TWI768513B - Artificial intelligence training system and artificial intelligence training method - Google Patents
Artificial intelligence training system and artificial intelligence training method Download PDFInfo
- Publication number
- TWI768513B TWI768513B TW109136193A TW109136193A TWI768513B TW I768513 B TWI768513 B TW I768513B TW 109136193 A TW109136193 A TW 109136193A TW 109136193 A TW109136193 A TW 109136193A TW I768513 B TWI768513 B TW I768513B
- Authority
- TW
- Taiwan
- Prior art keywords
- artificial intelligence
- training
- intelligence model
- test
- keyword
- Prior art date
Links
Images
Landscapes
- Feedback Control In General (AREA)
- Cable Transmission Systems, Equalization Of Radio And Reduction Of Echo (AREA)
- Machine Translation (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Description
本揭露是有關於一種人工智慧模型訓練系統及人工智慧模型訓練方法,且特別是有關於一種有效率地強化同義字詞學習的人工智慧模型訓練系統及人工智慧模型訓練方法。The present disclosure relates to an artificial intelligence model training system and an artificial intelligence model training method, and in particular, to an artificial intelligence model training system and an artificial intelligence model training method that effectively strengthen the learning of synonyms.
一個好的機器學習模型必需要能夠做到一定程度的泛化(Generalization)。也就是說,即使某些字詞並未經過特別訓練,機器學習模型也能夠將這些未經訓練的字詞辨識出來。然而,要訓練出一個高度泛化的機器學習模型並不容易。因此,如何提供一個訓練方法來有效地訓練出高度泛化的機器學習模型是本領域技術人員應致力的目標。A good machine learning model must be able to achieve a certain degree of generalization. That is, the machine learning model can recognize certain words even if they have not been specially trained. However, it is not easy to train a highly generalized machine learning model. Therefore, how to provide a training method to effectively train a highly generalized machine learning model is the goal of those skilled in the art.
有鑑於此,本揭露提供一種人工智慧模型訓練系統及人工智慧模型訓練方法,能夠有效率地強化同義字詞的學習。In view of this, the present disclosure provides an artificial intelligence model training system and an artificial intelligence model training method, which can effectively strengthen the learning of synonyms.
本揭露提出一種人工智慧模型訓練系統,包括人工智慧模型及處理器耦接到人工智慧模型。處理器將多個同義字從第一訓練資料移除以獲得第二訓練資料,並根據第二訓練資料來訓練人工智慧模型。處理器將根據同義字擴展第一測試資料以獲得第二測試資料,並在測試操作中根據第二測試資料來測試人工智慧模型。當對應關鍵字詞的測試結果產生錯誤時,處理器根據關鍵字詞及第二訓練資料獲得第三訓練資料,並根據第三訓練資料來訓練人工智慧模型。The present disclosure proposes an artificial intelligence model training system, including an artificial intelligence model and a processor coupled to the artificial intelligence model. The processor removes the plurality of synonyms from the first training data to obtain second training data, and trains the artificial intelligence model according to the second training data. The processor will expand the first test data according to the synonym to obtain the second test data, and test the artificial intelligence model according to the second test data in the test operation. When an error occurs in the test result corresponding to the key word, the processor obtains third training data according to the key word and the second training data, and trains the artificial intelligence model according to the third training data.
本揭露提出一種人工智慧模型訓練方法,適用於人工智慧模型訓練系統。人工智慧模型訓練系統包括人工智慧模型及處理器耦接到人工智慧模型。人工智慧模型訓練方法包括藉由處理器將多個同義字從第一訓練資料移除以獲得第二訓練資料,並根據第二訓練資料來訓練人工智慧模型。人工智慧模型訓練方法還包括藉由處理器將根據同義字擴展第一測試資料以獲得第二測試資料,並在測試操作中根據第二測試資料來測試人工智慧模型。人工智慧模型訓練方法還包括當對應關鍵字詞的測試結果產生錯誤時,藉由處理器根據關鍵字詞及第二訓練資料獲得第三訓練資料,並根據第三訓練資料來訓練人工智慧模型。The present disclosure proposes an artificial intelligence model training method, which is suitable for an artificial intelligence model training system. The artificial intelligence model training system includes an artificial intelligence model and a processor coupled to the artificial intelligence model. The artificial intelligence model training method includes removing a plurality of synonyms from the first training data by a processor to obtain second training data, and training the artificial intelligence model according to the second training data. The artificial intelligence model training method further includes that the processor expands the first test data according to the synonyms to obtain the second test data, and tests the artificial intelligence model according to the second test data in the test operation. The artificial intelligence model training method further includes obtaining third training data by the processor according to the key word and the second training data, and training the artificial intelligence model according to the third training data when the test result corresponding to the keyword is wrong.
基於上述,本揭露的人工智慧模型訓練系統及人工智慧模型訓練方法會從第一訓練資料移除同義字以獲得第二訓練資料並根據第二訓練資料來訓練人工智慧模型。接著,第一測試資料會根據同義字擴展成第二測試資料且第二測試資料會用於測試人工智慧模型。當對應一個關鍵字詞(即,無法被泛化處理的字詞)的測試結果產生錯誤時,第三訓練資料會根據關鍵字詞及第二訓練資料來獲得且人工智慧模型會再次根據第三訓練資料來訓練。Based on the above, the artificial intelligence model training system and the artificial intelligence model training method of the present disclosure remove the synonyms from the first training data to obtain the second training data and train the artificial intelligence model according to the second training data. Then, the first test data is expanded into second test data according to the synonym, and the second test data is used to test the artificial intelligence model. When an error occurs in the test result corresponding to a keyword (ie, a word that cannot be generalized), the third training data will be obtained according to the keyword and the second training data, and the artificial intelligence model will again be based on the third training data. training material to train.
在本揭露一實施例的語意模型中,使用者問題中重要的單詞可被稱作實體(Entity),而使用者的目的可被稱作意圖(Intention)。舉例來說,在問句「where is the repair center in Tokyo」中,語意模型可理解語意並進而將資訊簡化成意圖「Find Repair Center」及實體「Tokyo」,並根據解析出的意圖及實體到系統後端資料庫尋找對應的答案提供給使用者。由於相同的意圖可能會發生在不同的實體上,因此語意模型必需要有能力學習並辨認不同實體。In the semantic model of an embodiment of the present disclosure, the important words in the user's question may be referred to as an entity, and the user's purpose may be referred to as an intent. For example, in the question "where is the repair center in Tokyo", the semantic model can understand the semantics and then simplify the information into the intent "Find Repair Center" and the entity "Tokyo", and according to the parsed intent and entity to The back-end database of the system looks for the corresponding answer and provides it to the user. Since the same intent may occur on different entities, the semantic model must be able to learn and recognize different entities.
表一為使用者詢問的句子及其意圖的範例。Table 1 is an example of sentences and their intentions that users ask about.
表一
表一列舉了使用者問句及對應意圖的範例。句子中的「Tokyo」、「Taipei」、「wifi」、「system backup」、「Acer Recovery」則為實體。Table 1 lists examples of user questions and corresponding intentions. "Tokyo", "Taipei", "wifi", "system backup", and "Acer Recovery" in the sentence are entities.
在一實施例的語意模型中,「wifi」可透過人工替換成「WI-FI」或「wireless」等同義字,然而這會大幅提升人力成本。在另一實施例的語意模型中,可利用同義字詞應用程式介面(Application Program Interface,API)來替換句子中的同義字來增加語意模型對句子的意圖及/或實體中的同義字的識別能力。In the semantic model of an embodiment, "wifi" can be manually replaced with synonyms such as "WI-FI" or "wireless", but this will greatly increase labor costs. In the semantic model of another embodiment, a synonym application program interface (API) can be used to replace the synonyms in the sentence to increase the semantic model's recognition of the intent and/or the synonyms in the entity ability.
圖1為根據本揭露一實施例的人工智慧模型訓練系統的方塊圖。FIG. 1 is a block diagram of an artificial intelligence model training system according to an embodiment of the present disclosure.
請參照圖1,本揭露一實施例的人工智慧模型訓練系統100包括人工智慧模型110及處理器120。處理器120耦接到人工智慧模型110。人工智慧模型訓練系統100例如是伺服器、個人電腦、筆記型電腦、智慧型手機、平板電腦或其他類似電子裝置。處理器120例如是中央處理單元(Central Processing Unit,CPU)或其他類似裝置。在一實施例中,人工智慧模型110可為聊天機器人模型,用於接收使用者問題並作出相關的回應。處理器120可對人工智慧模型110進行訓練、測試及分析等操作。Referring to FIG. 1 , an artificial intelligence
在一實施例中,處理器120將多個同義字從第一訓練資料移除以獲得第二訓練資料,並根據第二訓練資料來訓練人工智慧模型110。處理器120將根據同義字擴展第一測試資料以獲得第二測試資料,並在測試操作中根據第二測試資料來測試人工智慧模型110。當對應關鍵字詞的測試結果產生錯誤時,處理器120根據關鍵字詞及第二訓練資料獲得第三訓練資料,並根據第三訓練資料來訓練人工智慧模型110。In one embodiment, the
圖2為根據本揭露一實施例的人工智慧模型訓練方法的流程圖。FIG. 2 is a flowchart of an artificial intelligence model training method according to an embodiment of the present disclosure.
請同時參照圖1及圖2,本揭露一實施例的人工智慧模型訓練方法包括前處理流程210、訓練流程220、測試流程230及分析流程240。Referring to FIG. 1 and FIG. 2 at the same time, an artificial intelligence model training method according to an embodiment of the present disclosure includes a
在前處理流程210中,處理器120可將多個同義字213從第一訓練資料211中移除,而產生不具有同義字213的第二訓練資料212。由於第二訓練資料212不包括同義字213,因此可大幅增加人工智慧模型110的訓練速度。In the
在訓練流程220中,第二訓練資料212會被輸入人工智慧模型110中進行訓練操作221以產生已訓練模型222。In the
在測試流程230中,處理器120可根據同義字213來擴展第一測試資料231以獲得擴展後測試資料(或稱為第二測試資料),並用擴展後測試資料對人工智慧模型110進行測試操作232以產生測試結果233。測試操作232的細節會在下文中詳細說明。In the test process 230, the
在分析流程240中,處理器120會對測試結果233進行分析操作241以產生分析結果242。當分析結果242記錄測試結果233的錯誤時,處理器120會根據產生錯誤的關鍵字詞及第二訓練資料212進行訓練資料調整操作244來獲得第三訓練資料243。訓練資料調整操作244的細節會在下文中詳細說明。In the
在一實施例中,測試操作232可包括第一測試及第二測試。在第一測試中,人工智慧模型110可接收句子以進行測試操作並產生對應句子的意圖及實體,且處理器120判斷意圖及實體是否與對應句子的預定意圖及預定實體相同。若判斷出的意圖及實體符合人工標註的正確答案則第一測試成功。在第二測試中,人工智慧模型110透過意圖及實體到資料庫中搜尋對應句子的答案(或稱為第一答案)。若人工智慧模型110透過意圖及實體到資料庫中正確地搜尋對應句子的第一答案則第二測試成功。In one embodiment, the
圖3為根據本揭露一實施例的分析流程的流程圖。FIG. 3 is a flowchart of an analysis process according to an embodiment of the present disclosure.
圖3,在步驟S301中,讀取測試結果。Fig. 3, in step S301, the test result is read.
在步驟S302中,從測試結果判斷錯誤類型。In step S302, the error type is determined from the test result.
在步驟S303中,判斷測試結果是否發生第一類錯誤或第二類錯誤。第一類錯誤及第二類錯誤會在下文中詳細說明。In step S303, it is determined whether the first type of error or the second type of error occurs in the test result. Type 1 and Type 2 errors are described in detail below.
若測試結果沒發生第一類錯誤或第二類錯誤,則在步驟S305中,產生分析結果。If the first type of error or the second type of error does not occur in the test result, in step S305, an analysis result is generated.
若測試結果發生第一類錯誤或第二類錯誤,在步驟S304中,根據產生錯誤的關鍵字詞產生重新訓練資料(或稱為第三訓練資料)。If the first type of error or the second type of error occurs in the test result, in step S304, re-training data (or referred to as third training data) is generated according to the key words that generate the error.
在步驟S305中,產生分析結果。In step S305, an analysis result is generated.
在一實施例中,當第一測試失敗且第二測試失敗則處理器120判斷發生第一類錯誤。第一類錯誤又可稱為缺同義字的錯誤,代表人工智慧模型110無法識別預先定義好的字詞且這個字詞不在同義字集中,也無法用預測出的意圖及實體到資料庫找出答案。當第一類錯誤發生時,處理器120可根據句子中的關鍵字詞搜尋同義字詞資料庫(例如,同義字詞應用程式介面)以獲得對應關鍵字詞的多個同義字詞,並根據同義字詞替換該句子中的關鍵字詞以獲得第三訓練資料。舉例來說,當人工智慧模型110無法識別「why my WI-FI is not working」時,處理器120可搜尋「WI-FI」的同義字「wifi」並以「wifi」替換問句中的「WI-FI」以將「why my wifi is not working」加入第三訓練資料中。In one embodiment, when the first test fails and the second test fails, the
當第一測試成功且第二測試失敗則判斷發生第二類錯誤。第二類錯誤又可稱為缺答案的錯誤,代表人工智慧模型110可以識別預先定義好的字詞但這個字詞不在同義字集裡面,由於這組同義字詞沒有與資料庫中的答案關聯,因此人工智慧模型也無法用識別出的字詞到資料庫找出答案。當第二類錯誤發生時,處理器120可根據句子中的關鍵字詞搜尋同義字詞資料庫以獲得對應關鍵字詞的多個同義字詞,並將同義字詞關聯到資料庫中對應關鍵字詞的答案(或稱為第二答案)。舉例來說,當人工智慧模型110成功識別出「why my WI-FI is not working」但「WI-FI」並不在同義字集裡面時,處理器120可透過同義字詞資料庫將「WI-FI」對應的答案與「wifi」對應的答案關聯,以將系統後端資料庫的答案補充完整。When the first test succeeds and the second test fails, it is determined that the second type of error occurs. The second type of error can also be called the error of lack of answer, which means that the
綜上所述,本揭露的人工智慧模型訓練系統及人工智慧模型訓練方法會從第一訓練資料移除同義字以獲得第二訓練資料並根據第二訓練資料來訓練人工智慧模型。接著,第一測試資料會根據同義字擴展成第二測試資料且第二測試資料會用於測試人工智慧模型。當對應一個關鍵字詞(即,無法被泛化處理的字詞)的測試結果產生錯誤時,第三訓練資料會根據關鍵字詞及第二訓練資料來獲得且人工智慧模型會再次根據第三訓練資料來訓練。To sum up, the artificial intelligence model training system and the artificial intelligence model training method of the present disclosure remove synonyms from the first training data to obtain the second training data and train the artificial intelligence model according to the second training data. Then, the first test data is expanded into second test data according to the synonym, and the second test data is used to test the artificial intelligence model. When an error occurs in the test result corresponding to a keyword (ie, a word that cannot be generalized), the third training data will be obtained according to the keyword and the second training data, and the artificial intelligence model will again be based on the third training data. training material to train.
雖然本揭露已以實施例揭露如上,然其並非用以限定本揭露,任何所屬技術領域中具有通常知識者,在不脫離本揭露的精神和範圍內,當可作些許的更動與潤飾,故本揭露的保護範圍當視後附的申請專利範圍所界定者為準。Although the present disclosure has been disclosed above with examples, 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 the present disclosure shall be determined by the scope of the appended patent application.
100:人工智慧模型訓練系統 110:人工智慧模型 120:處理器 210:前處理流程 211:第一訓練資料 212:第二訓練資料 213:同義字 220:訓練流程 221:訓練操作 222:已訓練模型 230:測試流程 231:第一測試資料 232:測試操作 233:測試結果 240:分析流程 241:分析操作 242:分析結果 243:第三訓練資料 244:訓練資料調整操作 S301~S305:分析流程的步驟 100: Artificial Intelligence Model Training System 110: Artificial Intelligence Models 120: Processor 210: Pre-processing flow 211: First training material 212: Second training material 213: Synonyms 220: Training Process 221: Training operations 222: Trained model 230: Test Process 231: First test data 232: Test operation 233: Test Results 240: Analysis Process 241: Analytical Operations 242: Analysis Results 243: Third training material 244: Training data adjustment operation S301~S305: Steps of the analysis process
圖1為根據本揭露一實施例的人工智慧模型訓練系統的方塊圖。 圖2為根據本揭露一實施例的人工智慧模型訓練方法的流程圖。 圖3為根據本揭露一實施例的分析流程的流程圖。 FIG. 1 is a block diagram of an artificial intelligence model training system according to an embodiment of the present disclosure. FIG. 2 is a flowchart of an artificial intelligence model training method according to an embodiment of the present disclosure. FIG. 3 is a flowchart of an analysis process according to an embodiment of the present disclosure.
210:前處理流程 211:第一訓練資料 212:第二訓練資料 213:同義字 220:訓練流程 221:訓練操作 222:已訓練模型 230:測試流程 231:第一測試資料 232:測試操作 233:測試結果 240:分析流程 241:分析操作 242:分析結果 243:第三訓練資料 244:訓練資料調整操作 210: Pre-processing flow 211: First training material 212: Second training material 213: Synonyms 220: Training Process 221: Training operations 222: Trained model 230: Test Process 231: First test data 232: Test operation 233: Test Results 240: Analysis Process 241: Analytical Operations 242: Analysis Results 243: Third training material 244: Training data adjustment operation
Claims (10)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW109136193A TWI768513B (en) | 2020-10-20 | 2020-10-20 | Artificial intelligence training system and artificial intelligence training method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW109136193A TWI768513B (en) | 2020-10-20 | 2020-10-20 | Artificial intelligence training system and artificial intelligence training method |
Publications (2)
Publication Number | Publication Date |
---|---|
TW202217627A TW202217627A (en) | 2022-05-01 |
TWI768513B true TWI768513B (en) | 2022-06-21 |
Family
ID=82558792
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
TW109136193A TWI768513B (en) | 2020-10-20 | 2020-10-20 | Artificial intelligence training system and artificial intelligence training method |
Country Status (1)
Country | Link |
---|---|
TW (1) | TWI768513B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI802459B (en) * | 2022-07-01 | 2023-05-11 | 中華電信股份有限公司 | A system and method for recommendation q&a based on data-enhanced |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107301227A (en) * | 2017-06-21 | 2017-10-27 | 北京百度网讯科技有限公司 | Search information analysis method and device based on artificial intelligence |
CN109635117A (en) * | 2018-12-26 | 2019-04-16 | 零犀(北京)科技有限公司 | A kind of knowledge based spectrum recognition user intention method and device |
CN110377911A (en) * | 2019-07-23 | 2019-10-25 | 中国工商银行股份有限公司 | Intension recognizing method and device under dialogue frame |
CN111382231A (en) * | 2020-03-05 | 2020-07-07 | 苏州思必驰信息科技有限公司 | Intention recognition system and method |
-
2020
- 2020-10-20 TW TW109136193A patent/TWI768513B/en active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107301227A (en) * | 2017-06-21 | 2017-10-27 | 北京百度网讯科技有限公司 | Search information analysis method and device based on artificial intelligence |
CN109635117A (en) * | 2018-12-26 | 2019-04-16 | 零犀(北京)科技有限公司 | A kind of knowledge based spectrum recognition user intention method and device |
CN110377911A (en) * | 2019-07-23 | 2019-10-25 | 中国工商银行股份有限公司 | Intension recognizing method and device under dialogue frame |
CN111382231A (en) * | 2020-03-05 | 2020-07-07 | 苏州思必驰信息科技有限公司 | Intention recognition system and method |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI802459B (en) * | 2022-07-01 | 2023-05-11 | 中華電信股份有限公司 | A system and method for recommendation q&a based on data-enhanced |
Also Published As
Publication number | Publication date |
---|---|
TW202217627A (en) | 2022-05-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10586155B2 (en) | Clarification of submitted questions in a question and answer system | |
US9460085B2 (en) | Testing and training a question-answering system | |
WO2021159632A1 (en) | Intelligent questioning and answering method and apparatus, computer device, and computer storage medium | |
WO2020233380A1 (en) | Missing semantic completion method and apparatus | |
US9621601B2 (en) | User collaboration for answer generation in question and answer system | |
CN110276071B (en) | Text matching method and device, computer equipment and storage medium | |
US20160321541A1 (en) | Information processing method and apparatus | |
JP2017097882A (en) | Machine translation evaluation method and device, and machine translation method and device | |
US20150178623A1 (en) | Automatically Generating Test/Training Questions and Answers Through Pattern Based Analysis and Natural Language Processing Techniques on the Given Corpus for Quick Domain Adaptation | |
US11526512B1 (en) | Rewriting queries | |
CN110428822A (en) | A kind of speech recognition error correction method and interactive system | |
WO2020143186A1 (en) | Recommendation system training method and apparatus, and computer device and storage medium | |
Ravichander et al. | NoiseQA: Challenge set evaluation for user-centric question answering | |
WO2020143301A1 (en) | Training sample validity detection method, computer device, and computer non-volatile storage medium | |
TWI768513B (en) | Artificial intelligence training system and artificial intelligence training method | |
CN112988212A (en) | Method, apparatus, system and storage medium for online incremental update of neural network model | |
WO2021239078A1 (en) | Field recognition method, interaction method, electronic device, and storage medium | |
US20180150747A1 (en) | Enhancing Time-to-Answer for Community Questions in Online Discussion Sites | |
JP2019036210A (en) | FAQ registration support method using machine learning, and computer system | |
CN114492397B (en) | Artificial intelligence model training system and artificial intelligence model training method | |
US9146918B2 (en) | Compressing data for natural language processing | |
RU2693328C2 (en) | Methods and systems for generating a replacement request for a user input request | |
TWI745878B (en) | Chat robot system and chat robot model training method | |
JP2018163586A (en) | Learning program, learning method and learning device | |
Long | A Grammatical Error Correction Model for English Essay Words in Colleges Using Natural Language Processing |