Thematic Analysis of Elementary Contexts

N.B.: The pictures shown in this section have been obtained by using a previous version of T-LAB. These pictures look slightly different in T-LAB 10. Also: a) there is a new button (TREE MAP PREVIEW) which allows the user to create dynamic charts in HTML format; b) the DENDROGRAM button has been replaced by the Graph Maker tool; c) a new table that shows in different columns the typical words of each cluster is available; d) when analysing a corpus which includes variable attributes, it is now possible to build and analyse tables which cross the themes and the attributes of each variable; e) a quick access gallery of pictures which works as an additional menu allows one to switch between various outputs with a single click.
Some of these new features are highlighted in the below image.

This T-LAB tool allows you to obtain and explore a representation of corpus contents through few and significant thematic clusters (from 3 to 50), each of which:

a) consists of a set of elementary contexts (i.e. sentences, paragraphs or short texts like responses to open-ended questions) characterized by the same patterns of key-words;

b) is described through the lexical units (i.e. words, lemmas or categories) and the variables (if present) most characteristic of the context units from which it is composed.

In many ways, analysis results can be considered as an isotopy (iso = same; topoi = places) map where each of them, as generic or specific theme (Rastier, 2002: 204), is characterized by the co-occurrences of semantic traits.

The analysis process can be performed through an unsupervised clustering (i.e. bottom-up approach), which is the default option, or a supervised classification (i.e. top-down approach). When choosing the latter (i.e. supervised classification), a dictionary of categories must be imported, either created by means of a previous T-LAB analysis or made up by the user.

A T-LAB dialog box (see above) allows the user to set some analysis parameters.
In particular:
- the (A) parameter allows the user to fix the maximum number of cluster partitions to be included in T-LAB outputs. Nonetheless, the clustering algorithm used stops when any further partition doesn't match statistical criteria;
- the (B) parameter allows the user to exclude from the analysis any context unit that doesn't contain a minimum number of key-words included in the list which is being used.

N.B.:
- When selecting the 'supervised classification' option, as the number of clusters to be obtained coincides with the number of categories present in the dictionary, the (A) parameter is not available.
- Both the above parameters produce significant changes in the analysis results only when the number of context units is very large and/or when they are short texts.

In the case of unsupervised clustering (default option), the analysis procedure consists of the following steps:

a - construction of a data table context units x lexical units (up to 300,000 rows x 5,000 columns), with presence/absence values;
b - TF-IDF normalization and scaling of row vectors to unit length (Euclidean norm);
c - clustering of the context units (measure: cosine coefficient; method: bisecting K-means; references: Steinbach, Karypis, & Kumar, 2000; Savaresi, Booley, 2001);
d - filing of the obtained partitions and, for each of them:
e- construction of a contingency table lexical units x clusters (n x k);
f- chi square test applied to all the intersections of the contingency table;
g- correspondence analysis of the contingency table lexical units x clusters (references: Benzécri, 1984; Greenacre, 1984; Lebart, Salem, 1994).

N.B.: Starting from T-LAB Plus 2016, the unsupervised clustering of the context units (see step 'c' above) can be performed either by using the bisecting K-means algorithm (1) or by using a not centered version of PDDP (i.e. Principal Direction Divisive Partitioning) proposed by D. Boley (1998) for selecting the seeds of each K-means bisection.
The main differences between the above methods relies on how the two seeds of each bisection are computed; in fact, in the (1) case they result from an iterative procedure whereas in the (2) case they are computed through SVD (i.e. Singular Value Decomposition), i.e. through a 'one-shot' algorithm (see Savaresi, S.M., & Boley, D.L. , 2004).

This procedure therefore performs a type of co-occurrence analysis (steps a-b-c) and, subsequently, a type of comparative analysis (steps e-f-g). In particular, comparative analysis uses the categories of the "new variable" derived from the co-occurrence analysis (categories of the new variable = thematic clusters) to form the contingency table columns.

In the case of supervised classification the steps of comparative analysis are the same (see e-f-g above), whereas co-occurrence analysis is performed as follows:
a - normalization of the seed vectors (i.e. co-occurrence profiles) corresponding to the 'k' categories of the dictionary used;
b - computation of Cosine similarity and of Euclidean distance between each 'i' context unit and each 'k' seed vector;
c - assignment of each 'i' context unit to the 'k' class or category for which the corresponding seed is the closest (In this case, maximum Cosine similarity and minimum Euclidean distance must coincide, otherwise T-LAB consider the 'i' context unit as unclassified).

N.B.: When the user decides to repeat/apply the results of a previous analysis (i.e. a Thematic Analysis of Elementary Contexts or a Modeling of Emerging Themes), T-LAB performs a comparative analysis only (steps e-f-g).

On completion of the analysis you can easily perform the following operations:

1 - explore the characteristics of the clusters;
2 - explore the relationships between the clusters;
3 - explore the relationships between clusters and variables;
4 - explore the various cluster partitions (from 3 to 50);
5 - refine the results of the chosen partition and, if necessary, repeat the above steps (1,2,3);
6 - assign labels to the clusters;
7 - verify which elementary contexts belong to each cluster;
8 - verify the score of each elementary context within the cluster to which it belongs;
9 - export a thematic document classification (only provided when the corpus is made up of at least 2 primary documents and when they are not short texts like responses to open ended questions);
10 - save the selected partition for exploration with other T-LAB tools;
11 - export a dictionary of categories;
12 - validate the chosen partition and assess the semantic coherence of each theme;
13 - moreover, when the corpus is structured like a discourse or like a conversation (i.e. the context units follow each other in a temporal order), it is possible to explore theme sequences by means of animated charts (see below, the final part of this section).

In details:

1 - Explore the characteristics of the clusters

Clicking on the CHARACTERISTICS button shows the lexical units and the variable values which characterize each cluster: Chi-square values and the sums of the elementary contexts in which it is found, both in the selected cluster ("IN CLUST") and in the analysed total ("IN TOT"). The "CAT" column also indicates whether the characteristic has been selected by the user ("A") with the Customized Settings function or has been suggested by T-LAB as a "supplementary" description ("S").

In the case of the chi square test the structure of the analysed table is the following:

Where:
nij refers to occurrences of word (a) within the selected cluster (A)
Nj refers to all occurrences of word (a) within the corpus (or the corpus sub-set) analysed
Ni refers to all word occurrences within the selected cluster (A)
N refers to all word occurrences of the contingency table word by cluster.

An HTML report (see below) is generated to permit detailed analysis of the cluster characteristics. In the report, in addition to the list of typical words, the most characteristic elementary contexts of the selected cluster are shown in descending order according to their respective score.

Pie charts and bar charts are used to verify the percentage of context units (i.e. elementary contexts) that belong to each cluster.

2 - Explore the relationships between the clusters

Some of the graphs obtained by Correspondence Analysis enable you to explore the relationships between clusters in bidimensional spaces.
More specifically:
- You can explore the various combinations of factorial axes, simply by selecting them in the appropriate boxes ("X axis", " Y axis");
- For each of the combinations (X-Y), you can display various types of elements (clusters, lemmas and variables).

All the graphs can be maximized and customized by using the appropriate dialog box (just right click on the chart). Moreover, when thematic clusters are 4 or more, their relationships can be explored through 3d moving (see below).

Moreover, for every factorial axis, T-LAB supplies tables that facilitate the interpretation.
These are shown after every selection in the appropriate boxes (see below).

By selecting the Complete Results option it is possible to check all the results of the Correspondence Analysis lexical units x clusters.

A specific option (see below) allows us to visualise/export the contingency table and to create charts showing the distribution of each word within the clusters and their corresponding chi-square value.
Moreover, by clicking on specific cells of the table, it is possible to create a HTML file including all elementary contexts where the word in row is present in the corresponding cluster.

N.B.: Such a table includes both active ('A') and supplementary ('S') key-words.

3 - Explore the relationships between clusters and variables

Bar charts allow you to verify the relationships between clusters and variables.

You can explore additional relationships between clusters and variables using the functions provided in the Factor Analysis section (see above).

4 - Explore the various cluster partitions

Because the algorithm used (bisecting K-means) produces a hierarchical clustering, the user can explore various analysis solutions: partitions from 3 to 50 clusters.

For each partition obtained, a specific table (see below) lists the following values:
- "Index", obtained by dividing the between cluster variance by the total variance;
- "Gap", corresponding to the difference between the index value and the value of the immediately previous partition:
- the number of the "child" cluster obtained from the bisection of the corresponding "parent".

The Partition option allows you to easily explore the characteristics of the available clustering solutions (just click on a table item).

Moreover the dendrogram option (see below) allows two possibilities:

a) to check the tree structure of the various bisections;

b) to check the tree with the characteristic words of each cluster.

5 - Refine the results of the chosen partition

After having explored different solutions, the user can refine the results of the chosen partition and, if necessary, repeat some of the three operations above illustrated.

For this purpose two methods are available (see the picture below):

When the 'A' method (i.e. Naïve Bayes Classifier) is chosen, this step allows the user to delete from the analysis all context units of which cluster membership doesn't fit either of the following criteria:
a) the cluster memberships of the i-context unit, determined by the bisecting K-means first (unsupervised clustering) and by a Naïve Bayes Classifier later (supervised clustering), must be the same;
b) the maximum posterior value (see below) corresponding to the i-context unit cluster membership must be, in percentage terms, at least 50% higher than its remaining values (i.e. posterior value in other clusters).

Whereas, in the case of 'B' method (i.e. Reclassification Based on Typical Words) T-LAB considers the cluster characteristics - i.e. the words with a significant Chi-Square value - like items of a category dictionary and performs the three steps of 'supervised classification' described at the beginning of this section. So, when the user is interested in re-using dictionaries and in comparing the analysis results, this method is highly recommended.

All the results of this computation are in the following table exported by T-LAB (see below), where the posteriori values for each cluster are in percentage format.

6 - Assign labels to the clusters

A specific T-LAB function allows you to assign labels to clusters.
(N.B: The software proposes a number of labels automatically the first time you use this function.)

Labels assigned to clusters can be displayed in the various graphs available (see below).

7 - Verify which elementary contexts belong to each cluster
8 - Verify the score of each elementary context within the cluster to which it belongs
9 - Obtain a thematic document classification

In fact the Cluster Membership button lets you export three types of tables (see below) in MS Excel format:

a - "Cluster_Partitions.xls" listing all the context unit correspondence for each cluster within the various partitions;

b - "Themes-Contexts.xls" (see below) listing the context unit correspondences for each cluster within the selected partition.

In particular, the relevance value (Score) assigned to each elementary context (j) belonging to the cluster (k) comes from the following formula:

Where:

Scorej = relevance value assigned to the elementary context (j);
SXij = sum of the Chi-square values assigned to the key-words (i) found in the elementary context in question (j) which are "typical" of the cluster (k);
nj = number of key-words (distinct words), typical of the cluster (k), found in the elementary context (j);
N = number of key-words (distinct words) typical of the cluster (k).

c - "Ec_Document_Classification.xls" (only provided when the corpus is made up of at least 2 primary documents at least and when they are not short texts like responses to open ended questions) listing the mixed cluster membership of each document (see below).

In this case the values come from the above formula (see "b") by summing the scores of elementary contexts belonging to each document and by applying a percentage calculation.

10 - Save the selected partition for exploration with other T-LAB tools

When you exit the Thematic Analysis of Elementary Contexts function, the software displays messages to remind you that you can use other T-LAB tools to explore the clusters obtained.

If you select Save, the < CONT_CLUST > variable (clusters of elementary contexts) remains available only for certain types of analysis (e.g. Sequences of Themes, Word Associations, Comparison between Pairs of Key-Words, Co-Word Analysis and Concept Mapping) and until the user modifies his word list.

11 - Export a dictionary of categories

When this option is selected, T-LAB allows the user to create two files:

- a dictionary file with the .dictio extension which is ready to be imported by any T-LAB tool for thematic analysis. In such a dictionary each cluster is a category described by its characteristic words, i.e. by all words with a significant Chi-Square value within it;

- a MyList.diz file ready to be imported via the Customized Settings tool. Since this file contains a list of all words with a significant chi-square value (i.e. all words that determine the differences between thematic clusters), its use may allow the user to repeat some analyses in a 'more selective' way.

12 - Validate the chosen partition and assess the semantic coherence of each theme

When clicking the Quality Indices button (see the picture above), T-LAB creates a HTML file listing various measures.
The first ones, i.e. the measures of 'cluster quality', refer to the quality of the chosen partition.
The second ones, i.e. the measure of 'semantic quality', refer to the similarities between the top 10 words of each theme.
More specifically:
- the top 10 words are those with the highest chi-square values over themes;
- the average similarity is computed using the cosine index;
- the cosine index of each word pairs, like the Word Association tool, is computed at the text segment (i.e. elementary context) level.

13 - Sequences of Themes

Unlike the Sequences of Themes tool included in the co-occurrence analysis sub-menu, this one has been specifically designed to complement the thematic analysis of elementary contexts. More specifically its use makes sense only when the entire corpus can be considered like a discourse and/or its various sections (e.g. chapters of a book, parts of an interview, turns in a conversation or a debate, etc.) follow each other in a temporal order.

In fact this tool deals with the relationships between elementary contexts (up to 100,000) along the linear chain of the corpus, by considering each of them - either predecessor or successor - as an analysis unit belonging to a thematic cluster (or as unclassified). Accordingly, all available outputs allow the user to explore sequential relationships between 'themes', either by means of static charts and tables or by means of animated charts showing changes over time. This way the user can check either when people are engaged in specific themes (e.g. by looking at a diagonal of the matrix below) or when they shift from a dominant theme to another.

Step by step, here is a short description of how to proceed.

(N.B.: All the following outputs refer to a thematic analysis of the book The Politics of Climate Change by Antony Giddens published in the T-LAB website).

When the Sequence of themes button is enabled, by clicking it the following 'player' becomes visible and active in the T-LAB working window.

Option '1' (see matrix / space above) refers to the type of chart for visualizing theme sequences, either within the entire corpus or within a part of it (see option 2 above).

When checking 'matrix', a 3d chart is available which summarizes the relationships between predecessors and successors. In this case, while exploring 3d animated charts the bar dimensions are continuously readjusted to indicate how the occurrences of each sequence (i.e. two way relationship between a 'predecessor' and a 'successor') increases (see below).

When checking the 'space', a 2d scatter chart is available which summarizes both the dimensions (i.e. percentages) and the relationships between thematic clusters. In this case, while exploring 2d animated charts the bubble dimensions - which are continuously readjusted to a total equal to 100% - indicate how the percentage of each thematic cluster changes over time; meanwhile the moving arrows indicate how themes follow each other (see below).

Moreover, at each step - after stopping the video (see the 'pause' button) - it is possible to obtain two further outputs:

A - html tables which summarize the relationships between predecessors and successors (see below);

B - graph files which can be imported by software for network analysis.

N.B.: The above graph, which refers to the third chapter of Giddens' book (i.e. 'The Greens and After') has been created by means of Gephi (see https://gephi.org/).

.