ir_all_i_know.txt

	Precision "точність": What fraction of the returned results are relevant to the information need?
	tp / len(chosen) = tp / (tp + fp)

Recall - "повнота": What fraction of the relevant documents in the collection were returned by the system?
	== sensitivity:
	tp / (tp + fn)

accuracy - "правильність":
	(tp + tn) / (tp + tn + fp + fn)

f measure == balanced f measure: ?:
	2pr/(p+r),  є [0, 1]


rocchio:
	user point of view:
		query
		user marks relevant documents
		query is expanded (new words are added)
		new results are sent

	internally:
		vector q is made distant from non-relevant docs and close to relevant docs

Pseudo relevance feedback:
	assume top k retrieved documents are relevant
	use rocchio

Indirect relevance feedback:
	treat user clicks, etc as an evidence of relevance


champion lists:
	r [top] docs to each term
	>> so r is chosen in advance
	r can be longer for rare terms

high and low lists:
	disjoint sets
	sorted by static quality
	high: m docs with highest tf for t


0		1		2		3		4		5
0		10		110		1110  	11110	111110		// unary code:
		0		100		101		11000	11001		// gamma code


permuterm index:
	hello
	hello$
	ello$h
	llo$he
	lo$hel
	o$hell
	$hello

	he*o
	$he*o
	o$he*


SPIMI:
	map for terms to termIDs

MapReduce:
	master, parsers, inverter
	master identifies idle machines and assigns role to them


???: каппа статистика: 


				    документ 	Релев 	Нерелев
термін присутній	Xt = 1		Pt		Ut
турмін відсутній	Xt = 0		1-Pt	1-Ut


Ціпф - приблизна к-сть термінів в колекції