TWI802459B - A system and method for recommendation q&a based on data-enhanced - Google Patents

Abstract

A system and method for recommendation Q&A based on data-enhanced are provided, the system includes a transceiver, a storage medium and a processor. The storage medium stores a plurality of modules and database. The processor is coupled to the storage medium and the transceiver, accesses and executes a pre-training model selection module, a data enhancement module, an intent recognition module and a reply retrieval module. The pre-training model selection module correspondingly generates a specific domain training subset and a general domain training subset. The data enhancement module is used to generate augmented training data sets. The intent recognition module receives data input by an user and outputs category labels and keyword combinations. The reply retrieval module retrieves a best recommendation from the database and recommends the best recommendation to the user based on category labels and keyword combinations.

Description

System and method for data-based enhanced recommendation question answering

The present invention relates to a system and method for recommending questions and answers, and in particular to a system and method for recommending questions and answers based on data enhancement.

Supervised Learning relies on a large amount of labeled data for training. When training a classification model, if the number of training data sets is insufficient, the regularization ability of the model will be reduced, which may lead to overfitting in machine learning. This is because the limitation of the data range reduces the regularization ability of the model, so that it is difficult to effectively exert the effectiveness of the forward-looking algorithm. Training a model with good regularization ability is highly dependent on sufficient and good-quality data, but the cost of data collection is often very expensive.

The present invention provides a system and method for recommending questions and answers based on data enhancement, which expands the training data set by generating a high-quality enhanced data set, so as to avoid the problem of overfitting in machine learning.

A system for recommending questions and answers based on data enhancement of the present invention includes a transceiver, a storage medium and a processor. The storage medium stores multiple modules and databases. The processor is coupled to the storage medium and the transceiver, and accesses and executes the modules, wherein the modules include a pre-training model selection module, a data enhancement module, an intention recognition module and a reply retrieval module. The pre-training model selection module classifies the training data into domain-specific data and general domain data according to the training data, selects the domain-specific pre-training model and the general domain pre-training model, and generates domain-specific training subsets corresponding to the domain-specific data and general domain data and general domain training subsets. The data enhancement module is electrically connected to the pre-training model selection module, and candidate keywords are extracted respectively according to the specific field training subset and the general field training subset, and a category keyword list is generated, and the category keyword list is combined with the specific field pre-training Models and general domain pre-trained models are augmented to generate enhanced training datasets. The intent recognition module is electrically connected with the data enhancement module, receives the data input by the user to identify the category label of the data based on the classification model trained by the enhanced training data set and outputs the category label, and extracts and outputs the category keyword list and data based on the category Keyword combination of data. The reply retrieval module is electrically connected with the intent recognition module, and retrieves the best recommended information from the database based on the combination of category tags and keywords and recommends them to users.

In an embodiment of the present invention, in the above-mentioned data-based enhanced recommendation question answering system, the pre-training model selection module is further used to classify the training data into specific domain data and general domain data according to the probability threshold.

In an embodiment of the present invention, in the above-mentioned data-based enhanced question-and-answer system, wherein the domain-specific training subset is training data related to a specific domain, and the general domain training subset is training data that is not related to a specific domain .

In an embodiment of the present invention, in the above-mentioned system for recommending questions and answers based on data enhancement, the data enhancement module is further used to calculate the category vector of each category label in the specific domain training subset and the general domain training subset, and calculate the candidate keywords The similarity with the category vector of each category label to generate a category keyword list, generate similar words based on the category keyword list, specific domain pre-training model and general domain pre-training model, use similar words to expand the training data set and generate enhanced training data set.

In an embodiment of the present invention, the above-mentioned system for recommending questions and answers based on data enhancement, wherein the data enhancement module is further used to filter the stop words (Stop Words) in the specific field training subset and the general field training subset word and output word segmentation result information, and extract candidate keywords from the word segmentation result information based on statistical methods.

In an embodiment of the present invention, in the above-mentioned system for recommending questions and answers based on data enhancement, the intent recognition module is further used to train the classification model based on the enhanced training data set, verification data set and neural network architecture, so as to analyze the user's The category label prediction is performed on the input data after the dialogue is converted from speech to text, so as to obtain the category label of the data.

In an embodiment of the present invention, in the above-mentioned system for recommending question-and-answer based on data enhancement, the answer retrieval module includes a conversational semantic analysis module and a dialogue state tracking module.

In an embodiment of the present invention, in the above-mentioned system for recommending questions and answers based on data enhancement, the answer retrieval module is further used to store the user's dialogue after combining the category labels and keywords of the data through the conversational semantic analysis module In the dialog state tracking module, and calculate the similarity between the data and the recommended data stored in the database, and recommend the best recommended data to the user through the conversational semantic analysis module.

In an embodiment of the present invention, the above-mentioned system for recommending questions and answers based on data enhancement, wherein the answer retrieval module is further used based on the Top-N recommendation algorithm, through the conversational semantic analysis module according to the semantic strength, dialogue time and number of clicks Sort the recommendation data stored in the database to recommend the best recommendation data to the user.

A method for recommending questions and answers based on data enhancement of the present invention, comprising: classifying the training data into specific domain data and general domain data according to the training data, selecting a specific domain pre-training model and a general domain pre-training model to separate the specific domain data and general domain data Domain-specific training subsets and general domain training subsets are correspondingly generated from domain data; keywords are extracted respectively according to domain-specific training subsets and general domain training subsets, and a category keyword list is generated, and the category keyword list is combined with domain-specific pre-training The model and the general domain pre-training model perform data enhancement to generate an enhanced training data set; receive the data input by the user to identify the category label of the data based on the classification model trained on the enhanced training data set and output the category label, and based on the list of category keywords and Extracting data and outputting a keyword combination of the data; and retrieving the best recommendation data from a database based on the category label and the keyword combination and recommending to the user.

Based on the above, the present invention provides a system and method for recommending questions and answers based on data enhancement. The sentences in the training data set are divided into a specific domain training subset and a general domain training subset according to domain relevance. By extracting category keywords in the training subset And based on the pre-training model in the corresponding field, the keywords are generated into similar words. The generated similar words can retain the key information of the category label, and expand the training data set by generating high-quality enhanced data sets to avoid machine learning from fitting The problem of overfitting.

In order to make the above-mentioned features and advantages of the present invention more comprehensible, the following specific embodiments are described in detail together with the accompanying drawings.

Parts of the embodiments of the present invention will be described in detail with reference to the accompanying drawings. For the referenced reference symbols in the following description, when the same reference symbols appear in different drawings, they will be regarded as the same or similar components. These embodiments are only a part of the present invention, and do not reveal all possible implementation modes of the present invention. More precisely, these embodiments are just examples of the method, electronic device and computer-readable storage medium within the scope of the patent application of the present invention.

FIG. 1 is a schematic diagram of a system for recommending questions and answers based on data enhancement according to an embodiment of the present invention.

Referring to FIG. 1 , the system 10 for recommending question-answer based on data enhancement includes a transceiver 110 , a storage medium 120 and a processor 130 .

The transceiver 110 transmits and receives signals in a wireless or wired manner. The transceiver 110 may also perform operations such as low noise amplification, impedance matching, frequency mixing, up or down frequency conversion, filtering, amplification, and the like.

The storage medium 120 is, for example, any type of fixed or removable random access memory (random access memory, RAM), read-only memory (read-only memory, ROM), flash memory (flash memory) , hard disk drive (hard disk drive, HDD), solid state drive (solid state drive, SSD) or similar components or a combination of the above components, the storage device 102 is used to record multiple instructions that can be executed by the processor 130, and is also used to store A plurality of modules or various application programs executable by the processor 130 .

The processor 130 is, for example, a central processing unit (central processing unit, CPU), or other programmable general purpose or special purpose micro control unit (micro control unit, MCU), microprocessor (microprocessor), digital signal processing Digital signal processor (DSP), programmable controller, application specific integrated circuit (ASIC), graphics processing unit (graphics processing unit, GPU), arithmetic logic unit (arithmetic logic unit, ALU) , complex programmable logic device (complex programmable logic device, CPLD), field programmable logic gate array (field programmable gate array, FPGA) or other similar components or a combination of the above components. The processor 130 is coupled to the storage medium 120 and the transceiver 110 , and accesses and executes a plurality of modules and various application programs stored in the storage medium 120 to control the overall operation of the system 10 based on data enhanced recommendation question answering.

In one embodiment, the storage medium 120 can store multiple modules including a pre-training model selection module 1201, a data enhancement module 1202, an intention recognition module 1203, and a reply retrieval module 1204, as well as a database 1205, wherein the database A plurality of recommendation materials are stored in 1205 , and the reply retrieval module 1204 includes a dialog semantic analysis module 113 and a dialog status tracking module 114 , the functions of which will be described later.

FIG. 2 is a schematic diagram of selecting module categories using a pre-trained model according to an embodiment of the present invention.

Please refer to FIG. 2 , the pre-training model selection module 1201 classifies the original training data set 101 into specific domain data and general domain data by the domain classifier 102 according to the training data and the probability threshold, and selects the specific domain pre-training model and the general domain pre-training model. The training model generates domain-specific training subset 104 and general domain training subset 103 by corresponding domain-specific data and general domain data, wherein domain-specific training subset 104 is training data that is more relevant to a specific domain, and the general domain training subset 103 is the training data that is less related to the specific domain. The original training data set 101 includes the data of the specific domain training subset 104 and the data of the general domain training subset 103 . In this embodiment, the domain classifier 102 is a pre-trained binary classification model, usually through pre-collected data in a specific domain and outside a specific domain or based on web crawler technology (Web Crawler) through pre-collected domain key The word list obtains a large amount of data in the domain (ie, the content in the keyword list) and outside the domain (ie, the content outside the keyword list) to train a binary classification model.

FIG. 3 is a schematic diagram of a usage data enhancement module according to an embodiment of the present invention.

Please refer to FIG. 3 , the data enhancement module 1202 is electrically connected to the pre-training model selection module 1201 , and the data enhancement module 1202 pre-processes the domain-specific training subset 104 and the general domain training subset 103 . Specifically, the data enhancement module 1202 filters the stop words (Stop Words) in the domain-specific training subset 104 and the general domain training subset 103 to perform word segmentation, and outputs word segmentation result information, based on statistical methods and according to categories Extract candidate keywords from word segmentation result information, calculate the category vector of the category label corresponding to each training data in the specific field training subset 104 and the general field training subset 103, and calculate the similarity between the candidate keyword and the category vector of each category label Degree to generate category keyword list 111, based on category keyword list 111, specific domain pre-training model and general domain pre-training model to generate similar words 105, in this embodiment, similar words 105 is based on the category label has representative this category The keywords extracted from the key value of the label can be used to retain the key information of the category label. And use similar words 105 to expand the original training data set 101 and generate an enhanced training data set 106 .

FIG. 4 is a schematic diagram of a usage intention recognition module according to an embodiment of the present invention.

Please refer to FIG. 4 , the intent recognition module 1203 is electrically connected to the data enhancement module 1202. The intent recognition module 1203 converts the user's dialogue into text, receives the data 110 input by the user, and classifies the data 110. Prediction, to obtain the category label of the data 110, and the intent recognition module 1203 trains the classification model 109 based on the enhanced training data set 106, the verification data set 107 and the neural network architecture 108, so as to obtain the key words based on the category keyword list 111 and the data 110 The category label and keyword combination 112 of the word extraction output data 110 .

FIG. 5 is a schematic diagram of using a reply retrieval module according to an embodiment of the present invention.

Please refer to FIG. 5 , the reply retrieval module 1204 is electrically connected to the intent recognition module 1203 , the reply retrieval module 1204 combines the category tags and keywords 112 of the data 110 through the conversational semantic analysis module 113 and converts the corresponding data 110 The user's dialogue is stored in the dialogue state tracking module 114, and the similarity calculation is performed on the data 110 and the recommended data stored in the database 1205. Based on the Top-N recommendation algorithm, the conversational semantic analysis module 113 is based on the semantic strength, The conversation time and the number of hits sort the recommendation data stored in the database 1205 to recommend the best recommendation data to the user.

FIG. 6 is a flow chart of a method for recommending questions and answers based on data enhancement according to an embodiment of the present invention.

Please refer to FIG. 6 , in step S601, the pre-training model selection module 1201 classifies the training data into domain-specific data and general domain data according to the training data, and selects the domain-specific pre-training model and the general domain pre-training model to separate the domain-specific data Corresponding to the general domain data, a domain-specific training subset and a general domain training subset are generated.

In step S602, the data enhancement module 1202 extracts candidate keywords respectively according to the domain-specific training subset and the general domain training subset, and generates a category keyword list 111, and combines the category keyword list 111 with the domain-specific pre-training model and the general domain Domain pre-trained models are augmented to generate augmented training datasets.

In step S603, the intent identification module 1203 receives the data input by the user to identify the category label of the data based on the classification model trained in the enhanced training data set and outputs the category label, and extracts and outputs the category label based on the category keyword list 111 and the data keyword combination.

In step S604 , the reply retrieval module 1204 retrieves the best recommendation information from the database 1205 based on the combination of category tags and keywords and recommends them to the user.

In one embodiment, the scenario of the user operating the speaker is used for illustration. The user interacts with the system 10 based on data-enhanced recommendation questions and answers through terminal equipment (such as speakers, mobile phones, etc.). The data set is supplemented by labor insurance-related question and answer data to illustrate. This data-enhanced and recommended Q&A system can provide Q&A services in related fields of labor insurance, such as: labor pension payment slip supplement method, agricultural insurance maternity benefits application method, etc.

In conjunction with Figure 1, in this embodiment, when the user enters a question "Description of the calculation of the age of labor insurance benefits for the elderly", after the speaker receives the question, it will convert the question through the network through the voice-to-text operation. The information is transmitted to the system 10 for recommending questions and answers based on data enhancement. At this time, the system 10 for recommending questions and answers based on data enhancement will search for the best recommended information from the database 120 through the intent identification module 1203 and the answer retrieval module 1204 in sequence (i.e. Answer) and the answer is also transmitted to the user's speaker device through the text-to-speech operation, so the user will hear the answer through the speaker: "The age of labor insurance old-age benefits is calculated according to the household registration record, and it is fully calculated from the date of birth." The pre-training model selection and data enhancement are completed in the pre-training stage, and do not participate in this question-and-answer process. The following describes in detail how the system 10 for recommending question and answer based on data enhancement processes the process of question answering.

First, the original training data set 101 is divided into a specific domain training subset 104 and a general domain training subset 104 based on the probability threshold, selecting a specific domain pre-training model and a general domain pre-training model through the pre-trained domain classifier 102. Training subset 103, for example: the original training data set 101 contains two training sentences of "the income amount and calculation method of individual labor pension special account" and "why I did not receive the electronic bill", and the domain classifier 102 The probability threshold is set to 0.7, and the probability of the sentence "the amount and calculation method of individual labor pension special accounts" is 0.73 after being predicted by the field classifier 102, and the sentence "why I did not receive the electronic bill" is classified through the field The probability obtained after prediction by the machine 102 is 0.42. Based on the foregoing results, the sentence "the amount and calculation method of the income of individual labor pension special accounts" will be divided into the specific field training subset 104. In addition, the sentence "why did I not receive the electronic account Single" will be divided into general domain training subset 103. Subsequently, the obtained domain-specific training subset 104 and the general domain training subset 103 are selected from the corresponding domain-specific pre-training models and general domain pre-training models, and then sent to the data enhancement module 1202 after training. The domain-specific pre-training model is usually trained based on a deep neural network through a domain-specific text data set containing hundreds of millions of words, and the general-purpose domain pre-training model is usually trained through a general-purpose domain text data containing hundreds of millions of words collected and trained based on a deep neural network.

After the data enhancement module 1202 receives the domain-specific training subset 104 , the general domain training subset 103 , the domain-specific pre-training model, and the general domain pre-training model from the above-mentioned pre-training model selection module 1201 , data enhancement begins. Firstly, basic text preprocessing is performed on the domain-specific training subset 104 and the general domain training subset 103 . Specifically, first delete the stagnant words in the sentence and give appropriate word breaks, for example: the original training data set 101 contains the sentence "the amount of income and calculation method of individual labor pension special account", which is processed through text It will be converted into ["individual" "labor" "pension" "special account" "income" "amount" "calculation" "method"]. Then, based on the Term Frequency–Inverse Document Frequency (TF-IDF) in the statistical method, candidates are extracted from the segmentation results of the two training subsets (specific domain training subset 104, general domain training subset 103) keywords, as shown in Figure 7.

FIG. 7 is a schematic diagram of candidate keywords according to an embodiment of the present invention. The two training subsets (the domain-specific training subset 104 and the general domain training subset 103) are respectively used to calculate the category of each category label in the two training subsets through the selected domain-specific pre-training model and the general domain pre-training model using formula 1 vector, whose Equation 1 is shown below.

Formula 1:

為各訓練子集中類別標籤的數量、

為類別標籤、

為各類別標籤中包含的資料集數量以及

為向量的維度大小，

為第

個句子word embedding向量的第

維。隨後將上述候選關鍵詞與各類別標籤的類別向量透過餘弦相似度 (Cosine Similarity) 計算相似性（即公式2），從而獲得類別關鍵詞清單111，其公式2如下所示。 in

is the number of category labels in each training subset,

for the category labels,

is the number of data sets contained in each category label and

is the dimension size of the vector,

for the first

The first sentence word embedding vector

dimension. Then, the above candidate keywords and the category vectors of each category label are used to calculate the similarity through cosine similarity (ie formula 2), so as to obtain the category keyword list 111, the formula 2 of which is shown below.

Formula 2:

和

分別代表

和

的向量，相似性範圍為

，

表示

和

兩個向量指向的方向完全相反，即相似性完全不同，另外

表示兩個向量指向的方向完全相同，即相似性完全相同。得到類別關鍵詞清單111後，基於上述對應的特定通用領域預訓練模型及通用領域預訓練模型將圖7中各類別標籤的關鍵詞生成相似詞，透過其生成的相似詞可以保留該類別標籤的關鍵訊息，如圖8所示。圖8是依照本發明的一實施例的類別關鍵詞的示意圖。生成相似詞後，基於相似詞來擴充原始訓練資料集101以得到增強訓練資料集106。重複此步驟至完成對所有的特定領域訓練子集104、通用領域訓練子集103以進行資料增強，合併所有擴充的增強訓練資料集106以作為最後的增強訓練資料集，如圖9所示，圖9是依照本發明的一實施例的增強訓練資料集的示意圖。 one of them

and

Representing

and

A vector with a similarity range of

,

express

and

The two vectors point in completely opposite directions, that is, the similarities are completely different, and in addition

Indicates that the two vectors point in exactly the same direction, that is, the similarity is exactly the same. After obtaining the category keyword list 111, based on the above-mentioned corresponding specific general domain pre-training model and general domain pre-training model, generate similar words for the keywords of each category label in Figure 7, and the generated similar words can retain the category label. Key information, as shown in Figure 8. FIG. 8 is a schematic diagram of category keywords according to an embodiment of the present invention. After the similar words are generated, the original training data set 101 is expanded based on the similar words to obtain an enhanced training data set 106 . Repeat this step to complete all domain-specific training subsets 104 and general domain training subsets 103 for data enhancement, and merge all expanded enhanced training data sets 106 as the final enhanced training data set, as shown in FIG. 9 , FIG. 9 is a schematic diagram of an augmented training data set according to an embodiment of the present invention.

The classification model 109 is trained by setting softmax as the activation function based on the neural network framework 108 through the above-mentioned enhanced training data set 106, and the current classification model is evaluated based on the verification data set 107 to modify the classification model. The output of softmax in Equation 3 represents the relative probability between different category labels, and its Equation 3 is as follows.

Formula 3:

表示類別標籤，

表示資料

為類別標籤

的機率。另外為了預防過擬合，在模型訓練中加入了

正則化，其增加一個索引來描述公式4中損失函數 (Loss Function) 中模型的複雜度，其公式4如下。 in

represents the category label,

Indicate data

for category labels

probability. In addition, in order to prevent overfitting, the model training is added

Regularization, which adds an index to describe the complexity of the model in the loss function (Loss Function) in formula 4, and its formula 4 is as follows.

Formula 4:

為模型的權重，做

正則化後當作損失函數

。 in

For the model weights, do

Regularized as a loss function

.

After the user enters the question sentence with the text content of "Calculation Instructions for the Age of Labor Insurance Elderly Benefits" obtained through voice conversion, the question sentence is input into the classification model 109 to predict the user's category label, and then the predicted question sentence can be obtained has a category label of "Labor Insurance". In addition, through the aforementioned category keyword list 111, keyword extraction is carried out on this question sentence, and the keywords whose contents are "labor insurance" and "old age benefit" can be obtained, and then the above-mentioned category label and keyword combination 112 is submitted to the reply retrieval module Group 1204.

The conversational semantic analysis module 113 uses the above-mentioned category label and keyword combination 112 to calculate the similarity based on the minimum edit distance (Minimum Edit Distance) and other indicators, from the database 1205 (in this embodiment, it can be a question and answer knowledge base) 115) to find similar results. Among them, the minimum edit distance we use Levenshtein's definition. If a character string is edited into another character string, the following three actions can be performed:

及

，依序去計算長度為

字串的最小編輯距離。 Starting from the initial state, by

and

, in order to calculate the length of

The minimum edit distance of a string.

Finally, various indicators are combined, such as: keywords in the dialogue state tracking module 114, dialogue time and number of clicks, sorted, so as to improve the accuracy of the recommendation results, and finally recommend the best recommendation information to the user through the Top-N algorithm .

Based on the above, the present invention provides a system and method for recommending questions and answers based on data enhancement. The sentences in the training data set are divided into a specific domain training subset and a general domain training subset according to domain relevance. By extracting category keywords in the training subset And based on the pre-training model in the corresponding field, the keywords are generated into similar words. The generated similar words can retain the key information of the category label, and expand the training data set by generating high-quality enhanced data sets to avoid machine learning from fitting The problem of overfitting.

Although the present disclosure has been disclosed above with embodiments, it is not intended to limit the present disclosure. Anyone with ordinary knowledge in the technical field may make some changes and modifications without departing from the spirit and scope of the present disclosure. The scope of protection of this disclosure should be defined by the scope of the appended patent application.

10: A system based on data enhancement recommendation question answering 110: Transceiver 120: storage media 130: Processor 1201: Pre-training model selection module 1202: Data enhancement module 1203: Intent recognition module 1204:Reply search module 1205: database 101: Raw training data set 102:Domain Classifier 103:General Domain Training Subsets 104:Domain-Specific Training Subsets 105: Similar words 106: Enhanced training data set 107:Validation dataset 108: Neural Network Architecture 109: Classification Models 110: Information 111:Category keyword list 112:Category label and keyword combination 113: Conversational Semantic Analysis Module 114:Dialogue state tracking module 115: Question and answer knowledge base S601, S602, S603, S604: steps

FIG. 1 is a schematic diagram of a system for recommending questions and answers based on data enhancement according to an embodiment of the present invention. FIG. 2 is a schematic diagram of selecting module categories using a pre-trained model according to an embodiment of the present invention. FIG. 3 is a schematic diagram of a usage data enhancement module according to an embodiment of the present invention. FIG. 4 is a schematic diagram of a usage intention recognition module according to an embodiment of the present invention. FIG. 5 is a schematic diagram of using a reply retrieval module according to an embodiment of the present invention. FIG. 6 is a flow chart of a method for recommending questions and answers based on data enhancement according to an embodiment of the present invention. FIG. 7 is a schematic diagram of candidate keywords according to an embodiment of the present invention. FIG. 8 is a schematic diagram of category keywords according to an embodiment of the present invention. FIG. 9 is a schematic diagram of an augmented training data set according to an embodiment of the present invention.

10: A system based on data enhancement recommendation question answering

110: Transceiver

120: storage media

130: Processor

1201: Pre-training model selection module

1202: Data enhancement module

1203: Intent recognition module

1204:Reply search module

1205: database

Claims (10)

A system for enhancing recommendation questions and answers based on data, including: transceiver; Storage media, storing multiple modules and databases; and The processor is coupled to the storage medium and the transceiver, and accesses and executes the modules, wherein the modules include: The pre-training model selection module classifies the training data into domain-specific data and general domain data according to the training data, and selects the domain-specific pre-training model and the general domain pre-training model to generate corresponding domain-specific data and general domain data respectively. domain training subset and general domain training subset; The data enhancement module is electrically connected with the pre-training model selection module, extracts candidate keywords respectively according to the specific field training subset and the general field training subset, and generates a category keyword list, and the category keyword list performing data augmentation with the domain-specific pre-training model and the general domain pre-training model to generate an enhanced training data set; The intention recognition module is electrically connected to the data enhancement module, receives the data input by the user, recognizes the category label of the data based on the classification model trained by the enhanced training data set and outputs the category label, and based on the category keyword A list and combination of keywords for extracting the data and outputting the data; and The reply retrieval module is electrically connected with the intent recognition module, and retrieves the best recommended information from the database based on the category tag and the keyword combination and recommends it to the user. The system for recommending questions and answers based on data enhancement as described in Claim 1, wherein the pre-training model selection module is further used to classify the training data into the specific domain data and the general domain data according to a probability threshold. The system for recommending questions and answers based on data enhancement as described in claim 1, wherein the domain-specific training subset belongs to the training data related to a specific domain, and the general domain training subset belongs to the training data not related to a specific domain . The system for recommending questions and answers based on data enhancement as described in claim 1, wherein the data enhancement module is further used to calculate the category vector of each category label in the specific domain training subset and the general domain training subset, and calculate the candidate keywords With the similarity of the category vector of each category label to generate a category keyword list, generate similar words based on the category keyword list, the specific domain pre-training model and the general domain pre-training model, use the similar words to expand the training data set and Generate the augmented training data set. The system for recommending questions and answers based on data enhancement as described in claim 1, wherein the data enhancement module is further used to filter the stop words (Stop Words) in the specific domain training subset and the general domain training subset to segment words And output word segmentation result information, and extract the candidate keyword from the word segmentation result information based on a statistical method. The system for recommending questions and answers based on data enhancement as described in claim 1, wherein the intent recognition module is further used to train a classification model based on the enhanced training data set, verification data set, and neural network architecture, so as to respond to the user's dialogue The class label prediction is performed on the input data after the voice-to-text conversion, so as to obtain the class label of the data. The system for recommending questions and answers based on data enhancement as described in Claim 6, wherein the answer retrieval module includes a conversational semantic analysis module and a dialogue state tracking module. The system for recommending questions and answers based on data enhancement as described in claim item 7, wherein the answer retrieval module is further used to combine the category label of the data and the keyword through the conversational semantic analysis module to combine the user's dialogue Stored in the dialog state tracking module, and calculate the similarity between the data and the recommended data stored in the database, and recommend the best recommended data to the user through the dialog semantic analysis module. The system for recommending questions and answers based on data enhancement as described in claim 8, wherein the answer retrieval module is further used to use the Top-N recommendation algorithm based on the conversational semantic analysis module according to semantic strength, dialogue time and number of clicks. The recommended data stored in the database is sorted to recommend the best recommended data to the user. A method for enhancing recommendation question answering based on data, comprising: According to the training data, the training data is classified into specific domain data and general domain data, and the specific domain pre-training model and the general domain pre-training model are selected to generate specific domain training subsets and general domain correspondingly to the specific domain data and the general domain data. training subset; Extracting keywords respectively according to the specific domain training subset and the general domain training subset, and generating a category keyword list, and performing data enhancement on the category keyword list with the specific domain pre-training model and the general domain pre-training model to Generate an augmented training dataset; receiving user-input data to identify a class label of the data based on the classification model trained on the enhanced training data set and outputting the class label, and extracting and outputting a keyword combination of the data based on the class keyword list and the data; and Based on the category tag and the keyword combination, the best recommended information is retrieved from the database and recommended to the user.

Images