Orange3 Text Mining Documentation

Biolab

Nov 27, 2019

 Contents

1 Widgets 1

2 Scripting 47

3 Indices and tables 49

i
ii
 CHAPTER 1

Widgets

1.1 Corpus

Load a corpus of text documents, (optionally) tagged with categories.

Inputs
• None
Outputs
• Corpus: A collection of documents.
Corpus widget reads text corpora from files and sends a corpus instance to its output channel. History of the most
recently opened files is maintained in the widget. The widget also includes a directory with sample corpora that come
pre-installed with the add-on.
The widget reads data from Excel (.xlsx), comma-separated (.csv) and native tab-delimited (.tab) files.

1
Orange3 Text Mining Documentation

1. Browse through previously opened data files, or load any of the sample ones.
2. Browse for a data file.
3. Reloads currently selected data file.
4. Information on the loaded data set.
5. Features that will be used in text analysis.
6. Features that won’t be used in text analysis and serve as labels or class.
You can drag and drop features between the two boxes and also change the order in which they appear.

1.1.1 Example

The first example shows a very simple use of Corpus widget. Place Corpus onto canvas and connect it to Corpus
Viewer. We’ve used book-excerpts.tab data set, which comes with the add-on, and inspected it in Corpus Viewer.

2 Chapter 1. Widgets
 Orange3 Text Mining Documentation

The second example demonstrates how to quickly visualize your corpus with Word Cloud. We could connect Word
Cloud directly to Corpus, but instead we decided to apply some preprocessing with Preprocess Text. We are again
working with book-excerpts.tab. We’ve put all text to lowercase, tokenized (split) the text to words only, filtered out
English stopwords and selected a 100 most frequent tokens.

1.1. Corpus 3
Orange3 Text Mining Documentation

1.2 Import Documents

Import text documents from folders.

Inputs
• None
Outputs
• Corpus: A collection of documents from the local machine.
Import Documents widget retrieves text files from folders and creates a corpus. The widget reads .txt, .docx, .odt,
.pdf and .xml files. If a folder contains subfolders, they will be used as class labels.

1. Folder being loaded.

2. Load folder from a local machine.
3. Reload the data.
4. Number of documents retrieved.
If the widget cannot read the file for some reason, the file will be skipped. Files that were successfully retrieved will
still be on the output.

1.2.1 Example

To retrieve the data, select the folder icon on the right side of the widget. Select the folder you wish to turn into corpus.
Once the loading is finished, you will see how many documents the widget retrieved. To inspect them, connect the
widget to Corpus Viewer. We’ve used a set of Kennedy’s speeches in a plain text format.

4 Chapter 1. Widgets
 Orange3 Text Mining Documentation

Now let us try it with subfolders. We have placed Kennedy’s speeches in two folders - pre-1962 and post-1962. If
I load the parent folder, these two subfolders will be used as class labels. Check the output of the widget in a Data
Table.

1.3 The Guardian

Fetching data from The Guardian Open Platform.

Inputs
• None
Outputs
• Corpus: A collection of documents from the Guardian newspaper.

1.3. The Guardian 5

Orange3 Text Mining Documentation

Guardian retrieves articles from the Guardian newspaper via their API. For the widget to work, you need to provide
the API key, which you can get at their access platform.

1. Insert the API key for the widget to work.

2. Provide the query and set the time frame from which to retrieve the articles.
3. Define which features to retrieve from the Guardian platform.
4. Information on the output.
5. Press Search to start retrieving the articles or Stop to stop the retrieval.

1.3.1 Example

Guardian can be used just like any other data retrieval widget in Orange, namely NY Times, Wikipedia, Twitter or
PubMed.
We will retrieve 240 articles mentioning slovenia between september 2017 and september 2018. The text will include
article headline and content. Upon pressing Search, the articles will be retrieved.
We can observe the results in the Corpus Viewer widget.

6 Chapter 1. Widgets
 Orange3 Text Mining Documentation

1.4 NY Times

Loads data from the New York Times’ Article Search API.
Inputs
• None
Outputs
• Corpus: A collection of documents from the New York Times newspaper.
NYTimes widget loads data from New York Times’ Article Search API. You can query NYTimes articles from
September 18, 1851 to today, but the API limit is set to allow retrieving only a 1000 documents per query. Define
which features to use for text mining, Headline and Abstract being selected by default.
To use the widget, you must enter your own API key.

1.4. NY Times 7
Orange3 Text Mining Documentation

1. To begin your query, insert NY Times’ Article Search API key. The key is securely saved in your system keyring
service (like Credential Vault, Keychain, KWallet, etc.) and won’t be deleted when clearing widget settings.

2. Set query parameters:

• Query
• Query time frame. The widget allows querying articles from September 18, 1851 onwards. Default is set
to 1 year back from the current date.
3. Define which features to include as text features.
4. Information on the output.
5. Produce report.
6. Run or stop the query.

8 Chapter 1. Widgets
 Orange3 Text Mining Documentation

1.4.1 Example

NYTimes is a data retrieving widget, similar to Twitter and Wikipedia. As it can retrieve geolocations, that is geo-
graphical locations the article mentions, it is great in combination with Document Map widget.

First, let’s query NYTimes for all articles on Slovenia. We can retrieve the articles found and view the results in Corpus
Viewer. The widget displays all the retrieved features, but includes on selected features as text mining features.
Now, let’s inspect the distribution of geolocations from the articles mentioning Slovenia. We can do this with Docu-
ment Map. Unsurprisingly, Croatia and Hungary appear the most often in articles on Slovenia (discounting Slovenia
itself), with the rest of Europe being mentioned very often as well.

1.5 Pubmed

Fetch data from PubMed journals.

Inputs
• None
Outputs
• Corpus: A collection of documents from the PubMed online service.
PubMed comprises more than 26 million citations for biomedical literature from MEDLINE, life science journals,
and online books. The widget allows you to query and retrieve these entries. You can use regular search or construct

1.5. Pubmed 9
Orange3 Text Mining Documentation

advanced queries.

1. Enter a valid e-mail to retrieve queries.

2. Regular search:
• Author: queries entries from a specific author. Leave empty to query by all authors.
• From: define the time frame of publication.
• Query: enter the query. Advanced search: enables you to construct complex queries. See PubMed’s
website to learn how to construct such queries. You can also copy-paste constructed queries from the
website.
3. Find records finds available data from PubMed matching the query. Number of records found will be displayed
above the button.
4. Define the output. All checked features will be on the output of the widget.
5. Set the number of record you wish to retrieve. Press Retrieve records to get results of your query on the output.

10 Chapter 1. Widgets
 Orange3 Text Mining Documentation

Below the button is an information on the number of records on the output.

1.5.1 Example

PubMed can be used just like any other data widget. In this example we’ve queried the database for records on
orchids. We retrieved 1000 records and kept only ‘abstract’ in our meta features to limit the construction of tokens
only to this feature.

We used Preprocess Text to remove stopword and words shorter than 3 characters (regexp \b\w{1,2}\b). This will
perhaps get rid of some important words denoting chemicals, so we need to be careful with what we filter out. For the
sake of quick inspection we only retained longer words, which are displayed by frequency in Word Cloud.

1.6 Twitter

Fetching data from The Twitter Search API.

Inputs
• None
Outputs
• Corpus: A collection of tweets from the Twitter API.
Twitter widget enables querying tweets through Twitter API. You can query by content, author or both and accumulate
results should you wish to create a larger data set. The widget only supports REST API and allows queries for up to
two weeks back.

1.6. Twitter 11
Orange3 Text Mining Documentation

1. To begin your queries, insert Twitter key and secret. They are securely saved in your system keyring service
(like Credential Vault, Keychain, KWallet, etc.) and won’t be deleted when clearing widget settings. You must
first create a Twitter app to get API keys.

2. Set query parameters:

• Query word list: list desired queries, one per line. Queries are automatically joined by OR.
• Search by: specify whether you want to search by content, author or both. If searching by author, you must
enter proper Twitter handle (without @) in the query list.
• Language: set the language of retrieved tweets. Any will retrieve tweets in any language.
• Max tweets: set the top limit of retrieved tweets. If box is not ticked, no upper bound will be set - widget
will retrieve all available tweets.
• Allow retweets: if ‘Allow retweets’ is checked, retweeted tweets will also appear on the output. This might
duplicate some results.
• Collect results: if ‘Collect results’ is ticked, widget will append new queries to the previous ones. Enter
new queries, run Search and new results will be appended to the previous ones.
3. Define which features to include as text features.
4. Information on the number of tweets on the output.
5. Run query.

12 Chapter 1. Widgets
 Orange3 Text Mining Documentation

1.6.1 Examples

First, let’s try a simple query. We will search for tweets containing either ‘data mining’ or ‘machine learning’ in the
content and allow retweets. We will further limit our search to only a 100 tweets in English.

First, we’re checking the output in Corpus Viewer to get the initial idea about our results. Then we’re preprocessing
the tweets with lowercase, url removal, tweet tokenizer and removal of stopword and punctuation. The best way to see
the results is with Word Cloud. This will display the most popular words in field of data mining and machine learning
in the past two weeks.
Our next example is a bit more complex. We’re querying tweets from Hillary Clinton and Donald Trump from the
presidential campaign 2016.

1.6. Twitter 13
Orange3 Text Mining Documentation

Then we’ve used Preprocess Text to get suitable tokens on our output. We’ve connected Preprocess Text to Bag of
Words in order to create a table with words as features and their counts as values. A quick check in Word Cloud gives
us an idea about the results.
Now we would like to predict the author of the tweet. With Select Columns we’re setting ‘Author’ as our target
variable. Then we connect Select Columns to Test & Score. We’ll be using Logistic Regression as our learner,
which we also connect to Test & Score.
We can observe the results of our author predictions directly in the widget. AUC score is quite ok. Seems like we can
to some extent predict who is the author of the tweet based on the tweet content.

1.7 Wikipedia

Fetching data from MediaWiki RESTful web service API.

Inputs
• None
Outputs
• Corpus: A collection of documents from the Wikipedia.
Wikipedia widget is used to retrieve texts from Wikipedia API and it is useful mostly for teaching and demonstration.

14 Chapter 1. Widgets
 Orange3 Text Mining Documentation

1. Query parameters:
• Query word list, where each query is listed in a new line.
• Language of the query. English is set by default.
• Number of articles to retrieve per query (range 1-25). Please note that querying is done recursively and
that disambiguations are also retrieved, sometimes resulting in a larger number of queries than set on the
slider.
2. Select which features to include as text features.
3. Information on the output.
4. Produce a report.

1.7. Wikipedia 15
Orange3 Text Mining Documentation

5. Run query.

1.7.1 Example

This is a simple example, where we use Wikipedia and retrieve the articles on ‘Slovenia’ and ‘Germany’. Then we
simply apply default preprocessing with Preprocess Text and observe the most frequent words in those articles with
Word Cloud.

Wikipedia works just like any other corpus widget (NY Times, Twitter) and can be used accordingly.

1.8 Preprocess Text

Preprocesses corpus with selected methods.

Inputs
• Corpus: A collection of documents.
Outputs
• Corpus: Preprocessed corpus.
Preprocess Text splits your text into smaller units (tokens), filters them, runs normalization (stemming, lemmatiza-
tion), creates n-grams and tags tokens with part-of-speech labels. Steps in the analysis are applied sequentially and
can be turned on or off.

16 Chapter 1. Widgets
 Orange3 Text Mining Documentation

1. Information on preprocessed data. Document count reports on the number of documents on the input. Total
tokens counts all the tokens in corpus. Unique tokens excludes duplicate tokens and reports only on unique
tokens in the corpus.
2. Transformation transforms input data. It applies lowercase transformation by default.
• Lowercase will turn all text to lowercase.
• Remove accents will remove all diacritics/accents in text. naïve → naive
• Parse html will detect html tags and parse out text only. <a href. . . >Some text</a> → Some text
• Remove urls will remove urls from text. This is a http://orange.biolab.si/ url. → This is a url.
3. Tokenization is the method of breaking the text into smaller components (words, sentences, bigrams).
• Word & Punctuation will split the text by words and keep punctuation symbols. This example. → (This),
(example), (.)
• Whitespace will split the text by whitespace only. This example. → (This), (example.)

1.8. Preprocess Text 17

Orange3 Text Mining Documentation

• Sentence will split the text by full stop, retaining only full sentences. This example. Another example. →
(This example.), (Another example.)
• Regexp will split the text by provided regex. It splits by words only by default (omits punctuation).
• Tweet will split the text by pre-trained Twitter model, which keeps hashtags, emoticons and other special
symbols. This example. :-) #simple → (This), (example), (.), (:-)), (#simple)
4. Normalization applies stemming and lemmatization to words. (I’ve always loved cats. → I have alway love
cat.) For languages other than English use Snowball Stemmer (offers languages available in its NLTK imple-
mentation).
• Porter Stemmer applies the original Porter stemmer.
• Snowball Stemmer applies an improved version of Porter stemmer (Porter2). Set the language for normal-
ization, default is English.
• WordNet Lemmatizer applies a networks of cognitive synonyms to tokens based on a large lexical database
of English.
5. Filtering removes or keeps a selection of words.
• Stopwords removes stopwords from text (e.g. removes ‘and’, ‘or’, ‘in’. . . ). Se-
lect the language to filter by, English is set as default. You can also load your
own list of stopwords provided in a simple *.txt file with one stopword per line.

Click ‘browse’ icon to select the file containing stopwords. If the file was properly loaded, its name will
be displayed next to pre-loaded stopwords. Change ‘English’ to ‘None’ if you wish to filter out only the
provided stopwords. Click ‘reload’ icon to reload the list of stopwords.
• Lexicon keeps only words provided in the file. Load a *.txt file with one word per line to use as lexicon.
Click ‘reload’ icon to reload the lexicon.
• Regexp removes words that match the regular expression. Default is set to remove punctuation.
• Document frequency keeps tokens that appear in not less than and not more than the specified number /
percentage of documents. If you provide integers as parameters, it keeps only tokens that appear in the
specified number of documents. E.g. DF = (3, 5) keeps only tokens that appear in 3 or more and 5 or less
documents. If you provide floats as parameters, it keeps only tokens that appear in the specified percentage
of documents. E.g. DF = (0.3, 0.5) keeps only tokens that appear in 30% to 50% of documents. Default
returns all tokens.
• Most frequent tokens keeps only the specified number of most frequent tokens. Default is a 100 most
frequent tokens.
6. N-grams Range creates n-grams from tokens. Numbers specify the range of n-grams. Default returns one-
grams and two-grams.
7. POS Tagger runs part-of-speech tagging on tokens.
• Averaged Perceptron Tagger runs POS tagging with Matthew Honnibal’s averaged perceptron tagger.
• Treebank POS Tagger (MaxEnt) runs POS tagging with a trained Penn Treebank model.
• Stanford POS Tagger runs a log-linear part-of-speech tagger designed by Toutanova et al. Please download
it from the provided website and load it in Orange. You have to load the language-specific model in Model
and load stanford-postagger.jar in the Tagger section.
8. Produce a report.
9. If Commit Automatically is on, changes are communicated automatically. Alternatively press Commit.

18 Chapter 1. Widgets
 Orange3 Text Mining Documentation

Note! Preprocess Text applies preprocessing steps in the order they are listed. This means it will first transform the
text, then apply tokenization, POS tags, normalization, filtering and finally constructs n-grams based on given tokens.
This is especially important for WordNet Lemmatizer since it requires POS tags for proper normalization.

1.8.1 Useful Regular Expressions

Here are some useful regular expressions for quick filtering:

\bword\b: matches exact word \w+: matches only words, no punctuation \b(B|b)\w+\b: matches words be-
ginning with the letter b \w{4,}: matches words that are longer than 4 characters\b\w+(Y|y)\b: matches words
ending with the letter y

1.8.2 Examples

In the first example we will observe the effects of preprocessing on our text. We are working with book-excerpts.tab
that we’ve loaded with Corpus widget. We have connected Preprocess Text to Corpus and retained default pre-
processing methods (lowercase, per-word tokenization and stopword removal). The only additional parameter we’ve
added as outputting only the first 100 most frequent tokens. Then we connected Preprocess Text with Word Cloud to
observe words that are the most frequent in our text. Play around with different parameters, to see how they transform
the output.

The second example is slightly more complex. We first acquired our data with Twitter widget. We quired the internet
for tweets from users @HillaryClinton and @realDonaldTrump and got their tweets from the past two weeks, 242 in
total.

1.8. Preprocess Text 19

Orange3 Text Mining Documentation

In Preprocess Text there’s Tweet tokenization available, which retains hashtags, emojis, mentions and so on. However,
this tokenizer doesn’t get rid of punctuation, thus we expanded the Regexp filtering with symbols that we wanted to
get rid of. We ended up with word-only tokens, which we displayed in Word Cloud. Then we created a schema for
predicting author based on tweet content, which is explained in more details in the documentation for Twitter widget.

1.9 Bag of Words

Generates a bag of words from the input corpus.

Inputs
• Corpus: A collection of documents.
Outputs
• Corpus: Corpus with bag of words features appended.
Bag of Words model creates a corpus with word counts for each data instance (document). The count can be either
absolute, binary (contains or does not contain) or sublinear (logarithm of the term frequency). Bag of words model is
required in combination with Word Enrichment and could be used for predictive modelling.

20 Chapter 1. Widgets
 Orange3 Text Mining Documentation

1. Parameters for bag of words model:

• Term Frequency:
– Count: number of occurrences of a word in a document
– Binary: word appears or does not appear in the document
– Sublinear: logarithm of term frequency (count)
• Document Frequency:
– (None)
– IDF: inverse document frequency
– Smooth IDF: adds one to document frequencies to prevent zero division.
• Regulariation:
– (None)
– L1 (Sum of elements): normalizes vector length to sum of elements
– L2 (Euclidean): normalizes vector length to sum of squares
2. Produce a report.
3. If Commit Automatically is on, changes are communicated automatically. Alternatively press Commit.

1.9.1 Example

In the first example we will simply check how the bag of words model looks like. Load book-excerpts.tab with Corpus
widget and connect it to Bag of Words. Here we kept the defaults - a simple count of term frequencies. Check what
the Bag of Words outputs with Data Table. The final column in white represents term frequencies for each document.

1.9. Bag of Words 21

Orange3 Text Mining Documentation

In the second example we will try to predict document category. We are still using the book-excerpts.tab data set,
which we sent through Preprocess Text with default parameters. Then we connected Preprocess Text to Bag of
Words to obtain term frequencies by which we will compute the model.

22 Chapter 1. Widgets
 Orange3 Text Mining Documentation

Connect Bag of Words to Test & Score for predictive modelling. Connect SVM or any other classifier to Test &
Score as well (both on the left side). Test & Score will now compute performance scores for each learner on the
input. Here we got quite impressive results with SVM. Now we can check, where the model made a mistake.
Add Confusion Matrix to Test & Score. Confusion matrix displays correctly and incorrectly classified documents.
Select Misclassified will output misclassified documents, which we can further inspect with Corpus Viewer.

1.10 Similarity Hashing

Computes documents hashes.

Inputs
• Corpus: A collection of documents.
Outputs
• Corpus: Corpus with simhash value as attributes.
Similarity Hashing is a widget that transforms documents into similarity vectors. The widget uses SimHash method
from from Moses Charikar.

1.10. Similarity Hashing 23

Orange3 Text Mining Documentation

1. Set Simhash size (how many attributes will be on the output, corresponds to bits of information) and shingle
length (how many tokens are used in a shingle).
2. Commit Automatically output the data automatically. Alternatively, press Commit.

1.10.1 Example

We will use deerwester.tab to find similar documents in this small corpus. Load the data with Corpus and pass it to
Similarity Hashing. We will keep the default hash size and shingle length. We can observe what the widget outputs
in a Data Table. There are 64 new attributes available, corresponding to the Simhash size parameter.

1.10.2 References

Charikar, M. (2002) Similarity estimation techniques from rounding algorithms. STOC ‘02 Proceedings of the thirty-
fourth annual ACM symposium on Theory of computing, p. 380-388.

1.11 Sentiment Analysis

Predict sentiment from text.

Inputs
• Corpus: A collection of documents.

24 Chapter 1. Widgets
 Orange3 Text Mining Documentation

Outputs
• Corpus: A corpus with information on the sentiment of each document.
Sentiment Analysis predicts sentiment for each document in a corpus. It uses Liu Hu and Vader sentiment modules
from NLTK. Both of them are lexicon-based. For Liu Hu, you can choose English or Slovenian version.

1. Method:
• Liu Hu: lexicon-based sentiment analysis (supports English and Slovenian)
• Vader: lexicon- and rule-based sentiment analysis
2. Produce a report.
3. If Auto commit is on, sentiment-tagged corpus is communicated automatically. Alternatively press Commit.

1.11.1 Example

Sentiment Analysis can be used for constructing additional features with sentiment prediction from corpus. First, we
load Election-2016-tweets.tab in Corpus. Then we connect Corpus to Sentiment Analysis. The widget will append
4 new features for Vader method: positive score, negative score, neutral score and compound (combined score).
We can observe new features in a Data Table, where we sorted the compound by score. Compound represents the
total sentiment of a tweet, where -1 is the most negative and 1 the most positive.

Now let us visualize the data. We have some features we are currently not interested in, so we will remove them with
Select Columns.

1.11. Sentiment Analysis 25

Orange3 Text Mining Documentation

Then we will make our corpus a little smaller, so it will be easier to visualize. Pass the data to Data Sampler and
retain a random 10% of the tweets.

Now pass the filtered corpus to Heat Map. Use Merge by k-means to merge tweets with the same polarity into one
line. Then use Cluster by rows to create a clustered visualization where similar tweets are grouped together. Click on
a cluster to select a group of tweets - we selected the negative cluster.

26 Chapter 1. Widgets
 Orange3 Text Mining Documentation

To observe the selected subset, pass the tweets to Corpus Viewer.

1.11. Sentiment Analysis 27

Orange3 Text Mining Documentation

1.11.2 References

Hutto, C.J. and E. E. Gilbert (2014). VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social
Media Text. Eighth International Conference on Weblogs and Social Media (ICWSM-14). Ann Arbor, MI, June 2014.
Hu, Minqing and Bing Liu (2004). Mining opinion features in customer reviews. In Proceedings of AAAI Conference

28 Chapter 1. Widgets
 Orange3 Text Mining Documentation

on Artificial Intelligence, vol. 4, pp. 755–760. Available online.

Kadunc, Klemen and Marko Robnik-Šikonja (2016). Analiza mnenj s pomočjo strojnega učenja in slovenskega lek-
sikona sentimenta. Conference on Language Technologies & Digital Humanities, Ljubljana (in Slovene). Available
online.

1.12 Tweet Profiler

Detect Ekman’s, Plutchik’s or Profile of Mood States’ emotions in tweets.

Inputs
• Corpus: A collection of tweets (or other documents).
Outputs
• Corpus: A corpus with information on the sentiment of each document.
Tweet Profiler retrieves information on sentiment from the server for each given tweet (or document). The widget
sends data to the server, where a model computes emotion probabilities and/or scores. The widget support three
classifications of emotion, namely Ekman’s, Plutchik’s and Profile of Mood States (POMS).

1. Options:
• Attribute to use as content.
• Emotion classification, either Ekman’s, Plutchik’s or Profile of Mood States. Multi-class will output one
most probable emotion per document, while multi-label will output values in columns per each emotion.
• The widget can output classes of emotion (categorical), probabilities (numeric), or embeddings (an emo-
tional vector of the document).
2. Commit Automatically automatically outputs the result. Alternatively, press Commit.

1.12.1 Example

We will use election-tweets-2016.tab for this example. Load the data with Corpus and connect it to Tweet Profiler.
We will use Content attribute for the analysis, Ekman’s classification of emotion with multi-class option and we will
output the result as class. We will observe the results in a Box Plot. In the widget, we have selected to observe the
Emotion variable, grouped by Author. This way we can see which emotion prevails by which author.

1.12. Tweet Profiler 29

Orange3 Text Mining Documentation

1.12.2 References

Colnerič, Niko and Janez Demšar (2018). Emotion Recognition on Twitter: Comparative Study and Training a Unison
Model. In IEEE Transactions on Affective Computing. Available online.

1.13 Topic Modelling

Topic modelling with Latent Dirichlet Allocation, Latent Semantic Indexing or Hierarchical Dirichlet Process.
Inputs
• Corpus: A collection of documents.
Outputs
• Corpus: Corpus with topic weights appended.
• Topics: Selected topics with word weights.
• All Topics: Topic weights by tokens.
Topic Modelling discovers abstract topics in a corpus based on clusters of words found in each document and their
respective frequency. A document typically contains multiple topics in different proportions, thus the widget also
reports on the topic weight per document.

30 Chapter 1. Widgets
 Orange3 Text Mining Documentation

1. Topic modelling algorithm:

• Latent Semantic Indexing
• Latent Dirichlet Allocation
• Hierarchical Dirichlet Process
2. Parameters for the algorithm. LSI and LDA accept only the number of topics modelled, with the default set to
10. HDP, however, has more parameters. As this algorithm is computationally very demanding, we recommend
you to try it on a subset or set all the required parameters in advance and only then run the algorithm (connect
the input to the widget).
• First level concentration (𝛾): distribution at the first (corpus) level of Dirichlet Process
• Second level concentration (𝛼): distribution at the second (document) level of Dirichlet Process
• The topic Dirichlet (𝛼): concentration parameter used for the topic draws
• Top level truncation (T): corpus-level truncation (no of topics)
• Second level truncation (K): document-level truncation (no of topics)
• Learning rate (𝜅): step size
• Slow down parameter (𝜏 )
3. Produce a report.
4. If Commit Automatically is on, changes are communicated automatically. Alternatively press Commit.

1.13.1 Example

In the first example, we present a simple use of the Topic Modelling widget. First we load grimm-tales-selected.tab
data set and use Preprocess Text to tokenize by words only and remove stopwords. Then we connect Preprocess Text
to Topic Modelling, where we use a simple Latent Semantic Indexing to find 10 topics in the text.

1.13. Topic Modelling 31

Orange3 Text Mining Documentation

LSI provides both positive and negative weights per topic. A positive weight means the word is highly representative
of a topic, while a negative weight means the word is highly unrepresentative of a topic (the less it occurs in a text, the
more likely the topic). Positive words are colored green and negative words are colored red.
We then select the first topic and display the most frequent words in the topic in Word Cloud. We also connected
Preprocess Text to Word Cloud in order to be able to output selected documents. Now we can select a specific word
in the word cloud, say little. It will be colored red and also highlighted in the word list on the left.
Now we can observe all the documents containing the word little in Corpus Viewer.
In the second example, we will look at the correlation between topics and words/documents. Connect Topic Modelling
to Heat Map. Ensure the link is set to All Topics - Data. Topic Modelling will output a matrix of topic weights by
words from text (more precisely, tokens).
We can observe the output in a Data Table. Tokens are in rows and retrieved topics in columns. Values represent how
much a word is represented in a topic.

32 Chapter 1. Widgets
 Orange3 Text Mining Documentation

To visualize this matrix, open Heat Map. Select Merge by k-means and Cluster - Rows to merge similar rows into one
and sort them by similarity, which makes the visualization more compact.
In the upper part of the visualization, we have words that highly define topics 1-3 and in the lower part those that
define topics 5 and 10.
We can similarly observe topic representation across documents. We connect another Heat Map to Topic Modelling
and set link to Corpus - Data. We set Merge and Cluster as above.
In this visualization we see how much is a topic represented in a document. Looks like Topic 1 is represented almost
across the entire corpus, while other topics are more specific. To observe a specific set of document, select either a
clustering node or a row in the visualization. Then pass the data to Corpus Viewer.

1.14 Corpus Viewer

Displays corpus content.

Inputs
• Corpus: A collection of documents.
Outputs
• Corpus: Documents containing the queried word.
Corpus Viewer is meant for viewing text files (instances of Corpus). It will always output an instance of corpus. If
RegExp filtering is used, the widget will output only matching documents.

1.14. Corpus Viewer 33

Orange3 Text Mining Documentation

1. Information:
• Documents: number of documents on the input
• Preprocessed: if preprocessor is used, the result is True, else False. Reports also on the number of tokens
and types (unique tokens).
• POS tagged: if POS tags are on the input, the result is True, else False.
• N-grams range: if N-grams are set in Preprocess Text, results are reported, default is 1-1 (one-grams).
• Matching: number of documents matching the RegExp Filter. All documents are output by default.
2. RegExp Filter: Python regular expression for filtering documents. By default no documents are filtered (entire
corpus is on the output).
3. Search Features: features by which the RegExp Filter is filtering. Use Ctrl (Cmd) to select multiple features.
4. Display Features: features that are displayed in the viewer. Use Ctrl (Cmd) to select multiple features.
5. Show Tokens & Tags: if tokens and POS tag are present on the input, you can check this box to display them.
6. If Auto commit is on, changes are communicated automatically. Alternatively press Commit.

1.14.1 Example

Corpus Viewer can be used for displaying all or some documents in corpus. In this example, we will first load book-
excerpts.tab, that already comes with the add-on, into Corpus widget. Then we will preprocess the text into words,
filter out the stopwords, create bi-grams and add POS tags (more on preprocessing in Preprocess Text. Now we want
to see the results of preprocessing. In Corpus Viewer we can see, how many unique tokens we got and what they
are (tick Show Tokens & Tags). Since we used also POS tagger to show part-of-speech labels, they will be displayed
alongside tokens underneath the text.

34 Chapter 1. Widgets
 Orange3 Text Mining Documentation

Now we will filter out just the documents talking about a character Bill. We use regular expression \bBill\b to find the
documents containing only the word Bill. You can output matching or non-matching documents, view them in another
Corpus Viewer or further analyse them.

1.15 Word Cloud

Generates a word cloud from corpus.

Inputs
• Topic: Selected topic.
• Corpus: A collection of documents.
Outputs
• Corpus: Documents that match the selection.
• Word: Selected word that can be used as query in Concordance.
Word Cloud displays tokens in the corpus, their size denoting the frequency of the word in corpus. Words are listed
by their frequency (weight) in the widget. The widget outputs documents, containing selected tokens from the word
cloud.

1.15. Word Cloud 35

Orange3 Text Mining Documentation

1. Information on the input.

• number of words (tokens) in a topic
• number of documents and tokens in the corpus
2. Adjust the plot.
• If Color words is ticked, words will be assigned a random color. If unchecked, the words will be black.
• Word tilt adjust the tilt of words. The current state of tilt is displayed next to the slider (‘no’ is the default).
• Regenerate word cloud plot the cloud anew.
3. Words & weights displays a sorted list of words (tokens) by their frequency in the corpus or topic. Clicking on a
word will select that same word in the cloud and output matching documents. Use Ctrl to select more than one
word. Documents matching ANY of the selected words will be on the output (logical OR).
4. Save Image saves the image to your computer in a .svg or .png format.

1.15.1 Example

Word Cloud is an excellent widget for displaying the current state of the corpus and for monitoring the effects of
preprocessing.
Use Corpus to load the data. Connect Preprocess Text to it and set your parameters. We’ve used defaults here, just to
see the difference between the default preprocessing in the Word Cloud widget and the Preprocess Text widget.

36 Chapter 1. Widgets
 Orange3 Text Mining Documentation

We can see from the two widgets, that Preprocess Text displays only words, while default preprocessing in the Word
Cloud tokenizes by word and punctuation.

1.16 Concordance

Display the context of the word.

Inputs
• Corpus: A collection of documents.
Outputs
• Selected Documents: Documents containing the queried word.
• Concordances: A table of concordances.
Concordance finds the queried word in a text and displays the context in which this word is used. Results in a single
color come from the same document. The widget can output selected documents for further analysis or a table of
concordances for the queried word. Note that the widget finds only exact matches of a word, which means that if you
query the word ‘do’, the word ‘doctor’ won’t appear in the results.

1.16. Concordance 37
Orange3 Text Mining Documentation

1. Information:
• Documents: number of documents on the input.
• Tokens: number of tokens on the input.
• Types: number of unique tokens on the input.
• Matching: number of documents containing the queried word.
2. Number of words: the number of words displayed on each side of the queried word.
3. Queried word.
4. If Auto commit is on, selected documents are communicated automatically. Alternatively press Commit.

1.16.1 Examples

Concordance can be used for displaying word contexts in a corpus. First, we load book-excerpts.tab in Corpus.
Then we connect Corpus to Concordance and search for concordances of a word ‘doctor’. The widget displays all
documents containing the word ‘doctor’ together with their surrounding (contextual) words.
Now we can select those documents that contain interesting contexts and output them to Corpus Viewer to inspect
them further.

38 Chapter 1. Widgets
 Orange3 Text Mining Documentation

In the second example, we will output concordances instead. We will keep the book-excerpts.tab in Corpus and the
connection to Concordance. Our queried word remains ‘doctor’.
This time, we will connect Data Table to Concordance and select Concordances output instead. In the Data Table,
we get a list of concordances for the queried word and the corresponding documents. Now, we will save this table
with Save Data widget, so we can use it in other projects or for further analysis.

1.17 Document Map

Displays geographic locations mentioned in the text.

Inputs
• Data: Data set.
Outputs

1.17. Document Map 39

Orange3 Text Mining Documentation

• Corpus: Documents containing mentions of selected geographical regions.

Document Map widget shows geolocations from textual (string) data. It finds mentions of geographic names (coun-
tries and capitals) and displays distributions (frequency of mentions) of these names on a map. It works with any
Orange widget that outputs a data table and that contains at least one string attribute. The widget outputs selected data
instances, that is all documents containing mentions of a selected country (or countries).

1. Select the meta attribute you want to search geolocations by. The widget will find all mentions of geolocations
in a text and display distributions on a map.
2. Select the type of map you wish to display. The options are World, Europe and USA. You can zoom in and out
of the map by pressing + and - buttons on a map or by mouse scroll.
3. The legend for the geographic distribution of data. Countries with the boldest color are most often mentioned in
the selected region attribute (highest frequency).
To select documents mentioning a specific country, click on a country and the widget will output matching documents.
To select more than one country hold Ctrl/Cmd upon selection.

1.17.1 Example

Document Map widget can be used for simply visualizing distributions of geolocations or for a more complex in-
teractive data analysis. Here, we’ve queried NY Times for articles on Slovenia for the time period of the last year
(2015-2016). First we checked the results with Corpus Viewer.

40 Chapter 1. Widgets
 Orange3 Text Mining Documentation

Then we sent the data to Document Map to see distributions of geolocations by country attribute. The attribute already
contains country tags for each article, which is why NY Times is great in combinations with Document Map. We
selected Germany, which sends all the documents tagged with Germany to the output. Remember, we queried NY
Times for articles on Slovenia.
We can again inspect the output with Corpus Viewer. But there’s a more interesting way of visualizing the data.
We’ve sent selected documents to Preprocess Text, where we’ve tokenized text to words and removed stopwords.
Finally, we can inspect the top words appearing in last year’s documents on Slovenia and mentioning also Germany
with Word Cloud.

1.18 Word Enrichment

Word enrichment analysis for selected documents.

Inputs
• Corpus: A collection of documents.
• Selected Data: Selected instances from corpus.
Outputs
• None
Word Enrichment displays a list of words with lower p-values (higher significance) for a selected subset compared to
the entire corpus. Lower p-value indicates a higher likelihood that the word is significant for the selected subset (not

1.18. Word Enrichment 41

Orange3 Text Mining Documentation

randomly occurring in a text). FDR (False Discovery Rate) is linked to p-value and reports on the expected percent of
false predictions in the set of predictions, meaning it account for false positives in list of low p-values.

1. Information on the input.

• Cluster words are all the tokens from the corpus.
• Selected words are all the tokens from the selected subset.
• After filtering reports on the enriched words found in the subset.
2. Filter enables you to filter by:
• p-value
• false discovery rate (FDR)

1.18.1 Example

In the example below, we’re retrieved recent tweets from the 2016 presidential candidates, Donald Trump and Hillary
Clinton. Then we’ve preprocessed the tweets to get only words as tokens and to remove the stopwords. We’ve
connected the preprocessed corpus to Bag of Words to get a table with word counts for our corpus.

42 Chapter 1. Widgets
 Orange3 Text Mining Documentation

Then we’ve connected Corpus Viewer to Bag of Words and selected only those tweets that were published by Donald
Trump. See how we marked only the Author as our Search feature to retrieve those tweets.
Word Enrichment accepts two inputs - the entire corpus to serve as a reference and a selected subset from the corpus
to do the enrichment on. First connect Corpus Viewer to Word Enrichment (input Matching Docs → Selected Data)
and then connect Bag of Words to it (input Corpus → Data). In the Word Enrichment widget we can see the list of
words that are more significant for Donald Trump than they are for Hillary Clinton.

1.19 Duplicate Detection

Detect & remove duplicates from a corpus.

Inputs
• Distances: A distance matrix.
Outputs
• Corpus Without Duplicated: Corpus with duplicates removed.
• Duplicates Cluster: Documents belonging to selected cluster.
• Corpus: Corpus with appended cluster labels.
Duplicate Detection uses clustering to find duplicates in the corpus. It is great with the Twitter widget for removing
retweets and other similar documents.

1.19. Duplicate Detection 43

Orange3 Text Mining Documentation

To set the level of similarity, drag the line vertical line left or right in the visualization. The further left the line, the
more similar the documents have to be in order to be considered duplicates. You can also set the threshold manually
in the control area.

1. Information on unique and duplicate documents.

2. Linkage used for clustering (Single, Average, Complete, Weighted and Ward).
3. Distance threshold sets the similarity cutoff. The lower the value, the more similar the data instances have to be
to belong to the same cluster. You can also set the cutoff by dragging the vertical line in the plot.
4. Cluster labels can be appended as attributes, class or metas.
5. List of clusters at the selected threshold. They are sorted by size by default. Click on the cluster to observe its
content on the output.

1.19.1 Example

This simple example uses iris data to find identical data instances. Load iris with the File widget and pass it to
Distances. In Distances, use Euclidean distance for computing the distance matrix. Pass distances to Duplicate
Detection.
It looks like cluster C147 contain three duplicate entries. Let us select it in the widget and observe it in a Data Table.
Remember to set the output to Duplicates Cluster. IThe three data instances are identical. To use the data set without
duplicates, use the first output, Corpus Without Duplicates.
The same procedure can be used also for corpora. Remember to use the Bag of Words between Corpus and Distances.

44 Chapter 1. Widgets
 Orange3 Text Mining Documentation

1.19. Duplicate Detection 45

Orange3 Text Mining Documentation

46 Chapter 1. Widgets
 CHAPTER 2

Scripting

2.1 Corpus

2.2 Preprocessor

2.3 Twitter

2.4 New York Times

2.5 The Guardian

2.6 Wikipedia

2.7 Bag of Words

2.8 Topic Modeling

2.9 Tag

47
Orange3 Text Mining Documentation

48 Chapter 2. Scripting
 CHAPTER 3

Indices and tables

• genindex
• modindex
• search

49