What? Investigating what a corpus is about

Max Kemman
University of Luxembourg
October 25, 2015

Online slides optimised for Full-HD screens in full-screen mode
Download PDF here

Doing Digital History: Introduction to Tools and Technology

Recap from last time

What is distant reading?

What is an n-gram?

What do the Y-axis and X-axis show?

Recap - Assignment

How did the assignment go?

What did you think of the tools used?

Could this be useful for your research?

One more thing on HTML: special characters

http://www.ascii.cl/htmlcodes.htm

Find the symbol and the HTML number

é & ü -> � & �

é & ü -> é & ü

In your HTML, write longue durée to write longue durée

One more thing: what is an algorithm?

A set of rules to follow to solve a problem

Pretty much like a cooking recipe

a = 0 
while(a < 10) { a = a + 1
}

Today

  • The W's of research
  • What a corpus is about
  • The entities in a corpus
  • Another look at our emails
  • Voyant Tools
  • Next time
    • Assignment

The W's of research

Thus far:

  1. Abundance of sources
  2. Writing for the Web
  3. Digitisation and Digital Libraries
  4. Big Data
  5. Distant Reading

Now: we have a digital corpus, what to do with it?

Research the corpus

Now come the W's of research:

  1. What - Investigating what a corpus is about
  2. Where - Investigating the spatial entities in a corpus
  3. When - Investigating the temporal entities in a corpus
  4. Who - Investigating the social entities in a corpus

What?

The first W of interest, what is this corpus actually about?

Different methods are possible

  • Find a description of the corpus to read
  • Select a sample of documents to read
  • Visualize the used words

What a corpus is about

What is this conference about?

Word clouds

Advantages of word clouds

  • Very easy to create
  • Visually pleasing
  • Gives a quick overview

What does a word cloud do?

Put very simply, a word cloud does the following:

  1. Count the number of occurrences per word
  2. Size each word by its frequency
  3. Layout the words to form a shape
  4. Optional: colorize words for distinguishing and better readability

Layout

Unlike the Ngram viewer: no X or Y axes

The position of each word is meaningless

The meaning is in the size of the words

Counting

Word clouds visualize the frequency of words

But how to count words that vary in spelling?

  • E.g. "Digital" and "digital" and "digitally", "digitize" and "digitization"

Normalization:

  • Lowercase
  • Tokenize
  • Stemming or lemmatizing
  • Stopwords

Lowercase

We were on vacation in France in August 2015

we were on vacation in france in august 2015

Tokenize

we were on vacation, in france, in august 2015

we|were|on|vacation|in|france|in|august|2015

Stemming or lemmatizing

digitized|digital|digitization|digitizing

Stemming: digit

Lemmatizing: digitiz|digital

Could be very useful especially with Latin texts

Stopwords

Most common words in the language: and, or, the

Sometimes: remove numbers

Not of interest (usually)

we|were|on|vacation|in|france|in|august|2015

we|were|vacation|france|august|

What are these grants about? (normalized)

Comparing between different parts of the corpus

Sources separated by their citation behaviour

Representing a model of the text

What if we do not know how to separate sources?

Or if we want to know what other words are related to our keywords?

Topic modelling

Documents and words can be directly observed, but topics are latent

How to represent the topics in a corpus?

  • Statistics to find topics represented by groups of words
  • Document is a mix of topics
  • Topic is a mix of words

(Slides on topic modelling from Pim Huijnen and Marijn Koolen)

Topic modelling

Assumption: two documents with the same topics will have overlap in words

For a given corpus, modelling process does:

  1. Create word probability distribution for topics
  2. Create topic probability distribution for documents

Topic modelling

In short: a corpus is represented by statistical topics

This allows us to:

  • Separate sources by topics
  • Find related keywords

Comparing different parts of the corpus

Mendeley Research Maps

Comparing the topical similarity

Assigned documents to disciplines to map disciplines by topics

Which form of machine learning would this be?

What is the corpus about?

We can now represent the words or the topics of a corpus

But, remember: World War I ≠ "World War I"

The entities in a corpus

Thus far we know the frequencies of all the words

But what are we interested in?

What do we need for the other W's?

  • Where - places
  • When - dates
  • Who - people

People in the corpus

Ter Braake & Fokkens - Fairly easy to discover famous people (with biographical dictionaries and Ngram viewers)

Ngrams help top-down: when you know who to search for

But how to discover who did not become famous, while prominent in their own time?

Need to find all people bottom-up by identifying all the names

Bottom-up proces

Ter Braake & Fokkens

  1. Identify all names in the corpus
  2. Give all names an identifier
  3. Disambiguate names referring to the same person
  4. Compare results with a non-digital corpus
  5. Visualize the results
  6. Interpret!

Identifying names

Combinations of words that start with a capital

This won't work for German

Their algorithm allows for two sequential lower case words: Johan van der Capellen

Note: built for recall, not precision

Recall & Precision

Recall: retrieve all relevant entities

Precision: do no retrieve irrelevant entities

For algorithms usually a choice what to optimize

  • Recall of people referred to with single name (Erasmus, Rembrandt) would lead to too much noise = lower precision

Difficulties

Spelling of names (especially before 19th century)

People with the same name

Nicknames and changing names

People with the same title

Context matters!

Named Entity Recognition

We want to identify the entities

We were on vacation in France in August 2015. I went to shop at the Intermarche. The area around Apt is really nice. Max also bought icecream, which cost €2.

We were on vacation in France in August 2015. I went to shop at the Intermarche. The area around Apt is really nice. Max also bought icecream, which cost €2.

Named Entities

Or we want to see:

We were on vacation in France in August 2015. I went to shop at the Intermarche. The area around Apt is really nice. Max also bought icecream, which cost €2.

  • People: Max
  • Places: France, Apt
  • Organizations: Intermarche
  • Dates: August 2015
  • Currencies: €2

Another look at our emails

For all 30k emails, we performed text normalisation and named entity recognition

Let's take a look at https://www.wikileaks.org/clinton-emails/emailid/8

Exercise 1: try to normalise the text

Exercise 2: try to discover the named entities: People, Places, Organisations

Normalised

See Email8-normalised.txt in Moodle under "Emails"

unclassife, us, department, state, case, f--, doc, date, release, full, hrod, clintonemailcom, sent, friday, july, pm, sullivanjj, stategov, subject, re, pakistan, bomb, ok, go, original, message, sullivan, jacob, sullivanjj, stategov, sent, fri, jul, subject, pakistan, bomb, fyi, put, follow, statement, statement, secretary, clinton, bomb, shrine, sy, ali, hujviri, lahore, shock, sadden, yesterday, attack, one, pakistan, popular, place, worship, shrine, sy, ali, hujviri, data, ganjbakhsh, lahore, claime, live, many, innocent, pakistane, extremist, shown, respect, neither, human, dignity, fundamental, religious, value, pakistani, society, violact, sanctity, rever, shrine, particularly, sinister, attempt, destabilize, pakistan, intimidate, people, attacker, will, succeed, pakistani, public, refuse, cow, violence, condemn, brutal, crime, reaffirm, commitment, support, pakistani, people, effort, defend, democracy, violent, extremist, seek, destroy, thought, prayer, family, victim, people, pakistan

Named Entities

Try to do it by hand

NER tool: http://nlp.stanford.edu:8080/ner/

PeoplePlacesOrganisations
Sullivan
Jacob
CLINTON
Ali Hujviri
Pakistan
Pakistan
Lahore
Pakistan
Lahore
Pakistan
Pakistan
U.S. Department of State Case No
Shrine of Syed Ali Hujviri

Visualise the email

Go to http://tagcrowd.com/

Compare with and without stopwords

Compare normal and normalised text

What?

So, what's the email about? Do we get different perspectives?

Voyant Tools

Go to www.voyant-tools.org/

Use Mozilla Firefox, it doesn't work in Chrome (that's what went wrong during lecture)

From Moodle: download the files for emails 6000-6019 f6-20-raw.txt and f6-20-normalised.txt

You can paste in text, or upload the file

Continue by hitting reveal

Saving the Voyant session

It might be a good idea to copy the URL early on, as this will allow you to refresh the page if the tool crashes, or to open the tool again later on using the data and stopwords you already had

Share the session: mousehover the top blue bar, and click the third icon in the topright (see image), you can then choose to share the URL: this will open a new browser window where you can copy the address from

Voyant windows

Look at all the windows in Voyant and see if you understand them

  1. Cirrus (word cloud)
  2. Reader
  3. Summary
  4. Trends
  5. Contexts

Voyant Word Clouds

  • In the Cirrus, hold mouse on the title bar, and click 3rd icon
  • Select the stopword list you need
  • Or Edit List to add more words: 1 word per line, click Save
  • Check apply globally to activate in all windows
  • Use the word cloud to detect common words we're not interest in: unclassified, department, subject, etc
  • Hit Confirm
  • When editing again, the stopwords are ordered alphabetically, so you might not see them at the end anymore

Voyant Summary

What is the longest email?

What are distinctive words?

Distinctive words calculated by TF-IDF: what was that again?

Update: the distinctive words feature doesn't work now that we combined all the emails in a single text-file

Searching specific words

In the Cirrus window, you can click Terms in the top bar to get the list of words ordered by count

You can see immediately per word how it develops over time in the emails

From this list you can select a word by checking the box left to it

Alternatively, you can search for words per window. For example, in the Contexts window (lower-right), at the bottom is a search box where you can search for words

Interpreting with Voyant

What are the biggest words?

How do they develop throughout the emails?

Does this tell what the emails are about and how it goes?

If not: what is different?

Sharing the Voyant

You can either

  • Take screenshots of what you want to show
  • Share the session: mousehover the top blue bar, and click the third icon in the topright (see image), you can then choose to share the URL: this will open a new browser window where you can copy the address from
  • The HTML snippet will give an HTML code that you can embed in your report.
  • Share specific windows: for example, in the top bar of trends, click the first icon (see image), and select to export a url, a HTML snippet for embedding, or a PNG for including in your report

Next time

1 November: No class

8 November

When? Temporal entities and timelines

Assignment

Perform Voyant analysis of HC emails

Compare (see next slide for all the available files):

  • f6-100-raw.txt vs f6-100-normalised.txt to see how text normalisation gives different perspective
  • For further comparisons, choose either the raw or the normalised text:
  • f6-1000-*.txt vs f7-1000-*.txt to see how the emails are different
  • If Voyant or your computer has difficulty with 1000 emails, compare f6-100-*.txt vs f7-100-*.txt

Do comparisons in separate Voyant windows

Download files from Moodle:

EmailsRawNormalised
6000-6099f6-100-raw.txtf6-100-normalised.txt
7000-7099f7-100-raw.txtf7-100-normalised.txt
6000-6999f6-1000-raw.txtf6-1000-normalised.txt
7000-7999f7-1000-raw.txtf7-1000-normalised.txt

Assignment

Work in pairs of two or three

Use the tools discussed today to try and find something you find interesting. Document your steps and choices and discuss why a finding is of interest, and whether you can be certain of this finding.

Hand in the assignment in HTML, include your name and a decent profile photo

500-1000 words, in English

Possible questions you might ask of your corpora

  • What are these emails about?
  • Do we need to further clean the data?
  • How are these corpora different?
  • Does text normalisation lead to different results?

Grading

  • 1pt for free
  • 3pts for HTML
  • 3pts for documentation of your process
  • 3pts for critical reflection on your finding

Do note: the finding itself is not the most important part

Email to max.kemman@uni.lu before the start of the next lecture