Tutorial 8: Combining widgets from the Story Navigator
This tutorial is part of a series demonstrating the use of StoryNavigator widgets. These tutorials show how to use StoryNavigator widgets with other pre-existing widgets available within the Orange platform, and how to generate output via tables or figures. Each tutorial addresses a research question related to the narrative structure and contents of the corpus of stories.
Research Question
In this tutorial, we will try to answer a research question which requires the results of multiple Story Navigator widgets. The research question is:
What actors are related to actions that make use of a means?
For this purpose, we will combine the analyses of the Story Navigator in a single table and then extract the required results from this table with widgets from the Orange module group Transform.
We use the following workflow:
This workflow can be downloaded here, and it uses a dataset of Dutch fairytales which can be found here.
Step 1: Load the data with the Corpus widget
The workflow starts with the Corpus widget. It loads the dataset containing Dutch fairytales. No action needs to be taken here unless you want to change the input text collection of the workflow.
Step 2: Perform language analysis with the Elements widget
Next, the Elements widget needs to perform the language analysis on the dataset:
Open the Elements widget by double-clicking on the widget icon
Make sure that the right language is selected (’nl’ for the fairytales dataset)
Verify that the number of segments of each story is 1
Do not remove stopwords: this setting should be ‘no’
Click on ‘Extract story elements’
Step 4: Concatenate the analyses with the Concatenate widget
Next, the analyses from the five widgets are combined in a single table by Orange’s Concatenate widget. If you want to see what is in this table, you can link an extra Data Table to the output of the Concatenate widget. The five widgets’ output tables have five columns in common:
text: the relevant phrase that was analyzed
text_id: the number of the story containing the phrase in the text collection
segment_id: the number of the segment containing the phrase
sentence_id: the number of the sentence containing the phrase
character_id: the position of the first character of the phrase in that sentence
Some widgets have extra columns which are also included in the combined table. These columns do not contain values for the other widgets (marked by ‘?’).
The names of the five widgets will appear in a new output column of the Concatenate widget: Source ID. This makes it possible to later check which widget generated which row in the combined table.
Step 5: Combine analyses with the Group By widget
Our research question is to find out what actors are related to actions that make use of a means. Both the Actions widget and Means Analysis widget identify actions. In this step we use the Orange’s Group By widget to combine all analyses of phrases by different widgets on single rows.
For example: if the Actions widget and the Means Widget have both analyzed the verb ‘aanraakte’, this verb will appear on two rows in the combined table, once for each widget. In this step, the Group By widget will combine these two rows on a single row. A new column ‘Source ID - Concatenated’ will contain the names of the widgets involved in the combined row, here ‘actions means’. A widget name can appear more than one time in this column, if the widget generated more than one analysis for the phrase.
Since the same phrase may appear in multiple places in the texts, we also check the relevant ids of the text, segment, sentence and first character before making combinations.
In order to avoid the workflow becoming more complex, we refrain from involving the analysis of the Actors widget in this step.
Step 6: Select relevant rows
Now the Select Rows widget can be used for selecting the information we are interested in: rows containing both action and means in the Source ID - Concatenated column and nsubj, the tag for sentence subject, in the column spacy_dependency - Concatenated.
The results can be inspected in the next Data Table widget (double-click on the widget): the workflow found 17 verbs with both a subject and a means. The subjects can be found in the column text_role - Concatenate. In general, the first word in each cell is the subject but there is one exception where the subject is the second word (text 23, sentence 35). The exception can be identified by the order of the labels in column spacy_dependency - Concatenated.
Step 7: Prepare data for merging
Optional: currently the required subject words appear in a column next to other words. We can achieve having them in a column by their own by merging the results of the Select Rows widget with the results of the Concatenate widget. However, Orange is a bit picky about merging information which includes duplicates and therefor we first need to remove irrelevant information before making the merge.
In this step we use the Select Columns widget to select the columns specifying our results (four position columns and means_id Concatenated) and we apply the Select Rows widget for retrieving the nsubj rows from the Concatenate widget.
Step 8: Merge
The two selected results can now be merged with Orange’s Merge widget. The results can be inspected with the following Data Table. The subjects can be found in the column text_role. Note that we have lost the exception from text 23. Retrieving this one as well will probably require writing task-specific Python code.
Conclusion
The Story Navigator widgets can jointly be applied to the same texts and there analyses can be combined with the general Orange widgets, as shown in this tutorial. This approach enables complex analyses of texts. However more complicated inquiries may require using Orange’s Python Script widget in combination with task-specific programming.