Preparing data for natural language interaction in Ask Data

Data sources with row-level permissions have no minimum performance requirements and can be used in Ask Data. However, Ask Data does not profile or index such data sources or store metadata for the fields in the semantic model. With the absence of metadata, Ask Data cannot provide filter defaults (as shown in Figure 2), recognise currency concepts such as “cheap” or “expensive”, or show profile data in tooltips in the data pane.

If you already publish certified data sources for your organisation, teams can use these sources in Ask Data. But you may want to open up additional sources to your end users or reshape existing sources for easier analysis in Ask Data. To get the most out of the Ask Data feature, curate your data with the following factors in mind.

Curate your data with the end user in mind

Ask Data users are far more successful at answering questions when the data has been carefully curated by a data steward or analyst who both understands how to prepare data for analysis and can anticipate the kinds of questions users will ask of the data using natural language.

Read more about best practices for curating a published data source.

When curating data for Ask Data, start by simplifying your published data sources as much as possible. That means keeping a minimal set of fields that users will query in their interaction with Ask Data and removing (or hiding) unnecessary fields from the data source. Ask Data supports data sources with up to 1,000 fields, but the less ambiguity, the better. This helps with the overall system performance for faster initialisation time and parsing of the natural language utterances. If the data source is slow, use a data extract with datasource filters (if needed) to improve performance.

Consider the following factors when curating data sources for querying in Ask Data:

Prepare your data. Try to anticipate the types of questions users will want to answer with the data source. Data shaping, join operations and related data prep functions may be needed to get the data into a suitable shape for answering these anticipated questions.

Set up appropriate field defaults. Give each field the correct data type (e.g. string, number, geographic, date, datetime, boolean) and data field role (e.g. discrete vs continuous, measure vs dimension). For each measure, assign default aggregation functions. For example, SUM may be an appropriate default for “Sales”, but AVERAGE might be a better default for “Test Score”.

Set up percent and currency number formats. To support common, colloquial concepts that people tend to ask of their data, Ask Data includes concepts such as “low”, “high”, “lowest”, “highest” and currency-based concepts such as “cheap” and “expensive”, along with their corresponding synonyms. To facilitate utterances like “show me the cheapest wineries in France”, set up measures in the data source with the appropriate currency format (as shown in Figure 3 and 4).

Set up logical hierarchies. This allows users to drill up and down within visualisations produced using Ask Data. This applies to geographic dimensions (e.g. City, State, Country), dates and times (e.g. year, quarter, month), and functionally dependent dimensions (e.g. Category and Sub-category).

Create meaningful binned fields (with appropriate bin sizes) for quantitative variables in the following scenarios:

Viewing a binned version of a field that is not a measure in the data source.
For example, “Age” is a numerical dimension that cannot be represented as a histogram in Tableau, and therefore Ask Data. However, creating a binned field for Age in the data source allows a user to ask a question using its binned version (as shown in Figure 5).

Eliciting histogram responses in Ask Data with custom bin settings.
Creating binned fields of measures with custom bin sizes allows for more control over how these fields are displayed in Ask Data. In the example below (Figure 7), the user can type “Fare as a histogram” and Ask Data will use the custom bin settings from the binned field “Fare (bin)” to generate a histogram.

Give fields unique and meaningful names

To make your data source more understandable by end users – and increase the chances that people get the answers they need from Ask Data – you should audit the field names in your data source.

Take the following steps for an ideal user experience:

Create meaningful aliases for field values. In Tableau Desktop, you can create user-friendly field names with aliases (e.g. “CustID” is “Customer ID”). This is a standard recommendation for data curation. For Ask Data, you can take this a step further by adding synonyms. For example, people may also use “Customer Number” to refer to “Customer ID” internally. In these circumstances, you could add these additional synonyms to Ask Data to support users’ questions.

Differentiate your attributes. Assigning unique names to attributes in your data source can provide a better user experience in Ask Data. If there is ambiguity in the expression, Ask Data finds strings that approximately match a pattern in your data up to one character difference, also known as fuzzy string matching. Ask Data then shows these multiple matches as options. In the example below (Figure 8), multiple attributes containing the word “Sales” exist in the data source. However, just typing “sales” only matches three attributes as the input word “sales” and the attribute “Sales Foo” differ by more than one character.

It is also recommended to name attributes that are semantically meaningful to the domain of the data source so that the questions posed in Ask Data are intuitive to the analytical task. For example, “Number of Records” is renamed as “Number of Earthquakes”, where each record is an earthquake in the data source (Figure 9).

Check field names. Ask Data filters values by the fields in the data source. Avoid naming fields as values to prevent Ask Data from misinterpreting data fields as values (e.g. numbers, dates or boolean values such as “true” or “false”). Also, for performance reasons, Ask Data doesn’t index fields that overlap with the supported analytical expressions. For example, avoid using field names like “Average”, “Sales in 2015” or “Most Products Sold”.

Geocode geographic fields. For fields with geographic roles, be sure that the data values are successfully geocoded. Fields whose values are successfully geocoded are recognised as geographic attributes in Ask Data, indicated by an icon in the data pane (Figure 10). Such attributes support utterances like “where are the highest fire fatalities?” with the system recognising the token “where” to indicate a map and inferring a valid geographic attribute like “County”.

Add relevant calculated fields

Ask Data cannot create on-the-fly calculations, so you need to add anticipated calculations to the data source in advance. For example, creating a calculated field called “Total Compensation” that adds the minimum base salary from “Base (Variable)” with the total commission from “Commission (Variable)” allows users to type questions like “what is the total compensation for each sales person?” (Figure 11 and 12).

Enrich data with user synonyms

You can define synonyms for fields within a data source through the Ask Data fields pane. For example, let’s assume you have a data source on car purchases and one of the fields is “New Vehicle Model”. Adding “vehicle purchased” and “car” as synonyms for the “New Vehicle Model” field can support natural language utterances such as “vehicles purchased by city” (Figure 13).