Watch a machine-learning system parse the grammatical structure of sentences
foxtype.comSo FoxType is just Spacy[0]? I honestly thought this might be something different until I checked out the Spacy page and saw FoxType listed as a customer. Kind of a dick move simply repackaging other people's work and passing. It off as your own. (Yes, this is literally just the "displacy" demo.)
Spacy is a pretty good system. It's fast, way faster thank NLTK and the Stanford stuff and it's pretty accurate. Also it's licensing is pretty flexible. The only knock I have against it is the lack of bindings for languages other than Python, but whatever.
There's more to their product it seems than just the front end for sentence diagramming. While a shout out wouldn't hurt, I don't think it's really a dick move.
There's also more to Spacy than just dependency parsing as well, but that's a big part of it.
I would certainly hope there's more to FoxLab, but I have literally seen this demo before.
I have also been around the block enough times to know that putting some ill fitting duct tape around a bunch of other libraries that do all the work to make a quick buck in a hot space is a very common trick as well.
Whilst an impressive demonstration, how is it an example of machine learning? It looks like lexical analysis to me.
Same thoughts as mine. Seems like ML was just thrown in the title to get more hits.
It uses machine learning to parse the sentence, all modern lexical parsers are based on machine learning.
Aren't most of them rule based?
Not for years. State of the art has moved on from Eric Brill's dissertation 25 years ago.
Seems to line up with Stanford's part of speech tagger: http://nlp.stanford.edu:8080/parser/
Seems to fall for a lot of easy traps. Garden path sentences ("The old man the boats" or "The horse raced past the barn fell") are parsed wrong, and it parses the two sentences/phrases "When he did that I laughed" and "When he said that I laughed" identically - by interpreting "that" as a preposition in both cases, which it isn't.
It's cool that it parses on the fly. But unless it parses correctly I don't see how it actually serves a purpose other than just looking cool.
I tried
> I heard you like phrase structure, so I put a phrase in your phrase so you can parse a phrase while you parse a phrase.
It misinterpreted "like" as 'in the manner of; akin to' ("I heard you as though [I were hearing] phrase structure"?), so I tried
> I heard you enjoy phrase structure, so I put a phrase in your phrase so you can parse a phrase while you parse a phrase.
That got parsed correctly.
Edit: well, it says "enjoy" is a base form rather than 3rd person singular, which is wrong in this context; it's an inflected form, like "I heard she enjoys phrase structure".
These require a system which either backtracks or recognizes it's made a mistake and corrects it. I know how to do that with a rule-based parser.
It also fails on "British left waffles on Falklands." While there are two syntactic parsers for that, one of them borders on the absurd and should be rejected. The problem is that to parse some sentences requires understanding of the words and phrases it reads, which in turn requires common sense.
Well, that is just choosing the semantically correct parse from multiple syntactically correct parses. This parser isn't even finding syntactically correct parses.
It is clearly not up to reading P G Wodehouse. The following sentence is parsed as having two subject nouns.
Jeeves, dark forces are drawing us to Totleigh.
One does not even dare to try Marx Brothers quotes.
My favorite is "Fruit flies like honey." which trips this parser up as well.
Type in "Fruit Flies like honey"
So, to get correct output, you need to give it incorrect input? You don't capitalize names of animal or plant species in English.
The dogs bark at the tree bark.
It doesn't do well on the classic: https://en.wikipedia.org/wiki/Buffalo_buffalo_Buffalo_buffal...
To be fair, if someone actually utters that sentence, I think most English speakers would reply with "I'm sorry?" or "Do you need help?"
See the video TED-Ed Buffalo buffalo buffalo: One-word sentences and how they work - Emma Bryce [1]
But it does try it's very best on: https://en.wikipedia.org/wiki/James_while_John_had_had_had_h...
Very interesting demo. I am not being critical, but it does not do anaphora resolution (resolving pronouns to noun phrases, etc).
It got "colorless green ideas sleep furiously," wrong.
I don't see how it got it wrong (and it would be surprising if it did given the sentence is not syntactically ambiguous). It does use a visual notation which, AFAIK, differs from the standard in linguistcs [0].
[0]: https://en.wikipedia.org/wiki/File:Syntax_tree_for_Colorless...
This site is producing dependency parses, which have become popular in the last 10 years or so. The linked figure is a phrase structure parse, which used to be the thing everyone worked on but have since turned out to be both harder to get and less useful than dependency parses.
It decided that "sleep," was a noun, rather than a verb (as in the diagram you note). If your parts of speech are wrong, it's not a successful parse.
"The cat ate the farty bird"
Sorry machine, "farty" is an adjective, not a noun.
Someone here in a research field related to NLP has to add this to their corpus.