In the Heart of Silicon Valley
Reprinted from Bill’s LUI News, April 2018 issue
Just the answer, please!
Bill Meisel, President, TMA Associates
Natural Language Processing (NLP) technology generally focuses on interpreting the “intent” (basic information or action being requested) of text or speech and any “entities” (further information) required to address that intent. The next step is, of course, creating an action that satisfies that request. That may be a “simple matter of software,” e.g., using an Application Programming Interface (API) that turns up the temperature on your thermostat or places a call to a contact in your smartphone. In general, the action step is considered outside of the NLP realm.
However, if the request is for information, attacking it by general programming is not a general or efficient approach. The user wants an answer, and defining what is an answer to a natural-language inquiry isn’t a simple matter of software.
When there is a screen available, one approach to providing an answer is to display information that can contain an answer or at least a guide to sources that may answer the question. The default of many digital assistants such as Siri is to conduct a web search and display a list of web sites when a specific answer is outside the capabilities of the assistant. The processing of the request could be nothing other than simply delivering the text of the inquiry to the search engine and let it do any NLP intrinsic to the search process.
Search engines can in some cases provide a single view that may contain the answer in formats other than a list of web sites. This is an intermediate approach that is likely to be increasingly common.
In fact, Google recently announced a new format of its “featured snippets” that will try to answer multiple different interpretations of a vague search query. Featured snippets are the boxed results that Google puts at the top of the page, based on an algorithmic determination of the best answer to a query. The question, “Is coffee bad for you?,” for example, produced the following at the top of search results:
Google’s new “multi-faceted snippets” will provide several actionable answers to a broad query. A search for “garden needs full sun?” will lead to two different answers, one for “What garden plants need full sun?” and another for “What counts as full sun?”
When a person asks the Google Home smart speaker a question, the Assistant pulls an answer from featured snippets or its Knowledge Graph (a database of facts that Google has built from sources like Wikipedia). For example, searching for “Amazon” produces the following Knowledge Graph:
With a voice-only interface such as Google Home, a visual response isn’t an option. Thus, being able to extract an answer from sources such as a web site, snippet, or knowledge graph is important. The alternative is “I’m sorry, I can’t help you with that.” Even with a visual interface, a specific guess as to the best answer, with other options displayed, helps maintain a conversational mode.
Dialog—a conversation to clarify the question in order to come up with a single answer—is in fact a subset of natural language technology that can contribute to delivering a specific answer. Maintaining a dialog is often not an intrinsic capability of most tools that help build digital assistants today, but expect progress in this area.
Technologies such as machine learning can help map answers to natural-language inquiries as a significant amount of data on what people ask and what answer satisfies them becomes increasingly available. Unsupervised learning can “cluster” inquiries that are essentially asking the same question, and the answer or answers that seem to satisfy requests within a cluster can possibly be used to label a cluster with the best answer. Or perhaps similar clusters should be defined by similar answers, making the answers part of the clustering process.
In terms of supplying reliable news answers, Google has also launched the Google News Initiative. The Initiative is aimed at distinguishing reliable sources from unreliable sources.
One aspect of language technologies seems to be that the better they get, the more demands we make of them. “Just say or type what you want to know” may evolve toward “Just answer my question!”