Why make conversational macros?

For ergonomic reasons, I use free, open-source stenography software to type. This means that I press anywhere from 1 to about 12 keys at once, and get a word. To learn more about it, you can read my post about the mechanics of stenography and my motivation for using it.

In short, machine stenography in the U.S. is a chording, mnemonic keyboarding system. When I press a chord of keys, the software looks it up in a dictionary file and outputs a pre-defined word for me. The dictionary file that comes with the software has many tens of thousands of entries, covering most of English.

Some of these entries are phrases instead of words. For example, I can type “I don’t know” in a single seven-key chord. These are very useful for me, because most of my typing is instant messaging, and these phrases come up a lot.

But this set of phrases is somewhat limited, because the author of the dictionary works in areas like medical transcription rather than instant messaging. Instead of conversationally-useful phrases, there are lots of phrases like “non-small cell lung cancer” which can be produced in one chord.

I’d like to improve the breadth of conversational entries in my dictionary for use in instant messaging. So how do I know which entries to add? The best way is to actually analyze all of my instant messages.

How do I get all of my Messenger conversations?

Facebook has an option to let you download all of your Facebook data. (Note: it takes a while to prepare the archive, so if you plan to download your data, start the process beforehand.)

Unfortunately, all of that data is in HTML form. This is not very useful for programmatic analysis. My messages page is 100 MB, and my add-ons choke the life out of Firefox when trying to open it.

All my personal data, held hostage by the tyranny of web standards.

All my personal data, held hostage by the tyranny of web standards.

Somebody has already written a script to extract the message data, so I just ran the script and got an 80 MB JSON file containing all of my messages. Below is an excerpt from the beginning of the JSON file:

  "threads": [
      "people": [
        "David Redacted",
        "Waleed Khan"
      "messages": [
          "sender": "Waleed Khan",
          "date_time": "Monday, March 09, 2015 at 12:16PM ",
          "text": "This Wednesday?"
          "sender": "David Redacted",
          "date_time": "Monday, March 09, 2015 at 12:16PM ",
          "text": "it's going out wednesday"

Identifying the candidate macros

How do I determine the phrases which I use most? Essentially, I need to look at the top n-grams among my messages. That is, I need to look at the most frequent sequences of words of length 1, 2, 3, …, etc.

Looking at the top sequences of single words would just tell me words like “the” are most common and that I ought to macro them. Of course, I have the word “the” macroed already. Thus, I’m going to start with n = 10 and work my way down until, say, n = 2.

So I need to extract my messages from the JSON and produce the n-grams, and then I need to determine which n-grams appear frequently enough that it’s worth adding a macro for them.

Parsing JSON

I could write a script to parse my messages out of the JSON, but perhaps better would be to take a opportunity to get practice with jq, which is a command-line tool to manipulate JSON.

After some finagling, I write the following query. I’ve told jq to produce “raw” output rather than JSON output so that I don’t have to parse JSON again to make use of this data.

$ jq <messages.json --raw-output '.threads[].messages[] | select(.sender == "Waleed Khan") | .text'
This Wednesday?
So I have plenty of time to do my bit, yeah? It's not like we're pushing it out before this project is done

Extracting useful n-grams

Now that I have a text stream of all of the messages I’ve ever said, I need to determine the n-grams for any given n; and furthermore, I only want to output the ones that appear frequently enough.

How do I determine if an n-gram appears frequently enough? Initially, I looked at the n-grams that appeared more often than the average n-gram appeared. But as it turns out, the average n-gram appears between 1 and 2 times for most values of n, so this qualification includes too many n-grams. Arbitrarily, I decided that an n-gram has to appear at least 5 times for it to be worth looking at.

I wrote a short script to find the n-grams:

import collections
import sys

n = int(sys.argv[1])
ngrams = collections.Counter()

for line in sys.stdin:
    words = line.split()
    line_ngrams = zip(*[tuple(words[i:]) for i in range(n)])
    for ngram in line_ngrams:
        ngrams[ngram] += 1

average = sum(ngrams.values()) / len(ngrams)
print("Average: {}\n".format(average))
most_frequent_ngrams = sorted(ngrams.items(), key=lambda x: x[1], reverse=True)
for ngram, count in most_frequent_ngrams:
    if count >= 5:
        print("{}: {}".format(count, " ".join(ngram)))


I’ve assembled a selection of the most interesting n-grams that I found. There weren’t useful n-grams beyond about n = 6.


Here are the top few results for n = 3. I’ve also annotated which of these were already in my dictionary.

Note that the entries with “<3” and “tsk” were part of very large copy-pasted messages. (That is, there was one very large message with many “<3”s or “tsks” that skewed the counts.) They also appear as top 10-grams, but they’re not really representative of my messages.

Average: 1.3638674842177816

623: <3 <3 <3
618: I don't think (already in dictionary)
594: I have to
568: I don't know (already in dictionary)
437: I love you
434: you want to
385: tsk tsk tsk
313: you have to
289: I'm going to
282: I'm not sure
265: I don't have (already in dictionary)
254: Do you want (already in dictionary)
245: I want to
238: a lot of
218: don't want to
216: I don't want (already in dictionary)
210: to go to
205: I thought you
187: Do you have (already in dictionary)
185: I have a
182: I think I
181: Love you too
181: be able to
178: you have a
176: want me to
176: go to bed
164: don't know what
163: was going to
160: I'm pretty sure

Some of these already appear in the dictionary in a lesser form. For example, “I have”, “have to”, “love you”, “you want”, and “I want”. By being 2-grams, they can be a little bit more flexible than if they were part of a 3-gram. Then I can also say that “I want food” in addition to “I want to get food”.

I had long suspected that “to go to” and variations were part of my most frequently used phrases. It turns out that “going to” already has entries in the dictionary, but I never thought to look for them.

Apparently, I’m usually “not sure” or “pretty sure” about things. This is probably to give myself an out for the cases when I’m wrong about things. Maybe I should reflect on that aspect of my personality.


Here are some of my top 4-grams. These are quite insightful, as none of these phrases are in my dictionary, but some evidently ought to be.

Average: 1.0939080011918132

447: <3 <3 <3 <3
382: tsk tsk tsk tsk
144: I was going to
142: I don't want to
138: I don't know what
130: you want me to
124: Do you want to
91: I don't know if
86: I have to go
82: I don't think I
79: I don't know how
75: I thought you were
72: Are you going to
71: Do you want me
70: I don't think so
69: do you want to
63: to go to bed
63: if you want to
62: Did you know that
58: so that I can
55: I have to get
54: I thought it was
52: I don't get it
52: I don't think it's
51: I have no idea
51: are you going to
50: have to go to
49: I'm gonna go to
48: don't know how to
48: I don't have a

Some of my favorites are “I was going to”, “are you going to”, “so that I can”, and “I have no idea”.


Here are some of my top 6-grams, none of which were already in my dictionary. There are a couple of anomalies with missed calls on Messenger, because I have never wanted to use Messenger to make a call and every instance of that was an accident.

Average: 1.0098786652250653

413: <3 <3 <3 <3 <3 <3
377: tsk tsk tsk tsk tsk tsk
20: nom nom nom nom nom nom
15: I have to go to bed
13: Anna missed a call from Waleed.
13: So what you're saying is that
12: don't know what you're talking about
12: I want to go to bed
12: I'm going to go to bed
11: I don't know what to do
10: Talk to you in a bit
10: missed a call from Waleed Khan.
10: I don't know what you're talking
9: What am I supposed to do
9: I don't think you know what
9: Do you want me to come
9: I thought you were going to
9: Why would you want to do
9: I was going to go to
8: would you want to do that
7: I thought you were talking about
7: I don't know how to get
7: What do you want me to
7: I was gonna go to bed
7: me know when you get here
7: Do you want to go to
7: I just don't want you to
7: I don't know if I can

One thing that you can see from this data is that I really like going to bed.

Some phrases from here which might be useful to add to my dictionary are “I don’t know what you’re talking about”, “so what you’re saying is”, and “let me know when you get here”.

Phrases like “I don’t know if I can” and “Do you want to go to” can be formed by two 3-grams (once I update my dictionary with the 3-gram results) so it might not be worth adding those phrases directly.