
S3.2.1, in order to promote the relevance between the entity identification and the relation detection two modules, the invention adopts pooling operation to select the part which has important effect on the extraction task. The new feature representation is then passed through a max pooling operation. The calculation process is as follows:
wherein V is A And the average information of the text features is represented, and V represents the context representation after the topic features and the text features are spliced.
S3.2.2 the new feature vectors obtained by the concatenation are maximally pooled to obtain the final context feature vector. The purpose of the maximum pooling is to reduce the sensitivity of the model to noise, enhance the characterizability of the context information, and furthermore use of the maximum pooling reduces the training parameters and increases the training speed for the whole model.
h t =max pooling(V)
Wherein h is t Representing the enhanced representation of contextual features.
S4, based on a tri-affine mechanism, the context is expressed as auxiliary information, and the capability of the model for sensing the relation between the aspect words and the viewpoint words is enhanced. The following description is made in connection with fig. 4:
s4.1, respectively projecting the query vector, the key vector and the value vector into different low-dimensional spaces by adopting a three-affine mechanism, and calculating the association relation among the three by introducing additional linear transformation. Obtaining aspect vector representation a by position coding t ={a i ,a i+1 ,...,a i+q Sum-view vector representation o t ={o i ,o i+1 ,...,o i+p }。
And S4.2, taking the aspect representation, the viewpoint representation and the context characteristic representation as TriAffine input. The specific process is as follows:
wherein W represents a parameter that can be learned,and->Respectively representing an aspect representation and a perspective representation from an entity identification module, h c Representing context feature vectors, Φ a,o,t Representing the probability of word pair matching, r a,o The classification result is represented.
Furthermore, the foregoing embodiments are merely illustrative of specific embodiments of the present invention and are not intended to be limiting, and it will be understood by those skilled in the art that some of the techniques may be equally substituted, and such modifications and substitutions are intended to be within the scope of the present invention.
Claims (6)
1. The context enhancement-based new energy automobile complaint text aspect-viewpoint pair extraction method is characterized by comprising the following steps of:
step 1: acquiring new energy automobile complaint data and preprocessing a text; acquiring new energy automobile complaint data, and performing part-of-speech tagging and text screening treatment on the new energy automobile complaint data;
step 2: acquiring an aspect word set and a viewpoint word set in a complaint text based on an entity relation extraction method of a conditional random field; the text word vector is used as the input of a bidirectional long-short-time memory network, the weight of the feature vector is updated through a multi-head attention mechanism, and CRF is used for sequence labeling, so that an entity set is obtained;
step 3: fusing text features and theme features to obtain enhanced text context representation; taking the subject of the complaint text as the input of the BERT pre-training model, acquiring a subject feature representation, and fusing the subject feature and the text feature to acquire an enhanced text context representation;
step 4: based on a three affine mechanism, using context representation as auxiliary information, and calculating the probability of matching between aspect words and viewpoint words; the context representation, the aspect word representation and the viewpoint word representation are used as inputs of a three affine mechanism, word pair matching scores are calculated, and correct aspect-viewpoint pairs are screened out.
2. The context-enhanced new energy automobile complaint text aspect-viewpoint pair extraction method according to claim 1, wherein step 1 comprises:
step 1.1, screening data; the nonsensical word groups which only have the structural function but do not provide useful information are removed, so that the interference of redundant information on the relation between words and viewpoint words in the perception of a model is reduced, complaint texts with the word numbers of less than 30 and more than 200 in crawling data are removed, and the balance of a data set is ensured;
step 1.2, part-of-speech tagging; and marking the parts of speech of the complaint text of the new energy automobile by using an NITK part of speech marking tool, and marking the data by using a BIO marking method, so that the model can learn more vivid parts of speech characteristics.
3. The method for extracting entity relation based on conditional random field according to claim 1, wherein the step 2 comprises:
acquiring text forward hidden state through bi-directional long-short-time memory network BiLSTMWith a backward hidden stateAnd splicing to obtain a complete sequence to capture long-term dependency in the sequence;
step 2.1 Consulting text data P= { C 1 ,C 2 ,...,C n N is greater than or equal to 2, each sentence contains m words, C p Represents the p-th sentence, w p,j Representing a j-th word in the p-th sentence;
and 2.2, adopting a BERT pre-training model as an encoder to encode sentences to obtain text word vectors with semantic characterization, wherein the text word vectors are expressed as follows:
W S ={w 1 ,w 2 ,w 3 ,...,w n },w k a vector representation representing a kth word;
step 2.3 capturing the bi-directional information through the bi-directional long-short-term memory network can transfer the context information containing the candidate aspect to the corresponding aspect, and simultaneously transfer the context information containing the candidate aspect to the corresponding aspect to promote the interactivity between the target entities.
4. The method for extracting entity relation based on conditional random field according to claim 1, wherein the step 2 comprises:
step 2.4, using a multi-head attention mechanism, and obtaining richer text feature representation while increasing weights occupied by aspect words and viewpoint words by considering multiple groups of attention scores so as to reduce redundant information interference in new energy automobile complaint texts; the attentiveness mechanism process is expressed as follows:
wherein Q is a query vector, K is a key vector, V is a value vector matrix, d k The term vector dimension Attention (Q, K, V) represents a single self-Attention score, head j Representing the calculation result of the j-th attention head, W being a weight matrix, MHA (Q, K, V) representing the final attention score;
step 2.5, inputting the obtained feature vector representation of the complaint text into a CRF for sequence labeling, labeling aspect words and viewpoint words, and obtaining an aspect word set and a viewpoint word set; let the current observation sequence be x= { x 1 ,x 2 ,x 3 ,..,x n Tag sequence y= { y } 1 ,y 2 ,y 3 ,..,y n The calculation process is as follows:
wherein W is 1 、W 2 Are trainable weight vectors, and F (y, x) is expressed as a combination of a transfer characteristic function and a state characteristic function.
5. A method of fusing text features with topic features to obtain an enhanced text context representation as in claim 1, wherein step 3 comprises:
step 3.1 Using BERT Pre-training model as encoder to obtain sentence representation of subject by adding special tags [ CLS ] to the head and tail of subject sentence respectively]And [ SEP ]]As input, the hidden state representation of each position is obtained after the coding layerThe acquisition process is as follows:
wherein the CLS vector is the hidden state representation of the first position in the input sequence, and represents the sentence characteristic of the complaint subject, and is marked as h CLS ;
Step 3.2 in order to promote the relevance between the entity identification and the relation detection two modules, the invention adopts pooling operation to select the part which has important effect on the extraction task, and then the new characteristic representation is processed through the maximum pooling operation, and the process is as follows:
step 3.2.1 representing the text feature as h t The method comprises the steps of carrying out averaging pooling to obtain vectors with the same dimension as the topic feature representation, and splicing new text feature vectors and topic feature vectors to obtain new feature representation, wherein the purpose is to uniformly consider each feature vector and better express semantic information of an entity, and the calculation process is as follows:
wherein V is A Average information representing text features, V representing a context representation after the subject features are spliced with the text features;
step 3.2.2 maximize pooling of the new feature vectors spliced to obtain final context feature representation, the purpose of the maximize pooling is to reduce the sensitivity of the model to noise, enhance the representation capability of the context information, and furthermore, the use of the maximize pooling reduces the training parameters and improves the training speed for the entire model, wherein h t Representing the enhanced representation of contextual features.
h t =maxpooling(V) (5)。
6. The method for calculating the probability of matching between terms and terms based on a three affine mechanism and using context representation as side information according to claim 1, wherein the step 4 comprises:
step 4.1, respectively projecting the query vector, the key vector and the value vector into different low-dimensional spaces by adopting a three-affine mechanism, and calculating the association relation among the three by introducing additional linear transformation; obtaining aspect vector representation a by position coding t ={a i ,a i+1 ,...,a i+q Sum-view vector representation o t ={o i ,o i+1 ,...,o i+p };
Step 4.2 takes aspect representation, viewpoint representation and context feature representation as input of a affine mechanism, and the specific process is as follows:
wherein W represents a parameter that can be learned,and->Respectively representing an aspect representation and a perspective representation from an entity identification module, h c Representing context feature vectors, Φ a,o,t Probability score, r, representing word pair match a,o The classification result is represented.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311024888.8A CN117056451A (en) | 2023-08-12 | 2023-08-12 | New energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311024888.8A CN117056451A (en) | 2023-08-12 | 2023-08-12 | New energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement |
Publications (1)
Publication Number | Publication Date |
---|---|
CN117056451A true CN117056451A (en) | 2023-11-14 |
Family
ID=88660340
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202311024888.8A Pending CN117056451A (en) | 2023-08-12 | 2023-08-12 | New energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117056451A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117520786A (en) * | 2024-01-03 | 2024-02-06 | 卓世科技(海南)有限公司 | Large language model construction method based on NLP and cyclic neural network |
-
2023
- 2023-08-12 CN CN202311024888.8A patent/CN117056451A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117520786A (en) * | 2024-01-03 | 2024-02-06 | 卓世科技(海南)有限公司 | Large language model construction method based on NLP and cyclic neural network |
CN117520786B (en) * | 2024-01-03 | 2024-04-02 | 卓世科技(海南)有限公司 | Large language model construction method based on NLP and cyclic neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112115238B (en) | Question-answering method and system based on BERT and knowledge base | |
CN112069811B (en) | Electronic text event extraction method with multi-task interaction enhancement | |
CN110633409A (en) | Rule and deep learning fused automobile news event extraction method | |
CN110866542B (en) | Depth representation learning method based on feature controllable fusion | |
WO2018028077A1 (en) | Deep learning based method and device for chinese semantics analysis | |
CN108932342A (en) | A kind of method of semantic matches, the learning method of model and server | |
CN110298043B (en) | Vehicle named entity identification method and system | |
CN112163089B (en) | High-technology text classification method and system integrating named entity recognition | |
CN114548099B (en) | Method for extracting and detecting aspect words and aspect categories jointly based on multitasking framework | |
CN113743119B (en) | Chinese named entity recognition module, method and device and electronic equipment | |
CN114648031A (en) | Text aspect level emotion recognition method based on bidirectional LSTM and multi-head attention mechanism | |
CN111651973A (en) | Text matching method based on syntax perception | |
CN111400478A (en) | Knowledge graph question-answering method and device based on word embedding matching technology | |
CN110297986A (en) | A kind of Sentiment orientation analysis method of hot microblog topic | |
CN117056451A (en) | New energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement | |
CN114818717A (en) | Chinese named entity recognition method and system fusing vocabulary and syntax information | |
CN112800184A (en) | Short text comment emotion analysis method based on Target-Aspect-Opinion joint extraction | |
CN113297375B (en) | Document classification method, system, device and storage medium based on label | |
Tarride et al. | A comparative study of information extraction strategies using an attention-based neural network | |
CN111309849B (en) | Fine-grained value information extraction method based on joint learning model | |
CN112862569A (en) | Product appearance style evaluation method and system based on image and text multi-modal data | |
CN116958677A (en) | Internet short video classification method based on multi-mode big data | |
CN116758558A (en) | Cross-modal generation countermeasure network-based image-text emotion classification method and system | |
CN116562291A (en) | Chinese nested named entity recognition method based on boundary detection | |
CN115906824A (en) | Text fine-grained emotion analysis method, system, medium and computing equipment |
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 |