Generate answers automatically for IBM watsonx Assistant
The NeuralSeek extension enables watsonx Assistant to automatically answer users’ questions using content from your documents
*Watson Assistant has been rebranded to watsonx Assistant. Any reference to it in this article refers to its new name.
When building a virtual assistant, it is usually sensible to build your first version quickly and launch right away. A popular approach to launching quickly is to build an assistant that returns search results from your existing content. Search-only assistants can be useful, but they can’t provide complete, direct answers to questions using natural language.
This article describes a way to use your existing content to actually generate complete answers to questions. With this approach, you can quickly build and launch an assistant that directly answers questions. As your assistant gets more usage, this approach can also help you optimize the assistant to be increasingly effective.
You can build a simple search-only assistant using the IBM Watson Assistant product via the built-in-search integration powered by IBM Watson Discovery or by bringing your own search. However, a solution of that sort can only find answers that are already present in your content and return the results — it can’t use your content to generate answers in natural, conversational language. Furthermore, a search-only assistant can provide multiple search results, but it can’t combine information from one search result with information from another to provide a single coherent answer.
Until now, assistant builders have worked around these limitations by manually authoring responses and programming in applicability conditions. However, when you are just getting started with a new assistant, you can guess what questions your users will ask it, but you can’t know until the assistant is deployed on a live channel.
Spending a lot of time and effort to craft responses before you have any real users can be very wasteful. In this article we present an alternative that lets you launch right away without knowing exactly what your users will ask and still provide direct, conversational answers that respond to the precise question being asked.
NeuralSeek
NeuralSeek by Cerebral Blue is a combined search and natural-language generation system designed to make conversational AI feel more conversational. NeuralSeek takes queries from a Watson Assistant chat and uses them to retrieve content in Watson Discovery. It then employs natural-language generation technology to generate a response based on the retrieved content, the query, and the full context of the conversation. The result? Answers that are both informed by your content and directly respond to the user query.
Here is a video from NeuralSeek’s creators that provides an overview of its capabilities:
Curating responses with NeuralSeek
As discussed in the video, once you start getting queries from users, NeuralSeek will keep track of the queries that it answered and help you curate answers to those queries. It can group similar queries into intents and generate additional sample queries to train the intent recognition within Watson Assistant. See the “Curate” section of the NeuralSeek documentation for more details.
Risks and challenges of generated answers
Whenever you let an automated system generate answers for you, there is a risk that they’ll produce answers that are incomplete or incorrect. Of course, human agents can occasionally give incorrect answers because their knowledge is incomplete. No approach to customer service is going to be successful every single time.
We recommend that you test any virtual agent carefully, but we also recommend that you consider the costs and benefits of an imperfect but often very useful virtual assistant with the costs and benefits of other approaches that are also imperfect.
NeuralSeek uses your content as background for generating answers, which limits the amount of misinformation it produces, but doesn’t eliminate it completely. NeuralSeek also provides a confidence score for each response so you can prevent your assistant from answering or provide a warning when the confidence is low. This can be a very effective tool for reducing the impact of incorrect answers — answers that are not well justified by your content often get very low confidence scores.
NeuralSeek has “filtration” capability that is intended to prevent outputs that are potentially sensitive. It is useful, but not perfect, so there is still some risk of producing sensitive content from the answer generation technology. For more details on the kinds of risks and available mitigation strategies, see the NeuralSeek EULA.
Connect NeuralSeek to Watson Assistant
If you have a NeuralSeek instance, you can get started connecting it to Watson Assistant by creating a new action, selecting “quick start with templates,” and then selecting the “Do more with starter kits” section of the template catalog:
If you click on “NeuralSeek starter kit” in this catalog, you will see additional details about this template entry:
The ‘set up this extension’ link in the detail screen takes you to detailed instructions to help you quickly integrate NeuralSeek with Watson Assistant.
The template and starter kit can help you get a virtual assistant running that directly answers queries from users based on the content you already have and the full context of the conversation with the user. Once you have the assistant running, NeuralSeek can also monitor the queries you get and enable you to curate responses to important ones.
Once you have used the template and starter kit to connect to Watson Assistant, you can try it out in the preview panel. Here is an example of a conversation using NeuralSeek:
Notice how the answers generated by NeuralSeek respond directly to the question being asked. Also notice how NeuralSeek uses the conversation context to respond to the second question: that question does not explicitly mention “SPSS” but NeuralSeek responds with a correct answer about SPSS because it was mentioned in the previous question.
NeuralSeek provides conversational answers to questions directly from content. You don’t need to know what questions will be asked or do any training to set up NeuralSeek, and you don’t need to do any coding. Just point Watson Discovery at your content, point NeuralSeek at Watson Discovery, and connect them all together with the Watson Assistant template and starter kit.
Acknowledgements
Special thanks to Garrett Rowe, who wrote Making Conversational AI feel more… conversational (which explains the architecture behind NeuralSeek). Also thank you to Mary Swift, James Walsh, and Arnesh Batlaw for reviewing this post and providing helpful input.
