Lecture 7- Text Statistics and Document Parsing

Text Statistics & Document Parsing
Lecture 7
Sean A. Golliher

- Only collect URLs on www.montana.edu
- List out total
- number of unique URLs
- No sub-domains (www.cs.montana.edu) etc.
- Assume page rank of home page is pr7 during 1st
iteration
- Calculate Google Matrix of all sub URLs.
- Use Google Matrix Formula
- Deliver list of Sub URLs and page ranks
- Read:
http://static.googleusercontent.com/external_conten
t/untrusted_dlcp/research.google.com/en/us/pubs/a
rchive/37043.pdf

- TREC: Test collections of documents. Evaluation
corpus
- Text REtreival conference http://trec.nist.gov
- Tweets2011 http://trec.nist.gov/data/tweets/
- Twitter corpus tools:
https://github.com/lintool/twitter-corpus-tools
- Papers from 2011:
http://trec.nist.gov/pubs/trec20/t20.proceedings.html
- Linguistics Data Constortium:
http://www.ldc.upenn.edu/
- http://googleresearch.blogspot.com/2006/08/all-our-
n-gram-are-belong-to-you.html (Not exactly. $24k)

 Converting documents to index terms (Last
time)
 Why?
 Matching the exact string of characters typed by
the user is too restrictive
○ i.e., it doesn‟t work very well in terms of
effectiveness
 Not all words are of equal value in a search
 Sometimes not clear where words begin and end
○ Not even clear what a word is in some languages
 e.g., Chinese, Korean

 Huge variety of words used in text but
 Many statistical characteristics of word
occurrences are predictable
 e.g., distribution of word counts
 Retrieval models and ranking algorithms
depend heavily on statistical properties of
words
 e.g., important words occur often in documents
but are not high frequency in collection

 Distribution of word frequencies is very
skewed
 a few words occur very often, many words hardly
ever occur
 e.g., two most common words (“the”, “of”) make
up about 10% of all word occurrences in text
documents
 Zipf‟s “law”:
 observation that rank (r) of a word times its
frequency (f) is approximately a constant (k)
○ assuming words are ranked in order of decreasing
frequency

 r = rank = position in frequency table ranked
by highest 1st
 f = % of word occurrence in corpus
 r = k/f
 Or, the most frequent word will appear ~ 2x
as often as the 2nd most frequent word.

 Probability of occurrence of a word
 Pr = f/n
○ n is total number of word occurrences in text
○ f is frequency of word

 Zipf‟s law becomes r *Pr = c
 For English c ~0.1

Total documents
84,678
Total word occurrences 39,749,179
Vocabulary size 198,763
Words occurring > 1000 times
4,169
Words occurring once
70,064
Word Freq. r Pr(%) r.Pr
assistant 5,095 1,021 .013 0.13
sewers 100 17,110 2.56 × 10−4 0.04
toothbrush 10 51,555 2.56 × 10−5 0.01
hazmat 1 166,945 2.56 × 10−6 0.04

• Note problems at high and low frequencies

 As corpus grows, so does vocabulary size
 Fewer new words when corpus is already large
 Observed relationship (Heaps’ Law):

v = k.nβ
where v is vocabulary size (number of unique
words), n is the number of words in
corpus, k, β are parameters that
vary for each corpus (typical values given are
10 ≤ k ≤ 100 and β ≈ 0.5)

 Predictions for TREC collections are
accurate for large numbers of words
 e.g., first 10,879,522 words of the AP89
collection scanned
 prediction is 100,151 unique words
 actual number is 100,024
 Predictions for small numbers of words
(i.e. < 1000) are much worse

 Heaps‟ Law works with very large
corpora
 new words occurring even after seeing 30
million!
New words come from a variety of sources
○ spelling errors, invented words (e.g. product,
company names), code, other languages, email
addresses, etc.
 Search engines must deal with these
large and growing vocabularies

 How many pages contain all of the query terms?
 For the query “a b c”:
 P(a ∩ b ∩ c) = P(a)*P(b)*P(c) (joint probability)
 P(a) = fa/N .....
 fabc = N · fa/N · fb/N · fc/N = (fa · fb · fc)/N2
○ Since f = N*P
○ Assuming that terms occur independently
○ fabc is the estimated size of the result set
○ fa, fb, fc are the number of documents that terms a, b, and c
occur in
○ N is the number of documents in the collection

Collection size (N) is
25,205,179

 Poor estimates because words are not
independent
 Better estimates possible if co-occurrence
information available
P(c|(a ∩ b))
Probability that the words occurs if a and b occur.

 Def: Forming words from sequence of
characters
 Surprisingly complex in English, can be
harder in other languages
 Early IR systems:
 any sequence of alphanumeric characters of
length 3 or more
 terminated by a space or other special
character
 upper-case changed to lower-case

 Example:
 “Bigcorp's 2007 bi-annual report showed profits
rose 10%.” becomes
 “bigcorp 2007 annual report showed profits rose”
 Too simple for search applications or even
large-scale experiments
 Why? Too much information lost
 Small decisions in tokenizing can have major
impact on effectiveness of some queries

 Small words can be important in some
queries, usually in combinations
○ xp, ma, pm, ben e king, el paso, master p, gm, j
lo, world war II
 Both hyphenated and non-hyphenated forms
of many words are common
 Sometimes hyphen is not needed
○ e-bay, wal-mart, active-x, cd-rom, t-shirts
 At other times, hyphens should be considered
either as part of the word or a word separator
○ winston-salem, mazda rx-7, e-cards, pre-
diabetes, t-mobile, spanish-speaking

 Special characters are an important part of
tags, URLs, code in documents
 Capitalized words can have different
meaning from lower case words
 Bush, Apple
 Apostrophes can be a part of a word, a part
of a possessive, or just a mistake
 rosie o'donnell, can't, don't, 80's, 1890's, men's
straw hats, master's degree, england's ten largest
cities, shriner's

 Numbers can be important, including
decimals
 nokia 3250, top 10 courses, united 93, quicktime
6.5 pro, 92.3 the beat, 288358
 Periods can occur in numbers, abbreviations,
URLs, ends of sentences, and other
situations
 I.B.M., Ph.D., cs.umass.edu, F.E.A.R.

 First step is to use parser to identify appropriate
parts of document to tokenize
 Defer complex decisions to other components
 word is any sequence of alphanumeric
characters, terminated by a space or special
character, with everything converted to lower-case
 everything indexed
 example: 92.3 → 92 3 but search finds documents
with 92 and 3 adjacent
 incorporate some rules to reduce dependence on
query transformation components

 Not that different than simple tokenizing
process used in past
 Examples of rules used with TREC
 Apostrophes in words ignored
○ o‟connor → oconnor bob‟s → bobs
 Periods in abbreviations ignored
○ I.B.M. → ibm Ph.D. → ph d

 Function words (determiners,
prepositions) have little meaning on their
own
 High occurrence frequencies
 Treated as stopwords (i.e. removed)
 reduce index space, improve response time,
improve effectiveness
 Can be important in combinations
 e.g., “to be or not to be”

 Stopword list can be created from high-
frequency words or based on a standard
list
 Lists are customized for
applications, domains, and even parts of
documents
 e.g., “click” is a good stopword for anchor text
 Best policy is to index all words in
documents, make decisions about which
words to use at query time

 Many morphological variations of words
 inflectional (plurals, tenses)
 derivational (making verbs nouns etc.)
 In most cases, these have the same or very
similar meanings
 Stemmers attempt to reduce morphological
variations of words to a common stem
 usually involves removing suffixes
 Can be done at indexing time or as part of
query processing (like stopwords)

 User searching for news about Michael
Phelps. Types “Michael Phelps Swimming”

 News articles appear with latest events
containing events that already happened like
“swam”.

 SE needs to match swimming to swam.

 Generally a small but significant
effectiveness improvement
 Finding words that are semantically similar.
 can be crucial for some languages
 e.g., 5-10% improvement for English, up to 50%
in Arabic

Words with the Arabic root ktb

 Two basic types
 Dictionary-based: uses lists of related words
 Algorithmic: uses program to determine
related words
 Algorithmic stemmers
 suffix-s: remove „s‟ endings assuming plural
○ e.g., cats → cat, lakes → lake, wiis → wii
○ Many false negatives: supplies → supplie
○ Some false positives: ups → up

 Algorithmic stemmer used in IR
experiments since the 70s
 Consists of a series of rules designed to
the longest possible suffix at each step
 Effective in TREC
 Produces stems not words
 Makes a number of errors and difficult to
modify

 Porter2 stemmer addresses some of these issues
 Approach has been used with other languages

 Hybrid algorithmic-dictionary
 Word checked in dictionary
○ If present, either left alone or replaced with “exception”
○ If not present, word is checked for suffixes that could be
removed
○ After removal, dictionary is checked again

 Produces words not stems
 Comparable effectiveness
 Lower false positive rate, somewhat higher false
negative

 Many queries are 2-3 word phrases
 Phrases are
 More precise than single words
○ e.g., documents containing “black sea” vs. two words
“black” and “sea”
 Less ambiguous
○ e.g., “big apple” vs. “apple”
 Can be difficult for ranking
○ e.g., Given query “fishing supplies”, how do we score
documents with
 exact phrase many times, exact phrase just once, individual
words in same sentence, same paragraph, whole
document, variations on words?

 Text processing issue – how are phrases
recognized?
 Three possible approaches:
 Identify syntactic phrases using a part-of-
speech (POS) tagger
 Use word n-grams
 Store word positions in indexes and use
proximity operators in queries

 POS taggers use statistical models of text to
predict syntactic tags of words
 Example tags:
○ NN (singular noun), NNS (plural noun), VB (verb),
VBD (verb, past tense), VBN (verb, past participle),
IN (preposition), JJ (adjective), CC (conjunction,
e.g., “and”, “or”), PRP (pronoun), and MD (modal
auxiliary, e.g., “can”, “will”).
 Phrases can then be defined as simple noun
groups, for example

 POS tagging too slow for large collections
 Simpler definition – phrase is any sequence of n
words – known as n-grams
 bigram: 2 word sequence, trigram: 3 word
sequence, unigram: single words
 N-grams also used at character level for applications
such as OCR
 N-grams typically formed from overlapping
sequences of words
 i.e. move n-word “window” one word at a time in
document
 http://googleresearch.blogspot.com/2006/08/all-our-n-
gram-are-belong-to-you.html?z

 Frequent n-grams are more likely to be
meaningful phrases
 N-grams form a Zipf distribution
 Better fit than words alone
 Could index all n-grams up to specified
length
 Much faster than POS tagging
 Uses a lot of storage
○ e.g., document containing 1,000 words would
contain 3,990 instances of word n-grams of length 2
≤n≤5

 Some parts of documents are more
important than others
 Document parser recognizes structure
using markup, such as HTML tags
 Headers, anchor text, bolded text all likely to
be important
 Metadata can also be important
 Links used for link analysis

 Links are a key component of the Web
(Page Rank Lecture)
 Important for navigation, but also for
search
 e.g., <a href="http://example.com" >Example
website</a>
 “Example website” is the anchor text
 “http://example.com” is the destination link
 both are used by search engines

 Used as a description of the content of the
destination page
 i.e., collection of anchor text in all links pointing to
a page used as an additional text field
 Anchor text tends to be
short, descriptive, and similar to query text
 Retrieval experiments have shown that
anchor text has significant impact on
effectiveness for some types of queries
 i.e., more than PageRank

 Preliminaries:
 1) Extract links from the source text. You'll also want to extract
the URL from each document in a separate file. Now you have
all the links (source-destination pairs) and all the source
documents
 2) Remove all links from the list that do not connect two
documents in the corpus. The easiest way to do this is to sort all
links by destination, then compare that against the corpus URLs
list (also sorted)
 3) Create a new file I that contains a (url, pagerank) pair for each
URL in the corpus. The initial PageRank value is 1/#D (#D =
number of urls)
 At this point there are two interesting files:
 [L] links (trimmed to contain only corpus links, sorted by source
URL)
 [I] URL/PageRank pairs, initialized to a constant

Lecture 7- Text Statistics and Document Parsing

More Related Content

What's hot

Similar to Lecture 7- Text Statistics and Document Parsing

More from Sean Golliher

Recently uploaded

Lecture 7- Text Statistics and Document Parsing

Editor's Notes