AI INTERVIEW
This project is an artificial intelligence interview APP designed for back-end development engineers during my work at Ping An Technology.
Role： UX Designer
Our Team： XDC
Partner：HR, face recognition team, voiceprint recognition team, intelligent voice team, data analysis team
Duration: 20 Days
Tools: Axure; Sketch; Ilustrator; AE
BACKGROUND
In HR work, the entire recruitment process include resume screening, evaluation, initial interview, second interview, confirmed salary and entry.
The development of AI interviews can simulate real interviews, and generate interview reports through the analysis of the underlying model to accurately screen candidates and save human resources.
At present, it is only designed for back-end development engineer. In the future, it will be expanded to multiple positions and used for campus recruitment.
TARGET
In UX, we should find suitable interaction for online interview. Ensure interview process and voice interaction are fluent.
Design suitable appearance, voice, and animation for Little I;
Work with other technical teams to evaluate technical capabilities to ensure our design can work;
Our business partner want a sense of science and technology producrt. It must be smart and intelligence.
Construct background evaluation module page (this article doesn't introduce);
CHALLENGE
Our business partner doesn't have a certain idea of the interaction form, and the designer needs to explore and guide them;
It's my first time to participate in VUI design. There is no previous experience. And I need to learn a lot of basic knowledge;
We must make a reasonable design in a short time.
COMPETITIVE ANALYSIS
Other designers and I work together and find competitive products. But it's difficult to find a AI interview product in market. Finally we find three types of products.
SAME TARGET USER
The interview bar and AI all-round interviewer are the only two intelligent interview applications that can be experienced at present. The overall experience is that the two are still satisfactory interview question-making products, which are difficult to give people a smart feeling.
CHAT UI ASSISTANT
People can use many interaction such as voice, buttons, and input box. It supports multiple rounds of dialogue in a single task line. It has the ability to complete the interview process and It seems more intelligence.
VOICE INTERACTIVE ASSISTANT
The form of dialogue is closer to face-to-face communication. It can support multiple rounds of dialogue and a single round of dialogue. It supports a lot of interactive forms such as voice, buttons, inputbox, and templates.
FINAL SELECTION
After evaluating the requirements of the intelligent interview and communicating with business partner, we determined that the interview process uses non-conversational interactions. It focus on the current scenario, without the context. Although the dialogue uses semiduplex, the voice assistant should be able to simulate face-to-face communication. It should let users immerse themselves in the current scene and get a better experience.
AI PERSONA
Role definition
In order to establish more accurate robot characteristics, we made a simple questionnaire for business partners:
Based on the survey results, we customized characters for AI assistants.
Robot Design
After the character definition, according to the business partner's requirement, our team designed two robots. their gender are a male and a female. If the interviewer is a man, it shows the female robot; If the interviewer is a woman, it shows the male robot. We don't change their character, just adding a interesting idea.
WORKFLOW
According to the requirements and the results of the research, we defined the main content into five parts: opening, identity verification, introduction, interview, evaluation and feedback. Then, I draw the workflow.
AI CONVERSATION DESIGN
We don't choose multiple rounds of dialogue during the AI interview because we use the non-conversational interaction. Only the last subject needs to answer users' question. The backend system can recognize the intent and slot and answer users through TTS. Therefore, we focus on the dialogue design to ensure that the dialogue is natural and smooth.
Conversation ability
We encountered several issues when designing conversation.
QUESTION 1: Robot: Talk like a human; Human: Answer like a dog.
We want to use natural language that could simulate human-to-human conversations to design dialogues. However, under technical constraints, people may not be able to answer with natural words.
Our business partner prefers to use voice submission. The user needs to actively make some choices. But the answers is fixed, the words other than the answer can not be identified. It can increase recognition rate of fixed answer, but the experience is not good.
After the improvement, the user answered Submit and automatically entered the next question. It felt better than before, but it wasn't perfect. In a normal dialogue, People usually say "I'm done, that's all for my answer". Then, the robot use fuzzy recognition to submit. Of course, we should consider development costs, time, and accuracy. Finally, in version 1.0, we temporarily give up the fuzzy recognition.
Question 2：The system is too talkative
In our design, we hope the voice simple and clear, let the users dominate the rhythm of dialogues, and let the user feel control over the entire interview process.
In Plan A, when there is no voice input, the robot talk most of time. It seems to lead user to the next question, but actually it interrupts users' thinking and disturbs the user.
After modification, We can see Plan B. The robot only makes a small tip at the key points, and they submit or not is determined by the user.
We want to return the initiative to users and let users feel a sense of control over the entire process.
Conversation style
Define the style of dialogue and uniform the whole style.
Our business partner gave us some dialogue to reference, but the style was not uniform. The dialogues both have serious dialogue and lively dialogue. Our responsibility is defining the style of dialogue according to Little I's character and write dialogue.
Interactive Form
We need to consider the scenario when we choice the interactive form. We need to know which scenarios are suitable for voice, and which scenarios are not suitable.
For examplle, some content is difficult to input such as fomulas, is more suitable for screen input. Some complex information which has too many words, is not suitable for voice input.
We have a subject that the robot need to answer user's question. It usually need to show a large piece of information to user. User use their voice to ask robot a question, and the robot answers user using both text and voice.
Privacy and ease of use
There is a problem in voice interaction: everything will be heard by others. In our design, we are not allowed the robot read private information. We choose other way to output information such as text.
In usability, a noisy environment and a poor network will seriously affect automatic speech recognition. So we design effective alert such as toast.
Dialogue test
At the end of dialogue design, we make a reading test and a real user test.
Through testing, we found some problems:
Some dialogues that are much different from the real conversation can be found in the reading test;
The system kept asking users when to submit, which would also interrupt users;
Through real user testing, we uniform the style of dialogue for Little i: keep the brief and professional in asking questions. And add some lively words on screen tips instead of reading them.
WIREFRAME
After determining all the details of the dialogues, I draw the wireframe, and write notes in some key point for developers.
UI DESIGN
USABILITY TEST
Before the product launch, I invited 5 development engineers in the company to do usability test. Users are asked to simulate real interviews as much as possible. They use AI interviews to complete the entire interview process. And I listen and record their suggestions.
Usability test helps us to discover some development problems in a time. We report problems to developers and they can fix it in time; it also helps us find many imperfect details in our design.
Test summary
Don't assume user behaviors too much: We give Little i a hint function and some animations. We hope to give users effective prompts and increase interest. But during the test, we found that the user was nervous at this time. They focused on the interview, and rarely noticed the prompts and animations;
We have made a animation design of submit button that can fluctuate follow the voice. We think is cool, actually, the fluctuation rate is slow and not obvious enough. Most of users still don't feel it;
Toast prompts are too strong and sometimes obstruct the view;
Some users suggest us that we can design a bigger windows that they can see themselves clearly. We will consider this suggestion in the next version.
SUMMARIZE
Quick learning and collaboration: Nobody had experience in VUI design before. Everyone collected a lot of information in a short time and learned from each other. We make brainstorming, defined styles, design dialogues. We use the design power to reflect and promote the development the project.
Cross-team collaboration: More and more teams involved in this project. We can communicated with many teams and got more knowledges such as face recognition technology, live detection technology, back-end corpora. However, there are some regrets. Because we have little communication with the TTS team, although we have provided the character of Little i, the Little i's voice sounds like older than we hope. Especially the male Little i, the voice age is a lot older. Cross-team collaboration brings us more knowledge exchange outside the field which benefits us a lot.
Continue learning: The end of the project is not the end of learning. VUI is a start, I will continue to learn and improve my knowledge tree.