Shulou(Shulou.com)06/02 Report--

With Siri as a precedent, personification has become a necessary capability of natural language interaction tools. Whether it is the AI voice assistant serving individual users, the intelligent customer service provided by enterprises, and even all kinds of household appliances with voice functions, they all have to do IP and artificial devices, which is almost perfect.

Most of the time, we think that the personification of natural language interaction tools can reduce the "horror valley effect" of users and make users prefer to communicate with them. But the latest findings suggest that this may not be the case.

Become the thousand tricks of human beings.

First of all, let's take a look at the anthropomorphic "thousand tricks" of natural language interaction tools.

The first step is to give yourself a harmless name.

We often say that when you pick up a small animal and give it a name, nine times out of ten it will become your pet. The same is true of AI, when a natural language interaction tool has a name, it is basically destined to go further and further along the road to refinement. The name of natural language interaction tools is usually "small", which is both weak and harmless and politically correct regardless of gender.

The second step is to use speech generation technology to imitate human tone.

Once you have a name, you can no longer use cold electronic sounds, and even the pattern of human recording + rule matching, which used to be useful in voice generation technology, is a little rigid. At this time, there is neural network voice generation represented by Google WaveNet. By grasping a variety of features of human speech, and comprehensively considering the parameters such as semantics, part of speech, grammar, including context, and so on, Google Assistant finally produces a human-like speech pause and thoughtful tone.

The third step is to make the dialogue more humane.

In the process of natural language interaction, speech generation needs to be based on text content. To meet the anthropomorphism of "speech tone", it is also necessary to make "speech content" more humanized. At this time, the maturity of technologies such as semantic understanding, multi-round dialogue, natural language generation and so on becomes very important. For example, the full-duplex natural language interaction applied by Microsoft on Microsoft Xiaoice can realize "listening while thinking" and "rhythm control"-- understanding the user's intention through the whole conversation process and reducing the user's waiting time. And can take the initiative to initiate new topics to break the silence and adjust the content and timing of the answer. The content of such a conversation is "displayed" through voice generation technology, which can make people think that they are really talking to human beings.

The last step is to put on "human skin".

In addition to technology, there are also some peripheral patterns to make natural language interaction tools more anthropomorphic. For example, designing a lovely cartoon character for them, adding a few instructions to let them learn some colloquial language of acting cute, adding some details to the interactive interface so that people do not realize that they are talking to the machine, and so on.

With these routines, we can basically create a natural language interaction tool that "turns into a human form".

The more human, the cuter? Expected value Management of Natural language interaction tools

But one question that we have never thought about is, in practical application, is it true that the more anthropomorphic natural language interaction tools are, the better? Recently, the Media effects Research Laboratory of Pennsylvania State University conducted such an experiment.

The researchers told volunteers that they would buy digital cameras on the e-commerce platform and need to talk to online customer service. Behind these customer services are intelligent natural language interaction systems, but researchers distinguish them in terms of humanization and responsiveness. Different groups of volunteers will come into contact with different online customer service systems, some directly tell each other that they are machine customer service during the conversation, some only show the contents of the dialog box, and some "disguise" as human beings through real-life avatars and names.

At the same time, these intelligent customer service with different degrees of personification have different degrees of response. Some can answer users' questions quickly and accurately, while others don't understand what people are saying.

When the experimenter's satisfaction was investigated after the interaction, the results were surprising.

In general logic, we would think that the higher the degree of response of intelligent customer service during interaction, the higher people's satisfaction will naturally be. But the actual situation is that, in the same degree of response, the degree of user satisfaction is related to the degree of humanization of intelligent customer service. For example, for the same interactive content, experimenters who clearly knew that the other person was machine customer service would give a satisfaction rating of 80 points, while those who pretended to be human could only get a satisfaction rating of 60 points. The reason is that when machine customer service shows high human characteristics, users' expectations of them will rise, hoping that they can help themselves solve problems like human beings, and will magnify the sense of disappointment if they don't get the answers they want.

In fact, we have the same feeling when we use natural language interaction. When products such as voice assistants and intelligent customer service can't solve the problem and force us to act cute and tell jokes, our grumpiness index tends to rise in a straight line.

In the final analysis, whether the natural language interaction is humanized or not is a problem of "user expectation management". Sometimes raising users' expectations too much will be self-defeating.

It is easy to be a man, but difficult to make tools

But at present, we can see that an important trend is that the degree of development of human nature and instrumentality of natural language interaction is uneven.

From the point of view of the difficulty of technological development, it is much easier to make natural language interaction tools closer to human beings than to make natural language interaction tools more effective.

Whether it is Google's WaveNet or Microsoft's full-duplex natural language interaction, it is enough to make the pronunciation patterns, conversation rhythm and other details of natural language interaction infinitely close to human beings. In the future, combined with the ability of computer vision and even robot production technology, we can create a human-like interlocutor.

In fact, today, we can see visually humanized "AI speakers" such as AI anchor or Sophia launched by harmony.

However, the interactive problem-solving ability of these natural languages has not been improved. Specifically, there are still some gaps in the understanding of human corpus, especially the relatively unpopular corpora such as small languages, the elderly, children and so on; the understanding of vocabulary in different fields is not comprehensive enough, and AI often enters the knowledge blind area when it comes to some vertical industries.

In this way, the "instrumentality" that helps natural language interaction to catch up with "human nature" may become an industry boom for a long time in the future. For example, the establishment of knowledge graphs in various industrial segments, the accumulation of vocabulary, or the collection of corpus of different dialects and languages of different groups for AI training.

With the continuous catch-up of technology, it is inevitable for people to have higher expectations of natural language interaction tools. In order to avoid the "short board effect", we should perhaps devote more energy to pursuing something other than "human nature".

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.