TWI805008B - Customized intent evaluation system, method and computer-readable medium - Google Patents
Customized intent evaluation system, method and computer-readable medium Download PDFInfo
- Publication number
- TWI805008B TWI805008B TW110136881A TW110136881A TWI805008B TW I805008 B TWI805008 B TW I805008B TW 110136881 A TW110136881 A TW 110136881A TW 110136881 A TW110136881 A TW 110136881A TW I805008 B TWI805008 B TW I805008B
- Authority
- TW
- Taiwan
- Prior art keywords
- intent
- plural
- intention
- instructions
- candidate
- Prior art date
Links
Images
Landscapes
- Measuring Or Testing Involving Enzymes Or Micro-Organisms (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Investigating Or Analysing Biological Materials (AREA)
- Machine Translation (AREA)
Abstract
Description
本發明係關於一種意圖評選技術,尤其指一種客製化意圖評選系統、方法及電腦可讀媒介。 The present invention relates to an intention selection technology, in particular to a customized intention selection system, method and computer readable medium.
現今的使用者習慣以語音對智慧型互動裝置(如智慧型手機、智慧型音箱等)下達指令,俾透過智慧型互動裝置完成使用者需求的服務,但使用者下達的指令可能同時包含多重語意,導致智慧型互動裝置容易產生錯誤回應。 Today's users are accustomed to giving instructions to smart interactive devices (such as smart phones, smart speakers, etc.) by voice, so as to complete the services required by users through smart interactive devices, but the instructions issued by users may contain multiple meanings at the same time , causing the smart interactive device to easily generate an error response.
舉例來說,當使用者的指令為[我要聽故事]時,由於智慧型互動裝置中可能存在有歌名為「故事」的歌曲以及書名為「故事」的書。在這種情境下,雖習知技術例如Google Home或Amazon Alexa,還是能針對多重語意的指令直接進行觸發裝置行為,但此行為不一定滿足使用者的真正意圖。因而,如何快速且精準地回應出使用者滿意的答案,將是智慧型互動裝置的重要任務。 For example, when the user's instruction is "I want to listen to a story", there may be a song titled "Story" and a book titled "Story" in the smart interactive device. In this situation, although conventional technologies such as Google Home or Amazon Alexa can still directly trigger device actions for multilingual commands, this action may not necessarily satisfy the user's true intention. Therefore, how to quickly and accurately respond to the user's satisfactory answer will be an important task of the intelligent interactive device.
因此,如何提供一種創新之意圖評選技術,以克服上述習知技術之缺失或提供相關之功能,已成目前亟欲解決的課題。 Therefore, how to provide an innovative intention selection technology to overcome the lack of the above-mentioned conventional technology or provide related functions has become an urgent problem to be solved.
為了解決上述問題或提供相關之功能,本發明提供一種客製化意圖評選系統,係包括:一自然語言處理模組,係將一文字訊息進行斷詞處理以產生複數斷詞後,將該複數斷詞進行詞性及命名實體標註,以使該複數斷詞均具有詞性標註及識別標籤,再由該自然語言處理模組將該複數具有詞性標註及識別標籤之斷詞與一語意清單中之複數意圖模板進行匹配,得到該複數具有詞性標註及識別標籤之斷詞之相關意圖模板,俾由該自然語言處理模組透過該複數具有詞性標註及識別標籤之斷詞之該相關意圖模板得到相對應之該語意清單中之複數候選意圖指令;以及一意圖評選模組,係接收來自該自然語言處理模組之該語意清單中之該複數候選意圖指令,以由該意圖評選模組依據一客製邏輯決策樹判斷該複數候選意圖指令之執行順序,進而根據該執行順序經計算後產生一執行意圖指令。 In order to solve the above problems or provide related functions, the present invention provides a customized intent selection system, which includes: a natural language processing module, which performs segmentation processing on a text message to generate plural segmentation words, and the plural segmentation Words are tagged with part-of-speech and named entities so that the plural segmented words have part-of-speech tagging and identification tags, and then the natural language processing module combines the plural segmented words with part-of-speech tagging and identification tags with the plural intentions in a semantic list Templates are matched to obtain the relevant intent templates of the plurality of segmented words with part-of-speech tags and identification tags, so that the natural language processing module can obtain the corresponding intent templates through the plurality of segmented words with part-of-speech tags and identification tags. A plurality of candidate intent commands in the semantic list; and an intent selection module that receives the plurality of candidate intent commands in the semantic list from the natural language processing module, so that the intent selection module is based on a custom logic The decision tree judges the execution order of the plurality of candidate intended instructions, and generates an execution intention instruction after calculation according to the execution order.
本發明提供一種客製化意圖評選方法,係包括:由一自然語言處理模組將一文字訊息進行斷詞處理以產生複數斷詞後,將該複數斷詞進行詞性及命名實體標註,以使該複數斷詞均具有詞性標註及識別標籤;由該自然語言處理模組將該複數具有詞性標註及識別標籤之斷詞與一語意清單中之複數意圖模板進行匹配,得到該複數具有詞性標註及識別標籤之斷詞之關聯意圖模板;由該自然語言處理模組透過該複數具有詞性標註及識別標籤之斷詞之該關聯意圖模板得到相對應之該語意清單中之複數候選意 圖指令;由一意圖評選模組接收來自該自然語言處理模組之該語意清單中之該複數候選意圖指令,以由該意圖評選模組依據一客製邏輯決策樹判斷該複數候選意圖指令之執行順序;以及由該意圖評選模組根據該執行順序經計算後產生一執行意圖指令。 The present invention provides a method for selecting customized intentions, which includes: after a natural language processing module performs segmentation processing on a text message to generate plural segmentations, the plural segmentations are tagged with part of speech and named entities, so that the Plural segmented words have part-of-speech tagging and identification tags; the natural language processing module matches the plural segmented words with part-of-speech tagging and identification tags with the plural intent template in a semantic list, and obtains the plural with part-of-speech tagging and identification The association intention template of the tagged word segmentation; the natural language processing module obtains the corresponding plural candidate meanings in the semantic list through the association intent template of the plural segmentation words with part-of-speech tags and identification tags Graph instruction: an intent selection module receives the plurality of candidate intent instructions from the semantic list of the natural language processing module, so that the intent selection module judges the plurality of candidate intent instructions according to a custom logic decision tree an execution order; and the intention selection module generates an execution intention instruction after calculation according to the execution order.
於一實施例中,該自然語言處理模組包括一詞性標註單元及一命名實體單元,其中,該詞性標註單元對該複數斷詞進行詞性標註,以使該複數斷詞均具有該詞性標註,而該命名實體單元再將該複數具有詞性標註之斷詞進行命名實體標註,以使該複數斷詞均進一步具有該識別標籤。 In one embodiment, the natural language processing module includes a part-of-speech tagging unit and a named entity unit, wherein the part-of-speech tagging unit performs part-of-speech tagging on the plural segmented words, so that the plural segmented words all have the part-of-speech tagging, The named entity unit then performs named entity tagging on the plural segmented words with part-of-speech tagging, so that the plural segmented words further have the identification tag.
於一實施例中,該自然語言處理模組包括一邏輯與語意編輯單元,係提供一服務提供端編輯該複數意圖模板及其相對應之複數意圖指令及複數服務行為,以依據該複數意圖模板、該複數意圖指令及該複數服務行為形成該語意清單後,該邏輯與語意編輯單元復提供該服務提供端設定該複數意圖指令之執行順序,以依據該複數意圖指令之執行順序形成該客製邏輯決策樹。 In one embodiment, the natural language processing module includes a logic and semantic editing unit, which is to provide a service provider to edit the plural intent templates and corresponding plural intent instructions and plural service behaviors, so as to base on the plural intent templates After the semantic list is formed by the multiple intent commands and the multiple service behaviors, the logic and semantic editing unit further provides the service provider to set the execution order of the multiple intent commands, so as to form the custom system according to the execution order of the multiple intent commands Logical decision tree.
於一實施例中,該意圖評選模組包括一訂閱資訊單元,係依據一訂閱資訊篩選該複數候選意圖指令中與該訂閱資訊無關的候選意圖指令。 In one embodiment, the intention selection module includes a subscription information unit, which is used to screen candidate intention instructions among the plurality of candidate intention instructions that are not related to the subscription information according to a subscription information.
於一實施例中,該意圖評選模組包括一支援向量機分類器,係將該文字訊息轉成向量後,透過機器學習語意理解模型進行意圖預測以產生一預測意圖指令,其中,當該意圖評選模組無法依據該客製邏輯決策樹判斷出該複數候選意圖指令中何者之執行順序為最高者時,該支援向量機分類器將該預測意圖指令比對該複數候選意圖指令,進而從該複數候選 意圖指令中比對出與該預測意圖指令一致之意圖指令,作為該執行意圖指令。 In one embodiment, the intent selection module includes a support vector machine classifier, which converts the text message into a vector, and performs intent prediction through a machine learning semantic understanding model to generate a predicted intent instruction, wherein, when the intent When the selection module cannot determine which of the plurality of candidate intent instructions has the highest execution order according to the custom logic decision tree, the support vector machine classifier compares the predicted intent instruction with the plurality of candidate intent instructions, and then selects from the plurality of candidate intent instructions plural candidates The intention instruction consistent with the predicted intention instruction is compared among the intention instructions as the execution intention instruction.
於一實施例中,該意圖評選模組包括一特徵工程單元,係將使用者之服務歷程與個人資訊進行特徵萃取以得到所需要的複數特徵後,該特徵工程單元將該複數特徵建置成特徵向量,以提供給該支援向量機分類器對該機器學習語意理解模型進行訓練。 In one embodiment, the intent selection module includes a feature engineering unit, which extracts the features of the user's service history and personal information to obtain the required multiple features, and the feature engineering unit builds the multiple features into The feature vector is provided to the support vector machine classifier to train the machine learning semantic understanding model.
此外,本發明之電腦可讀媒介係應用於計算裝置或電腦中,並儲存有指令,以執行上述之客製化意圖評選方法。 In addition, the computer-readable medium of the present invention is applied in a computing device or a computer, and stores instructions to execute the above-mentioned customizing intent selection method.
由上述可知,本發明係提供一種客製化意圖評選系統、方法及電腦可讀媒介,藉由預設的客製邏輯決策樹或經由機械學習得到的意圖預測結果來分析使用者的候選意圖指令,進而將正確的意圖指令提供給使用端裝置,以執行使用者需要的服務。因此,相較於無法正確地判斷使用者所下達具有多重語意之指令的習知技術,本發明不但能提供服務提供端依據目標客群設定客製邏輯決策樹以進行服務導引,更能藉由使用著過去的使用歷程及個人資料快速且精準地分析出使用者真正的意圖。 As can be seen from the above, the present invention provides a system, method, and computer-readable medium for customizing intention selection, which analyzes the user's candidate intention instructions by using a preset custom logical decision tree or intention prediction results obtained through machine learning. , and then provide the correct intent command to the user device to execute the service required by the user. Therefore, compared with the conventional technology that cannot correctly judge the instruction with multiple semantics issued by the user, the present invention can not only provide the service provider to set a customized logical decision tree according to the target customer group for service guidance, but also can use Quickly and accurately analyze the user's true intentions from the user's past usage history and personal data.
1:客製化意圖評選系統 1: Customized intent selection system
10:自然語言處理模組 10: Natural language processing module
11:斷詞單元 11: Segmentation unit
12:詞性標註單元 12: Part-of-speech tagging unit
13:命名實體單元 13: Named Entity Units
14:語意理解單元 14: Semantic Comprehension Unit
15:邏輯與語意編輯單元 15: Logic and Semantic Editing Unit
16:語意理解資料庫 16:Semantic understanding database
20:意圖評選模組 20: Intent Selection Module
21:訂閱資訊單元 21:Subscription information unit
22:邏輯執行單元 22: Logic execution unit
23:支援向量機分類器 23: Support Vector Machine Classifier
24:特徵工程單元 24: Feature engineering unit
25:使用者歷程資料庫 25: User Journey Database
26:個人資料庫 26: Personal database
30:使用端裝置 30: Use end device
40:服務提供端 40: Service Provider
C1:服務歷程之特徵向量 C1: Feature vector of service history
C2:個人資料之特徵向量 C2: Feature vector of personal data
S31至S36:步驟 S31 to S36: Steps
圖1係為本發明之客製化意圖評選系統之架構示意圖; FIG. 1 is a schematic diagram of the structure of the customized intention evaluation system of the present invention;
圖2係為本發明之使用者之服務歷程及個人資料的特徵向量之示意圖;以及 Figure 2 is a schematic diagram of the user's service history and personal data feature vectors of the present invention; and
圖3係為本發明之客製化意圖評選方法之流程示意圖。 FIG. 3 is a schematic flow chart of the method for selecting customized intentions according to the present invention.
以下藉由特定的具體實施例說明本發明之實施方式,熟悉此技藝之人士可由本說明書所揭示之內容輕易地瞭解本發明之其他優點及功效。 The implementation of the present invention is described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification.
須知,本說明書所附圖式所繪示之結構、比例、大小等,均僅用以配合說明書所揭示之內容,以供熟悉此技藝之人士之瞭解與閱讀,並非用以限定本發明可實施之限定條件,故不具技術上之實質意義,任何結構之修飾、比例關係之改變或大小之調整,在不影響本發明所能產生之功效及所能達成之目的下,均應仍落在本發明所揭示之技術內容得能涵蓋之範圍內。同時,本說明書中所引用之如「一」、「第一」、「第二」、「上」及「下」等之用語,亦僅為便於敘述之明瞭,而非用以限定本發明可實施之範圍,其相對關係之改變或調整,在無實質變更技術內容下,當視為本發明可實施之範疇。 It should be noted that the structures, proportions, sizes, etc. shown in the drawings attached to this specification are only used to match the content disclosed in the specification, for the understanding and reading of those familiar with this technology, and are not used to limit the implementation of the present invention Therefore, it has no technical substantive meaning. Any modification of structure, change of proportional relationship or adjustment of size shall still fall within the scope of this invention without affecting the effect and purpose of the present invention. The technical content disclosed by the invention must be within the scope covered. At the same time, terms such as "a", "first", "second", "upper" and "lower" quoted in this specification are only for the convenience of description and are not used to limit the scope of the present invention. The scope of implementation, the change or adjustment of its relative relationship, without substantial change in technical content, shall be regarded as the scope of implementation of the present invention.
圖1係為本發明之客製化意圖評選系統1之架構示意圖。如圖1所示,該客製化意圖評選系統1係包括一自然語言處理模組10及一意圖評選模組20,亦可進一步包括至少一(如複數)使用端裝置30及至少一(如複數)服務提供端40,其中,「複數」代表二個以上(如二、三、四或五個以上)。自然語言處理模組10與意圖評選模組20係分別通訊連接使用端裝置30及服務提供端40,且自然語言處理模組10接收由使用端裝置30進行語音轉換所產生之一文字訊息。在一實施例中,使用端裝置30係利用現有技術之語音辨識技術將語音轉換成文字訊息,再傳送文字訊息至自然語言處理模組10,於此不限定採用何種語音辨識技術。
FIG. 1 is a schematic diagram of the architecture of a customized
自然語言處理模組10包括一斷詞單元11、一詞性標註單元12、一命名實體單元13、一語意理解單元14、一邏輯與語意編輯單元15及一語意理解資料庫16,其中,所述之自然語言處理模組10係用以接收文字訊息,且對文字訊息進行斷詞處理,以將斷詞處理之結果進行文字訊息之詞性標註與命名實體標註,最後進行語意理解,進而得到複數候選意圖指令。
The natural
意圖評選模組20包括一訂閱資訊單元21、一邏輯執行單元22、一支援向量機(Support Vector Machine,SVM)分類器23、一特徵工程單元24、一使用者歷程資料庫25及一個人資料庫26,其中,所述之意圖評選模組20係通訊連接自然語言處理模組10,且用以將自然語言處理模組10所理解出之複數候選意圖指令進行評選,以從複數候選意圖指令中找出最適合使用者的執行意圖指令,再執行此執行意圖指令所對應之服務行為。
The
具體而言,自然語言處理模組10與意圖評選模組20係建立於伺服器(如通用型伺服器、檔案型伺服器、儲存單元型伺服器等)及電腦等具有適當演算機制之電子設備中,且自然語言處理模組10與意圖評選模組20均可為軟體、硬體或韌體;若為硬體,則可為具有資料處理與運算能力之處理單元、處理器、電腦或伺服器;若為軟體或韌體,則可包括處理單元、處理器、電腦或伺服器可執行之指令,且可安裝於同一硬體裝置或分布於不同的複數硬體裝置。再者,使用端裝置30可為智慧型音箱、智慧型手機、平板電腦、個人電腦或其他電子裝置等,但不限於上述,而服務提供端40係指提供各項服務(如影音串流服務等)之服務提供商或服務提供設備。
Specifically, the natural
所述之斷詞單元11係接收文字訊息後進行斷詞處理,以將文字訊息切分成複數斷詞。在一實施例中,斷詞之目的是將輸入的文字訊息切分成具有最小意涵的單位,且斷詞方式有統計式、辭典式或結合兩者之混合式,於此不限定斷詞方式。例如,斷詞單元11接收到使用者之文字訊息的指令(例如[我要聽周杰倫的稻香]),並將文字訊息的指令進行斷詞以得到複數斷詞(例如[我,要,聽,周杰倫,的,稻香]),其中,各個斷詞之間使用符號(如逗點)分隔。
The
所述之詞性標註單元12係分別對複數斷詞進行詞性標註,以使複數斷詞均具有詞性標註,其中,詞性標註之方式有統計式、辭典式或結合兩者之混合式。再者,目前主要使用統計式,而本實施例係採用基於Chinese Treebank(中文樹庫)3.0定義下的統計式。例如,詞性標註單元12將複數斷詞(例如[我,要,聽,周杰倫,的,稻香])進行詞性標註,以產生複數具有詞性標註之斷詞(例如[我_PN,要_VV,聽_VV,周杰倫_NR,的_DEG,稻香_NR])。
The part-of-
所述之命名實體單元13係透過命名實體識別(Named Entity Recognizer,NER)將複數具有詞性標註之斷詞進行命名實體標註,以使複數斷詞均具有詞性標註及識別標籤,其中,命名實體識別具有統計式、辭典式或結合兩者之混合式。再者,目前主要使用統計式,而本實施例之識別標籤係由訓練資料所標註的標籤來決定,代表可自定義。例如,命名實體單元13係透過命名實體識別將[我_PN,要_VV,聽_VV,周杰倫_NR,的_DEG,稻香_NR]進行命名實體標註,以產生[我_PN_o,要_VV_o,聽_VV_o,周杰倫_NR_SINGER,的_DEG_o,稻香_NR_SONG],其中,o代表非特別之命名實體、SINGER代表歌手以及SONG代表歌曲,且SINGER及SONG係為自定義的標籤。
The named
所述之語意理解單元14係利用複數具有識別標籤之斷詞匹配語意清單中之意圖模板進行語意理解,以匹配出複數候選意圖指令。詳言之,語意理解單元14採用模板基礎(Pattern-based)之語意理解方式透過意圖模板來進行語意理解。是以,語意理解單元14將複數具有詞性標註及識別標籤之斷詞與語意清單中之複數意圖模板進行匹配,進而得到與複數具有詞性標註及識別標籤之斷詞相關聯之複數意圖模板,且語意理解單元14再透過複數具有詞性標註及識別標籤之斷詞相關聯之複數意圖模板得到相對應之語意清單中之複數候選意圖指令。此外,若語意理解單元14未能匹配出相對應之意圖模板,則語意理解單元14透過使用端裝置30回應無法執行服務;或是,若語意理解單元14僅匹配出一個相對應之意圖模板,則將此意圖模板所對應之一候選意圖指令提供至意圖評選模組20,以由意圖評選模組20將此候選意圖指令直接作為執行意圖指令。
The
例如,語意理解單元14將[我_PN_o,要_VV_o,聽_VV_o,周杰倫_NR_SINGER,的_DEG_o,稻香_NR_SONG]依序比對所有的意圖模板,可以比對出[我_PN_o,要_VV_o,聽_VV_o,周杰倫_NR_SINGER的_DEG_o,稻香_NR_SONG]與[我,要,聽,SINGER,的,SONG]之意圖模板一致或相同,進而得到候選意圖指令。
For example, the
所述之邏輯與語意編輯單元15係提供服務提供端40能依據其目標客群預先編輯複數意圖模板及其相對應之複數意圖指令及複數服務行為,以依據複數意圖模板、複數意圖指令及複數服務行為形成語意清單(如表1所示)。例如,邏輯與語意編輯單元15提供介面(如網頁或智慧型手機應用程式,於此不限)給服務提供端40建置[我,要,聽,SINGER,的,SONG]之意圖模板,並設定意圖指令係為[I1]以及服務行為係為[開啟音樂服務S1]。
The logic and
表1:語意清單
在一實施例中,邏輯與語意編輯單元15亦可提供服務提供端40設定當複數具有詞性標註及識別標籤之斷詞觸發不同意圖指令之服務時的執行順序,以形成客製邏輯決策樹,例如,服務提供端40(如第一服務提供端)可依據其目標客群設定當同時觸發第一服務行為與第二服務行為時,第一服務行為永遠會優先於第二服務行為被執行,另一服務提供端40(如第二服務提供端)也可依據其目標客群設定第二服務行為永遠優先於第三服務行為被執行。
In one embodiment, the logic and
所述之語意理解資料庫16係儲存服務提供端40自定義的語意清單,且提供語意清單中之意圖模板給語意理解單元14進行相似度比對。
The
所述之訂閱資訊單元21係依據使用者之訂閱資訊篩選複數候選意圖指令中與該訂閱資訊無關的候選意圖指令,以排除與訂閱資訊無關的候選意圖指令。在一實施例中,訂閱資訊係指使用者於使用端裝置30中能夠使用的服務列表,例如,使用端裝置30可提供給使用者具有複數服務行為之服務清單,由使用者選擇要訂閱(取用)之服務清單中之服務行為,藉此形成服務列表(亦即訂閱資訊)。例如,當觸發候選意圖指令時,若訂閱
資訊中沒有訂閱音樂類別的服務,則令使用端裝置30回應無訂閱音樂服務予使用者。
The
所述之邏輯執行單元22係自訂閱資訊單元21接收與訂閱資訊有關之複數候選意圖指令,且依據邏輯與語意編輯單元15之客製邏輯決策樹來判斷複數候選意圖指令之執行順序,進而將複數候選意圖指令中執行順序為最高者作為一執行意圖指令。在另一實施例中,該執行順序亦可經計算,以產生一執行意圖指令。
The
所述之支援向量機(SVM)分類器23係透過機器學習(Machine Learning)中之支援向量機(SVM)並依據使用者過往使用之服務歷程與個人資訊所建置的特徵向量進行客製化的機器學習語意理解模型訓練,且將文字訊息轉成向量後,透過機器學習語意理解模型進行意圖預測以產生一預測意圖指令(即意圖預測結果)。在一實施例中,當邏輯執行單元22能透過客製邏輯決策樹將複數候選意圖指令中執行順序為最高者作為一執行意圖指令,且將執行意圖指令傳送給支援向量機分類器23時,雖支援向量機分類器23一樣會產生預測意圖指令,但仍然直接觸發由邏輯執行單元22傳來的執行意圖指令。
The support vector machine (SVM)
所述之特徵工程單元24係將使用者過往使用之服務歷程與個人資訊進行特徵萃取以得到所需要的複數特徵後,將複數特徵建置成特徵向量提供給支援向量機分類器23對機器學習語意理解模型進行訓練。
The
所述之使用者歷程資料庫25係記錄使用者之服務歷程。例如,使用者每周一晚上七點都會使用指令「我要聽周杰倫的稻香」來聽周杰倫的音樂,使用者歷程資料庫25就會記錄此服務歷程。
The
所述之個人資料庫26係記錄個人資訊。例如,使用者的性別、職業、教育程度、所在地與年齡等個人資訊,但不限於上述。此外,此個人資訊會在一開始使用使用端裝置30時進行登錄。
The
下列實施例係為本發明之客製化意圖評選系統1之第一實施例,且此第一實施例之主要內容如下,其餘內容相同於上述圖1之說明。
The following embodiment is the first embodiment of the customized
於本實施例中,使用者透過使用端裝置30訂閱了音樂服務S1、音樂服務S2及音樂服務S3,且當使用者對使用端裝置30講了[我要聽劉德華的冰雨這首歌]之語音訊息後,由使用端裝置30將語音訊息轉換為文字訊息以傳送至客製化意圖評選系統1。
In this embodiment, the user subscribes to music service S1, music service S2, and music service S3 through the
具體而言,客製化意圖評選系統1之自然語言處理模組10接收來自使用端裝置30之文字訊息,由斷詞單元11將文字訊息(例如[我要聽劉德華的冰雨這首歌])切分成複數斷詞(例如[我,要,聽,劉德華,的,冰雨,這,首,歌]),且詞性標註單元12將複數斷詞進行詞性標註,以產生複數具有詞性標註之斷詞(例如[我_PN,要_VV,聽_VV,劉德華_NR,的_DER,冰雨_NR,這_PN,首_M,歌_NN]),其中,PN為代名詞、VV為動詞、NR為專有名詞、DER是「的」這個字的特有詞性,M為量詞,NN則是名詞。再者,由命名實體單元13將複數具有詞性標註之斷詞進行命名實體標註,以產生複數具有詞性標註及識別標籤之斷詞(例如[我_PN_o,要_VV_o,聽_VV_o,劉德華_NR_SINGER,的_DER_o,冰雨_NR_SONG,這_PN_,首_M_o,歌_NN_o]),其中,o代表該詞彙不是命名實體,SINGER代表歌手,SONG代表歌曲。
Specifically, the natural
接著,語意理解單元14將複數具有詞性標註及識別標籤之斷詞(例如[我_PN_o,要_VV_o,聽_VV_o,劉德華_NR_SINGER,的_DER_o,冰雨_NR_SONG,這_PN_,首_M_o,歌_NN_o])匹配語意清單中之意圖模板
(如上述表1所示),以透過匹配出的意圖模板得到兩個候選意圖指令I1,I2,如表2所示。
Next, the
表2:候選意圖指令
意圖評選模組20之訂閱資訊單元21依據使用者之訂閱資訊篩選候選意圖指令I1,I2,以確認使用者有訂閱相關服務後,由邏輯執行單元22依據服務提供端40於邏輯與語意編輯單元15中所設定之客製邏輯決策樹(如表3所示)來決定候選意圖指令之執行順序,其中,邏輯執行單元22依據規則R1將候選意圖指令I1作為執行意圖指令。
The
表3:客製邏輯決策樹
之後,支援向量機分類器23將原先的[我要聽劉德華的冰雨這首歌]之文字訊息轉成向量後,透過機器學習語意理解模型進行意圖預測以產生一預測意圖指令,且將預測意圖指令與執行意圖指令I1進行比對,其中,不論預測意圖指令與執行意圖指令I1比對後是否相同,皆觸發執行意圖指令I1,以令使用端裝置30開啟音樂服務S1。
Afterwards, the support
此外,特徵工程單元24會從使用者歷程資料庫25與個人資料庫26對使用者之服務歷程及個人資訊做特徵萃取,以產生服務歷程之特徵向量C1及個人資料之特徵向量C2,俾提供給支援向量機分類器23進行意圖
預測。舉例而言,如圖2所示,係為使用者過去在周一晚上七點,曾用意圖指令觸發開啟音樂服務S1所產生的特徵向量C1,C2。
In addition, the
甚者,如圖2及表4所示,服務歷程之特徵向量C1之文字向量中之詞袋(Bag of Words)模型會因使用者累積使用的意圖指令多寡而有變化,且使用者歷程資料庫25儲存有服務歷程之特徵向量C1中各特徵的對應表,例如,表5之時間特徵對應表、表6之星期特徵對應表、表7之意圖指令特徵對應表,其中,意圖指令特徵對應表會由服務提供端40於語意理解資料庫16中預先進行編號,例如,表1之語意清單可對應出表7之意圖指令特徵對應表。又,個人資料庫26儲存有個人資料之特徵向量C2中各特徵的對應表,例如,表8之性別特徵對應表、表9之行業別特徵對應表、表10之地區特徵對應表與表11之教育程度特徵對應表,其中,年齡係以原始年齡不計月當成特徵值。是以,前述該些特徵向量C1,C2之對應表(如表4至表11)可為機器學習之特徵向量表,用以提供給支援向量機分類器23對文字訊息進行機器學習。
What's more, as shown in Figure 2 and Table 4, the Bag of Words model in the text vector of the feature vector C1 of the service history will change due to the number of intention instructions accumulated by the user, and the user history data The
表4:詞袋模型
表5:時間特徵對應表
表6:星期特徵對應表
表7:意圖指令特徵對應表
表8:性別特徵對應表
表9:行業別特徵對應表(按照行政院主計處資訊分類)
表10:地區特徵對應表
表11:教育程度特徵對應表
另一方面,使用者歷程資料庫25記錄有使用者情境特徵(如表12所示),以提供支援向量機分類器23的訓練資料。
On the other hand, the
表12:使用者情境特徵表
又,個人資料庫26也記錄有使用者之個人資料特徵(如表13所示),以提供支援向量機分類器23的訓練資料。
Moreover, the
表13:個人資料特徵表
下列實施例係為本發明之客製化意圖評選系統1之第二實施例,且此第二實施例之主要內容如下,其餘內容相同於上述圖1及第一實施例之說明。
The following embodiment is the second embodiment of the customized
於本實施例中,使用者透過使用端裝置30訂閱了音樂服務S2、故事服務S4及廣播服務S5,且當使用者對使用端裝置30講了[我要聽故事]之語音訊息後,將語音訊息轉換為文字訊息以傳送至客製化意圖評選系統1。
In this embodiment, the user subscribes to the music service S2, the story service S4 and the broadcast service S5 through the
具體而言,客製化意圖評選系統1之自然語言處理模組10接收來自使用端裝置30之文字訊息,由斷詞單元11將文字訊息(例如[我要聽故事])切分成複數斷詞(例如[我,要,聽,故事]),且詞性標註單元12將複數斷詞進行詞性標註,以產生複數具有詞性標註之斷詞(例如[我_PN,要_VV,聽_VV,故事_NN])。再者,由命名實體單元13將複數具有詞性標註之斷詞進行命名實體標註,以產生複數具有詞性標註及識別標籤之斷詞(例如[我_PN_o,要_VV_o,聽_VV_o,故事_NN_SONG/STORY]),其中,「故事」可能代表歌名為故事的歌曲,或是單純表示要聽故事。
Specifically, the natural
語意理解單元14將複數具有詞性標註及識別標籤之斷詞(例如[我_PN_o,要_VV_o,聽_VV_o,故事_NN_SONG/STORY])匹配語意清單中之意圖模板(如上述表1所示),以透過匹配出的意圖模板得到兩個候選意圖指令12,I4,如表14所示。
The
表14:候選意圖指令
意圖評選模組20之訂閱資訊單元21依據使用者之訂閱資訊篩選候選意圖指令I2,I4,以確認使用者有訂閱相關服務後,由邏輯執行單元22依據服務提供端40於邏輯與語意編輯單元15中所設定之客製邏輯決策樹(如表3所示)來判斷候選意圖指令的執行順序。然而,當客製邏輯決策樹並未有設定候選意圖指令I2,I4之執行順序時,邏輯執行單元22會將候選意圖指令I2,I4直接傳送給支援向量機分類器23。
The
此時,支援向量機分類器23將[我要聽故事]之文字訊息轉成向量後進行意圖預測以產生一預測意圖指令,再將預測意圖指令與候選意圖指令I2,I4進行比對,俾取得預測意圖指令與候選意圖指令I2,I4之間的交集;換言之,支援向量機分類器23可取得與預測意圖指令一致或相同之候選意圖指令以得到候選意圖指令I2,最後將候選意圖指令I2作為執行意圖指令,進而由使用端裝置30開啟音樂服務S2。
At this time, the support
圖3係為本發明之客製化意圖評選之方法流程示意圖,且此客製化意圖評選之方法流程之主要內容如下,其餘內容相同於上述圖1說明,於此不再重覆敘述,其中,該方法流程包含下列步驟S31至步驟S36。 Fig. 3 is a schematic diagram of the method flow chart of the customization intention selection method of the present invention, and the main content of the method flow chart of the customization intention selection method is as follows, and the rest of the content is the same as that described in Fig. 1 above, and will not be repeated here. , the method flow includes the following steps S31 to S36.
於步驟S31中,自然語言處理模組10將使用者之文字訊息進行斷詞處理以產生複數斷詞。
In step S31 , the natural
於步驟S32中,自然語言處理模組10對斷詞進行詞性標註及命名實體標註,以使複數斷詞均具有詞性標註及識別標籤。
In step S32 , the natural
於步驟S33中,自然語言處理模組10利用複數具有識別標籤之斷詞匹配語意清單中之複數意圖模板以進行語意理解,進而匹配出複數候選意圖指令。
In step S33 , the natural
於步驟S34中,意圖評選模組20接收來自自然語言處理模組10的複數候選意圖指令後,依據使用者之訂閱資訊篩選複數候選意圖指令,以排除與訂閱資訊無關的候選意圖指令。
In step S34 , after receiving the plurality of candidate intent commands from the natural
於步驟S35中,意圖評選模組20依據客製邏輯決策樹判斷複數候選意圖指令之執行順序,進而將複數候選意圖指令中執行順序為最高者作為一執行意圖指令。
In step S35 , the
於步驟S36中,意圖評選模組20將原先的文字訊息轉成向量後進行意圖預測,以產生一預測意圖指令,其中,當無法依據客製邏輯決策樹判斷複數候選意圖指令中何者的執行順序為最高者時,意圖評選模組20將預測意圖指令比對複數候選意圖指令,進而從複數候選意圖指令中比對出與預測意圖指令一致或相同之意圖指令以作為一執行意圖指令。
In step S36, the
此外,本發明還揭示一種電腦可讀媒介,係應用於具有處理器(例如,CPU、GPU等)及/或記憶體的計算裝置或電腦中,且儲存有指令,並可利用此計算裝置或電腦透過處理器及/或記憶體執行此電腦可讀媒介,以於執行此電腦可讀媒介時執行上述之方法及各步驟。 In addition, the present invention also discloses a computer-readable medium, which is applied to a computing device or computer having a processor (for example, CPU, GPU, etc.) and/or memory, and stores instructions, and can be used by this computing device or The computer executes the computer-readable medium through the processor and/or memory, so as to execute the above-mentioned method and each step when executing the computer-readable medium.
綜上所述,本發明係提供一種客製化意圖評選系統、方法及電腦可讀媒介,藉由預設的客製邏輯決策樹或經由機器學習得到的意圖預測結果來分析使用者的候選意圖指令,進而將正確的意圖指令提供給使用端裝置,以執行使用者需要的服務,故相較於無法正確地判斷使用者所下達具有多重語意 之指令的習知技術,本發明不但能提供服務提供端依據目標客群設定客製邏輯決策樹以進行服務導引,更能藉由使用著過去的使用歷程及個人資料快速且精準地分析出使用者真正的意圖。 In summary, the present invention provides a system, method, and computer-readable medium for customizing intention selection, which analyzes the user's candidate intentions by using a preset custom logical decision tree or intention prediction results obtained through machine learning. Instructions, and then provide the correct intent instructions to the user device to perform the services required by the user, so compared with the inability to correctly judge the multi-semantics issued by the user According to the conventional technology of command, the present invention can not only provide the service provider to set a customized logical decision tree according to the target customer group for service guidance, but also quickly and accurately analyze it by using past usage history and personal data the user's true intent.
是以,相較於習知技術,本發明具有以下技術特徵及其功效: Therefore, compared with the prior art, the present invention has the following technical features and effects thereof:
1.本發明能提供服務提供端調整滿足目標客群的服務優先順序,藉此形成客製邏輯決策樹,並讓使用者可以依照服務提供端所設定之客製邏輯決策樹進行服務導引。 1. The present invention can provide the service provider to adjust the service priority to meet the target customer group, thereby forming a customized logical decision tree, and allowing the user to conduct service guidance according to the customized logical decision tree set by the service provider.
2.本發明利用使用者之服務歷程與個人資訊進行機器學習語意理解模型的訓練,以針對使用者之個人習慣提供客製化的機器學習語意理解模型,使機器學習語意理解模型能依據使用者之習慣產生意圖預測結果,且進行正確的服務回應,以更符合使用者真正的意圖,進而避免提供錯誤的服務給使用者。 2. The present invention uses the user's service history and personal information to train the machine learning semantic understanding model to provide a customized machine learning semantic understanding model for the user's personal habits, so that the machine learning semantic understanding model can be based on the user's The habit of generating intention prediction results, and the correct service response, in order to better meet the user's true intentions, and thus avoid providing wrong services to users.
3.本發明將使用者之文字訊息進行斷詞後進行詞性標註及命名實體標註,以透過匹配複數具有詞性標註及識別標籤之斷詞與意圖模板,進而快速地找出相對應的候選意圖指令。 3. The present invention performs part-of-speech tagging and named entity tagging after segmenting the user's text message, so as to quickly find the corresponding candidate intent commands by matching plural segmented words and intent templates with part-of-speech tagging and identification tags .
上述實施形態僅例示性說明本發明之原理及其功效,而非用於限制本發明。任何熟習此項技藝之人士均可在不違背本發明之精神及範疇下,對上述實施形態進行修飾與改變。因此,本發明之權利保護範圍應如申請專利範圍所列。 The above-mentioned embodiments are only illustrative to illustrate the principles and effects of the present invention, and are not intended to limit the present invention. Anyone skilled in the art can modify and change the above-mentioned embodiments without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the rights of the present invention should be listed in the scope of the patent application.
1:客製化意圖評選系統 1: Customized intent selection system
10:自然語言處理模組 10: Natural language processing module
11:斷詞單元 11: Segmentation unit
12:詞性標註單元 12: Part-of-speech tagging unit
13:命名實體單元 13: Named Entity Units
14:語意理解單元 14: Semantic Comprehension Unit
15:邏輯與語意編輯單元 15: Logic and Semantic Editing Unit
16:語意理解資料庫 16:Semantic understanding database
20:意圖評選模組 20: Intent Selection Module
21:訂閱資訊單元 21:Subscription information unit
22:邏輯執行單元 22: Logic execution unit
23:支援向量機分類器 23: Support Vector Machine Classifier
24:特徵工程單元 24: Feature engineering unit
25:使用者歷程資料庫 25: User Journey Database
26:個人資料庫 26: Personal database
30:使用端裝置 30: Use end device
40:服務提供端 40: Service Provider
Claims (11)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW110136881A TWI805008B (en) | 2021-10-04 | 2021-10-04 | Customized intent evaluation system, method and computer-readable medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW110136881A TWI805008B (en) | 2021-10-04 | 2021-10-04 | Customized intent evaluation system, method and computer-readable medium |
Publications (2)
Publication Number | Publication Date |
---|---|
TW202316415A TW202316415A (en) | 2023-04-16 |
TWI805008B true TWI805008B (en) | 2023-06-11 |
Family
ID=86943179
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
TW110136881A TWI805008B (en) | 2021-10-04 | 2021-10-04 | Customized intent evaluation system, method and computer-readable medium |
Country Status (1)
Country | Link |
---|---|
TW (1) | TWI805008B (en) |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TW201027516A (en) * | 2009-01-09 | 2010-07-16 | Aibelive Co Ltd | Indication method of voice recognition system |
CN1815554B (en) * | 2005-02-03 | 2010-11-17 | 台达电子工业股份有限公司 | Vocabulary generating device, method and voice identifying system with same |
TW201115370A (en) * | 2009-10-28 | 2011-05-01 | Ind Tech Res Inst | Systems and methods for capturing and managing collective social intelligence information |
CN105869640A (en) * | 2015-01-21 | 2016-08-17 | 上海墨百意信息科技有限公司 | Method and device for recognizing voice control instruction for entity in current page |
CN108227932A (en) * | 2018-01-26 | 2018-06-29 | 上海智臻智能网络科技股份有限公司 | Interaction is intended to determine method and device, computer equipment and storage medium |
TWI639997B (en) * | 2017-09-28 | 2018-11-01 | 大仁科技大學 | Dialog understanding method based on probabilistic rule |
US20210035575A1 (en) * | 2019-08-01 | 2021-02-04 | Samsung Electronics Co., Ltd. | Electronic apparatus, and method of controlling to execute function according to voice command thereof |
TW202121230A (en) * | 2019-11-20 | 2021-06-01 | 中央研究院 | Natural language processing method and computing apparatus thereof |
-
2021
- 2021-10-04 TW TW110136881A patent/TWI805008B/en active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1815554B (en) * | 2005-02-03 | 2010-11-17 | 台达电子工业股份有限公司 | Vocabulary generating device, method and voice identifying system with same |
TW201027516A (en) * | 2009-01-09 | 2010-07-16 | Aibelive Co Ltd | Indication method of voice recognition system |
TW201115370A (en) * | 2009-10-28 | 2011-05-01 | Ind Tech Res Inst | Systems and methods for capturing and managing collective social intelligence information |
CN105869640A (en) * | 2015-01-21 | 2016-08-17 | 上海墨百意信息科技有限公司 | Method and device for recognizing voice control instruction for entity in current page |
TWI639997B (en) * | 2017-09-28 | 2018-11-01 | 大仁科技大學 | Dialog understanding method based on probabilistic rule |
CN108227932A (en) * | 2018-01-26 | 2018-06-29 | 上海智臻智能网络科技股份有限公司 | Interaction is intended to determine method and device, computer equipment and storage medium |
US20210035575A1 (en) * | 2019-08-01 | 2021-02-04 | Samsung Electronics Co., Ltd. | Electronic apparatus, and method of controlling to execute function according to voice command thereof |
TW202121230A (en) * | 2019-11-20 | 2021-06-01 | 中央研究院 | Natural language processing method and computing apparatus thereof |
Also Published As
Publication number | Publication date |
---|---|
TW202316415A (en) | 2023-04-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11302337B2 (en) | Voiceprint recognition method and apparatus | |
US10657332B2 (en) | Language-agnostic understanding | |
US11106983B2 (en) | Intelligent interaction method and intelligent interaction system | |
CN107818781B (en) | Intelligent interaction method, equipment and storage medium | |
US10089299B2 (en) | Multi-media context language processing | |
CN109196496B (en) | Unknown word predictor and content integrated translator | |
US20190013011A1 (en) | Machine learning dialect identification | |
US8156060B2 (en) | Systems and methods for generating and implementing an interactive man-machine web interface based on natural language processing and avatar virtual agent based character | |
US10984027B2 (en) | Techniques for user-centric document summarization | |
US9202142B1 (en) | Automatic assessment of books to determine suitability for audio book conversion | |
US20200134398A1 (en) | Determining intent from multimodal content embedded in a common geometric space | |
US20160071510A1 (en) | Voice generation with predetermined emotion type | |
US10460731B2 (en) | Apparatus, method, and non-transitory computer readable storage medium thereof for generating control instructions based on text | |
US11538476B2 (en) | Terminal device, server and controlling method thereof | |
US9418058B2 (en) | Processing method for social media issue and server device supporting the same | |
US10817264B1 (en) | User interface for a source code editor | |
WO2023045184A1 (en) | Text category recognition method and apparatus, computer device, and medium | |
US20230214579A1 (en) | Intelligent character correction and search in documents | |
US11653071B2 (en) | Responsive video content alteration | |
CN114722832A (en) | Abstract extraction method, device, equipment and storage medium | |
WO2021174814A1 (en) | Answer verification method and apparatus for crowdsourcing task, computer device, and storage medium | |
JP2022076439A (en) | Dialogue management | |
TWI805008B (en) | Customized intent evaluation system, method and computer-readable medium | |
WO2021097629A1 (en) | Data processing method and apparatus, and electronic device and storage medium | |
US20190005949A1 (en) | Linguistic profiling for digital customization and personalization |