(Week 5) [T8] NLP

Question 1

Talk a little bit about the classical NLP approach.

Answer 1

Can be seen as a pipeline, and its beginning is with text documents.
Preprocessing → Section bound. detect. → Sentence b. d. → Tokenizer
After the tokenizer, it could happen more actions.

Question 2

There could be ambiguity and linguistic variation in NLP. Give 4 examples.

Answer 2

• Different words with the same meaning
• General polysemy.
• Acronyms and domain specific language.
• Negation.

Question 3

[1st method - Simple rule-based approach]
Pros?
Cons?

Answer 3

+ Simple and still a common approach.
+ Regular expression included in many programming languages.
+ Great for semi-structured targets.

• Patterns must consider all possible configurations.

Question 4

[2nd method - Symbolical/grammatical NLP (classical approach)]
Pros?
Cons?

Answer 4

+ Great for extracting and mapping large numbers of concepts.

• Complex, more steps and more opportunities for error.
• Can be slow.

Question 5

[3rd method - Machine learning]
Pros?
Cons?

Answer 5

+ Targeted approach = high accuracy.
+ Capable of learning from examples.

• Involves manual training.
• New target = new training effort.

Question 6

Calculate the TF-IDF (term frequency - inverse document frequency) value of the following.

The term t appears 20 times in a document that contains 100 words.
We have a collection of 10000 documents, the term t appears in 100.

Interpret the value.

Answer 6

TF = 20/100 = 0.2
IDF = log_10(10000/100) = 2

TF-IDF = TF * IDF = 0.2 * 2 = 0.4.

The higher the TF, more relevant a document is.
The lower the IDF, less discriminative the term is.
The higher the TF-IDF score the more important or relevant the term is.