Watson Discovery is a powerful, cloud insight engine that can be used to create unique experiences and interactions that have significant business impact. In this post, we’ll highlight one of the key features of Discovery, Passage Retrieval, and look at three example use cases for how it can be used in combination with other Watson capabilities to make unique user experiences.
Use Case 1 : Customer support bot
The first use case we’ll look at is a conversational support bot. In this use case, customers interact with a chat bot to help solve problems they may be having with a product or service. For many organizations, this use case can help reduce costs by deflecting support from live agents, and also create better experiences overall for customers by addressing their problems more quickly. In this example, the Watson Conversation Service is used to manage the interaction with the customer, and Watson Discovery using Passage Retrieval is used as the knowledge base to provide results. This kind of conversational interface allows the application to both engage and inform the customer.
Watson Discovery helps make this chat bot unique by leveraging a broad knowledge base from which to provide a response rather than relying solely on fixed, scripted responses. This ability to pull from content and data is important to creating a bot experience that truly addresses customer needs. Most traditional search solutions return a list of documents, which can be difficult for a user to interact with in a bot experience. With Passage Retrieval, Watson Discovery looks at results and applies intelligent scoring techniques at the passage level. This means that smaller sections of a large document can be found to be relevant and shown to an end user. In the case of a conversational bot, returning a list of these smaller passages allows a user to read the response quickly and stay engaged with the bot.
Let’s take a look at how this might work in an example use case:
This example is a virtual vehicle assistant. It helps customers answer questions about their vehicle, and present an overall better experience for vehicle service and maintenance. Watson Discovery has been loaded with documents from a complete car manual. Each section of the manual, usually about one or two pages in length, has been added as a separate document to Discovery. In a real use case, this data could be supplemented with additional sources of information like previously resolved customer questions or data from public forums.
When the customer enters a question, Conversation analyzes the input for intent and passes it to Discovery. Discovery then returns the results. Here, Discovery is returning passage results. Candidate passages have been selected from relevant documents, scored, re-ranked, and returned to the Conversation service to display along with the title of the document from which they appear. The passages returned by Discovery provide a quick way for the customer to find the answer they need without having to search through full documents.
What makes Watson unique, is the ability to extend this use case even further. Instead of just returning the results of Discovery and ending the interaction, Watson Conversation can allow the user to take action based on these results. This can drive valuable business activities. In this case, Conversation prompts the user for feedback on whether the passage results were relevant to the customer’s original question. If they indicate the result was useful, the passage text is fed back to Conversation to interpret the intent and take action. This Conversation instance has been trained to recognize an intent for passages that might indicate the user needs to make a service appointment, and when it’s recognized, Conversation drives a dialog flow to set one up. This creates a more natural and proactive experience for end customers that can drive real business goals.
Use Case 2: Agent Assist
The next use case is also related to customer support, but instead of an interactive chat bot, this example is a customer support agent assist solution. In this example, there is no real-time customer communication, instead a support agent is researching the problems a customer has submitted, and attempting to find a resolution. A variation of this use case could involve monitoring agent interactions in real time using a service like Watson Speech to Text, and proactively surfacing information to the agent to help them respond.
Resolutions can often be technical in nature and require specific knowledge and context to find a solution. Watson Discovery with Passage Retrieval can be used effectively in this situation to help surface relevant results and provide supporting information for agents to form into a response.
Here is a simple example of how this might work:
In this example, to mimic customer issues, we’re using a Stack Exchange data set for mobile phones where customers discuss issues and possible solutions. This data set is similar to searching past resolved issues to find appropriate resolution techniques.
This basic example resembles a more traditional search application, but highlights several unique Discovery capabilities.
- First the support agent enters the customer’s issue as it appears in natural language from the customer’s original complaint.
- Then Discovery returns a set of passages that may contain useful information. The passages allow an agent to quickly evaluate each result.
- To actually formulate a response the agent may need more context and information. The API response for Passage retrieval provides all the information needed to show the passage in the context of the full document including the document from which the passage was found and the offsets for where the passage appeared. Using this information the example application allows the passage to be expanded.
- When the agent expands a result, the passage is highlighted in context of the full document text. The context around a passage contains useful information for an agent, like a set of steps a customer could try. Combining passage and document results allows the agent can be both more productive and thorough in responding to customer issues.
This use case also has an extension that leverages the ability of Discovery to not only find passages and documents relevant to a customer’s natural language query, but also provide enrichments that demonstrate understanding of the content itself, and aid the agent in quickly finding the best response.
In this example, Concepts are being automatically extracted from each document in Discovery. These concepts are based on the Watson Natural Language Understanding service, and present the agent with a quick way to see the topics being addressed in the documents. Concepts don’t need to appear in the content itself, they are abstractions based on Watson’s understanding of the language. An agent can use the list to quickly identify the results that contain relevant concepts like “Uploading and Downloading”, without needing to read or even skim through the full document. This again helps businesses reduce support costs, and find better resolutions for their customers.
Use Case 3: Research Assistant
Finally we’ll look at a similar use case, but more focused towards a specialized knowledge worker. In this example, the user could be someone in research and development, looking for new opportunities to apply AI to complex problems. This example could be extended based on the data set to many applications like financial research, legal research, market research, and more.
In this use case, the user is more knowledgeable about the content, and therefore is less looking for a direct answer to a question or resolution to an issue, but instead looking to discover new insights that can spark more detailed research.
This example application again starts with a user entering a natural language query, but here the query is more open ended rather than looking for a particular answer.
Discovery returns a set of relevant passages and documents again using Passage Retrieval, and then using the enrichment capabilities of Discovery also returns a list of Entities that were identified from the documents. The passage and full document text provide the relevant information for the user to investigate her topic, while the entities provide a way to dig deeper.
In this example, clicking on an entity performs a Discovery Knowledge Graph query. The Discovery Knowledge Graph is built from entities and the relations between them that have been understood from the documents. The Knowledge graph API allows querying these connections to uncover new entities and relationships (You can learn more about the Discovery Knowledge Graph Beta here).
For the researcher looking at the results, the second document contains interesting information, and clicking on one of the entities allows her to start digging deeper.
Selecting the entity runs a Knowledge Graph query to identify related entities, and then those entities are used as the basis for a new passage retrieval query. The related entities to “Genetic algorithms” that were found through the Knowledge Graph are displayed along with the passage results corresponding to a new query using those terms. This new set of results reveal similar approaches to “Genetic algorithms” that are all part of a group of techniques.
Without prior knowledge, a user would never know to search for those terms, and be limited to only the information they already know. But with Discovery’s ability to understand content and language and build connections, the user can truly discover new information and new insights that would previously be hidden.
Note: The Discovery Knowledge Graph capability is currently in Beta. See release notes for more information.
Watson Discovery provides a unique way to find answers and insights quickly, helping improve the efficiency of many business tasks across customer support, research, and other areas. But Discovery is much more than a content retrieval service. By combining capabilities together within Discovery, and with other Watson services like Conversation, you can build experiences that drive not just increased efficiency but also increased value and opportunity across your business.