CN115392233A - Intelligent collection prompting auxiliary system based on central sentence recognition and Bert intention recognition - Google Patents
Intelligent collection prompting auxiliary system based on central sentence recognition and Bert intention recognition Download PDFInfo
- Publication number
- CN115392233A CN115392233A CN202211018187.9A CN202211018187A CN115392233A CN 115392233 A CN115392233 A CN 115392233A CN 202211018187 A CN202211018187 A CN 202211018187A CN 115392233 A CN115392233 A CN 115392233A
- Authority
- CN
- China
- Prior art keywords
- sentence
- intention
- model
- recognition
- client
- 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
- 239000013598 vector Substances 0.000 claims description 46
- 238000012549 training Methods 0.000 claims description 26
- 238000001514 detection method Methods 0.000 claims description 22
- 238000000034 method Methods 0.000 claims description 20
- 238000011068 loading method Methods 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 9
- 230000002457 bidirectional effect Effects 0.000 claims description 7
- 238000011049 filling Methods 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 5
- 238000012360 testing method Methods 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000012216 screening Methods 0.000 claims description 4
- 238000012544 monitoring process Methods 0.000 claims description 3
- 230000001502 supplementing effect Effects 0.000 claims description 3
- 238000002372 labelling Methods 0.000 claims description 2
- 230000011218 segmentation Effects 0.000 claims description 2
- 238000003306 harvesting Methods 0.000 claims 2
- 230000000694 effects Effects 0.000 abstract description 7
- 238000001914 filtration Methods 0.000 description 7
- 230000006870 function Effects 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000008451 emotion Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000036651 mood Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
Images
Classifications
-
- 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
- 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
- G06F16/35—Clustering; Classification
- G06F16/355—Class or cluster creation or modification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
-
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Machine Translation (AREA)
Abstract
The invention relates to the technical field of intelligent collection urging, in particular to an intelligent collection urging auxiliary system based on central sentence recognition and Bert intention recognition. The technical scheme is provided mainly aiming at the problems of low call-in efficiency and poor call-in effect of the existing telephone: the system comprises a call prompting system, a call prompting system and a call prompting system, wherein the call prompting system is an outbound system, is continuously communicated with a client through the interface and simultaneously displays labels of a call prompting person and the client; the intelligent collection assisting system comprises a standard sentence management module, an intention management module, a knowledge base management module, a label management module, an intention association module and a session summary management module. The invention is an assisted support system for collection, which can improve the collection transaction rate, automatically fill in the client label and automatically generate the client session summary, improve the client experience and increase the collection revenue; the intelligent payment prompting device is mainly applied to intelligent payment prompting.
Description
Technical Field
The invention relates to the technical field of intelligent collection urging, in particular to an intelligent collection urging auxiliary system based on central sentence recognition and Bert intention recognition.
Background
With the development of artificial intelligence technology, intelligent technologies such as text extraction, intention recognition, dialectical matching, text generation and the like provide a new engine for collection promotion. And in the process of customer collection, extracting the text central sentence answered by the customer. In the processes of information transmission, payment negotiation and the like in the collection urging process, various intentions of the client are extracted, and the reasons why the payment cannot be made are described by the client. In order to relieve the mood of customers and achieve the purpose of hastening receipts, appropriate conversational answers need to be displayed in real time according to different purposes.
The call collection is the most main collection method, and when the collector communicates with the client, if the emotion of the client fluctuates greatly, the experience of the collector is relatively short, and the collection fails under the condition that the call operation cannot be effectively controlled. After the collection is finished, the collector needs to efficiently fill in the label describing the client session and write the summary of the collection session, so that the client label can be automatically filled in and the summary of the client session can be automatically generated in order to improve the collection efficiency. In order to improve the customer experience, the income of hastening is increased; we propose an intelligent hastening assistance system based on central sentence recognition and Bert intent recognition.
Disclosure of Invention
The invention aims to provide an intelligent call collection assisting system based on central sentence recognition and Bert intention recognition, aiming at the problems of low call collection efficiency and poor call collection effect in the prior art.
The technical scheme of the invention is as follows: the intelligent collection assisting system based on the central sentence recognition and the Bert intention recognition is characterized in that the collection assisting system is an outbound system, continuously communicates with the client through the interface, and simultaneously displays labels of a collector and the client; the system comprises a standard sentence management module, an intention management module, a knowledge base management module, a label management module, an intention association module and a session summary management module;
the standard sentence management module displays standard sentences with different subdivision intentions;
the intention management comprises model management and regular management;
the knowledge base management module is used for carrying out preferred language configuration on different intentions;
the label management module is used for modifying the labels automatically filled through intention association and manually filling the content of the prefabricated labels according to the conversation content;
the intention association module is used for increasing, deleting, improving and checking the association relationship between the intention and the client prefabricated label and importing the association relationship between the intention and the client prefabricated label;
the session summary management module is used for generating summary of the collection session.
Preferably, the model management comprises model training, model testing and model publishing; the model management is used for monitoring the state of the model and calling the model;
the regular management comprises word slots, rules and model configuration; the regular management is used for configuring sentence patterns and word slots for model missing detection; configuring the content of word slots in the sentence pattern; configuring violation rules; and configuring the compliance rules.
Preferably, the main contents of the knowledge base management module comprise the selection of collection case types, intention category setting, knowledge base coding setting, knowledge base name display and display sequence setting.
Preferably, the label management module further comprises a label of whether to resell, provide other contact information, whether a repayment intention exists, a repayment scheme requirement, a delay period number, a complaint point, a risk point, a hang-up node and whether to own.
Preferably, the intention association management module is further configured to set an association relationship between the intention and the customer preformed label.
Preferably, the session summary module comprises case type selection, session template selection and session template name; the session summary module is also used for adding and deleting session models; setting sentence patterns and labels of the conversation template; modifying the automatically generated session summary; the session summary is filled in manually.
The invention also provides an implementation method of the intelligent collection assisting system based on the central sentence recognition and the Bert intention recognition, which comprises the following processing steps:
s1, marking through customer data in a collection urging system, and defining the category of a collection urging text;
s2, clustering texts by means of word vectors of the texts, and subdividing text categories;
s3, generating standard sentences classified in detail through the TextRank;
s4, after the standard sentence is obtained, judging whether the standard sentence is a text center sentence on line;
s5, training a Bert model, supplementing an optimization model of missed inspection rules, violation rules and compliance rules, judging whether the model is a text central sentence or not, and loading the model to perform intention identification on the text central sentence;
s6, automatically displaying the collection prompting and preferred dialect;
and S7, automatically generating a session summary.
Preferably, the step S1 includes collecting customer hasty history data; dividing the intention of the client into a plurality of intentions of chatting, answering inconveniently and agreeing to repayment; manually marking data according to the historical data which is received by the client and the defined intention;
s2, extracting the intents of the same type in the labeling data; carrying out word segmentation on each sentence of the intention, and removing stop words and word vectors to obtain a sentence vector; carrying out text clustering by using each sentence vector and Kmeans to obtain a fine classification of the intended question; selecting appropriate values of the detailed classes according to the distribution condition of the text;
the S3 comprises sentence vectors aiming at each subdivision intention category, and the TextRank performs sentence extraction; using a TextRank algorithm, taking the text data as nodes of the graph, and taking the similarity between the text and the text as an adjacency matrix; screening out 10 standard question sentences which can represent the detailed classification most by utilizing the principle that the value of a certain sentence is determined by the value of each sentence linked to the sentence and the corresponding weight; storing the sentence vectors of the standard sentences into a database;
and S4, carrying out model discrimination and carrying out no model discrimination: model discrimination is not performed: the dialogue content of the client contains keywords as a text central sentence; and (3) judging a model: calling standard question sentences of the fine classification, calculating similarity between sentence vectors of the sentences and the standard question sentences, and if the standard question sentences higher than a threshold value exist, determining the sentences as text center sentences;
the S5 training model and preferably comprises: word vectors, word embedding and a bidirectional long and short memory network model; word vectors, word embedding and a bidirectional long and short memory network model; a fine tuning depth model based on AlBert pre-training; fine tuning depth models based on Bert pre-training;
the supplementary missing detection rule and the illegal compliance rule comprise the following steps: determining the category of the sentence pattern using rule with obvious rules, namely the missing detection rule;
after the model is judged, the rule violation and the rule compliance of experience are judged, and a model result, namely rule violation and rule compliance, is optimized;
configuration of default rules: if hit, a reservation is required; miss, need to be excluded;
and (3) compliance rule configuration: if hit, it needs to be excluded; miss, need to be reserved;
loading a model, wherein a text center sentence meaning graph identifies whether the answer of a client is a text center sentence or not in the collection urging process, if so, automatically loading an optimal model, a missing detection rule and a violation compliance rule, and judging the intention type;
the S6 is used for configuring intention corresponding dialogues in a knowledge base configuration module; after the intention label is detected, automatically displaying the preferred conversation collection urging operation;
the S7 is used for configuring a session template required by generating a session summary according to the label; the client label and the conversation template automatically generate a conversation summary.
Compared with the prior art, the invention has the following beneficial technical effects:
1. the model training comprises character vectors, word embedding and a bidirectional long and short memory network model; pre-training a depth model of word features based on AlBert; fine tuning depth models based on Bert pre-training; and optimizing a model result by combining a supplement missing detection rule; the accurate recognition of the central sentence of the text is realized, and the effect of intelligently urging to receive and recognize the central sentence of the sentence is improved;
2. the system of the invention prefabricates some labels for customers, including whether the labels are resold, other contact ways are provided, whether repayment intention exists, repayment scheme requirements and the like, can set the associated display of the intention and the labels through intention associated management, can automatically fill the label contents after identifying the intention, and also provides the function of manually filling the labels; after the conversation between the collector and the client is finished, comprehensive description needs to be carried out on the conversation, the system provides a template for generating the conversation by the label, the summary of the conversation can be automatically generated, and meanwhile, the function of manual filling is also provided;
3. in conclusion, the invention is an auxiliary support system for collection prompting, which can improve the collection prompting rate, automatically fill in the client label and automatically generate the client session summary, improve the client experience and increase the collection prompting income.
Drawings
FIG. 1 is a schematic block diagram of an intelligent hastening system of the present invention;
FIG. 2 is a schematic diagram of a Bert model training structure of the present invention;
FIG. 3 is a flow chart of the model and missed detection rule of the present invention;
FIG. 4 is a flow chart of an implementation method of the intelligent collection system of the present invention.
Detailed Description
The technical solution of the present invention is further explained with reference to the accompanying drawings and specific embodiments.
Example one
As shown in fig. 1-3, the intelligent collection-urging auxiliary system based on central sentence recognition and Bert intention recognition provided by the present invention, wherein the collection-urging system is an outbound system, and continuously communicates with the customer through the interface, and simultaneously displays labels of the collector and the customer; the system comprises a standard sentence management module, an intention management module, a knowledge base management module, a label management module, an intention association module and a session summary management module;
the standard sentence management module displays standard sentences with different subdivision intentions; mainly comprises the steps of collecting case type selection, intention category setting, subdivided intention category setting and standard sentence display;
intention management includes model management and canonical management; the model management comprises model training, model testing and model publishing; model management is used for monitoring the state of the model and calling the model;
the regular management comprises word slot, rule and model configuration; regular management is used for configuring sentence patterns and word slots of model missing detection; configuring the content of word slots in the sentence pattern; configuring violation rules; and configuring the compliance rules.
The knowledge base management module is used for carrying out preferred language configuration on different intents; the main contents of the knowledge base management module comprise the selection of collection case types, the setting of intention categories, the setting of knowledge base codes, the display of knowledge base names and the setting of display sequences.
The label management module is used for modifying the labels automatically filled through intention association and manually filling the content of the prefabricated labels according to the conversation content; the label management module also comprises whether to resell, provide other contact information, whether a repayment intention exists, a repayment scheme requirement, a delay period number, a complaint point, a risk point, a hang-up node and whether a label is available.
The intention association module is used for increasing, deleting, improving and checking the association relationship between the intention and the client prefabricated label and importing the association relationship between the intention and the client prefabricated label; the intention association management module is also used for setting the association relationship between the intention and the client pre-made label.
The session summary management module is used for generating summary of the receiving session; the session summary module comprises case type selection, session template selection and session template name; the session summary module is also used for adding and deleting session models; setting sentence patterns and labels of the conversation template; modifying the automatically generated session summary; the session summary is filled in manually.
Example two
As shown in fig. 1 to 4, compared with the first embodiment, the implementation method of the intelligent collection prompting auxiliary system provided by the present invention based on central sentence recognition and Bert intention recognition includes the following processing steps:
s1, marking through customer data in a collection urging system, and defining the category of a collection urging text;
s2, clustering texts by means of word vectors of the texts, and subdividing text categories;
s3, generating standard sentences classified in detail through the TextRank;
s4, judging whether the standard sentence is a text center sentence on line or not after the standard sentence is obtained;
s5, training a Bert model, supplementing a missing detection rule, a violation rule and a compliance rule optimization model, judging whether the model is a text center sentence or not, and loading the model to perform intention identification on the text center sentence;
s6, automatically displaying the collection prompting and preferred dialect;
and S7, automatically generating a session summary.
In this embodiment, S1 includes collecting customer collection history data; dividing the intention of the client into a plurality of intentions of chatting, answering inconveniently and agreeing to repayment; manually marking data according to historical data which is urged to be received by a client and a defined intention;
s2, extracting the same type of intents in the annotation data; segmenting words of each sentence of the intention, and removing stop words and word vectors to obtain sentence vectors; carrying out text clustering by using each sentence vector and Kmeans to obtain a fine classification of the intended question; selecting appropriate values of the detailed classes according to the distribution condition of the text;
s3, sentence vectors of each subdivision intention category are extracted by the TextRank; using a TextRank algorithm, taking the text data as nodes of the graph, and taking the similarity between the text and the text as an adjacency matrix; screening out 10 standard question sentences which can represent the detailed classification most by using the principle that the value of a certain sentence is determined by the value of each sentence linked to the sentence and the corresponding weight; storing the sentence vectors of the standard sentences into a database;
s4, carrying out model discrimination and carrying out no model discrimination: model discrimination is not performed: the dialogue content of the client contains keywords and is used as a text central sentence; and (3) judging a model: calling standard question sentences of the fine classification, calculating similarity between sentence vectors of the sentences and the standard question sentences, and if the standard question sentences higher than a threshold value exist, determining the sentences as text center sentences;
s5 training the model and preferably comprises: word vectors, word embedding and a bidirectional long and short memory network model; word vectors, word embedding and a bidirectional long and short memory network model; a fine tuning depth model based on AlBert pre-training; fine tuning depth models based on Bert pre-training;
the supplementary missing detection rule and the illegal compliance rule comprise the following steps: determining the category of the sentence pattern using rule with obvious rules, namely the missing detection rule; and under the conditions of large number of samples and good quality, the model effect is better. However, under the condition of poor sample quality, the model has missing detection, and missing detection rules need to be supplemented to make up for model loss; after the model is judged, the misjudgment result of the model is made up, the rule violation and the rule compliance judgment of experience are needed, and the model result is optimized;
after the model is judged, the rule violation and the rule compliance of experience are judged, and a model result, namely rule violation and rule compliance, is optimized;
configuration of default rules: if hit, a reservation is needed; miss, need to be excluded;
and (3) compliance rule configuration: if hit, it needs to be excluded; miss, need to be reserved;
configuring default rules and compliance rules for each category with obvious rules;
in some specific categories, rules are preferred over models;
therefore, models and rules need to be lifted simultaneously, and the recognition effect is improved;
loading a model, automatically identifying whether the answer of a client is a text center sentence or not in the process of identifying and urging collection of the text center sentence and automatically loading an optimal model, a missing detection rule and a violation compliance rule to judge the intention category; if the rule is not hit, entering the published list model for prediction;
entering a label corresponding to the acquisition rule when the rule hits;
predicting a published model and judging whether to chat;
chatting, and intention is not processed;
non-chatting, and taking the intention that the model prediction score is greater than the threshold score;
querying a mapping label corresponding to the intention;
judging whether the label has a configuration filtering rule violation or not;
if the configuration exists, obtaining a filtering rule violation rule corresponding to the label;
judging an identification result;
hit the rule, judge whether the label has and disposes and filters the compliance rule;
configuring, namely acquiring a filtering compliance rule corresponding to the label;
judging an identification result;
hit rule, the tag does not output;
a miss rule, outputting the tag;
s6, configuring intention corresponding dialogs in a knowledge base configuration module; after the intention label is detected, automatically displaying the preferred conversation collection urging operation;
s7, a session template required by the generation of a session summary according to the label is configured; the client label and the conversation template automatically generate a conversation summary.
In this embodiment, a text-centric sentence is discriminated:
before the intention recognition, whether the sentence needs the intention recognition or not, namely whether the sentence is a text center sentence or not, needs to be judged.
It can be seen from the labeled data that if the sentence pattern difference in the intention is not large, the central sentence of the text can be obtained by directly performing the sentence similarity calculation.
If the intended sentence pattern type is different greatly. The sentences in the intention need to be subdivided, and a certain number of sentences are taken out as standard question sentences for the subdivision intention to judge whether the sentences are text central sentences or not.
The method comprises the following specific steps:
firstly, segmenting each sentence of the intention, removing stop words and forming a word list;
loading word2vec to obtain word vectors of words, wherein the dimensionality is 300 dimensions, and obtaining a matrix of the word vectors;
the importance degree of the word in the sentence can be obtained through tf-idf weighting, if the importance degree is higher, the word vector proportion of the word is higher when the average word vector is solved by using the part of speech quantity in the subsequent process, and the sentence vector is solved by using the weighted average;
summing and averaging the weighted word vector matrixes to obtain a sentence vector of the sentence;
performing Kmeans text clustering by using the sentence vector of each sentence;
observing the influence of the K values on the distribution result, and selecting proper K values as the number of the fine categories;
mutually calculating the cosine similarity of the sentences of each fine classification, and storing the cosine similarity in a matrix;
using a TextRank algorithm, taking the text data as nodes of a graph, and taking the similarity between texts as an adjacency matrix;
screening out 10 standard question sentences which can represent the detailed classification most by utilizing the principle that the value of a certain sentence is determined by the value of each sentence linked to the sentence and the corresponding weight;
and (3) judging a text center sentence:
after a standard question sentence is obtained by using text clustering and TextRank key sentence extraction, the similarity of the sentence is stored in a database as a model result;
the method comprises the following steps of obtaining a text center sentence in two modes, wherein one mode is to perform model discrimination, and the other mode is not to perform model discrimination;
carrying out model discrimination, calling standard question sentences of the fine classification, calculating similarity between sentence vectors of the sentences and the standard question sentences, and if the similarity higher than a threshold value exists, determining the sentences as text center sentences;
model discrimination is not carried out, and the dialogue content of the client contains keywords and can be used as a text central sentence;
intent recognition model and preferences:
the method comprises the following steps that various models are provided for intention recognition, firstly, a word vector + embedding + biLSTM model is selected, 5000 common Chinese characters are selected as the model, the vector dimension of an embedding word is 64 dimensions, and 3 layers of hidden layers are used as model training parameters;
because the word vector splits the semantics among the words, a fasttext word vector + embedding + biLSTM model is selected, 25000 commonly used words are selected in the model, the dimension of the embedding word vector is 300 dimensions, and 1 layer of hidden layers are used as model training parameters;
based on a tiny tuning depth model of AlBert, 4 layers of Encoders are stacked, each Encoder uses 12-head attention, the number of hidden layers is 312, and a linear + softmax network is connected externally;
in the training model selection and optimization process, considering that a transform is adopted by the Bert and the bottom layer of the improved model thereof, the Bert can generate a dynamic word vector in the fine tuning model to replace a static word vector of word2 vec; the bottom layer transform adopts a multi-head self-attention mechanism, and position coding is added to replace RNN; the self-attention model replaces a seq2seq coding model, so that the calculation efficiency is improved; a multi-head self-attention mechanism adopted by the transformer optimizes the resolution problem of the index;
optimally selecting a Bert + BilSTM model for training, and referring to FIG. 2, the method specifically comprises the following steps:
dividing text data into a training set test set, 80% training and 20% testing;
loading a trained Bert model, which comprises 18000 word vectors;
the pre-training model parameters are as follows:
12-layer stacked encoders, each using a 12-head attention, number of hidden layers 768;
connecting an LSTM network behind the Bert as a whole fine tuning model;
LSTM network parameters are as follows:
hidden layer 2, word vector 300, hidden layer neurons 128
Softmax layer with dimension of 1000, 43
Predicting characters by using Mask of the characters, predicting intervals between sentences by using [ CLS ] and [ SEP ], and synthesizing the two tasks to carry out Bert model fine tuning;
performing iterative optimization on the model through the difference between the prediction sample and the real sample to finally obtain an optimal model;
in order to improve the overall recognition effect, models and rules are required to be improved. And (4) configuring a missing detection rule for part of categories, configuring violation rules of the intention label, if the rule is hit, enabling the text to belong to the intention label, and entering the next judgment. Contract rules for the intent tag are configured, and if the rules are hit, the text does not belong to the intent tag.
Referring to fig. 3, the model and the missing detection process of the embodiment are as follows:
if the rule is not hit, predicting a corresponding label by using the published model;
the rule is hit, and the corresponding label is judged by using the rule;
predicting whether the published model is chatty or not;
if the intention belongs to the chatting category, the intention jumps out and the judgment processing is not carried out;
if the model does not belong to the non-chat category, the intention that the model prediction score is larger than the threshold score is taken as a judgment intention;
inquiring the mapping label corresponding to the intention;
judging whether the label has a configuration filtering rule violation rule or not;
if the rule violation is configured, acquiring a filtering rule violation corresponding to the label;
judging rule violation rules and judging identification results;
if the rule is hit, judging whether the tag has a configuration filtering compliance rule;
if the configuration exists, acquiring a filtering compliance rule corresponding to the label;
judging a compliance rule and judging a recognition result;
if the rule is hit, the tag is not output;
if the rule is not hit, the tag is output;
additional branching, as shown in the flow chart of FIG. 3
The average accuracy of models and class detection of missed detection reaches 84.6 percent;
after the intention recognition is completed, the preferred configuration of the knowledge base, namely the dialect base, can be carried out, and the preferred answer configuration is carried out on the intentions one by one. The preferred response may be displayed in real time during the on-line collection.
In order to describe the conversation behavior of the client, the system presets some labels for the client, including whether the client is resell or not, other contact ways are provided, whether a repayment intention exists or not, a repayment scheme requirement and the like, the associated display of the intention and the labels can be set through intention association management, after the intention is identified, the label content can be automatically filled, and meanwhile, the function of manually filling the labels is also provided.
After the conversation between the collector and the client is finished, comprehensive description needs to be carried out on the conversation, the system provides a template for generating the conversation by the label, the summary of the conversation can be automatically generated, and meanwhile, the function of manual filling is also provided.
The above embodiments are merely some preferred embodiments of the present invention, and those skilled in the art can make various alternative modifications and combinations of the above embodiments based on the technical solution of the present invention and the related teaching of the above embodiments.
Claims (8)
1. The intelligent collection-urging auxiliary system based on the central sentence recognition and the Bert intention recognition can realize the real-time outbound analysis of the seat end based on the semantics, and construct the client seat dialogue data and collection-urging strategy analysis closed loop; the method is characterized in that: the system comprises a standard sentence management module, an intention management module, a knowledge base management module, a label management module, an intention association module and a session summary management module;
the standard sentence management module displays standard sentences with different subdivision intentions;
the intention management comprises model management and canonical management;
the knowledge base management module is used for carrying out preferred language configuration on different intents;
the label management module is used for modifying the labels automatically filled through intention association and manually filling the content of the prefabricated labels according to the conversation content;
the intention association module is used for increasing, deleting, improving and checking the association relationship between the intention and the client prefabricated label and importing the association relationship between the intention and the client prefabricated label;
the session summary management module is used for generating summary of the session.
2. The intelligent collection assisting system based on central sentence recognition and Bert intention recognition as claimed in claim 1, wherein the model management comprises model training, model testing and model publishing; the model management is used for monitoring the state of the model and calling the model;
the regular management comprises word slots, rules and model configuration; the regular management is used for configuring sentence patterns and word slots for model missing detection; configuring the content of word slots in the sentence pattern; configuring violation rules; and configuring the compliance rules.
3. The intelligent collection assisting system based on central sentence recognition and Bert intention recognition as claimed in claim 1, wherein the knowledge base management module comprises a selection of collection case type, an intention category setting, a knowledge base coding setting, a knowledge base name display and a display sequence setting.
4. The intelligent collection assisting system based on central sentence recognition and Bert intention recognition as claimed in claim 1, wherein the tag management module further comprises a tag of whether to resell, provide other contact means, whether to have a repayment intention, a repayment scheme requirement, a delay period number, a complaint point, a risk point, a hang-up node, and whether to be self.
5. The intelligent collection assisting system based on central sentence recognition and Bert intention recognition as claimed in claim 1, wherein the intention association management module is further configured to set an association relationship between the intention and a client pre-made label.
6. The intelligent collection assisting system based on central sentence recognition and Bert intention recognition as claimed in claim 1, wherein the session summary module comprises case type selection, session template selection and session template name; the session summary module is also used for adding and deleting session models; setting sentence patterns and labels of the conversation template; modifying the automatically generated session summary; the session summary is filled in manually.
7. An implementation method of the intelligent hasty harvesting auxiliary system based on central sentence recognition and Bert intention recognition according to any one of claims 1-6, characterized by comprising the following processing steps:
s1, marking through customer data in a collection urging system, and defining the category of a collection urging text;
s2, clustering texts by means of word vectors of the texts, and subdividing text categories;
s3, generating standard sentences classified in detail through the TextRank;
s4, after the standard sentence is obtained, judging whether the standard sentence is a text center sentence on line;
s5, training a Bert model, supplementing a missing detection rule, a violation rule and a compliance rule optimization model, judging whether the model is a text center sentence or not, and loading the model to perform intention identification on the text center sentence;
s6, automatically displaying the preferred conversation of hastening and harvesting;
and S7, automatically generating a session summary.
8. The method for implementing an intelligent collection assisting system based on central sentence recognition and Bert intention recognition as claimed in claim 7, wherein the S1 comprises collecting historical collection data of customers; dividing the intention of the client into a plurality of intentions of chatting, answering inconveniently and agreeing to repayment; manually marking data according to the historical data which is received by the client and the defined intention;
s2, extracting the intents of the same type in the labeling data; carrying out word segmentation on each sentence of the intention, and removing stop words and word vectors to obtain a sentence vector; performing text clustering by using each sentence vector and Kmeans to obtain a fine classification of the intention question; selecting appropriate values of the detailed classes according to the distribution condition of the text;
the S3 comprises sentence vectors aiming at each subdivision intention category, and the TextRank performs sentence extraction; using a TextRank algorithm, taking the text data as nodes of the graph, and taking the similarity between the text and the text as an adjacency matrix; screening out 10 standard question sentences which can represent the detailed classification most by utilizing the principle that the value of a certain sentence is determined by the value of each sentence linked to the sentence and the corresponding weight; storing the sentence vectors of the standard sentences into a database;
and S4, carrying out model discrimination and carrying out no model discrimination: model discrimination is not performed: the dialogue content of the client contains keywords as a text central sentence; and (3) judging a model: calling standard question sentences of fine categories, calculating similarity between sentence vectors of the sentences and the standard question sentences, and if the standard question sentences higher than a threshold value exist, determining the sentences as text center sentences;
the S5 training model and preferably comprises: word vectors, word embedding and a bidirectional long and short memory network model; word vectors, word embedding and a bidirectional long and short memory network model; a fine tuning depth model based on AlBert pre-training; a fine-tuning depth model based on Bert pre-training;
the supplementary missing detection rule and the illegal compliance rule comprise the following steps: determining the category of the sentence pattern using rule with obvious rules, namely the missing detection rule;
after the model is judged, the rule violation and the rule compliance of experience are judged, and a model result, namely rule violation and rule compliance, is optimized;
configuration of default rules: if hit, a reservation is needed; miss, need to be excluded;
and (3) compliance rule configuration: if hit, need to exclude; miss, need to be reserved;
loading a model, wherein a text center sentence meaning graph identifies whether the answer of a client is a text center sentence or not in the collection urging process, if so, automatically loading an optimal model, a missing detection rule and a violation compliance rule, and judging the intention type;
the S6 is used for configuring intention corresponding dialogues in a knowledge base configuration module; after the intention label is detected, automatically displaying the preferred conversation collection urging operation;
the S7 is used for configuring a session template required by generating a session summary according to the label; and automatically generating a conversation summary by the client label and the conversation template.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211018187.9A CN115392233A (en) | 2022-08-24 | 2022-08-24 | Intelligent collection prompting auxiliary system based on central sentence recognition and Bert intention recognition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211018187.9A CN115392233A (en) | 2022-08-24 | 2022-08-24 | Intelligent collection prompting auxiliary system based on central sentence recognition and Bert intention recognition |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115392233A true CN115392233A (en) | 2022-11-25 |
Family
ID=84121560
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211018187.9A Pending CN115392233A (en) | 2022-08-24 | 2022-08-24 | Intelligent collection prompting auxiliary system based on central sentence recognition and Bert intention recognition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115392233A (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090136013A1 (en) * | 2007-11-19 | 2009-05-28 | Kuykendall Peter A | System for obtaining information regarding telephone calls |
| CN109949805A (en) * | 2019-02-21 | 2019-06-28 | 江苏苏宁银行股份有限公司 | Intelligent collection robot and collection method based on intention assessment and finite-state automata |
| CN111538821A (en) * | 2020-04-17 | 2020-08-14 | 北京智齿博创科技有限公司 | Method and device for solving cold start of knowledge base in intelligent customer service |
| CN111835921A (en) * | 2020-07-16 | 2020-10-27 | 普强时代(珠海横琴)信息技术有限公司 | Real-time automatic telephone traffic summary system and method |
| CN113515613A (en) * | 2021-06-25 | 2021-10-19 | 华中科技大学 | Intelligent robot integrating chatting, knowledge and task question answering |
| CN114218392A (en) * | 2022-02-22 | 2022-03-22 | 浙商期货有限公司 | Futures question-answer oriented user intention identification method and system |
| CN114860934A (en) * | 2022-05-09 | 2022-08-05 | 青岛日日顺乐信云科技有限公司 | Intelligent question-answering method based on NLP technology |
-
2022
- 2022-08-24 CN CN202211018187.9A patent/CN115392233A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090136013A1 (en) * | 2007-11-19 | 2009-05-28 | Kuykendall Peter A | System for obtaining information regarding telephone calls |
| CN109949805A (en) * | 2019-02-21 | 2019-06-28 | 江苏苏宁银行股份有限公司 | Intelligent collection robot and collection method based on intention assessment and finite-state automata |
| CN111538821A (en) * | 2020-04-17 | 2020-08-14 | 北京智齿博创科技有限公司 | Method and device for solving cold start of knowledge base in intelligent customer service |
| CN111835921A (en) * | 2020-07-16 | 2020-10-27 | 普强时代(珠海横琴)信息技术有限公司 | Real-time automatic telephone traffic summary system and method |
| CN113515613A (en) * | 2021-06-25 | 2021-10-19 | 华中科技大学 | Intelligent robot integrating chatting, knowledge and task question answering |
| CN114218392A (en) * | 2022-02-22 | 2022-03-22 | 浙商期货有限公司 | Futures question-answer oriented user intention identification method and system |
| CN114860934A (en) * | 2022-05-09 | 2022-08-05 | 青岛日日顺乐信云科技有限公司 | Intelligent question-answering method based on NLP technology |
Non-Patent Citations (1)
| Title |
|---|
| 吕诗宁;张毅;胡若云;沈然;江俊军;欧智坚;: "融合神经网络与电力领域知识的智能客服对话系统研究", 浙江电力, no. 08, pages 80 - 86 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112804400B (en) | Customer service call voice quality inspection method and device, electronic equipment and storage medium | |
| CN107329967B (en) | Question answering system and method based on deep learning | |
| US6944592B1 (en) | Interactive voice response system | |
| CN112185358A (en) | Intention recognition method, model training method, device, equipment and medium | |
| CN111858854B (en) | Question-answer matching method and relevant device based on historical dialogue information | |
| CN104462600A (en) | Method and device for achieving automatic classification of calling reasons | |
| CN111062220B (en) | End-to-end intention recognition system and method based on memory forgetting device | |
| CN115292461B (en) | Man-machine interaction learning method and system based on voice recognition | |
| CN112732871A (en) | Multi-label classification method for acquiring client intention label by robot | |
| CN112131358A (en) | Scene flow structure and intelligent customer service system applied by same | |
| CN112632244A (en) | Man-machine conversation optimization method and device, computer equipment and storage medium | |
| CN114398512A (en) | Big data-based voice portrait analysis method for communication operator business customer | |
| CN113806503A (en) | Dialog fusion method, device and equipment | |
| CN114818649A (en) | Service consultation processing method and device based on intelligent voice interaction technology | |
| CN112364622A (en) | Dialog text analysis method, dialog text analysis device, electronic device and storage medium | |
| CN115599894A (en) | Emotion recognition method and device, electronic equipment and storage medium | |
| CN113486174B (en) | Model training, reading understanding method and device, electronic equipment and storage medium | |
| CN113850387A (en) | Expert system knowledge base construction method, question and answer method, system, device and medium | |
| CN110795531B (en) | Intention identification method, device and storage medium | |
| CN115022471B (en) | Intelligent robot voice interaction system and method | |
| CN115392233A (en) | Intelligent collection prompting auxiliary system based on central sentence recognition and Bert intention recognition | |
| CN114372476B (en) | Semantic truncation detection method, device, equipment and computer readable storage medium | |
| CN115688758A (en) | Statement intention identification method and device and storage medium | |
| CN116186259A (en) | Session cue scoring method, device, equipment and storage medium | |
| CN115510213A (en) | Question answering method and system for working machine and working machine |
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 |