CN106997375B - Customer service reply recommendation method based on deep learning - Google Patents
Customer service reply recommendation method based on deep learning Download PDFInfo
- Publication number
- CN106997375B CN106997375B CN201710112855.7A CN201710112855A CN106997375B CN 106997375 B CN106997375 B CN 106997375B CN 201710112855 A CN201710112855 A CN 201710112855A CN 106997375 B CN106997375 B CN 106997375B
- Authority
- CN
- China
- Prior art keywords
- customer service
- vector
- dialog
- sentence
- reply
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 238000013135 deep learning Methods 0.000 title claims abstract description 7
- 238000012549 training Methods 0.000 claims abstract description 16
- 239000013598 vector Substances 0.000 claims description 77
- 230000008569 process Effects 0.000 claims description 13
- 238000012545 processing Methods 0.000 claims description 12
- 230000011218 segmentation Effects 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 7
- 238000003064 k means clustering Methods 0.000 claims description 3
- 239000013604 expression vector Substances 0.000 claims description 2
- 230000014509 gene expression Effects 0.000 claims description 2
- 238000010276 construction Methods 0.000 abstract description 6
- 239000003814 drug Substances 0.000 abstract description 2
- 230000008859 change Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
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
- G06F16/35—Clustering; Classification
- G06F16/355—Class or cluster creation or modification
-
- 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
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Human Computer Interaction (AREA)
- Computational Linguistics (AREA)
- Machine Translation (AREA)
Abstract
The invention discloses a customer service reply recommendation method based on deep learning, which is used for recommending customer service reply by directly learning a conversation model from a conversation record of customer service; the method carries out model construction through an end-to-end training mode, and is high in construction speed compared with the traditional method based on rules and artificial features; the method improves the coverage of the reply content in the customer service reply recommendation system; meanwhile, the invention can be applied to the customer service in various vertical fields, including but not limited to: e-commerce, medicine, law, etc.
Description
Technical Field
The invention belongs to the technical field of customer service assistance, and particularly relates to a customer service reply recommendation method based on deep learning.
Background
With the continuous development of the internet economy, the scale and the volume of the e-commerce platform for providing online goods and service shopping are increasing, and the changing trend puts higher requirements on the service efficiency of online customer service. Meanwhile, some new customer service scenes such as online health consultation and online legal consultation relate to more knowledge and have higher requirements on professional staff.
The existing solution ideas are two kinds: one is to use an intelligent customer service robot to directly take over a part of consultation tasks instead of manual work; the other method is to use a computer to assist manual customer service so as to improve the working efficiency of the customer service. The common intelligent customer service robot usually uses a manually constructed knowledge base as a core engine to replace manual work to complete some common simple consultation services. The disadvantages of this method are mainly: the robot can only process common standard problems generally, and is difficult to process some problems with strong individuation or low occurrence frequency; the establishment of a complex knowledge base requires a great deal of cost; the acceptance of the robot by the user is not as good as that of manual customer service. Computer-assisted human customer service, in general, can accomplish tasks that are not directly interactive with the user, such as: standard reply recommendations, frequently asked question and answer pair recommendations. The existing customer service auxiliary system is mainly designed aiming at scenes with relatively less required professional knowledge, such as trade disputes, commodity information consultation and the like.
The existing solution needs to manually establish and arrange a previous knowledge base or a corpus, and the cost is high. Meanwhile, the coverage rate problem exists in the knowledge base and the standard reply, the design is usually carried out aiming at the problems with more occurrence times, and the coverage is less for the problems with stronger individuation. The traditional online customer service scene problems are distributed more intensively, and the related professional knowledge is relatively less, so that the problems of construction cost and coverage rate of a knowledge base and corpora are relatively easy to solve. However, in a scene with higher professional requirements, as the depth and the breadth of knowledge are increased, the difficulty in constructing a knowledge base and corresponding linguistic data is increased, and the higher problem coverage rate is difficult to realize.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a customer service reply recommendation method based on deep learning, which can recommend reply texts which are possibly required by the customer service currently for the customer service according to a conversation record in a customer service consultation process, thereby reducing the input amount of the customer service in the working process and improving the working efficiency of the customer service.
The technical scheme of the invention is as follows:
the customer service reply recommendation method based on deep learning is characterized by comprising the following steps:
(1) clustering all customer service replies in the chat corpus and constructing a candidate customer service reply set;
(2) training word vectors on a customer service chat corpus;
(3) carrying out standardization processing on the conversation records of the customer service chat corpus;
(4) training a dialogue model on the processed corpus;
(5) and inputting the current conversation record of the customer service into the model, and acquiring the recommended reply at the moment.
The step of clustering all replies in the step (1) is as follows:
1.1, processing the speech by using a Chinese word segmentation device to obtain a word segmentation result;
1.2 running an LDA theme generation model on the word segmentation result and calculating the representation of each reply vector;
1.3 running a k-means clustering algorithm on all the replied vector representations, and completing clustering of replied texts by using Euclidean distance through a distance measurement function to obtain n clusters: { c1，…，cn}
Wherein: d () is a function of the calculated distance, X, Y are vector representations of the two replies, Xi，yiComponents X, Y, respectively;
1.4 setting a threshold m from { c }1,…,cnRandomly sampling m sentences in each corresponding class cluster to form a candidate reply set C;
the step of training the word vector in the step (2) is as follows:
2.1 using a Chinese word segmentation device to perform word segmentation processing on all corpora;
2.2 training word vectors on the corpus of the divided words by using a word vector training tool to obtain a word vector matrix.
The step (3) of normalizing the corpus comprises the following steps:
3.1 deleting meaningless format control symbols in all chat records and non-manually generated conversations;
3.2 dividing the conversation in the chat log into the following according to the role of generating the conversation: the method comprises the following steps that two types of conversations of a user and a customer service are adopted, and a plurality of conversation records of each service are processed as follows on the basis: { user, customer service, … }, i.e., each set of conversations begins with the user's conversation, with the user and customer service conversations occurring strictly in alternation;
3.3, carrying out truncation processing on the overlong dialogue single sentence dialogue and the field-crossing dialogue sequence.
The step of training the deep dialogue model in the step (4) is as follows:
4.1 randomly picking a set of samples from the dataset, a set of dialog sequences for each sample can be expressed as { (u)0,h0,u1,h1…), … } wherein uiI-th sentence, h, representing a user in a dialog sequenceiThe ith sentence of the dialogue representing the customer service in the dialogue sequence;
4.2 randomly selecting one customer service dialogue from the corpus as a negative sample for each customer service dialogue in the sample
4.3 for all dialogs in the sample:
the coding is carried out by using a single sentence coder (Utterance Encoder) provided by the invention, and the vector expressions of each sentence are respectively obtained:
4.4 represent one sample as a sequence of vectors:
on the basis, the sequence is coded by using a Context coder (Context Encoder) to obtain a vector output sequence which represents the coding result of the dialog history at each moment of the dialog:
4.5 at each moment of the customer service session, using the corresponding moment
As input, the local loss function value is calculated using the following formula:
wherein margin is a manually set threshold, Sim (x, y) is a similarity function, and the calculation method is shown in the following formula:
4.6 add all local losses in the current batch of samples to obtain the loss function value required by the update, and the calculation formula is as follows:
wherein: n is the number of samples contained in the current batch, miIs the number of customer service sessions, loss, contained in the ith samplei，jLocal loss values of the jth sentence in the ith sample at the moment corresponding to customer service are obtained;
4.7 update all parameters in the dialogue model using gradient descent;
4.8 if the iteration number reaches a threshold (the threshold can be set according to the requirement), the model is saved, the iteration is ended, and if not, the step 4.1 is returned.
The flow of encoding a single sentence by the single sentence encoder in said step 4.3 is as follows:
4.3.1 according to the different roles of the generated sentences, assigning a role mark to each sentence, setting the role mark of the sentence generated by the user to be 0, setting the role mark of the sentence generated by the customer service or the candidate reply to be 1, and using a symbol r to represent the role mark;
4.3.2 replacing each word in the sentence by the word vector representation form corresponding to the word vector matrix trained in the step (2) by searching the word vector matrix, and representing the word as
4.3.3 and color coordinates the corner with each word vector in the sentenceThe cascading is respectively carried out, and the process can be represented as follows: v. ofw′＝[vw；r]；
4.3.4, using a two-layer GRU network to read in the processed word vector sequence and obtain the output of the last moment as the encoding result of Utterance Encoder, wherein the updating mode of each GRU unit is shown as the following formula:
zt＝σ(Wz·[ht-1,xt])
rt＝(Wr·[ht-1,xt])
ht＝(1-zt)*ht-1+zt*ht
wherein xtRepresenting input at a certain moment of time, htIndicating the output at the corresponding time instant.
The flow of encoding the dialog record sequence by the context encoder in said step 4.4 is as follows:
4.4.1 reading in the vector representation of each sentence in the dialogue record:
4.4.2 insert an all zero vector of the same length at the head of the vector sequence:
4.4.3 use a sliding window of length 2 to splice the vectors in the sequence, resulting in an output of:
4.4.4 using the spliced vector sequence as input, encoding it using a two-layer GRU network, taking the output at each time and expressing as:
wherein
An output representing the ith time instant;
4.4.5 the context expression vector of each time is combined with the sentence vector corresponding to the time in a cascade manner, and the output obtained at the ith time is as follows:
wherein
Represents the output of the GRU network at time i,
a vector encoding of the ith utterance representing the user in the dialog sequence,
a vector code representing an ith utterance serviced in the dialog sequence;
4.4.6 remapping the stitched vector with the linear layer, the calculation is as follows
Vi′＝WT·Vi
Wherein WTIs a linear layer weight matrix obtained by training, ViFor the spliced vector, Vi' A vector representing a context for context at that time represents a result.
The process of selecting a reply according to the history of the current customer service conversation in the step (5) is as follows:
5.1 the candidate reply set constructed in step (1) is represented as: a ═ a0,a1,…,anIn which aiRepresenting a candidate reply text;
5.2 Using a single sentence encoder to represent each candidate reply sentence as a vector, the process is the same as in step 4.3, and its output is represented as:
5.3 using Utterance Encoder to code each dialog in the current dialog record, the processing steps are the same as 4.3, and the processing result is expressed as:
5.4 use
The encoding process is the same as step 4.4 as the input of the context encoder, and the output of the last moment is taken out as the encoding result of the current context vector, Vc；
5.5 traversing the vector codes of all the candidate replies, and calculating the matching degree of each candidate reply and the current context, wherein the calculation method is shown as the following formula:
The invention provides a customer service reply recommendation method based on a depth model by utilizing a chat conversation record generated in the service process of the customer service, and provides a method for applying the customer service reply recommendation method to an actual scene. Therefore, compared with the traditional method, the method has the following beneficial effects:
(1) the method carries out model construction in an end-to-end training mode, and the construction speed is higher;
(2) the method of the invention improves the coverage of the reply content in the customer service reply recommendation system;
(3) the method of the invention ensures that the construction of the customer service reply recommendation system does not need the participation of personnel with expert knowledge;
(4) the method of the invention can be applied to the customer service in various vertical fields, including but not limited to: e-commerce, medicine, law, etc.
Drawings
FIG. 1 is a schematic view of the main process of the present invention.
Fig. 2 is a schematic structural diagram of a single sentence encoder proposed in the present invention.
Fig. 3 is a schematic structural diagram of a context encoder proposed in the present invention.
Detailed Description
In order to more specifically describe the present invention, the following detailed description is provided for the technical solution of the present invention with reference to the accompanying drawings and the specific embodiments.
The method flow of the invention specifically comprises the following steps:
(1) extracting and screening a candidate reply set from the customer service chat corpus;
(2) training word vectors by using customer service chatting corpus;
(3) carrying out standardization processing on customer service chat corpora;
(4) training a customer service dialogue model;
(5) selecting recommended reply content using a conversation model based on a current conversation record
The step (1) is divided into the following steps:
1-1: screening all dialogues spoken by customer service from the corpus;
1-2: segmenting the selected dialogs by using a knot segmentation tool;
1-3: performing topic modeling on the dialog after word segmentation by using an LDA topic model, and acquiring vector representation of the dialog;
1-4: clustering the screened replies on a vector space by using a k-means clustering algorithm;
1-5: and randomly selecting equal replies in each cluster, and performing deduplication operation on the replies to form a candidate reply set.
The step (2) is divided into the following steps:
2-1, segmenting all conversation records by using a bus segmentation tool;
2-2 training word vectors on the segmented word conversation records, the specific tools used are: word2 vec.
The step (3) is divided into the following steps:
3-1 deleting all format control symbols in the chat records and non-artificially generated conversations;
3-2 dividing the conversations in the chat log into: the method comprises the following steps that two types of conversations of a user and a customer service are adopted, and a plurality of conversation records of each service are processed as follows on the basis: the sequence of user, host, user, host, …, i.e. each set of dialogs starts with a user's dialog and the dialogs of user and customer service appear strictly alternating, where user stands for user and host stands for customer service.
The step (4) is divided into the following steps:
4-1 initializing model parameters, wherein a word vector weight matrix is initialized by using pre-trained word vectors, and the rest parameters are initialized by using random numbers which accord with Gaussian distribution;
4-2, extracting n samples from the data by using a random sampling method to serve as the currently trained batch;
4-3, randomly selecting a customer service dialogue as a negative sample reply from the corpus for replying all customer services in the sample;
4-4 as shown in fig. 1, a single sentence encoder is used to encode all sentences in a sample to obtain their corresponding vector representations, and the specific process is as follows: setting character codes of sentences, converting all words into word vector representations, splicing the character codes and the word vectors, coding a vector sequence by using a double-layer GRU (generalized regression Unit), and acquiring the output of the last moment as a coding result;
4-5 as shown in fig. 2, for each sample, vector representations of all sentences are taken as input, and the vector sequences are encoded by using a context encoder respectively, so as to obtain dialog context representation results at different times, wherein the specific process is as follows: filling an all-zero vector with the same length in the head of a vector sequence, splicing adjacent vectors in the vector sequence by using a sliding window with the length of 2, reading the spliced vector sequence by using a double-layer GRU network, calculating the output at each moment, cascading the output of the GRU network and the vector code of an original sentence at each moment, performing linear transformation and dimension reduction by using a linear layer, and taking out the output as a context coding result at the moment;
4-6, respectively calculating the similarity degree of the original reply and the negative sample reply with the context code at the moment at all the time when the customer service replies for each sample, and calculating a local loss value by using a change loss function;
4-7, adding all local loss values of all samples in the current batch to obtain a global loss value;
4-8, updating all parameters in the model by using a gradient descent method, and setting the learning rate to be 0.01;
4-9, judging whether the change degree of the iteration times and the global loss value reaches a threshold value, if so, stopping training and storing the model, and if not, returning to the step 4-2.
The step (5) is divided into the following steps:
5-1, using a single sentence coder to code all sentences in the candidate set into a vector form;
5-2, coding each dialog in the current dialog record by using a single sentence coder;
5-3, reading in vector representation of each sentence in the current conversation by using a context encoder, and taking the output of the last moment as a current context representation result;
5-4, traversing all candidate replies, calculating the similarity degree of the candidate replies and the current context by using cosine similarity, and selecting the reply with the highest score from the candidate replies and recommending the reply to the customer service.
The above examples are not intended to limit the present invention, and the present invention is not limited to the above embodiments, and the present invention is within the scope of the present invention as long as the requirements of the present invention are met.
Claims (6)
1. A customer service reply recommendation method based on deep learning comprises the following steps:
(1) clustering all customer service replies in the chat corpus and constructing a candidate customer service reply set; training word vectors on a customer service chat corpus;
(3) carrying out standardization processing on the conversation records of the customer service chat corpus;
(4) training a dialogue model on the corpus after the standardization processing;
(5) inputting the current conversation record of the customer service into the conversation model to obtain a corresponding recommendation reply;
the specific steps of the dialogue model training in the step (4) are as follows:
4.1 randomly picking a batch of samples from the dataset, each sample being a multi-group dialog sequence that can be expressed as { (u)0,h0,u1,h1…), … }, where u isiI-th sentence, h, representing a user in a dialog sequenceiThe ith sentence of the dialogue representing the customer service in the dialogue sequence;
4.2 randomly selecting one customer service dialogue from the corpus as a negative sample for each customer service dialogue in the sample
4.3 for all dialogues of each set of dialog sequences in the sample:
all use single sentence encoder to encode, all encode each sentence into the fixed length vector, obtain the vector expression of each sentence respectively:
4.4 represent each set of dialog sequences as a sequence of vectors:
on the basis, the sequence is coded by using a context coder, the coding result of the dialog record generated when each dialog appears can be obtained, each dialog of the user and the customer service corresponds to a moment, and a vector output sequence is obtained after coding and represents the coding result of the dialog history at each moment:
wherein k is the number of dialogs contained in the set of dialog sequences;
4.5 at each moment of the customer service session, use the corresponding moment i
As input, the local loss function value is calculated using the following formula:
wherein margin is a manually set threshold, Sim (x, y) is a similarity function, and the calculation method is shown in the following formula:
4.6 add all local losses in the current batch of samples to obtain a loss function value required by the update, wherein the calculation formula is as follows:
wherein: n is the number of samples contained in the current batch, miNumber of customer service replies, loss, contained in the ith samplei,Replying a local loss value at a corresponding moment for the jth customer service in the ith sample;
4.7 update all parameters in the dialogue model using gradient descent;
4.8 if the iteration times reach the threshold value, the model is saved, the iteration is ended, and if not, the step 4.1 is returned.
2. The customer service reply recommendation method according to claim 1, wherein: the step (1) specifically comprises the following steps:
1.1, processing all customer service replies in the corpus by using a Chinese word segmentation device to obtain word segmentation results;
1.2, performing topic modeling on the dialog after word segmentation by using an LDA topic model, and acquiring vector representation of the dialog;
1.3 clustering the customer service replies on a vector space by using a k-means clustering algorithm;
1.4 randomly selecting equal replies in each cluster, and performing deduplication operation on the replies to form a candidate reply set.
3. The customer service reply recommendation method according to claim 1, wherein: the specific steps of the standardized processing of the dialogue records in the step (3) are as follows:
3.1 deleting all format control symbols and non-artificially generated dialogs in the dialog record;
3.2 the dialog is divided into, according to the role that generates it: the method comprises the following steps that two types of conversations of a user and a customer service are adopted, and a plurality of conversation records of each service are processed as follows on the basis: the sequence of user, host, user, host, …, i.e. each set of dialogs starts with a user's dialog and the dialogs of user and customer service appear strictly alternating, where user stands for user and host stands for customer service.
4. The customer service reply recommendation method according to claim 1, wherein: the specific steps of encoding a single sentence by using a single sentence encoder in the step 4.3 are as follows:
5.1 according to the different roles of the sentences, assigning a role mark to each sentence, setting the role mark of the sentence generated by the user to be 0, setting the role mark of the sentence generated by the customer service or the candidate reply to be 1, and using a symbol r to represent the role mark;
5.2 replacing each word in the sentence by a word vector representation form corresponding to the word vector representation form by searching the word vector matrix trained in the step (2) to represent as
5.3 and the role mark is respectively cascaded with each word vector in the sentence, and the process can express thatComprises the following steps: v. ofw′＝[vw；r]；
And 5.4, reading the processed word vector sequence by using a two-layer GRU network, and obtaining the output of the last moment as the coding result of the single sentence coder.
5. The customer service reply recommendation method according to claim 1, wherein: the specific steps of encoding the dialog sequence by using the context encoder in step 4.4 are as follows:
6.1 reading in the vector representation of each sentence in the dialogue record:
6.2 insert an all-zero vector of the same length at the head of the vector sequence:
6.3 use the length of 2 sliding window to splice the vectors in the sequence, get the output:
6.4 using the spliced vector sequence as input, using a double-layer GRU network to encode the vector sequence, and taking the output of each time and expressing as:
wherein
An output representing the ith time instant;
6.5 the context expression vector of each time is combined with the sentence vector corresponding to the time in a cascade way, and the output obtained at the ith time is as follows:
wherein
Represents the output of the GRU network at time i,
a vector code representing the ith utterance spoken by the user in the dialog sequence,
a vector code representing an ith utterance to be spoken by a customer in the dialog sequence;
6.6 remapping each spliced vector with a linear layer, the calculation is as follows
Vi′＝WT·Vi
Wherein WTIs a linear layer weight matrix, ViFor the spliced vector, Vi' is the vector representation result for the context at that time.
6. The customer service reply recommendation method according to claim 1, wherein: the specific steps of selecting the recommended reply according to the current conversation history in the step (5) are as follows:
7.1 the set of candidate replies constructed in step (1) is represented as: a ═ a0，a1，…，anIn which aiRepresenting a candidate reply text;
7.2 Using a single sentence encoder, each candidate reply sentence is represented in vector form, the output of which is represented as:
7.3 use the single sentence encoder to encode each dialog in the current dialog record, and express its processing result as:
7.4 use
Taking the last moment output as the current context vector coding result V as the input of the context coderc；
7.5 vector encoding V over all candidate repliesaCalculating the matching degree of each candidate reply and the current context;
7.6 comparing the matching degree scores of all the candidate replies, and taking the highest score as the recommended reply.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710112855.7A CN106997375B (en) | 2017-02-28 | 2017-02-28 | Customer service reply recommendation method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710112855.7A CN106997375B (en) | 2017-02-28 | 2017-02-28 | Customer service reply recommendation method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106997375A CN106997375A (en) | 2017-08-01 |
| CN106997375B true CN106997375B (en) | 2020-08-18 |
Family
ID=59431014
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710112855.7A Expired - Fee Related CN106997375B (en) | 2017-02-28 | 2017-02-28 | Customer service reply recommendation method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106997375B (en) |
Families Citing this family (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107368524B (en) * | 2017-06-07 | 2020-06-02 | 创新先进技术有限公司 | Dialog generation method and device and electronic equipment |
| CN107330130B (en) * | 2017-08-29 | 2020-10-20 | 北京易掌云峰科技有限公司 | Method for realizing conversation robot recommending reply content to manual customer service |
| CN110019700B (en) * | 2017-09-13 | 2023-01-17 | 阿里巴巴集团控股有限公司 | Data processing method and device |
| CN107679231A (en) * | 2017-10-24 | 2018-02-09 | 济南浪潮高新科技投资发展有限公司 | A kind of vertical field and the implementation method of Opening field mixed type intelligent Answer System |
| CN108038230B (en) * | 2017-12-26 | 2022-05-20 | 北京百度网讯科技有限公司 | Information generation method and device based on artificial intelligence |
| US11886823B2 (en) * | 2018-02-01 | 2024-01-30 | International Business Machines Corporation | Dynamically constructing and configuring a conversational agent learning model |
| CN108763477A (en) * | 2018-05-29 | 2018-11-06 | 厦门快商通信息技术有限公司 | A kind of short text classification method and system |
| CN108804611B (en) * | 2018-05-30 | 2021-11-19 | 浙江大学 | Dialog reply generation method and system based on self comment sequence learning |
| CN109062951B (en) * | 2018-06-22 | 2021-04-06 | 厦门快商通信息技术有限公司 | Dialogue flow extraction method, device and storage medium based on intention analysis and dialogue clustering |
| CN108920715B (en) * | 2018-07-26 | 2020-11-10 | 百度在线网络技术（北京）有限公司 | Intelligent auxiliary method, device, server and storage medium for customer service |
| CN109032381B (en) * | 2018-08-01 | 2022-05-17 | 平安科技（深圳）有限公司 | Input method and device based on context, storage medium and terminal |
| CN109189901B (en) * | 2018-08-09 | 2021-05-18 | 北京中关村科金技术有限公司 | Method for automatically discovering new classification and corresponding corpus in intelligent customer service system |
| CN109189931B (en) * | 2018-09-05 | 2021-05-11 | 腾讯科技（深圳）有限公司 | Target statement screening method and device |
| CN109543177B (en) * | 2018-10-19 | 2022-04-12 | 中国平安人寿保险股份有限公司 | Message data processing method and device, computer equipment and storage medium |
| CN111199325B (en) * | 2018-11-19 | 2023-12-26 | 阿里巴巴集团控股有限公司 | Online customer service distribution method and device and electronic equipment |
| CN109615009B (en) * | 2018-12-12 | 2021-03-12 | 广东小天才科技有限公司 | Learning content recommendation method and electronic equipment |
| CN111324704B (en) * | 2018-12-14 | 2023-05-02 | 阿里巴巴集团控股有限公司 | Method and device for constructing speaking knowledge base and customer service robot |
| CN109977194B (en) * | 2019-03-20 | 2021-08-10 | 华南理工大学 | Text similarity calculation method, system, device and medium based on unsupervised learning |
| CN110196930B (en) * | 2019-05-22 | 2021-08-24 | 山东大学 | Multi-mode customer service automatic reply method and system |
| CN110825852B (en) * | 2019-11-07 | 2022-06-14 | 四川长虹电器股份有限公司 | Long text-oriented semantic matching method and system |
| US11409965B2 (en) * | 2020-01-15 | 2022-08-09 | International Business Machines Corporation | Searching conversation logs of a virtual agent dialog system for contrastive temporal patterns |
| CN111368531B (en) * | 2020-03-09 | 2023-04-14 | 腾讯科技（深圳）有限公司 | Translation text processing method and device, computer equipment and storage medium |
| CN111985934B (en) * | 2020-07-30 | 2024-07-12 | 浙江百世技术有限公司 | Intelligent customer service dialogue model construction method and application |
| CN113239157B (en) * | 2021-03-31 | 2022-02-25 | 北京百度网讯科技有限公司 | Method, device, equipment and storage medium for training conversation model |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103425635A (en) * | 2012-05-15 | 2013-12-04 | 北京百度网讯科技有限公司 | Method and device for recommending answers |
| CN105068661A (en) * | 2015-09-07 | 2015-11-18 | 百度在线网络技术（北京）有限公司 | Man-machine interaction method and system based on artificial intelligence |
| CN105072173A (en) * | 2015-08-03 | 2015-11-18 | 谌志群 | Customer service method and system for automatically switching between automatic customer service and artificial customer service |
| CN105787560A (en) * | 2016-03-18 | 2016-07-20 | 北京光年无限科技有限公司 | Dialogue data interaction processing method and device based on recurrent neural network |
| CN105955965A (en) * | 2016-06-21 | 2016-09-21 | 上海智臻智能网络科技股份有限公司 | Question information processing method and device |
| CN106448670A (en) * | 2016-10-21 | 2017-02-22 | 竹间智能科技（上海）有限公司 | Dialogue automatic reply system based on deep learning and reinforcement learning |
-
2017
- 2017-02-28 CN CN201710112855.7A patent/CN106997375B/en not_active Expired - Fee Related
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103425635A (en) * | 2012-05-15 | 2013-12-04 | 北京百度网讯科技有限公司 | Method and device for recommending answers |
| CN105072173A (en) * | 2015-08-03 | 2015-11-18 | 谌志群 | Customer service method and system for automatically switching between automatic customer service and artificial customer service |
| CN105068661A (en) * | 2015-09-07 | 2015-11-18 | 百度在线网络技术（北京）有限公司 | Man-machine interaction method and system based on artificial intelligence |
| CN105787560A (en) * | 2016-03-18 | 2016-07-20 | 北京光年无限科技有限公司 | Dialogue data interaction processing method and device based on recurrent neural network |
| CN105955965A (en) * | 2016-06-21 | 2016-09-21 | 上海智臻智能网络科技股份有限公司 | Question information processing method and device |
| CN106448670A (en) * | 2016-10-21 | 2017-02-22 | 竹间智能科技（上海）有限公司 | Dialogue automatic reply system based on deep learning and reinforcement learning |
Non-Patent Citations (3)
| Title |
|---|
| Li J等.Deep Reinforcement Learning for Dialogue.《aiXiv preprint arXiv》.2016, * |
| SutskeverI等.Sequence to sequence learning with neural.《Advances in neural information processing systems》.2014, * |
| Xing C等.Topic Aware Neural Response Generation.《arXiv preprint arXiv》.2016, * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106997375A (en) | 2017-08-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106997375B (en) | Customer service reply recommendation method based on deep learning | |
| CN108874972B (en) | Multi-turn emotion conversation method based on deep learning | |
| CN112000791B (en) | Motor fault knowledge extraction system and method | |
| CN110309287B (en) | Retrieval type chatting dialogue scoring method for modeling dialogue turn information | |
| CN105723362B (en) | Naturally processing method, processing and response method, equipment and system are expressed | |
| CN110134946B (en) | Machine reading understanding method for complex data | |
| CN111966800B (en) | Emotion dialogue generation method and device and emotion dialogue model training method and device | |
| CN106448670A (en) | Dialogue automatic reply system based on deep learning and reinforcement learning | |
| CN111414476A (en) | Attribute-level emotion analysis method based on multi-task learning | |
| CN107862087A (en) | Sentiment analysis method, apparatus and storage medium based on big data and deep learning | |
| CN107798140A (en) | A kind of conversational system construction method, semantic controlled answer method and device | |
| CN108256968B (en) | E-commerce platform commodity expert comment generation method | |
| CN113205817A (en) | Speech semantic recognition method, system, device and medium | |
| CN110427616B (en) | Text emotion analysis method based on deep learning | |
| CN111339305A (en) | Text classification method and device, electronic equipment and storage medium | |
| CN111062220B (en) | End-to-end intention recognition system and method based on memory forgetting device | |
| CN114722839B (en) | Man-machine cooperative dialogue interaction system and method | |
| CN111930918B (en) | Cross-modal bilateral personalized man-machine social interaction dialog generation method and system | |
| CN110222184A (en) | A kind of emotion information recognition methods of text and relevant apparatus | |
| CN111368066B (en) | Method, apparatus and computer readable storage medium for obtaining dialogue abstract | |
| CN112364148B (en) | Deep learning method-based generative chat robot | |
| CN112349294B (en) | Voice processing method and device, computer readable medium and electronic equipment | |
| CN112287106A (en) | Online comment emotion classification method based on dual-channel hybrid neural network | |
| CN111523328B (en) | Intelligent customer service semantic processing method | |
| CN113159831A (en) | Comment text sentiment analysis method based on improved capsule network |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200818 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |