Moldable Emacs: let's make English easy to query!
Too long; didn't read
This is about an experiment to parse text with NLP software so that we can ask our text questions!
The problem
Lately I have been spending time querying my code. I have been asking mostly basic things: how many functions has this file? How complex are those? How many functions in the whole project? Now I am finally (and slowly) getting into more complex queries, but then I said to myself: "wait! Let's try with English now".
My thinking is: if I can recognize things like functions in code, in text I should find things like predicates. Or objects, or subjects or sentences that depend on others. Since we have fancy stuff like GT-3, I was hopeful to find some software I could plug in moldable-emacs to obtain some structural information for text.
So I wondered: how would this work?
It is a problem indeed
Imagine structural editing for English. I mean Emacs already lets you move words, paragraphs and sentences all over the place, if you like. But what about predicates? And what about finding me all the sentences that have "pear" as subject? By structural editing I mean to change the subject of those sentences from pear to orange.
And what if we could lint our text by coming up with rules like: "careful! 3 sentences contain recursive relative clauses: that is difficult to read".
There is more though! How cool would be to query a book? Imagine you have fond memories of a passage in a book. You can vaguely remember that passage is about a prince and a snake. How cool would it be if you can ask the book itself? Find me all those phrases that have as subject a prince and as object a snake. We would at least have a more fun at it than skimming through.
You could even filter your text to read only what you need!! How exciting would that be?
And there is a solution
Well, this is a pretty rough prototype that I haven't properly tested yet. I am pretty excited about it, so this blog is more like a spoiler of work in progress.
First things first, I have been looking for a low hanging fruit CLI to a NLP model. I am not an expert of machine learning (yet XD), so I wanted something plug&play.
Looking at packages available via pip, I fell onto Elit. Well, I don't know much about this project, but plug&play, right? Just playing around.
After downloading a few GigaBytes of models (it is pretty usable, Elit does all itself), I run elit in the command line for the first time:
andrea@andrea-pc:/tmp$ elit parse Moldable development is a way of programming through which you construct custom tools for each problem.
That returns the following:
{
"lem": [
["moldable", "development", "be", "a", "way", "of", "programming", "through", "which", "you", "construct", "custom", "tool", "for", "each", "problem", "."]
],
"pos": [
["JJ", "NN", "VBZ", "DT", "NN", "IN", "NN", "IN", "WDT", "PRP", "VBP", "JJ", "NNS", "IN", "DT", "NN", "."]
],
"ner": [
[]
],
"srl": [
[[["ARGM-LOC", 3, 7, "a way of programming"], ["R-ARGM-MNR", 7, 9, "through which"], ["ARG0", 9, 10, "you"], ["PRED", 10, 11, "construct"], ["ARG1", 11, 16, "custom tools for each problem"]], [["ARG1", 0, 2, "Moldable development"], ["PRED", 2, 3, "is"], ["ARG2", 3, 16, "a way of programming through which you construct custom tools for each problem"]]]
],
"dep": [
[[1, "attr"], [4, "nsbj"], [4, "cop"], [4, "det"], [-1, "root"], [6, "case"], [4, "ppmod"], [8, "case"], [10, "r-ppmod"], [10, "nsbj"], [4, "relcl"], [12, "attr"], [10, "obj"], [15, "case"], [15, "det"], [12, "ppmod"], [4, "p"]]
],
"con": [
["TOP", [["S", [["NP", [["JJ", ["Moldable"]], ["NN", ["development"]]]], ["VP", [["VBZ", ["is"]], ["NP", [["NP", [["DT", ["a"]], ["NN", ["way"]]]], ["PP", [["IN", ["of"]], ["NP", [["NP", [["NN", ["programming"]]]], ["SBAR", [["WHPP", [["IN", ["through"]], ["WHNP", [["WDT", ["which"]]]]]], ["S", [["NP", [["PRP", ["you"]]]], ["VP", [["VBP", ["construct"]], ["NP", [["NP", [["JJ", ["custom"]], ["NNS", ["tools"]]]], ["PP", [["IN", ["for"]], ["NP", [["DT", ["each"]], ["NN", ["problem"]]]]]]]]]]]]]]]]]]]]]], [".", ["."]]]]]]
],
"amr": [
[["c4", "ARG0", "c6"], ["c0", "ARG1", "c1"], ["c4", "ARG1", "c5"], ["c8", "ARG1", "c5"], ["c4", "beneficiary", "c7"], ["c9", "domain", "c7"], ["c0", "instance", "develop-02"], ["c1", "instance", "moldable"], ["c2", "instance", "way"], ["c3", "instance", "program-01"], ["c4", "instance", "construct-01"], ["c5", "instance", "tool-01"], ["c6", "instance", "you"], ["c7", "instance", "problem"], ["c8", "instance", "custom"], ["c9", "instance", "each"], ["c0", "manner", "c2"], ["c3", "manner", "c2"], ["c4", "manner", "c2"]]
],
"tok": [
["Moldable", "development", "is", "a", "way", "of", "programming", "through", "which", "you", "construct", "custom", "tools", "for", "each", "problem", "."]
]
}
Naturally I barely know what that means. Anyway, I seem to see a
relcl there: elit found a relative clause! That is encouraging.
Furthermore, that seem to be also in the "con" section of the JSON
output: WHNP seems to wrap just our relative clause (i.e., "which
you construct custom tools for each problem").
I discovered that the strange abbreviations (e.g, JJ and NN) are
types of nodes of the natural syntax tree.
I made a list to use later from an university lecture I found online:
(defvar me-natural-syntax-tree-labels ;; TODO may be incomplete!
'(
(:label "NN" :long-name "singular noun" :example "pyramid" )
(:label "NNS" :long-name "plural noun" :example "lectures" )
(:label "NNP" :long-name "proper noun" :example "Khufu" )
(:label "VBD" :long-name "past tense verb" :example "claimed" )
(:label "VBZ" :long-name "3rd person singular present tense verb" :example "is" )
(:label "VBP" :long-name "non-3rd person singular present tense verb" :example "have" )
(:label "VBN" :long-name "past participle" :example "found" )
(:label "VB" :long-name "base" :example "He may like/VB cookies. I heard her answer/VB the question. They may be/VB tired.")
(:label "VBG" :long-name "present participle, G-form" :example "Eating/VG cookies is unhealthy. He likes eating/VG cookies.")
(:label "VBN" :long-name "past participle, N-form" :example "He has eaten/VBN the cookies. She has ansered/VBN the questions. My question was not answered/VBN.")
:label "MD" :long-name "modal" :example "She will/MD prevail."
:label "TO" :long-name "auxiliary to" :example "She expects to/TO prevail."
(:label "PRP" :long-name "pronoun" :example "they" )
(:label "PRP$" :long-name "possessive pronoun" :example "their" )
(:label "JJ" :long-name "adjective" :example "public" )
(:label "IN" :long-name "preposition/complementizer" :example "in/that" )
(:label "DT" :long-name "determiner" :example "the" )
(:label "NP" :long-name "noun phrase" :example "their public lectures" )
(:label "VP" :long-name "verb phrase" :example "built the pyramid" )
(:label "PP" :long-name "prepositional phrase" :example "in the five chambers" )
(:label "S" :long-name "sentence " :example "Khufu built thepyramid" )
(:label "SBAR" :long-name "sbar" :example "that Khufu built the pyramid" )
)
"Natural syntax tree labels and examples according to https://www.cs.cornell.edu/courses/cs474/2004fa/lec1.pdf.")
Anyway, that seemed promising, so I decided to make a mold out of it. Let's try it!
You can see that I just mark some text and after a few seconds I get an Elisp list with some of elit's information. I took just the "con" bit for now, because it is an experiment. Also because I don't understand most of the output of elit yet (and I am not sure I should just yet).
The format in which I am showing the nodes is the same as the one for code. The idea is indeed to keep the interface the same, so maybe we could even ending up reusing things between the two.
In the video I also try a simple query! And you can see that I can jump to the original text in just the same way I can for code nodes.
Well I just finished implementing this, so it is all I can show you for now.
(Note that I took a lot of shortcuts for this: for example, I patched elit code to take a text file as input. Also the mold I have does not handle newlines well -- the positions I get for the node become imprecise).
All in all I am pretty happy that we are at a stage where I can hack something usable in a couple of days of evening work!
Conclusion
This was just a bite of machine learning in moldable-emacs for you. We can make Emacs understand English to let us query it. It is possible :D I am not sure where this will head to, but the start is exciting!
Please get in touch if you have suggestions for better ML software or ideas on how to use this feature. In the meanwhile I will try to experiment and refine it, and post some more experiments.
Happy machine-reading!