How a language model learns: guessing the next word, sentence 13

The text a bot learned from had good and bad parts; companies filter it and train the bot to be polite, but they do not catch everything.

Who acts: bot, text, companies · What happens: learn, have, filter, train, be, not catch · To what, where: text, good parts, bot, everything

This is a sentence of the lesson «How a language model learns: guessing the next word» of the ai course of Mova's primary school. The page gives its syntax tree — a dependency tree in the style of Universal Dependencies — and its pragmatics clause by clause: who acts (bot, text, companies), what happens, when and where, as Mova's frames read it from the tree.

The tree

The words, one row each (CoNLL-U)

#	word	lemma	UPOS	features	head	relation
1	The	the	DET	Definite=Def PronType=Art	2 text	det
2	text	text	NOUN	Number=Sing	7 had	nsubj
3	a	a	DET	Definite=Ind PronType=Art	4 bot	det
4	bot	bot	NOUN	Number=Sing	5 learned	nsubj
5	learned	learn	VERB	Mood=Ind Number=Sing Person=3 Tense=Past VerbForm=Fin	2 text	acl
6	from	from	ADP	_	2 text	case
7	had	have	VERB	Mood=Ind Tense=Past VerbForm=Fin	root	root
8	good	good	ADJ	Degree=Pos	11 parts	amod
9	and	and	CCONJ	_	10 bad	cc
10	bad	bad	ADJ	Degree=Pos	8 good	conj
11	parts	part	NOUN	Number=Plur	7 had	obj
12	;	;	PUNCT	_	14 filter	punct
13	companies	company	NOUN	Number=Plur	14 filter	nsubj
14	filter	filter	VERB	Mood=Ind Number=Plur Person=3 Tense=Pres VerbForm=Fin	7 had	parataxis
15	it	it	PRON	Case=Acc Gender=Neut Number=Sing Person=3 PronType=Prs	14 filter	obj
16	and	and	CCONJ	_	17 train	cc
17	train	train	VERB	Mood=Ind Number=Plur Person=3 Tense=Pres VerbForm=Fin	14 filter	conj
18	the	the	DET	Definite=Def PronType=Art	19 bot	det
19	bot	bot	NOUN	Number=Sing	17 train	obj
20	to	to	PART	_	22 polite	mark
21	be	be	AUX	VerbForm=Inf	22 polite	cop
22	polite	polite	ADJ	Degree=Pos	17 train	xcomp
23	,	,	PUNCT	_	28 catch	punct
24	but	but	CCONJ	_	28 catch	cc
25	they	they	PRON	Case=Nom Number=Plur Person=3 PronType=Prs	28 catch	nsubj
26	do	do	AUX	Mood=Ind Number=Plur Person=3 Tense=Pres VerbForm=Fin	28 catch	aux
27	not	not	PART	Polarity=Neg	28 catch	advmod
28	catch	catch	VERB	VerbForm=Inf	14 filter	conj
29	everything	everything	PRON	Number=Sing PronType=Tot	28 catch	obj
30	.	.	PUNCT	_	7 had	punct

Pragmatics

form: statement
act: narrate
voice: narrator
polarity: neg
teacher: kept the draft

#	verb	tense, mood	who	what, whom	other	roles by
1	learn	past	a bot	the text not said here		vn learn-14-2-1 «NP V NP»: agent=Agent text=Topic
2	have	past	the text	good parts		vn birth-28.2-1 «NP V NP»: agent=Agent part=Patient
3	filter	present	companies	bot		vn wipe_instr-10.4.2 «NP V NP.initial_location»: agent=Agent it=Initial_Location
4	train	present	companies not said here	the bot		vn hire-13.5.3 «NP V NP»: agent=Agent bot=Theme
5	is	to-infinitive	the bot not said here		polite property	tree rules
6	not catch	present · do-support	companies = company	everything		vn get-13.5.1 «NP V NP»: agent=Agent everything=Theme

One row per clause the reader finds. «not said here»: a role filled from the clause above. «roles by»: a VerbNet class and frame when one fits the verb, else the reader's rules on the tree.

Licence: CC BY-SA 4.0, as the whole of Mova corpus 0.4, of which this treebank (the agreed school, mova-school-agreed-en.conllu) is part. The lessons were written for Mova's primary school. Sentence school:ai:002-guessing-the-next-word.lesson:13.

The tree: Mova annot: Rust (en) → Haiku 5.5 (medium, Batches API) patches; gates UD 2.18; agreed by treeagree rules (Stanza, Trankit, spaCy trf): R8-pos-agree ×3. The frames are read from the tree by Mova's frames reader (prag) when this page is built, the lesson read in order, so a pronoun may be resolved from an earlier sentence.

See the release 0.4.

This page in Markdown → https://md.mova.elamur.ai/free/en/corpora/school/ai/2-guessing-the-next-word/13/