University Libraries

A topic model employs machine learning to aid the discovery of abstract topics or themes present in a large or unstructured dataset.

For this kind of analysis, the text is chunked into "documents", and stop words (frequently used words such as "the", "and", "if") are removed since they reveal little about the substance of a text. The computer treats the documents as bags of words, and guesses which words make up a "topic" based on their proximity to one another in the documents, with the idea the words that frequently co-occur are likely about the same thing. The results are groupings of words that the computer has statistically analyzed and determined are likely related to each other...

- HTRC, Digging Deeper, Reaching Further

1. Once you have logged in to HTRC, select the "Algorithms" tab to execute the InPho Topic Model Explorer.

2. Provide a name for the job and select the workset to be analyzed.

3. Define the parameters for the number of iterations. This determines the number of samples that the topic model will use to conduct its analysis. A lower number of iterations (i.e. 200 iterations) will process faster and is good for experimentation. A higher number will give higher quality results (i.e. 1000 iterations).

4. Indicate the number of topics you would like the algorithm to generate from your dataset. Multiple values, such as the default "20 40 60 80," are accepted. With this configuration, the model will run multiple times creating a list of 20, 40, 60, and 80 topics for comparison. An appropriate number will depend on your research inquiry and the size of the workset.

5. Click "Submit" and you'll be redirected to the algorithm jobs subpage. The status bar for active jobs indicates when they are first staging, then running. Processing times will vary depending on the size of the workset and amount of text.

6. Clicking the name of a completed job displays the results and data available for download.

Output & Interpretation

Bubble visualization: Displays each topic as a node or bubble. When hovering over a bubble, you can see the top words that were grouped into the topic. The colors of the bubbles are a loose representation of topics with similar themes. Click the "collision detection" box to minimize overlap between bubbles and improve readability. The numbers on the left side relate to the number of topics generated, as do the size of the bubbles. You can toggle the display of the topic clusters by clicking on the numbers.

Topic model visualization of the Cairns Collection of American Women Writers. Click image to open interactive visualization.

Topics json file: In addition to the bubble visualization, the algorithm generates several files including one called topics.json. This file allows researchers a detailed view of the word groupings that constitute a topic. Each word has a decimal number that represents the probability of its appearance in the topic. Topics.json can be viewed directly in the browser or pasted into a conversion tool such as json2table.com for easier viewing.

 Topics.json text showing grouped words and probabilities

Topic models require interpretation. They provide lists of words that may be related given the frequency with which they appear together throughout the corpus. It's up to the researcher to derive meaning from the word groupings and interpret the topic. The topic model generated for the Cairns Collection of American Women Writers reveals multiple word groupings that suggest overarching themes such as religion, romance, children's literature, travel, American history, and the role of women in society and the home. The dataset also contains some very specific word groupings that are less broadly represented, including those related to slavery, science, women's organizations, and book publishing.

The HTRC topic model explorer has limited parameters that often yield imprecise results. However it may prove useful for identifying themes and outliers within a large, unfamiliar text corpus.

"Generate a list of all of the names of people and places, as well as dates, times, percentages, and monetary terms, found in a workset. " - HTRC

Named Entity Recognition (NER) is a process by which parts of an unstructured text corpus are labeled and extracted as names. It relies on machine learning to structure components of the text and classify them using statistical models of word use. NER is a branch of Natural Language Processing (NLP) which utilizes computer systems to parse human language and extract data. 

1. Once you have logged in to HTRC, select the "Algorithms" tab to execute the Named Entity Recognizer.

2. Provide a name for the job, select the workset to be analyzed, and identify the predominant language.

3. Click "Submit" and you'll be redirected to the algorithm jobs subpage. The status bar for active jobs indicates when they are first staging, then running. Processing times will vary depending on the size of the workset and amount of text.

4. Clicking the name of a completed job displays the results and data available for download.

This video walks through the process of applying the NER algorithm to a workset.

The resulting CSV file can be used to create a map visualization (see instructions & video below.)

Output & Interpretation

Table of named entities: A downloadable CSV file that displays the entity name and type, as well as the volume ID and page sequence. The file can be opened using a spreadsheet program such as Microsoft Excel or Google Sheets. However, the size of the dataset could prohibit some software from opening the file. 

Entity types include: date, location, miscellaneous proper nouns, money, organization, percent, person, and time. The corresponding volume ID and page sequence can be used to identify the source of a given entity by plugging the vol_id and page_seq into the following URL template:

https://babel.hathitrust.org/cgi/pt?id=hvd.32044021584222&seq=33 

In this example the volume id is "hvd.32044021584222" and the page sequence is "33". 

The link points to the HathiTrust Digital Library volume and page on which the entity appears.