DATA MINING COURSE - LECTURE 6: TEXT MINING & NLP (PART 1)

and analyze because :

 It is often ambiguous (fuzzy)

 It is often noisy (contain mistakes in vocabulary or

grammar)

 It can be informal (e.g. social media comments)

 It can be multilingual

 These challenges can often be alleviated by the

use of linguistic knowledge

11/4/2015

3

NATURAL TEXT DATA EXAMPLES

 Web pages

 Email messages

 Social media

 SMS

 News articles

 Scientific articles

 Books

 …

TEXT MINING

 Text mining is a subfield of data mining in which

the input data consists of natural text documents

  automated discovery of novel, hidden, and

useful patterns

11/4/2015

4

NATURAL LANGUAGE PROCESSING

 Intelligent automated processing of natural

language text based on linguistic knowledge

Done by executing

computer programs

Text written in

some human

language

knowledge about phonetics,

structure of words, syntax

of single sentences,

semantics

9

TEXT MINING  NLP

Text mining =

automated

processing of

natural text data

to extract novel

useful patterns

Text mining  NLP =

automated processing

of natural text data

based on linguistic

knowledge to extract

novel useful patterns

NLP = automated

processing of

natural text data

based on linguistic

knowledge

11/4/2015

5

TYPES OF LINGUISTIC KNOWLEDGE

 There are several levels of linguistic knowledge :

 Morphological : structure of words – how single words

are constructed from smaller units (called morphemes)

 Lexical semantics : meaning of single words based on

their current context

 Syntactic : grammatical structure of sentences

 Sentential semantics : meaning of words and

sentences

 Discourse semantics : meaning across many sentences

(e.g. a paragraph or an entire document)

EXAMPLES (ENGLISH LANGUAGE)

 Morphology:

 going = go + ing

 dogs = plural of dog; rated is past tense of rate

 carried = carry + ed

 aren’t = are not

 Syntax:

 simple sentence = noun + verb

 Sentential semantics:

 X killed Y : X is the killer and Y is the killed person

 Discourse semantics:

 George woke up. He went to the kitchen.

11/4/2015

6

DOG-NLP?

 Can dogs process human language?

DOG-NLP?

 In this caricature, the dog can only recognize its name

 Dogs can recognize certain words, but cannot deduce

meaning of whole sentences

11/4/2015

7

NLP VS. NO-NLP

 Consider the simple problem of counting words in

an English text documents

 Non-NLP approach: find letter sequences separated

by whitespace or punctuation

 For example, the wc command in Linux works like this

 NLP approach:

1) Find letter sequences separated by whitespace characters;

however the letter sequence may contain 1-2 hyphens or one

apostrophe in the middle

2) Verify the extracted words against some electronic/online

English dictionary

Question: What type of linguistic

knowledge does each step use ?

APPLICATION EXAMPLE: SEMANTIC SEARCH

 Problem: given a user query, find all documents

that have similar meaning as the query

 Semantic search ≠ Traditional “dumb” search

 Traditional search = finding documents that contain all

(key) words of the user query

 grep in Linux is an example of dumb search

 Semantic search is difficult/challenging because :

 Words can have different meanings ;

 Meaning of words/sentences can vary with context ;

 User query can contain mistakes ;

 And of course because the repository is very large!

11/4/2015

8

11/4/2015

9

OTHER APPLICATIONS

 Place email messages into different folders (e.g. gmail)

 Analyze web sites of competitors to compare prices

 Translate a Web page (Google chrome) or a user

comment (on Facebook)

 Investigate public opinion on social media with regard

to some product, brand, company, political figure,

issue, etc. (called sentiment analysis)

 Analyze user reviews/comments in order to extract

implicit ratings for a recommender system (i.e. to

initialize the ratings matrix)

11/4/2015

10