Everyone in today's smartphone-saturated world has had their share of interactions with textual "autocomplete." You may have sometimes even wondered if autocomplete is worth the trouble, given the ridiculous completions it sometimes attempts.
But how would you actually make an autocomplete system?
In this project, CompleteMe we'll be exploring this idea by a simple textual autocomplete system. Perhaps in the process we will develop some sympathy for the developers who built the seemingly incompetent systems on our phones...
A common way to solve this problem is using a data structure called a Trie. The name comes from the idea of a Re-trie-val tree, and it's useful for storing and then fetching paths through arbitrary (often textual) data.
A Trie is somewhat similar to the binary trees you may have seen before,
but whereas each node in a binary tree points to up to 2 subtrees,
nodes within our retrieval tries will point to N subtrees, where N
is the size of the alphabet we want to complete within.
Thus for a simple latin-alphabet text trie, each node will potentially have 26 children, one for each character that could potentially follow the text entered thus far. (In graph theory terms, we could classify this as a Directed, Acyclic graph of order 26, but hey, who's counting?)
What we end up with is a broadly-branched tree where paths from the root to the leaves represent "words" within the dictionary.
Take a moment and read more about Tries:
Of course, our Trie won't be very useful without a good dataset
to populate it. Fortunately, our computers ship with a special
file containing a list of standard dictionary words.
It lives at /usr/share/dict/words
Using the unix utility wc (word count), we can see that the file
contains 235886 words:
$ cat /usr/share/dict/words | wc -l
235886
Should be enough for us!
To complete the project, you will need to build an autocomplete system which provides the following features:
- Insert a single word to the autocomplete dictionary
- Count the number of words in the dictionary
- Populate a newline-separated list of words into the dictionary
- Suggest completions for a substring
- Mark a selection for a substring
- Weight subsequent suggestions based on previous selections
We'll expect to interact with your completion project from an interactive pry session, following a model something like this:
# open pry from root project directory
require "./lib/complete_me"
completion = CompleteMe.new
completion.insert("pizza")
completion.count
=> 1
completion.suggest("piz")
=> ["pizza"]
dictionary = File.read("/usr/share/dict/words")
completion.populate(dictionary)
completion.count
=> 235886
completion.suggest("piz")
=> ["pizza", "pizzeria", "pizzicato", "pizzle", "pize"]The common gripe about autocomplete systems is that they give us suggestions that are technically valid but not at all what we wanted.
A solution to this problem is to "train" the completion dictionary over time based on the user's actual selections. So, if a user consistently selects "pizza" in response to completions for "pizz", it probably makes sense to recommend that as their first suggestion.
To facilitate this, your library should support a select method,
which takes a substring and the selected suggestion. You
will need to record this selection in your trie and use it
to influence future selections to make.
Here's what that interaction model should look like:
require "./lib/complete_me"
completion = CompleteMe.new
dictionary = File.read("/usr/share/dict/words")
completion.populate(dictionary)
completion.suggest("piz")
=> ["pizza", "pizzeria", "pizzicato"]
completion.select("piz", "pizzeria")
completion.suggest("piz")
=> ["pizzeria", "pizza", "pizzicato", "pizzle", "pize"]This is the first project where we'll use an automated spec harness to evaluate the correctness of your projects.
For this reason, you'll want to make sure to follow the top-level interface described in the previous sections closely.
You can structure the internals of your program however you like, but if the top level interface does not match, the spec harness will be unable to evaluate your work.
Spec harness available here.
Please make sure that, before your evaluation, your project has each of the following:
- SimpleCov reporting accurate test coverage statistics
In addition to the base features included above, you must choose one of the following to implement:
A simple approach to tracking selections would be to simply "count" the number of times a given word is selected (e.g. "pizza" - 4 times, etc). But a more sophisticated solution would allow us to track selection information per completion string.
That is, we want to make sure that when selecting a given word,
that selection is only counted toward subsequent suggestions against
the same substring. Here's an example:
require "./lib/complete_me"
completion = CompleteMe.new
dictionary = File.read("/usr/share/dict/words")
completion.populate(dictionary)
completion.select("piz", "pizzeria")
completion.select("piz", "pizzeria")
completion.select("piz", "pizzeria")
completion.select("pi", "pizza")
completion.select("pi", "pizza")
completion.select("pi", "pizzicato")
completion.suggest("piz")
=> ["pizzeria", "pizza", "pizzicato", "pizzle", "pize"]
completion.suggest("pi")
=> ["pizza", "pizzicato","pizzeria", "pizzle", "pize"]
In this example, against the substring "piz" we choose "pizzeria" 3 times, making it the dominant choice for this substring.
However for the substring "pi", we choose "pizza" twice and "pizzicato" once. The previous selections of "pizzeria" against "piz" don't count when suggesting against "pi", so now "pizza" and "pizzicato" come up as the top choices.
Let's add a feature that let's us delete words from the tree. When deleting a node, we'll need to consider a couple of cases.
First, make sure that we adjust our tree so that the node relating to the removed word is no longer seen as a valid word. This means that subsequent suggestions should no longer return it as a match for any of its substrings.
For "intermediate" nodes (i.e. nodes that still have children below them), this is all you need to do.
However, for leaf nodes (i.e. nodes at the end of the tree), we will also want to completely remove those nodes from the tree. Since the node in question no longer represents a word and there are no remaining nodes below it, there's no point in keeping it in the tree, so we should remove it.
Additionally, once we remove this node, we would also want to remove any of its parents for which it was the only child. That is -- if, once we remove our word in question, the node above it is now a path to nowhere, we should also remove that node. This process would repeat up the chain until we finally reach "word" node that we want to keep around.
The exact implementation of this process will depend on how your tree is built, so we likely won't include it in the spec harness. You will need to provide your own tests that demonstrate this functionality.
Working with words was interesting, but what about a bigger dataset? Check out this data file (you'll want the CSV version) that contains all the known addresses in the city of Denver. Use the full_address field that's last in the row. Can you make your autocomplete work with that dataset?
Could your word lookup possibly handle middle-of-the-word matches? So that com would list both the possibilities complete and incomplete? How does this change the memory requirements of your running program?
Can you create a graphical user interface for your code? Something that a "normal person" might plausibly use? Consider a toolkit like Shoes or Ruby Processing.
- Setup SimpleCov to monitor test coverage along the way