CN111061850A - Dialog state tracking method, system and device based on information enhancement - Google Patents
Dialog state tracking method, system and device based on information enhancement Download PDFInfo
- Publication number
- CN111061850A CN111061850A CN201911276031.9A CN201911276031A CN111061850A CN 111061850 A CN111061850 A CN 111061850A CN 201911276031 A CN201911276031 A CN 201911276031A CN 111061850 A CN111061850 A CN 111061850A
- Authority
- CN
- China
- Prior art keywords
- state
- dialog
- slot
- dialogue
- value
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 106
- 238000013528 artificial neural network Methods 0.000 claims abstract description 6
- 239000013598 vector Substances 0.000 claims description 36
- 238000011156 evaluation Methods 0.000 claims description 20
- 238000012549 training Methods 0.000 claims description 13
- 230000002457 bidirectional effect Effects 0.000 claims description 6
- 230000002093 peripheral effect Effects 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 230000014509 gene expression Effects 0.000 claims description 4
- 238000012935 Averaging Methods 0.000 claims description 2
- 230000003190 augmentative effect Effects 0.000 claims 1
- 238000003058 natural language processing Methods 0.000 abstract description 2
- 230000008569 process Effects 0.000 description 11
- 238000010586 diagram Methods 0.000 description 4
- 230000002787 reinforcement Effects 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 235000013305 food Nutrition 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 235000012054 meals Nutrition 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000000126 substance Substances 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
- G06F16/33—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
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (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)
- Mathematical Physics (AREA)
- Human Computer Interaction (AREA)
- Computational Linguistics (AREA)
- Databases & Information Systems (AREA)
- Machine Translation (AREA)
Abstract
The invention belongs to the technical field of natural language processing, and particularly relates to a dialog state tracking method, a dialog state tracking system and a dialog state tracking device based on information enhancement, aiming at solving the problem that the accuracy of an unknown slot value generated by only utilizing context information of a dialog text in the conventional dialog state tracking method is poor. The system method comprises the steps that based on a dialog text of a user at the t moment, a dialog state at the t moment is obtained through a dialog state tracking model; the dialogue state at the time t comprises one or more slot value pairs and corresponding probabilities; the slot value pair comprises a slot and a slot value; the dialog state tracking model is constructed based on a neural network of an encoder-decoder architecture. The invention improves the accuracy of the generation of the unknown slot value.
Description
Technical Field
The invention belongs to the technical field of natural language processing, and particularly relates to a dialogue state tracking method, system and device based on information enhancement.
Background
With the rapid development of the internet and the popularization of intelligent devices, more and more conversation systems appear in the daily life of people. These dialog systems tend to assist people in performing a number of tasks. For example, the device can help people to order food, play music and other tasks. In the intelligent dialogue system, how to understand the user intention and clarify the user request becomes the most important part of the dialogue system. Dialog state tracking is emerging to address this issue.
Dialog state tracking is an important part of task-oriented dialog systems and one of the difficult problems in dialog systems. The task is to automatically recognize structured user states from a plurality of rounds of dialogs with various expressions. These structured states are typically composed of some predefined slots and their slot values. For example, in a meal ordering task, there will be slots of 'food type', 'restaurant location' and 'price', as well as some known values (slot values in a predefined state library).
Conventional dialog state tracking methods are mostly based on one such assumption. All the slots and the slot values are predefined in advance, so that the dialogue state tracking task can be simplified into a classification task, model design can be facilitated through the simplification, and the performance of the model on a data set with a limit value is improved. However, these models have difficulty dealing with unknown bin values that are not in the predefined state library bin value pairs. To deal with unknown slot values, there have been methods that generate unknown slot values from similar context information by selecting words from the dialog as slot values, but this method can only utilize the context information in the dialog. Because of the variety of the utterance expression modes in the dialog, the context information often has no obvious regularity, so that a model which only utilizes the context information to generate an unknown slot value is often unreliable in a complex context, and the application of a dialog system in practice is influenced.
Disclosure of Invention
In order to solve the above-mentioned problem in the prior art, that is, to solve the problem that the accuracy of an unknown slot value generated by using only context information of a dialog text in the conventional dialog state tracking method is poor, in a first aspect of the present invention, a method for dialog state tracking based on information enhancement is provided, the method comprising:
based on the dialog text of the user at the time t, obtaining the dialog state at the time t through a dialog state tracking model; the dialogue state at the time t comprises one or more slot value pairs and corresponding probabilities; the slot value pair comprises a slot and a slot value;
the dialog state tracking model is constructed based on a neural network of an encoder-decoder architecture; the training method of the model comprises the following steps:
a100, acquiring a dialog text of a user at a time t based on a labeled dialog set, and acquiring the semantic association degree of each word in the dialog text with a slot and a slot value of each slot value pair in a predefined state library by a preset semantic association degree evaluation method;
step A200, obtaining the dialog state at the time t through a dialog state tracking model based on the dialog text at the time t of the user, the dialog state at the time t-1 and the semantic association degree; calculating a loss value of the dialogue state tracking model, and updating parameters of the dialogue state tracking model;
step A300, making t equal to t +1, and circularly executing the method of the steps A100-A200 until the session is ended;
step A400, based on the corresponding conversation states at all times, obtaining the clustered conversation states by a preset semantic clustering method, and adding the clustered conversation states into the predefined state library;
and step A500, updating the dialogue states at all times by adopting the method of the steps A100-A200, calculating the loss value of the dialogue state tracking model, and updating the parameters of the dialogue state tracking model to obtain the trained dialogue state tracking model.
In some preferred embodiments, the preset semantic relevance evaluation method includes a central relevance evaluation method and a peripheral relevance evaluation method;
the central association degree evaluation method obtains semantic association degree by solving the cosine similarity of each word in the dialog text and the word vector of each slot value to the middle slot;
the peripheral association degree evaluation method obtains semantic association degree by solving the cosine similarity between each word in the dialog text and the slot value in each slot value pair.
In some preferred embodiments, the dialog state tracking model includes a dialog utterance encoder, a previous dialog state encoder, a known value encoder, a current dialog state decoder;
the dialogue utterance encoder is used for replying the dialogue text of the user based on the dialogue text at the time t of the user and the dialogue text of the user at the time t-1, performing word vector splicing by a Glove word embedding method and an n-gram word embedding method, and obtaining dialogue utterance encoding expression by a spliced word vector sequence through a bidirectional long-short memory network;
the last step dialogue state encoder is used for splicing the slot value pairs in the dialogue state at the t-1 moment, and the spliced sequence is expressed by the last step dialogue state encoding through a bidirectional long-short memory network;
the known value encoder is used for splicing the slot values in each slot value pair of the predefined state library, and the spliced sequence is subjected to two-way long-short memory network to obtain the coded representation of the known slot value;
and the current conversation state decoder is used for decoding the conversation state at the time t through a bidirectional long-short memory network based on the conversation history representation formed by splicing the conversation utterance coding representation and the last conversation state coding representation and the slots in the value pairs of the slots of the current conversation state.
In some preferred embodiments, the word vector concatenation is performed by a Glove word embedding method or an n-gram word embedding method, and the method comprises the following steps:
based on a dialog text of a user at a time t and a dialog text of a system reply user at a time t-1, acquiring corresponding word vectors by a Glove word embedding method and an n-gram word embedding method respectively and splicing to obtain a user dialog word vector sequence and a system reply word vector sequence;
and splicing the user dialogue word vector sequence and the system reply word vector sequence.
In some preferred embodiments, the method of "obtaining the dialog state at time t by the dialog state tracking model" in step a200 is as follows:
obtaining the probability of each slot value pair in the conversation state at the time t, and if the slot values correspond to the middle slot value and the slot value in a one-to-one correspondence manner, selecting a known value with the highest probability and the probability greater than a preset probability threshold value as the slot value; the known value is a slot value in a predefined state library;
if the slot value is in a one-to-many relationship with the middle slot and the slot value, selecting a plurality of known values corresponding to the probability greater than a preset probability threshold value as the slot value corresponding to the slot;
otherwise, acquiring the reward value of each word in the conversation history representation according to a preset reward rule; and multiplying the probability corresponding to the slot value pair with the slot value in the conversation state, wherein the slot value is one-to-many, and taking the word with the highest probability value after multiplication as the conversation state.
In some preferred embodiments, the method of obtaining the reward value of each word in the dialog history representation according to a preset reward rule includes:
randomly acquiring a slot of a slot value to be acquired in a dialogue state of words in the dialogue history representation and at the time t and semantic association degrees of known values corresponding to the slot, and averaging all the semantic association degrees to obtain an average association degree;
if the selected word is the target word, the reward value is a preset reward value; if the average relevance of the selected words exceeds a preset threshold value, the reward value is the product of the preset reward value and the average relevance; otherwise, the reward value corresponding to the selected word is set to be 0.
In some preferred embodiments, in step a400, "obtaining the clustered dialog states by using a preset semantic clustering method" includes:
step A410, sorting each slot value pair in a first dialogue state according to the semantic association degree of each slot value pair in the first dialogue state and each slot value pair in a predefined state library, and selecting N slot value pairs as a second dialogue state; the first conversation state is a conversation state corresponding to all moments of the current conversation;
step A420, sorting the semantic association degrees of the slot values in the slot value pairs of the predefined state library according to the slot values of the second dialogue state, and selecting M slot value pairs as a third dialogue state;
and step A430, deleting the slot value pairs containing the set word part of speech and the stop word in the third dialogue state, and taking the rest slot value pairs as the dialogue state after clustering.
In a second aspect of the present invention, a system for session state tracking based on information enhancement is provided, the system includes an obtaining module and a session state tracking module;
the acquisition module is configured to acquire a dialog text of a user at the moment t as input information;
the dialogue state tracking module is configured to obtain a dialogue state at the time t through a dialogue state tracking model based on the input information; the dialogue state at the time t comprises one or more slot value pairs and corresponding probabilities; the slot value pair includes a slot and a slot value.
In a third aspect of the present invention, a storage device is provided, in which a plurality of programs are stored, and the program applications are loaded and executed by a processor to implement the above-mentioned dialog state tracking method based on information enhancement.
In a fourth aspect of the present invention, a processing apparatus is provided, which includes a processor, a storage device; a processor adapted to execute various programs; a storage device adapted to store a plurality of programs; the program is adapted to be loaded and executed by a processor to implement the above-described information-enhancement-based dialog state tracking method.
The invention has the beneficial effects that:
the invention improves the accuracy of the generation of the unknown slot value. The invention combines the semantic relation between each word and the slot-slot value in the user dialogue text as the enhanced information through two semantic association degree evaluation methods, and trains the dialogue state tracking model by adopting a reinforcement learning method. The method solves the technical problem of only utilizing context, can overcome the defects that the traditional method generates irrelevant values and can not generate rarely mentioned values, and improves the accuracy of state tracking.
Drawings
Other features, objects and advantages of the present application will become more apparent upon reading of the following detailed description of non-limiting embodiments thereof, made with reference to the accompanying drawings.
FIG. 1 is a schematic diagram of a training process of a dialog state tracking model based on an information-enhanced dialog state tracking method according to an embodiment of the present invention;
FIG. 2 is a block diagram of a dialog state tracking system based on information enhancement according to an embodiment of the present invention;
FIG. 3 is a diagram of training a conversation state tracking model based on a user's conversation in booking a restaurant, in accordance with one embodiment of the present invention;
fig. 4 is a schematic diagram of the effect of the invention compared to other methods according to an embodiment of the invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some, but not all embodiments of the present invention. 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 present application will be described in further detail with reference to the following drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the relevant invention and not restrictive of the invention. It should be noted that, for convenience of description, only the portions related to the related invention are shown in the drawings.
It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict.
The invention discloses a dialogue state tracking method based on information enhancement, which comprises the following steps:
based on the dialog text of the user at the time t, obtaining the dialog state at the time t through a dialog state tracking model; the dialogue state at the time t comprises one or more slot value pairs and corresponding probabilities; the slot value pair comprises a slot and a slot value;
the dialog state tracking model is constructed based on a neural network of an encoder-decoder architecture; the training method of the model comprises the following steps:
a100, acquiring a dialog text of a user at a time t based on a labeled dialog set, and acquiring the semantic association degree of each word in the dialog text with a slot and a slot value of each slot value pair in a predefined state library by a preset semantic association degree evaluation method;
step A200, obtaining the dialog state at the time t through a dialog state tracking model based on the dialog text at the time t of the user, the dialog state at the time t-1 and the semantic association degree; calculating a loss value of the dialogue state tracking model, and updating parameters of the dialogue state tracking model;
step A300, making t equal to t +1, and circularly executing the method of the steps A100-A200 until the session is ended;
step A400, based on the corresponding conversation states at all times, obtaining the clustered conversation states by a preset semantic clustering method, and adding the clustered conversation states into the predefined state library;
and step A500, updating the dialogue states at all times by adopting the method of the steps A100-A200, calculating the loss value of the dialogue state tracking model, and updating the parameters of the dialogue state tracking model to obtain the trained dialogue state tracking model.
In order to more clearly describe the dialog state tracking method based on information enhancement, the following describes in detail the steps of an embodiment of the method in conjunction with the accompanying drawings.
In the following preferred embodiment, the training method of the dialog state tracking model is first described in detail, and the dialog state obtained by the dialog state tracking method based on information enhancement is described in detail.
1. Training method of dialogue state tracking model
Step A100, obtaining a dialog text of a user at a time t based on a labeled dialog set, and obtaining the semantic association degree of each word in the dialog text with a slot and a slot value of each slot value pair in a predefined state library through a preset semantic association degree evaluation method.
In this embodiment, a dialog text at the time t of the user is obtained from a labeled dialog set, and the association between each word in the dialog and an ontology (a predefined state library including a plurality of slot value pairs), that is, the semantic association between each word in the dialog text and the slot and slot value of the predefined state library slot value pair, is evaluated. Regarding the evaluation of semantic relevance, the invention designs two methods: a center relevance evaluation method and a periphery relevance evaluation method.
The central relevance evaluation method is used for evaluating the semantic relation between the words in the dialog text and the slots of each slot value pair in the predefined state library. This degree of association is based on the assumption that values within the same slot in a conversation should have a significant semantic relationship with the slot.
The peripheral relevance evaluation method is used for evaluating the semantic relation between the words in the dialog text and the slot values of the slot value pairs of the predefined state library.
The semantic relationship is evaluated by cosine similarity of the word vectors. For each slot, we use the average of the semantic relationships of words to slots and slot values as their semantic relatedness.
Step A200, obtaining the dialog state at the time t through a dialog state tracking model based on the dialog text at the time t of the user, the dialog state at the time t-1 and the semantic association degree; and calculating the loss value of the dialogue state tracking model and updating the parameters of the dialogue state tracking model.
In the present embodiment, the dialog state tracking model is constructed based on a neural network of an encoder-decoder architecture, and the model includes a dialog utterance encoder, a previous dialog state encoder, a known value encoder, and a current dialog state decoder.
The conversational speech encoder is primarily used to encode the current user conversational text and the system response of the previous step (the system replies to the user's conversational text at the previous time) into a vector representation with context information for subsequent calculations. The method comprises the following specific steps:
firstly, based on a dialog text of a user at a time t and a dialog text of a system reply user at a time t-1, corresponding word vectors are obtained through a Glove word embedding method and an n-gram word embedding method respectively and are spliced to obtain a user dialog word vector sequence and a system reply word vector sequence. And splicing the user dialogue word vector sequence and the system reply word vector sequence to be used as the word vector input of the initial dialogue utterance.
The word vectors of the initial conversational utterances are then encoded cyclically using a bi-directional long-short memory network
Wherein, BilSTM is the long and short memory network coder used by us,
is the word vector of the initial conversational utterances we input, bi-1The word vector for the initial conversational utterance at the previous time instant, i is the subscript. Therefore, each coded word vector can be provided with context information, and a more accurate conversational speech coding representation can be obtained.
The previous dialog state encoder is mainly used for encoding the user state (the dialog state at the previous moment) obtained in the previous dialog. This is mainly due to the inheritance of user states. If the user expressed some intent in a previous conversation turn. These intentions will be known by the system by default in later sessions. We therefore need to encode the last dialog state in the current round. The slot value pairs in the previous dialog state are spliced into a sequence, and then the sequence is encoded by using another two-way long-short memory network to obtain the previous dialog state encoding representation containing the context information. This representation is concatenated with the conversational speech encoder representation to yield a complete conversational history representation. The user may select a word from the dialog history representation as the dialog state. I.e. the word is used as information for the slot.
The known value encoder is primarily used to encode each known value (the bin value in each bin value pair of the predefined state library) into a vector representation. Some of the known values are composed of multiple words and we need a representation of their entirety for generation. Similarly, we encode the known values using a two-way long-short memory network, and we use the last vector representation output by the encoder as the known bin encoded representation.
The dialog state decoder takes the dialog history representation with the slot and attention weight as the decoder input, and obtains the decoding vector qs。qs＝LSTM(cs,[zs,es]) Wherein e issIs the slot name, zsIs a dialog history representation with attention weight, csIs a hidden layer we get with the self-attention mechanism. c. CsThe obtaining process of (2) is shown in the formulas (1), (3):
Ps＝softmax(as) (2)
wherein, asFor non-normalized weight distribution, Ws、bsFor trainable parameters, M is a vector representation of the dialog history, PsFor normalized weight distribution, dmIs the length of M and i is subscript.
The process of obtaining a dialog history representation with attention weight is shown in equations (4) (5) (6):
Pz＝softmax(az) (5)
wherein v isz、We、WmFor trainable parameters, azFor non-normalized weight distribution, PzIs a normalized weight distribution.
Then we can get the decoded state q by using a long and short memory network as the decoders，qsProbability P that can be used to obtain a known valuevsThe calculation is shown in equation (7):
Pvs(v|s)＝sigmoid(qs*Vs) (7)
wherein, is the dot product operation, VsIs a vector representation of known values.
So that we can select a known value for each bin by this probability. For the selection of words in dialog text, we use the calculation method (8) (9) to obtain the probability Pws：
Pws(w|s)＝softmax(aw) (9)
Wherein v iso、WoFor trainable parameters, awRepresented as non-normalized weights.
This probability may be a word selected as the probability for the current slot. However, in the dialog state tracking task, there is often a case that one slot is composed of a plurality of slot values, and the probability is obtained by using the method (10):
Pws(w|s)＝sigmoid(aw) (10)
for a slot containing one slot value, we select a known value with the highest probability and exceeding a threshold (preferably 0.5 in the present invention), if there is no known value that satisfies the condition, we select as the value a word of such condition, for a slot containing multiple slot values, we select all values with probabilities exceeding the threshold, and if there is no word that satisfies the condition in the dialog history representation.
For the probability PvsWe train using the cross-entropy loss of label labels and probabilities in the dataset, for PwsWe use the pointer of the target word in the dialogue as the training target, and then train by the reinforced learning of the cross entropy and semantic information of the training target and the prediction target, namely, the training is carried outAnd selecting a slot value by acquiring the probability of the words in the dialog text through reinforcement learning. The reinforcement learning is described in detail below:
taking words in the conversation history as an action space, and randomly sampling a word from the conversation space;
and evaluating the semantic relevance of the sampled words and the current slot. Calculating the semantic relations between the selected words and the current decoding slot and the known values of the slot, wherein the semantic relations are evaluated through cosine similarity of word vectors, and the average value of the semantic relations is taken as the degree of association;
if the selected word is just the target word, the prize value is determined to be the preset prize value, the invention is preferably set to 5, if the association degree exceeds the threshold value (0.68 in the first training and 0.75 in the second training), the prize value is set to 5 times the association degree, otherwise, the prize value is set to 0 if the word is stopped.
Based on the obtained reward value and the selection probability of the word in the dialogue text, the loss of reinforcement learning of the probability of selecting the word satisfying the condition as the slot value in the dialogue history is obtained, as shown in equation (11):
wherein the content of the first and second substances,
for the prize value corresponding to the selected word, LrlTo reinforce the loss of learning.
And step a300, making t equal to t +1, and executing the method of steps a100-a200 in a loop until the session is ended.
In this embodiment, based on the current dialog between the user and the system, the dialog state tracking model is trained in a loop until the end.
And step A400, based on the corresponding conversation states at all times, obtaining the clustered conversation states by a preset semantic clustering method, and adding the clustered conversation states into the predefined state library.
In this embodiment, based on the dialog states at all times in the current dialog process, we perform semantic clustering to obtain a complete ontology. The method comprises the following specific steps:
step A410, sorting the semantic relations (semantic association degrees) of the slot values of the conversation states at all times and the slots of each slot value pair in a predefined state library, and removing noise according to the length set manually, namely selecting the first N slot value pairs as a second conversation state;
step A420, sorting the semantic relations of the slot values in the predefined state library based on the slot values in the second dialogue state, and removing more noises according to the length set manually, namely selecting the first M slot values as a third dialogue state;
step A430, judging the part of speech of the word in the multiple utterances according to the part of speech rules, removing words with incorrect part of speech, selecting correct values according to the context words of the word, and removing stop words to obtain a clustered conversation state, namely deleting the slot value pairs containing the part of speech of the set word and the stop words in the third conversation state. For example, the slot value pairs of the dialog state are only allowed to be nouns, and the slot value pairs of other parts of speech are deleted. And adding the clustered conversation states into a predefined state library.
As shown in FIG. 3, the present invention takes the example of a user booking a restaurant to interact with the system, and a complete training process of the model can be seen. Wherein, the speech coder is the abbreviation of the dialogue speech coder.
And step A500, updating the dialogue states at all times by adopting the method of the steps A100-A200, calculating the loss value of the dialogue state tracking model, and updating the parameters of the dialogue state tracking model to obtain the trained dialogue state tracking model.
In this embodiment, based on the updated predefined state library, the semantic association and the dialog state corresponding to the dialog text at all times of the user are obtained again. Namely, the generalization capability of the model is improved, the values which do not appear in the predefined state library are added into the library, and the model is trained again based on the dialog text of the user until the trained dialog state tracking model is obtained.
2. Dialog state tracking method based on information enhancement
Based on the dialog text of the user at the time t, obtaining the dialog state at the time t through a dialog state tracking model; the dialogue state at the time t comprises one or more slot value pairs and corresponding probabilities; the slot value pair includes a slot and a slot value.
In this embodiment, based on the dialog text at the time t of the user, the trained dialog state tracking model obtains the dialog state at the time t. Other modules fed into the dialog system obtain the dialog's responses and policies.
In order to prove the effectiveness of the Dialog state tracking model of the present invention on the Dialog system, the Dialog state tracking Challenge 2 and Wizard of Oz 2 (simulation scenario Dialog 2 data set) corpora are used to perform testing, and the testing effect of the method and other existing methods is shown in table 1:
TABLE 1
In table 1, Models is the name of the model method, Joint is the Joint target accuracy, and Request is the accuracy of the Request slot. The comparison model comprises: a de-lexical network (Delexi.), a de-lexical + semantic dictionary (Delexi. + semdit), a neural state tracker-deep neural network (NBT-DNN), a neural state tracker-convolutional neural network (NBT-CNN), a global local attention network (GLAD), an extensible tracker (Scalable-DST), a pointer network (PtrNet), a pointer network + ontology (PtrNet + ont.), a copy enhanced tracker (CEDST), a copy enhanced tracker + ontology (CEDST + ont.).
The last one in Table 1 is the accuracy of the model method of the present invention, and it can be seen that on Wizard of Oz 2, the method exceeds all methods, and on Dialog State Tracking Change 2, exceeds all similar methods, but is slightly lower than GLAD, but GLAD utilizes the results of multiple speech recognitions in Dialog State Tracking Change 2, and it cannot handle unknown values, and the method of the present invention is not a similar method. As also shown in fig. 4, it can be seen that the present invention (TS-DST) has a good processing effect on less-mentioned (Rare occurrences) and Irrelevant values (Irrelevant values) compared to the prine method and the CEDST method.
A dialog state tracking system based on information enhancement according to a second embodiment of the present invention, as shown in fig. 2, includes: the system comprises an acquisition module and a conversation state tracking module;
the acquisition module is configured to acquire a dialog text of a user at the moment t as input information;
the dialogue state tracking module is configured to obtain a dialogue state at the time t through a dialogue state tracking model based on the input information; the dialogue state at the time t comprises one or more slot value pairs and corresponding probabilities; the slot value pair includes a slot and a slot value.
It can be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working process and related description of the system described above may refer to the corresponding process in the foregoing method embodiment, and details are not described herein again.
It should be noted that, the dialog state tracking system based on information enhancement provided in the foregoing embodiment is only illustrated by the division of the foregoing functional modules, and in practical applications, the above functions may be allocated to different functional modules according to needs, that is, the modules or steps in the embodiment of the present invention are further decomposed or combined, for example, the modules in the foregoing embodiment may be combined into one module, or may be further split into multiple sub-modules, so as to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps, and are not to be construed as unduly limiting the present invention.
A storage device according to a third embodiment of the present invention stores therein a plurality of programs, which are adapted to be loaded by a processor and to implement the above-described information enhancement-based dialog state tracking method.
A processing apparatus according to a fourth embodiment of the present invention includes a processor, a storage device; a processor adapted to execute various programs; a storage device adapted to store a plurality of programs; the program is adapted to be loaded and executed by a processor to implement the above-described information-enhancement-based dialog state tracking method.
It can be clearly understood by those skilled in the art that, for convenience and brevity, the specific working processes and related descriptions of the storage device and the processing device described above may refer to the corresponding processes in the foregoing method examples, and are not described herein again.
Those of skill in the art would appreciate that the various illustrative modules, method steps, and modules described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that programs corresponding to the software modules, method steps may be located in Random Access Memory (RAM), memory, Read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. To clearly illustrate this interchangeability of electronic hardware and software, various illustrative components and steps have been described above generally in terms of their functionality. Whether these functions are performed in electronic hardware or software depends on the intended application of the solution and design constraints. Those skilled in the art may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The terms "first," "second," and the like are used for distinguishing between similar elements and not necessarily for describing or implying a particular order or sequence.
The terms "comprises," "comprising," or any other similar term are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
So far, the technical solutions of the present invention have been described in connection with the preferred embodiments shown in the drawings, but it is easily understood by those skilled in the art that the scope of the present invention is obviously not limited to these specific embodiments. Equivalent changes or substitutions of related technical features can be made by those skilled in the art without departing from the principle of the invention, and the technical scheme after the changes or substitutions can fall into the protection scope of the invention.
Claims (10)
1. A dialogue state tracking method based on information enhancement is characterized by comprising the following steps:
based on the dialog text of the user at the time t, obtaining the dialog state at the time t through a dialog state tracking model; the dialogue state at the time t comprises one or more slot value pairs and corresponding probabilities; the slot value pair comprises a slot and a slot value;
the dialog state tracking model is constructed based on a neural network of an encoder-decoder architecture; the training method of the model comprises the following steps:
a100, acquiring a dialog text of a user at a time t based on a labeled dialog set, and acquiring the semantic association degree of each word in the dialog text with a slot and a slot value of each slot value pair in a predefined state library by a preset semantic association degree evaluation method;
step A200, obtaining the dialog state at the time t through a dialog state tracking model based on the dialog text at the time t of the user, the dialog state at the time t-1 and the semantic association degree; calculating a loss value of the dialogue state tracking model, and updating parameters of the dialogue state tracking model;
step A300, making t equal to t +1, and circularly executing the method of the steps A100-A200 until the session is ended;
step A400, based on the corresponding conversation states at all times, obtaining the clustered conversation states by a preset semantic clustering method, and adding the clustered conversation states into the predefined state library;
and step A500, updating the dialogue states at all times by adopting the method of the steps A100-A200, calculating the loss value of the dialogue state tracking model, and updating the parameters of the dialogue state tracking model to obtain the trained dialogue state tracking model.
2. The information enhancement-based dialog state tracking method according to claim 1, wherein the preset semantic association degree evaluation method comprises a central association degree evaluation method and a peripheral association degree evaluation method;
the central association degree evaluation method obtains semantic association degree by solving the cosine similarity of each word in the dialog text and the word vector of each slot value to the middle slot;
the peripheral association degree evaluation method obtains semantic association degree by solving the cosine similarity between each word in the dialog text and the slot value in each slot value pair.
3. The information enhancement based dialog state tracking method of claim 1 wherein the dialog state tracking model comprises a dialog utterance encoder, a previous dialog state encoder, a known value encoder, a current dialog state decoder;
the dialogue utterance encoder is used for replying the dialogue text of the user based on the dialogue text at the time t of the user and the dialogue text of the user at the time t-1, performing word vector splicing by a Glove word embedding method and an n-gram word embedding method, and obtaining dialogue utterance encoding expression by a spliced word vector sequence through a bidirectional long-short memory network;
the last step dialogue state encoder is used for splicing the slot value pairs in the dialogue state at the t-1 moment, and the spliced sequence is expressed by the last step dialogue state encoding through a bidirectional long-short memory network;
the known value encoder is used for splicing the slot values in each slot value pair of the predefined state library, and the spliced sequence is subjected to two-way long-short memory network to obtain the coded representation of the known slot value;
and the current conversation state decoder is used for decoding the conversation state at the time t through a bidirectional long-short memory network based on the conversation history representation formed by splicing the conversation utterance coding representation and the last conversation state coding representation and the slots in the value pairs of the slots of the current conversation state.
4. The information enhancement-based dialog state tracking method according to claim 3, wherein word vector concatenation is performed by a Glove word embedding method and an n-gram word embedding method, and the method comprises:
based on a dialog text of a user at a time t and a dialog text of a system reply user at a time t-1, acquiring corresponding word vectors by a Glove word embedding method and an n-gram word embedding method respectively and splicing to obtain a user dialog word vector sequence and a system reply word vector sequence;
and splicing the user dialogue word vector sequence and the system reply word vector sequence.
5. The information-enhancement-based dialog state tracking method according to claim 3, wherein the method for obtaining the dialog state at time t through the dialog state tracking model in step A200 is as follows:
obtaining the probability of each slot value pair in the conversation state at the time t, and if the slot values correspond to the middle slot value and the slot value in a one-to-one correspondence manner, selecting a known value with the highest probability and the probability greater than a preset probability threshold value as the slot value; the known value is a slot value in a predefined state library;
if the slot value is in a one-to-many relationship with the middle slot and the slot value, selecting a plurality of known values corresponding to the probability greater than a preset probability threshold value as the slot value corresponding to the slot;
otherwise, acquiring the reward value of each word in the conversation history representation according to a preset reward rule; and multiplying the probability corresponding to the slot value pair with the slot value in the conversation state, wherein the slot value is one-to-many, and taking the word with the highest probability value after multiplication as the conversation state.
6. The information enhancement-based dialog state tracking method according to claim 5, wherein the method for obtaining the reward value of each word in the dialog history representation according to a preset reward rule comprises the following steps:
randomly acquiring a slot of a slot value to be acquired in a dialogue state of words in the dialogue history representation and at the time t and semantic association degrees of known values corresponding to the slot, and averaging all the semantic association degrees to obtain an average association degree;
if the selected word is the target word, the reward value is a preset reward value; if the average relevance of the selected words exceeds a preset threshold value, the reward value is the product of the preset reward value and the average relevance; otherwise, the reward value corresponding to the selected word is set to be 0.
7. The method for tracking dialog state based on information enhancement according to claim 1, wherein in step a400, "obtaining the clustered dialog state by a preset semantic clustering method" comprises:
step A410, sorting each slot value pair in a first dialogue state according to the semantic association degree of each slot value pair in the first dialogue state and each slot value pair in a predefined state library, and selecting N slot value pairs as a second dialogue state; the first conversation state is a conversation state corresponding to all moments of the current conversation;
step A420, sorting the semantic association degrees of the slot values in the slot value pairs of the predefined state library according to the slot values of the second dialogue state, and selecting M slot value pairs as a third dialogue state;
and step A430, deleting the slot value pairs containing the set word part of speech and the stop word in the third dialogue state, and taking the rest slot value pairs as the dialogue state after clustering.
8. A dialogue state tracking system based on information enhancement is characterized by comprising an acquisition module and a dialogue state tracking module;
the acquisition module is configured to acquire a dialog text of a user at the moment t as input information;
the dialogue state module is configured to obtain a dialogue state at the time t through a dialogue state tracking model based on the input information; the dialogue state at the time t comprises one or more slot value pairs and corresponding probabilities; the slot value pair includes a slot and a slot value.
9. A storage device having a plurality of programs stored therein, wherein the program applications are loaded and executed by a processor to implement the information-based augmented session state tracking method of any one of claims 1-7.
10. A processing device comprising a processor, a storage device; a processor adapted to execute various programs; a storage device adapted to store a plurality of programs; characterized in that said program is adapted to be loaded and executed by a processor to implement the information enhancement based dialog state tracking method of any of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911276031.9A CN111061850B (en) | 2019-12-12 | 2019-12-12 | Dialogue state tracking method, system and device based on information enhancement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911276031.9A CN111061850B (en) | 2019-12-12 | 2019-12-12 | Dialogue state tracking method, system and device based on information enhancement |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111061850A true CN111061850A (en) | 2020-04-24 |
| CN111061850B CN111061850B (en) | 2023-04-28 |
Family
ID=70300726
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911276031.9A Active CN111061850B (en) | 2019-12-12 | 2019-12-12 | Dialogue state tracking method, system and device based on information enhancement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111061850B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111651573A (en) * | 2020-05-26 | 2020-09-11 | 上海智臻智能网络科技股份有限公司 | Intelligent customer service dialogue reply generation method and device and electronic equipment |
| CN111753076A (en) * | 2020-08-12 | 2020-10-09 | 腾讯科技（深圳）有限公司 | Dialogue method, dialogue device, electronic equipment and readable storage medium |
| CN111859902A (en) * | 2020-07-16 | 2020-10-30 | 微医云(杭州)控股有限公司 | Text processing method, device, equipment and medium |
| CN112131372A (en) * | 2020-11-25 | 2020-12-25 | 中国科学院自动化研究所 | Knowledge-driven conversation strategy network optimization method, system and device |
| CN112766319A (en) * | 2020-12-31 | 2021-05-07 | 平安科技（深圳）有限公司 | Dialogue intention recognition model training method and device, computer equipment and medium |
| CN112988960A (en) * | 2021-02-09 | 2021-06-18 | 中国科学院自动化研究所 | Dialog state tracking method, device, equipment and storage medium |
| CN112990292A (en) * | 2021-03-10 | 2021-06-18 | 平安科技（深圳）有限公司 | Neural network-based dialog state generation method and device |
| CN113705652A (en) * | 2021-08-23 | 2021-11-26 | 西安交通大学 | Task type conversation state tracking system and method based on pointer generation network |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070148627A1 (en) * | 2005-12-22 | 2007-06-28 | Katherine Assimon Dusckas | Merchandising method and system |
| CN108282587A (en) * | 2018-01-19 | 2018-07-13 | 重庆邮电大学 | Mobile customer service dialogue management method under being oriented to strategy based on status tracking |
| CN109726396A (en) * | 2018-12-20 | 2019-05-07 | 泰康保险集团股份有限公司 | Semantic matching method, device, medium and the electronic equipment of question and answer text |
| US10387575B1 (en) * | 2019-01-30 | 2019-08-20 | Babylon Partners Limited | Semantic graph traversal for recognition of inferred clauses within natural language inputs |
| CN110245221A (en) * | 2019-05-13 | 2019-09-17 | 华为技术有限公司 | The method and computer equipment of training dialogue state tracking classifier |
-
2019
- 2019-12-12 CN CN201911276031.9A patent/CN111061850B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070148627A1 (en) * | 2005-12-22 | 2007-06-28 | Katherine Assimon Dusckas | Merchandising method and system |
| CN108282587A (en) * | 2018-01-19 | 2018-07-13 | 重庆邮电大学 | Mobile customer service dialogue management method under being oriented to strategy based on status tracking |
| CN109726396A (en) * | 2018-12-20 | 2019-05-07 | 泰康保险集团股份有限公司 | Semantic matching method, device, medium and the electronic equipment of question and answer text |
| US10387575B1 (en) * | 2019-01-30 | 2019-08-20 | Babylon Partners Limited | Semantic graph traversal for recognition of inferred clauses within natural language inputs |
| CN110245221A (en) * | 2019-05-13 | 2019-09-17 | 华为技术有限公司 | The method and computer equipment of training dialogue state tracking classifier |
Non-Patent Citations (2)
| Title |
|---|
| 任航等: "口语对话状态追踪的研究" * |
| 陈红燕: "面向任务的对话状态追踪方法和应用" * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111651573B (en) * | 2020-05-26 | 2023-09-05 | 上海智臻智能网络科技股份有限公司 | Intelligent customer service dialogue reply generation method and device and electronic equipment |
| CN111651573A (en) * | 2020-05-26 | 2020-09-11 | 上海智臻智能网络科技股份有限公司 | Intelligent customer service dialogue reply generation method and device and electronic equipment |
| CN111859902A (en) * | 2020-07-16 | 2020-10-30 | 微医云(杭州)控股有限公司 | Text processing method, device, equipment and medium |
| CN111753076A (en) * | 2020-08-12 | 2020-10-09 | 腾讯科技（深圳）有限公司 | Dialogue method, dialogue device, electronic equipment and readable storage medium |
| CN111753076B (en) * | 2020-08-12 | 2022-08-26 | 腾讯科技（深圳）有限公司 | Dialogue method, dialogue device, electronic equipment and readable storage medium |
| CN112131372A (en) * | 2020-11-25 | 2020-12-25 | 中国科学院自动化研究所 | Knowledge-driven conversation strategy network optimization method, system and device |
| CN112766319A (en) * | 2020-12-31 | 2021-05-07 | 平安科技（深圳）有限公司 | Dialogue intention recognition model training method and device, computer equipment and medium |
| CN112766319B (en) * | 2020-12-31 | 2024-06-21 | 平安科技（深圳）有限公司 | Dialogue intention recognition model training method, device, computer equipment and medium |
| CN112988960A (en) * | 2021-02-09 | 2021-06-18 | 中国科学院自动化研究所 | Dialog state tracking method, device, equipment and storage medium |
| CN112990292A (en) * | 2021-03-10 | 2021-06-18 | 平安科技（深圳）有限公司 | Neural network-based dialog state generation method and device |
| CN112990292B (en) * | 2021-03-10 | 2023-07-11 | 平安科技（深圳）有限公司 | Method and device for generating dialogue state based on neural network |
| CN113705652B (en) * | 2021-08-23 | 2024-05-28 | 西安交通大学 | Task type dialogue state tracking system and method based on pointer generation network |
| CN113705652A (en) * | 2021-08-23 | 2021-11-26 | 西安交通大学 | Task type conversation state tracking system and method based on pointer generation network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111061850B (en) | 2023-04-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111061850A (en) | Dialog state tracking method, system and device based on information enhancement | |
| CN108062388B (en) | Reply generation method and device for man-machine conversation | |
| CN111914067B (en) | Chinese text matching method and system | |
| CN108846077B (en) | Semantic matching method, device, medium and electronic equipment for question and answer text | |
| JP2021524623A (en) | Multitask learning as a question answering | |
| CN111026857B (en) | Conversation state tracking method, man-machine conversation method and system | |
| CN111930914B (en) | Problem generation method and device, electronic equipment and computer readable storage medium | |
| CN110069612B (en) | Reply generation method and device | |
| CN110570879A (en) | Intelligent conversation method and device based on emotion recognition and computer equipment | |
| CN109637527B (en) | Semantic analysis method and system for dialogue statement | |
| CN113553860A (en) | Reply diversity multi-round conversation generation method and system based on multi-task learning | |
| CN116737938A (en) | Fine granularity emotion detection method and device based on fine tuning large model online data network | |
| CN110084323A (en) | End-to-end semanteme resolution system and training method | |
| CN111400468B (en) | Conversation state tracking system and method, man-machine conversation device and method | |
| CN111859940A (en) | Keyword extraction method and device, electronic equipment and storage medium | |
| CN110597968A (en) | Reply selection method and device | |
| CN113239702A (en) | Intention recognition method and device and electronic equipment | |
| CN110992943A (en) | Semantic understanding method and system based on word confusion network | |
| CN111814479B (en) | Method and device for generating enterprise abbreviations and training model thereof | |
| CN116402064B (en) | Comment generation method, comment generation system, storage medium and electronic equipment | |
| CN115512692B (en) | Voice recognition method, device, equipment and storage medium | |
| CN114707518A (en) | Semantic fragment-oriented target emotion analysis method, device, equipment and medium | |
| CN114398875A (en) | Training method of error correction model, search word error correction method, device and medium | |
| CN113836273A (en) | Legal consultation method based on complex context and related equipment | |
| CN112966520B (en) | Natural language generation method and device |
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 |