Realtime database data structure modeling

The common recommendation is to store the data that you app needs to display. So if you want to display top sentences for today, for this week, and for this month, that means storing precisely those aggregates: the top sentences by day, week, and month.

A simple model for storing these is to keep your current, but then for each aggregation level, and each interval:

stats
   --topPhrases
     --keyA
        --phrase: "who are you"
        --count: 2
     --key
        --phrase: "hello"
        --count: 1
   --topPhrases_byDay
     --20190607
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
     --20190607
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
   --topPhrases_byWeek
     --201922
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
     --201923
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
   --topPhrases_byMonth
     --201905
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
     --201906
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1

Alternatively, store all aggregations as a single list, and use prefixes to indicate their aggregation level (and the format of the rest of the key):

stats
   --topPhrases
     --keyA
        --phrase: "who are you"
        --count: 2
     --key
        --phrase: "hello"
        --count: 1
     day_20190607
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
     day_20190608
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
     week_201922
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
     week_201923
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
     month_201905
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1
     month_201906
        --keyA
           --phrase: "who are you"
           --count: 2
        --key
           --phrase: "hello"
           --count: 1

You're definitely duplicating a lot of data here, but the advantage of these models is that displaying the stats to a user is now trivial. That's a common trade-off with NoSQL databases, writing of data is made more complex, and more (duplicate) data is stored, but it makes reading the data trivial, and thus very scalable.