Lecture 1: Introduction

DID NOT RESPECT MY BREAK. 8 min instead of 15.

Sticky note:

Solve the problem. Don't solve the data.

What is text (and multimedia) mining?

Text and multimedia mining is the automatic extraction of knowledge from text and multimedia data.

Where we understand knowledge as "reusable information" that potentially applies across several contexts. Knowledge provides answers to open questions.

Quite a broad and loose definition in my opinion, but this is what we work with within this lecture.

Multimedia indicates {image, audio, video}.

Examples:

Q: In which patterns do wildfires occur? A: Mine Flickr pictures of wildfires.

TMM Data:

  1. Reflects its creators
  2. Reflects the world
    • But not necessarily directly; reflects people's understanding and experience
  3. Can reflect how people use language.

Text mining principles

  • Respect the limitations of ethics.
    • Do not harm the people who wrote the text or the people about which the text is written.
  • State the limitations of the approach
    • Text mining never produces complete knowledge.

E.g. 1) is analyzing twitter data on gender, religious preferences etc. ok? And 2) what can you really conclude? (And would not realizing the limitations be harmful?)

Conceptual Framework

How does TMM relate to:

  1. Information Retrieval
  2. NLP
  3. Machine learning tasks like classification

Text mining vs. information retrieval

What is the difference between text mining and information retrieval?

Information retrieval is about efficiently and effectively accessing information.

Text mining is about knowledge acquisition.

They are heavily interrelated. You can first do IR, and then TM or the other way around.

REVIEW: WENT TOO FAST

  1. IR precedes TM

  2. TM precedes IR

Text mining vs. machine learning

Statement in lecture: Machine learning carries out tasks and text mining is used to extract knowledge.

I think this distinction may be too black and white.

Isayev 2019 equates text mining with unsupervised machine learning. The distinction of TM is that there the emphasis is on knowledge acquisition, even though machine learning is used as a tool.

(I didn't think personally Isayev equates TM with machine learning; it's a too strong characterization)

Text mining and data mining

Went too fast, again.

TODO review.

Validating a text mining approach

The goal is knowledge, specifically knowledge . So how do we validate we indeed extracted (reliable) knowledge

Internal validity:

  • Correctness of approach
  • No hidden or unjustified assumptions (such as confounding variables)
  • No implementation mistakes and bugs

External validity:

  • TM project answers a research question
  • Data collection is appropriate and data is representative
  • The output is confirmed by experiments, observations, or human judgments.

Without analysis of external validation your approach is worthless. Validity is key.

Characteristics of textual data

Text is different than other forms of data that you looked at, because it is:

  • Unstructured/semi-structured
  • Mixed source (multi-style; multi-lingual)
  • Noisy

Besides that, texts inherit the properties of language:

  • Language is infinite
  • Interconnected; meaning of words depends on textual elements that are not necessarily "local"
  • Requires contextual interpretation

Heaps' Law :

Any new document in your collection is very likely to add more words to your vocabulary. There is no upper bound to the size of the vocabulary. However, the larger your corpus, the slower the vocabulary increases. Vocabulary growth "levels".

Zipf's Law

The frequency of any word is inversely proportional to its rank in the frequency table. In the English language, about 10-15% of all word occurrences are accounted for by the top four most frequent words. The top 50 words account for about 35-40% of all word occurrences.