CN114417889A - Task type dialogue system and method based on seq2seq framework - Google Patents
Task type dialogue system and method based on seq2seq framework Download PDFInfo
- Publication number
- CN114417889A CN114417889A CN202210043220.7A CN202210043220A CN114417889A CN 114417889 A CN114417889 A CN 114417889A CN 202210043220 A CN202210043220 A CN 202210043220A CN 114417889 A CN114417889 A CN 114417889A
- Authority
- CN
- China
- Prior art keywords
- module
- task
- answer
- input
- intention
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 8
- 238000005457 optimization Methods 0.000 claims abstract description 12
- 238000000605 extraction Methods 0.000 claims abstract description 10
- 238000007635 classification algorithm Methods 0.000 claims description 6
- 235000013305 food Nutrition 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 230000000873 masking effect Effects 0.000 claims description 2
- 235000013361 beverage Nutrition 0.000 description 3
- 241001122767 Theaceae Species 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 235000015203 fruit juice Nutrition 0.000 description 1
- 239000008267 milk Substances 0.000 description 1
- 210000004080 milk Anatomy 0.000 description 1
- 235000013336 milk Nutrition 0.000 description 1
- 238000011144 upstream manufacturing Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a task-based dialogue system and method based on a seq2seq architecture, which comprises four parts, namely an intention and slot filling network module, a database operation module, a trust tracking module and an optimization and generation network module; the method comprises the following steps: the bert output of the classification task and the vector of the input label are mapped to the same vector space, and the groove value extraction task is trained through a CRF layer; the database operation module is mainly used for inquiring other slot values according to the identified slot values, trusts a tracking layer, takes the output of the intention and slot value network module and the database operation module as input, takes the output of the layer as input, takes the input of intention and generation network, optimizes and generates the network module to select the answer An with the highest score as the answer of the current round of conversation, and carries out subsequent prediction on the answer. The invention combines the task type dialogue system and the chatting type dialogue system, and combines the module of the task type dialogue system inside, thereby ensuring the function of context and carrying out integral optimization.
Description
Technical Field
The invention relates to a task type dialogue system and method based on a seq2seq framework, belonging to the technical field of intelligent question answering.
Background
A traditional task type dialogue system adopts a pipeline structure. Due to the fact that the structure adopts a modular mode, errors of the upstream task parameters are conducted to the downstream. The chatting dialogue system adopts a seq2seq structure and produces corresponding answers through an Encoder module and a Decoder module. But the chatty type dialogue system does not consider the role of context, and the generated NLG answer has uncontrollable property.
Disclosure of Invention
The invention aims to provide a task type dialogue system and method based on a seq2seq framework, which not only consider the context influence and avoid generating uncertain answers, but also can carry out overall optimization.
In order to achieve the purpose, the invention is realized by the following technical scheme:
a task-based dialogue system based on a seq2seq architecture comprises four parts, namely an intention and slot filling network module, a database operation module, a trust tracking module and an optimization and generation network module;
the intention and trough value network module is built by adopting a Bert + CRF structure, wherein the Bert is used for intention classification, the CRF is used for trough value extraction, and the module pre-trains three tasks, namely an intention classification task, a trough value extraction task and a masking word prediction task of the Bert;
the database operation module is mainly used for inquiring other slot values according to the identified slot values;
the trust tracking module (Belief Tracker) adopts a Roberta model, and outputs a vector S containing all information from the beginning to the current roundn+1;
And the optimization and generation network module adopts a classification algorithm, selects the highest score as the answer of the current round of conversation in the existing action, and takes the answer as the input of the trust tracking module for subsequent prediction.
A task-based dialog method based on seq2seq architecture comprises the following steps:
1) bert output __ CLS __ Token a of classification taskclsAnd the vector y of the input labelintentMapping to the same vector space by maximizing the similarity S of the predicted value and the label positive caseI +=hCLSh+ intentAnd minimizing the similarity S of the predicted value and the label negative caseI -=hCLSh- intentThe optimization is carried out, and the specific formula is as follows:
2) the groove value extraction task is trained through a CRF layer, and the formula is as follows:
LE=LCRF(a,yentity)
wherein y isentityInputting a hidden vector of a label for an entity, a being a corresponding vector Token, L of a data inputCRF(.) is the negative similarity of the CRF;
3) the overall loss of the intent and slot value fill layers is calculated as:
Ltotal=LI+LE+LM
in the formula LMMask prediction for Bert self;
4) the database operation module is mainly used for inquiring other slot values according to the identified slot values and searching other slot values through select, where food and drink according to the keywords provided by the system;
5) a trust tracking layer using the output of the intention and slot value network module and the database operation module as input UIOne, with the output Sn of the layer as one of the inputs, one of the inputs An of the intent and generation network, the Belief Tracker, as the final input
Ib=UI+Sn+An
Wherein U isIRepresenting the user intention and the corresponding entity in the current round, Sn representing all states from the beginning to the previous round, An representing the answer to the user question in the previous round;
6) and the optimizing and generating network module adopts a classification algorithm for a primary final result, selects the answer An with the highest score as the answer of the current conversation in the existing action, and takes the answer as the input of the Belief Tracker for subsequent prediction.
The invention has the advantages that: the present invention combines a task-based dialog system with a chat-based dialog system. The whole adopts a seq2seq structure, and the module of the task dialog system is combined inside, so that the context effect is ensured, and the whole optimization can be carried out.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention.
Fig. 1 is a schematic view of the overall structure of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The entry to this model is the intent and slot value network. The network is built by adopting a Bert + CRF structure. Bert is used for intent classification and CRF is used for bin value extraction. During pre-training, three tasks, namely an intention classification task, a groove value extraction task and a Bert self mask word prediction task, are mainly considered.
Bert output __ CLS __ Token a of classification taskclsAnd the vector y of the input labelintentMapping to the sameVector space by maximizing the similarity S of the predicted value and the label positive caseI +=hCLSh+ intentAnd minimizing the similarity S of the predicted value and the label negative caseI -=hCLSh- intentThe optimization is carried out, and the specific formula is as follows:
the groove value extraction task is trained through a CRF layer, and the formula is as follows:
LE=LCRF(a,yentity)
wherein y isentityInputting a hidden vector of a label for an entity, a being a corresponding vector Token, L of a data inputCRF(.) is the negative similarity of the CRF;
the overall loss of the intent and slot value fill layers is calculated as:
Ltotal=LI+LE+LM
in the formula LMMask prediction for Bert self;
and the database operation module is mainly used for inquiring other slot values according to the identified slot values and is used for answering and generating. For example:
human: giving me a cup of beverage.
The system comprises the following steps: ask what beverage you need?
Human: what do you have?
The system comprises the following steps: we have coffee, milk tea, and fruit juice.
In the above dialog, after the system extracts the "beverage" keyword, it can search through select where food is drink.
A trust tracking layer, which takes the output of the intention and slot value network module and the database operation module as input UIOne, with the output Sn of the layer as one of the inputs, one of the inputs An of the intent and generation network, the Belief Tracker, as the final input
Ib=UI+Sn+An
Wherein U isIRepresenting the user intention and the corresponding entity in the current round, Sn representing all states from the beginning to the previous round, An representing the answer to the user question in the previous round; the network considers the historical information and the influence of the information of the current round. The network adopts the Roberta model. The output is a vector S containing all information from the beginning to the current roundn+1。
The optimization and generation network module adopts a classification algorithm, and because the model is used for a task-based dialog system, the accuracy of output results is most important. And (4) adopting a classification algorithm as a final result, selecting the answer An with the highest score as the answer of the current conversation in the existing action, and taking the answer as the input of a Belief Tracker for subsequent prediction.
Claims (2)
1. A task-based dialogue system based on a seq2seq architecture is characterized by comprising four parts, namely an intention and slot filling network module, a database operation module, a trust tracking module and an optimization and generation network module;
the intention and trough value network module is built by adopting a Bert + CRF structure, wherein the Bert is used for intention classification, the CRF is used for trough value extraction, and the module pre-trains three tasks, namely an intention classification task, a trough value extraction task and a masking word prediction task of the Bert;
the database operation module is mainly used for inquiring other slot values according to the identified slot values;
the trust tracking module adopts a Roberta model and outputs a vector S containing all information from the beginning to the current roundn+1;
And the optimization and generation network module adopts a classification algorithm, selects the highest score as the answer of the current round of conversation in the existing action, and takes the answer as the input of the trust tracking module for subsequent prediction.
2. A task-based dialog method using the seq2seq architecture as claimed in claim 1, characterized in that it comprises the following steps:
1) of classification tasksbert output __ CLS __ Token aclsAnd the vector y of the input labelintentMapping to the same vector space by maximizing the similarity S of the predicted value and the label positive caseI +=hCLSh+ intentAnd minimizing the similarity S of the predicted value and the label negative caseI -=hCLSh- intentThe optimization is carried out, and the specific formula is as follows:
2) the groove value extraction task is trained through a CRF layer, and the formula is as follows:
LE=LCRF(a,yentity)
wherein y isentityInputting a hidden vector of a label for an entity, a being a corresponding vector Token, L of a data inputCRF(.) is the negative similarity of the CRF;
3) the overall loss of the intent and slot value fill layers is calculated as:
Ltotal=LI+LE+LM
in the formula LMMask prediction for Bert self;
4) the database operation module is mainly used for inquiring other slot values according to the identified slot values and searching other slot values through select, where food and drink according to the keywords provided by the system;
5) a trust tracking layer module using the output of the intention and slot value network module and the database operation module as input UIOne, with the output Sn of the layer as one of the inputs, one of the inputs An of the intent and generation network, the Belief Tracker, as the final input
Ib=UI+Sn+An
Wherein U isIRepresenting the user intention and the corresponding entity in the current round, Sn representing all states from the beginning to the previous round, An representing the answer to the user question in the previous round;
6) and the optimizing and generating network module adopts a classification algorithm for a primary final result, selects the answer An with the highest score as the answer of the current conversation in the existing action, and takes the answer as the input of the Belief Tracker for subsequent prediction.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210043220.7A CN114417889A (en) | 2022-01-14 | 2022-01-14 | Task type dialogue system and method based on seq2seq framework |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210043220.7A CN114417889A (en) | 2022-01-14 | 2022-01-14 | Task type dialogue system and method based on seq2seq framework |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114417889A true CN114417889A (en) | 2022-04-29 |
Family
ID=81273088
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210043220.7A Pending CN114417889A (en) | 2022-01-14 | 2022-01-14 | Task type dialogue system and method based on seq2seq framework |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114417889A (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111324744A (en) * | 2020-02-17 | 2020-06-23 | 中山大学 | Data enhancement method based on target emotion analysis data set |
US20210201079A1 (en) * | 2019-12-27 | 2021-07-01 | Fujitsu Limited | Training data generation method and information processing apparatus |
CN113705249A (en) * | 2021-08-25 | 2021-11-26 | 上海云从企业发展有限公司 | Dialogue processing method, system, device and computer readable storage medium |
-
2022
- 2022-01-14 CN CN202210043220.7A patent/CN114417889A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20210201079A1 (en) * | 2019-12-27 | 2021-07-01 | Fujitsu Limited | Training data generation method and information processing apparatus |
CN111324744A (en) * | 2020-02-17 | 2020-06-23 | 中山大学 | Data enhancement method based on target emotion analysis data set |
CN113705249A (en) * | 2021-08-25 | 2021-11-26 | 上海云从企业发展有限公司 | Dialogue processing method, system, device and computer readable storage medium |
Non-Patent Citations (1)
Title |
---|
李培芸;李茂西;裘白莲;王明文;: "融合BERT语境词向量的译文质量估计方法研究", 中文信息学报, no. 03, 15 March 2020 (2020-03-15) * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111931513B (en) | Text intention recognition method and device | |
You et al. | Image captioning at will: A versatile scheme for effectively injecting sentiments into image descriptions | |
US20230244879A1 (en) | Language representation model system, pre-training method and apparatus, device, and medium | |
CN111914085B (en) | Text fine granularity emotion classification method, system, device and storage medium | |
CN109977207A (en) | Talk with generation method, dialogue generating means, electronic equipment and storage medium | |
CN108563779B (en) | Neural network-based template-free natural language text answer generation method | |
US11475225B2 (en) | Method, system, electronic device and storage medium for clarification question generation | |
US20230395075A1 (en) | Human-machine dialogue system and method | |
CN112562669B (en) | Method and system for automatically abstracting intelligent digital newspaper and performing voice interaction chat | |
CN109933773B (en) | Multiple semantic statement analysis system and method | |
CN110472255A (en) | Neural network machine interpretation method, model, electric terminal and storage medium | |
Hani et al. | Image caption generation using a deep architecture | |
Zhu et al. | Prompt-based learning for unpaired image captioning | |
CN111128175B (en) | Spoken language dialogue management method and system | |
Zhu et al. | A prompt model with combined semantic refinement for aspect sentiment analysis | |
CN114328866A (en) | Strong anthropomorphic intelligent dialogue robot with smooth and accurate response | |
Zhang et al. | Chatbot design method using hybrid word vector expression model based on real telemarketing data | |
CN114417889A (en) | Task type dialogue system and method based on seq2seq framework | |
CN116644759B (en) | Method and system for extracting aspect category and semantic polarity in sentence | |
CN116403608A (en) | Speech emotion recognition method based on multi-label correction and space-time collaborative fusion | |
CN115841106A (en) | Aspect-level emotion analysis method based on graph convolution network | |
Lei et al. | Zero-shot state tracking and user adoption tracking on schema-guided dialogue | |
Guo et al. | Ernie-bilstm based Chinese text sentiment classification method | |
CN114330367A (en) | Semantic similarity obtaining method, device and equipment based on sentences | |
Mecik et al. | Neural language generation for a Turkish task-oriented dialogue system |
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 |