16
Lecture 2: Retrieval Models Maya Ramanath

Lecture 2: Retrieval Models

  • Upload
    leane

  • View
    41

  • Download
    0

Embed Size (px)

DESCRIPTION

Lecture 2: Retrieval Models. Maya Ramanath. QQ1. Vector space model: 0 for non-presence of a term, 1 for presence: Query: q1 AND q2 AND q3 Compare the set of results returned by the vector space model and boolean model. Term weighting (1/3): tf. Query: Cat. D2. D3. D1. D1. D3. - PowerPoint PPT Presentation

Citation preview

Page 1: Lecture 2: Retrieval Models

Lecture 2: Retrieval Models

Maya Ramanath

Page 2: Lecture 2: Retrieval Models

QQ1• Vector space model: 0 for non-

presence of a term, 1 for presence:• Query: q1 AND q2 AND q3

Compare the set of results returned by the vector space model and boolean model.

Page 3: Lecture 2: Retrieval Models

Term weighting (1/3): tf

The lion (Panthera leo) is one of four big cats …

Highly distinctive, the male lion is easily recognised by its mane…

The cat (Felis catus), also known as the domestic cat …

Cats are similar in anatomy to the other felids…

The New World monkeys are classified within the parvorder Platyrrhini, whereas the Old World monkeys form part of the parvorder Catarrhini, which also includes the hominoids…

DOC

COUNT

Lion

Cat

The

D1 2 1 2D2 0 3 2D3 0 0 5D1 D2 D3

Query: Cat

Query: Lion

Query: The Lion

D2 D1 D3

D1 D2 D3

D1 D2 D3

Page 4: Lecture 2: Retrieval Models

Term weighting (2/3): tf.idf

The lion (Panthera leo) is one of four big cats …

Highly distinctive, the male lion is easily recognised by its mane…

The cat (Felis catus), also known as the domestic cat …

Cats are similar in anatomy to the other felids…

The New World monkeys are classified within the parvorder Platyrrhini, whereas the Old World monkeys form part of the parvorder Catarrhini, which also includes the hominoids…

DOC

COUNT

Lion

Cat

The

D1 2 1 2D2 0 3 2D3 0 0 5D1 D2 D3

Query: Cat

Query: Lion

Query: The Lion

D2

D1

D3

D1

D2

D3

D1

D2

D3

DOC

WEIGHT

Lion

Cat

The

D1 2/1

1/2

2/3

D2 0 3/2

2/3

D3 0 0 5/3

Page 5: Lecture 2: Retrieval Models

Term weighting (3/3): doc length

• Shorter the text, more important the match

• Longer the text, more likely you “accidentally” fine a match

“Let me tell you about the cat, a domestic animal”

“Let me tell you about all the animals in the whole world (including the cat) !”

Page 6: Lecture 2: Retrieval Models

PROBABILISTIC RANKING

Page 7: Lecture 2: Retrieval Models

Binary Independence Model• Rank documents in decreasing

probability of relevance

Derivation is long, but not difficult!

Page 8: Lecture 2: Retrieval Models

Let

Let if

Page 9: Lecture 2: Retrieval Models
Page 10: Lecture 2: Retrieval Models

We still need relevant/irrelevant documents

• Sample of corpus, exhaustively judged

• Relevance feedback• Pseudo-relevance feedback

• 2-poisson model• …

Page 11: Lecture 2: Retrieval Models

LANGUAGE MODELS

Page 12: Lecture 2: Retrieval Models

Intuition (1/2)

Document D Document Q

This is the observationCan we figure out the source?

Page 13: Lecture 2: Retrieval Models

Intuition (2/2)PD PQ

Document D Document Q

These are the observations

Can we estimate PD and PQ?

Page 14: Lecture 2: Retrieval Models

Query as a sample

Estimated using MLE

Page 15: Lecture 2: Retrieval Models

References• For term weighting and the long derivation– Introduction to Information Retrieval:

Raghavan, Manning and Shuetze, Cambridge University Press, 2008. Also available from: http://nlp.stanford.edu/IR-book/html/htmledition/irbook.html

• Language Models– Statistical Language Models for Information

Retrieval: A Critical Review. ChengXiang Zhai, Foundations and Trends in IR 2(3), 2008

– Also available in the IR book above

Page 16: Lecture 2: Retrieval Models

QUESTIONS ?