top of page

人工智能 – 虛擬服務大使A.I. Ambassador的創作歷程 | Artificial Intelligence – The story of A.I. Ambassador


進入2020年,亦是人工智能的關鍵年。研究人工智能的公司都在尋找落地的位置,即所謂應用場景,以及他們能提供的應用方案。

Entering 2020, this is the milestone of artificial intelligence. Enterprises that develop A.I. are finding a place to enter the real business environment.


虛擬服務大使A.I. Ambassador是人工智能例子,它是一個怎樣的產品?它是在什麼原因下創造出來的?有興趣知道它的前世今生的讀者可以閱讀以下深度文。

A.I. Ambassador is one of the examples of Artificial Intelligence. What kind of product is it? Why is it created? For those who are interested to know more about its story behind could read our article.



 

筆者所在的人工智能公司於2017年在香港成立,最初業務以Chatbot為主, 創辦團隊最初已經非常關注用戶體驗, 不希望創造一個冷冰冰的對答系統,希望科技能夠解決問題之餘,亦相信人機互動需要溫度,畢竟可見的將來人類會花更多的時間在手機上。礙於當時的技術層面上,大多數Chatbot都是以按圖或者關鍵字為主,但是相信這類型的chatbot很快便無用武之地, 或者會淪為廣告信息另一個媒體而已。

The company I have joined is founded in 2017 in Hong Kong, focusing on developing Chatbot at first. Our founders pay special attention to user experience. Not only solving problems by technology, but creating a human-computer relationship with temperature, as people are spending more and more time on mobile phones. However, most of the Chatbots are click-based or keyword chatbot, these kinds of products are going to be fallen behind the times.


在這個關鍵位置需要一種技術讓電腦能夠理解人類的語言,即使今日所說的自然語言處理技術Natural Language Processing (NLP),最初很多人工智能公司都會利用Google的dialogue flow(前身api.ai) 或者 chatfuel等工具開發,但是很快便會遇上瓶頸,特別當對話intent比較多的時候,或者需要製作很複雜的Chatbot時制肘會非常多;而複雜的廣東話內容或中英夾雜等問題亦不好應付,所以自行開發的NLP 系統變得非常有需要。

At this point, allowing computer to understand human’s language become a key. Even there is Natural Language Processing (NLP) nowadays, many companies would still use Google’s dialogue flow (api.ai) or chatful to develop their chatbot, but the solution is not perfect. Especially when there are several intents, or a complicated chatbot is required. Cantonese and code-switching are not easy to handle, that is the reason why a self-developed NLP is important.


另外,Chatbot可以存在於手機等隨身電子媒介之上,為很多人處理日常所需,但是公司團隊很早便意識到Chatbot可以存在於更多的生活場景上面。早於2017年Asiabots 團隊已經嘗試把第一代的NLP系統連接在當時還未成熟的機械人身上。 那時候,與其說那是一個機械人,不如說那只是一個能在平地平衡地走路的顯示屏。第一次出席的場合是香港科學園人工智能會議亞太創新峰會 Apec innovation summit,當時NLP系統連接了一個與醫療為核心的人工智能系統clinicbot,能夠回答在場人士有關醫療的知識;除此之外,亦教曉了Chatbot一些與展會有關的問題,例如可以詢問演講嘉賓資料、回答和峰會主題有關的問題和會場設施等。是作為第一代Event Bot的技術展示, 讓在場的嘉賓感受和體驗機械人可塑造的角色和提供的服務。

Though chatbots are already integrated on mobile phones to handle daily enquiries, Asiabots soon realized that chatbots could be applied to more situations. In 2017, Asiabots attempted to integrate the first NLP onto robots, or maybe a walkable display. The first event it participated in was Apec Innovation Summit help in HKSTP. The Ambassador’s NLP system was connected to clinicbot, a medical chatbot, to answer visitors’ medical enquiry. The Chatbot also knew some more questions about the summit, for example, the information of guest speakers, event facilities, etc. This is the first showcase of Event Bot to the users.


第一代人工智能服務大使已經有懂得四處望的眼睛,用眼神與人互動 | A.I. Ambassador 1.0 has a pair of eyes that could look around

人工智能機械人

A.I. Ambassador


這一次的展示第一代看得到而且能夠互動的chatbot, 已經意識到聲音的重要性,因為當時的chatbot主要以文字作溝通媒介,但是當chatbot脫離手機界面,語音就變得舉足輕重了。那時候的解決方案主要利用Google以及Microsoft的TTS (Text-to-speech) Engine,英語的發音效果尚可,但是廣東話就遇到問題了,因為發音太像機械聲,而且選擇太少,基本上當你聽到語音,業界人士便知道來自哪一個Engine,全無個性可言。

The first chatbot with the interrelationship between human and robot, we had already realized the importance of voice, instead of using words only. While a chatbot ‘leaves’ a mobile, interacting by voice conversation becomes valuable. At that time, people mainly use Google or Microsoft Text-to-Speech Engine. English pronunciation is well-spoken, the problem is Cantonese. The pronunciation is robotic and not natural. Moreover, there are too few choices of voice, experts could identify which engine it is using.


第二個考量就是這種對話形式的人工智能應該放在哪一種「身體」上。Asiabots於2018年嘗試與大學合作,把NLP放在Humanoid身上,類近於Hanson Robotics的Sophia。當時的人形機械人士以氣壓帶動,動作反應雖快但是不太自然,比較之下一些商用型的機械人例如三寶 Sanbot,以及Pepper 則更好,起碼價錢會比較安心(笑) 這一類型的機械人都通常沒有自家NLP,取而言之採用Google 等大型機構的方案,但是當它們用作處理大量問題Intent時必定出現問題。市場上有一間零售集團,希望Pepper解答一千多條產品問題,結果準確度非常低,很快機械人便已經鋪塵不再使用。一方面原因可能是員工欠缺訓練人工智能的經驗,另外因為NLP engine並非自家開發,即使訓練後作答問題的準確度低亦不明所以,難以解決問題,久而久之便不再使用。


Another concern is which ‘body’ should we apply A.I. onto. Asiabots once cooperated with a university in 2018 to apply NLP onto Humanoid, a robot similar to Sophia of Hanson Robotics. At that time, the robot is not stable enough. Comparing with some business-oriented robots, like Sanbot or Pepper are a lot better, at least in terms of price :). These kinds of robots usually do not have their NLP Engine, but Google or other company’s solutions. However, when it is going to handle questions with various intent, it might not be that smooth. There was one retail group, set up a robot, hoping it could answer thousands of questions. It turned out to be useless with low accuracy. It might be caused because the staff lack experience, or because of the NLP engine is not developed by themselves.



加上自家 NLP Engine 後Pepper 變得更聰明 | Pepper becomes smarter after integration of self-developed NLP Engine




 

Comments


Explore how Asiabots's A.I. solution can evolve your customer communication
bottom of page