Home > Article > Technology peripherals > Nine Tips to Improve Your Conversational System Performance
Conversational AI is seen as a superior solution for providing users with the information they need, when they need it, and in a way that best suits their wishes. However, even so, conversational AI systems still have some common flaws.
The following are some tips that I have compiled over the years to help you improve the performance of conversational systems:
Conversational artificial intelligence The goal of the system is to obtain the information needed to complete the task, which information can come from context or user questions.
However, every question asked by the system may cause the conversation to fail. Therefore, our goal is to minimize the number of questions asked, or to make the questions simpler.
Take car insurance as an example. The user wants to check the status of the claim just submitted. We simply ask users to provide their claim's 11-digit ID with no complicated questions or explanations. Such a design is more in line with user needs and reduces the risk of conversation failure.
In a technology-driven approach, the first consideration is usually the required "claim details" API". The developer might even mention that an API already exists that can leverage the policy ID to get the details. This is excellent! We only need the user to provide 1 policy ID to obtain the required information. This design is diametrically opposed to the current technological approach.
While this is relatively easy for developers, it can pose certain challenges for users. Users must enter the exact 11-digit ID. In a speech context, even if the user knows the 11 digits, capturing them correctly can be challenging.
To make it easier for users to enter claim IDs, we can reduce the number of digits required, such as only requiring the last four digits number, not all 11 digits. Although the full claim ID is still required on the backend, this design makes it significantly easier for users to ask questions. There is less chance of making mistakes when entering or speaking small numbers.
There are many ways to “Ask Simpler Questions”:
However, there are limitations to asking simpler questions because you are still asking. For example, in the case of a car claim, we expect callers to be able to obtain detailed information about the claim, but this process can pose challenges for users. So, can we simplify things further?
Let’s think in a different way and shift the topic from a technical perspective to a user perspective. We should put the user first and understand why they are calling about a car claim. Maybe it's because they've just been in a car accident and are tracking a claim they've filed.
Although users may have multiple claims, they may call to inquire about information about their most recent claim. In such cases, we need to make reasonable assumptions and it is necessary to confirm these assumptions in order to better meet the needs of users.
The above figure shows a method that can reduce the burden on users. Now, instead of actively providing claim information, users can receive information provided by the system. This improvement not only reduces the user's interaction burden, but also meets the needs of back-end data retrieval. It is more convenient for users to confirm information than to actively provide information. In addition, this approach also makes the conversational artificial intelligence system look more intelligent and improves users' trust in the system.
However, this mode requires additional contextual information to ensure that the information is not leaked to malicious third parties. For example, in the example of claims search, the system facilitates this process by identifying the caller (perhaps tying the incoming phone number to the claims policy).
While it’s not always possible to skip questions and just make reasonable assumptions, this is one of the powerful techniques for building more efficient conversational AI systems. Next, let’s explore ways to improve the questions you need to ask.
In the previous section, we discussed how to improve the performance of artificial intelligence systems by reducing problems. However, sometimes, deleting a question is not an option and the question still needs to be asked. In this case, a powerful approach is to provide clear choices, i.e. asking the user to choose from an explicit list. Let's imagine that we ask the user to choose between apples and oranges. While this sounds relatively simple, how do we ensure that the user successfully makes a choice?
Implementing multiple choice questions may be more challenging than you think. Let’s examine some potential pitfalls together.
We need to beware of “yes/no” confusion. It is not always clear to users how to answer optional questions. For example, when you ask a user "Do you want an apple or an orange?" they might answer "Yes!" In this case, we need to reframe the question to make it more clear to the user:
Rewording of these questions may appear to be technically the same, But the effect in an actual conversation can be very different, so choosing conversation copy that is appropriate for the conversational medium is key to ensuring optimal performance of your conversational AI system.
Some users are reluctant to enter or voice answer questions. In a chat scenario, clicking a button may be more convenient than typing an answer. In a voice scenario, users usually have the ability to press buttons on the keyboard—for example, pressing "1" means selecting the first option, pressing "2" means selecting the second option, and so on. When evaluating user responses to the question "Please choose an apple or an orange," we can accept "apple" or "1" as equivalent answers. This way the operation can be completed more smoothly.
Buttons may not feel "high-tech," but if used correctly, they can make users more efficient.
We should have a certain degree of flexibility for small mistakes made by users during the selection process. When users are asked to make a choice, both the user and the artificial intelligence system are likely to make misoperations. For example, in a chat interface, a user might spell "apple" instead of "appel." At this point, the AI should not simply respond with “I don’t understand,” but should reasonably assume that the user intends to choose Apple, thereby providing the correct response. This fault tolerance improves the user-friendliness.
In a speech application, if your options contain domain-specific terms, you may need to train a custom model to ensure that these selection options are accurately recognized. Now that you know how your users are likely to react, whatever choices you offer, you have to make sure your voice service can accurately transcribe them. For this fruit selection example, be sure to include some entries for "apples" and "oranges" in the training of your custom model to ensure accuracy.
Even if you have trained a custom model, you may still encounter transcription errors. Your voice service may sometimes incorrectly transcribe "apple" as "appeal" or "hackle" etc. In this case, you may consider adding these potential mistranscription scenarios to your Watson Assistant implementation to further improve recognition accuracy.
Selectivity problems are common in conversational artificial intelligence systems. By applying these tips, you can ensure your selective questions are more efficient.
In the previous content, we have discussed how precise word choice can improve the performance of artificial intelligence, because more accurate dialogue text can produce better users response. More granular conversational text not only improves the user experience, but also helps increase user satisfaction and speed task completion. Together, this series of effects promotes a smoother user experience.
In a conversation, every word will increase the burden on the user. In a text chat, they need to read; in a voice conversation, they need to listen (and wait) for the words. Therefore, it is crucial to review conversations to exclude unnecessary text. Here are some examples:
Conversational AI systems use text-to-speech engines to converse with users, and these engines rely on punctuation clues to determine where in a sentence Add accents and pauses.
I remember a clever T-shirt that read “Let’s Eat, Grandma,” with a bold comma that inspired the slogan: “Punctuation Saves Lives.” In reality, punctuation marks may not save the "life" of your AI system, but they do make your conversational text sound smoother.
When checking your dialogue for punctuation, make sure:
Additionally, you may even consider using automation features to create audio files for all conversations in your virtual assistant.
This article explores a variety of technical methods to improve conversational artificial intelligence. Consider adopting these techniques when developing new conversational AI systems or improving existing ones. Whether it is a new system or an existing one, continuous improvements and iterations have a positive impact on the performance of AI systems. These practices have been proven to significantly improve the performance of conversational AI systems.
The above is the detailed content of Nine Tips to Improve Your Conversational System Performance. For more information, please follow other related articles on the PHP Chinese website!