Location:Home > News
Let an AI tell you, who is the conversation terminator? These are the 5 "high risk" factors
Edit:shenmin Date:2018-04-08 15:28

On social media, when there is a disagreement between parties or individuals, blacklisting has been the solution. Because of one time, a senior student at Zhejiang University of Computer Science and Technology Jiao Yun Hao, inadvertently caused a classmate of 16 years and a friend to put him on the black list. This just goes to show that because of one action, even a friendship of many years can be fleeting. The result of this incident has lead him to think up two problems: what is the “conversation killer” as well as what is the fuel that allowed this to continue. The engineering student’s solution to this problem is A.I.


Jian Yun Hao developed a software called ConverNet. It then analyzed a conversation which led to its own judgment: Conversation will not continue. The A.I. even went through a big data driven deep learning method to tell mankind that all types of language is at risk to the “conversation killer.”


The related paper Find the Conversation Killers: A Predictive Study of Thread-ending Posts, has been taken on by a high-level international conference in the internet field—The International Conference of World Wide Web, or WWW Conference for short. On April 22, Jian Yun Hao will go to France to share his research results with the academic community in Lyon, France.


100,000 Studied Dialogues


The software ConverNet is a deep learning neural network system that is capable of “self-studying” dialogues after dialogues, meaningful insight into the details of human conversation, and enhancing its own intellectual prowess. However, a question remains, and that is where should ConverNet study?


Under the guidance of the University of Michigan’s School of Computer Science Professor Mei Qiao Zhu, Jiao Yun Hao found two public massive databases for ConverNet to use: Reddit-Threads and Cornell University’s Movie Dialogs.


Two people chatting and talking and then nothing to say, is a type of problem. In group chats, when an individual sends a message and is responded with cold emptiness, this is also a type of problem. Jiao Yun Hao says, the A.I. is able to analyze both of these scenarios. Reddit-Threads is a media platform that is similar to China’s “Zhi Hu”; it is a place where many people discuss an assortment of topics, critic each other’s perspective, and share their own unique take on the topic. The concept of Reddit-Threads is similar in fashion to how group chats on social media operate, while Movie Dialogs provide a magnitude of movie dialogues which is similar to one on one messaging (also known as private/direct messaging).


Altogether, ConverNet self-learned a total of 83,097 discussions from Reddit-Threads, 100,000 dialogues from Movie Dialogs; involving 29797 individual sentences from Reddit-Threads and 107354 from Movie Dialogs. Jiao Yun Hao says, after deep method learning, ConverNet is able to identity unique elements in a couple “conversation killers.” It is readily able to identify red flags in new random set of dialogues and predict whether a conversation will be “killed” or not.


ConverNet began as a novice that was unable to understand the human language; however, after self-studying in massive data bases, its predicting accuracy is able to achieve approximately 80% and its indicators were able to surpass traditional methods. Jiao Yu Hao reveals the key steps of its operation:


First Step, when it faces information coding, it forms a computer that is sufficient in understanding the natural language. Second step, it uses the LSTM neural network unit which then digests the information extracted from the dialogues. Third step, it then outputs the probability of a conversation’s ending.


5 characteristics that leads to the end of Friendship


Computers use their own methods to predict the red flags that can end human dialogue, which also allows it to further refine its accuracy. This prompted Jiao Yun Hao to think further: how does one converse so that it will not become a conversation killer. In order to answer this, we must go to the backside of the A.I. and understand what it has learned from its self-study, or whether it can guide us to longer conversations and extend them. Jiao Yun Hao says, that through a series of methods such as deep learning and machine learning, and re-visiting the massive database, he discovered that there are 5 unique characteristics that leads to the end of conversations.


1. Caring. When words such as Mr., Mrs., talked, heard, seen, care, etc. are used, conversations are extended. This is because these words possess a positive connotation that reflects the desire to respect, and partake their own joy. When words such as YOU, THE, Christ, but, ass, etc. appears, it will increase the possibility of ending the conversation.


2. Responses. In one on one chat, response will be a lot more; thus, it is more likely for a conversation to be killed. However, in group chats, more responses can represent a more energetic conversation, which leads to prolonging too.

3. Long Short. When conversing one on one with another, sentences will be longer, conversations can be longer. The reason for this is that conversations held are more in-depth and possess more substance compared to group chats. Short responses are more readily welcomed, and if someone answers back with a long response, will lead to a silence.


4. Time. If you initiate a conversation during the late night when people are resting, it is easily to stir up disgust from the recipient, which ultimately can lead to conversations ending.


5. Mood. When talking in groups, negative emotions may allow new rounds of resonance and discussion to arise. In single conversations, the opposite will happen and conversations will end.


Consulting the Future of Conversation


Now I know why I am a conversation killer. Jiao Yun Hao’s dad is one of the beneficiaries of this study. On WeChat, his messages are all big, which compared to pile of short sentences is incompatible. This characteristic of the conversation killer is what the A.I. comprehended. Jiao Hao Yun says that after his dad read his research, his fashion of writing long messages shortened a lot. This changed his opinion and idea of writing sentences, and the chances to create silence in a chat is reduced.


Jiao Yun Hao’s undergraduate design instructor from the College of Computer Science is Professor Wu Fei. Jiao Yun Hao’s research concerns the phenomenon that has never been noticed before, and prioritize the use of A.I. to answer these problems. We hope that someone in the future can carry on and improve our proposed method. It is also possible that this system will learn more from massive data systems, thus become even more intelligent.


It can be expected that one day in the future, we will have an A.I. assist us in conversations. When we are about to make a mistake, it will immediately tell the individual to stop, so it will ensure the dialogues to come will be longer and longer.


If such an A.I. was able to arrive earlier, maybe the incident of Jiao Yun Hao being blacklisted by his classmate and friend of 16 years would not have happened. This occurred when Jiao Yun Hao and his classmate were playing the game King’s Glory, and he said: “With your skill level and you were someone else, I would never play with him again”.  Even though they added each other back as friend, the damage was already done, and they never played King’s Glory together again.