Text Summarization AI Blueprint

Batch-Predict

Text Summarization is a process of creating a fluent, concise, and continuous summary from large textual paragraphs and passages while preserving main ideas, key points, and important information.

Purpose

Use the PubMed Summarization Batch blueprint with a tailor-trained model to extract text summaries from PubMed papers. This blueprint uses the PubMed Connector to access the PubMed National Library of Medicine database and downloads the user-selected papers and creates their summaries.

NOTE

This blueprint and the associated PubMed Connector downloads only full papers. It does not summarize the PubMed database’s abstracts or titles.

Use the Text Summarization Batch blueprint with a tailor-trained model to extract text summaries from custom text and Wikipedia articles. To use this blueprint, provide one summarization folder in the S3 Connector with the training file containing the text to be summarized and the model/tokenizer files.

The main summarization folder has two subdirectories, namely:

• default_model − A folder with the base model used to fine tune to customize the model
• tokenizer − A folder with the tokenizer files used to assist in text summarization

NOTE

This documentation uses the Wikipedia connection and subsequent summarization as an example. The users of this blueprint can select any source of text and input it to the Batch Predict task.

Deep Dive

The following flow diagram illustrates the PubMed Text Summarization batch blueprint’s pipeline:

The following flow diagram illustrates the Text Summarization batch-predict blueprint’s pipeline:

Arguments/Artifacts

For information on these blueprints’ tasks, inputs, and outputs, click here.

PubMed Summarization Inputs

When the Batch task is used the pubmed_connector, ensure the dir value contains the location of the PubMed Connector outputs, such as /input/pubmed_connector/pdfs. The PubMed Connector requires the following inputs:

• --email (string, required) is the user’s email address as an input argument.
• --field (string, required) is the field to be extracted from the papers. For this blueprint, the only valid field is full for the full articles.
• --max_results (integer, required) is the maximum number of results to return. Default: 5.
• --query (string, required) is the text to query the PubMed database.

PubMed Summarization Outputs

• --output.csv is the name of the file containing the paper summaries.

Text Summarization Inputs

• --input_path is the name and path of the file that contains the articles. It is like the output_summaries_file described in the Wikipedia Connector. This CSV formatted file has two columns, one with the summary and the other with a dummy value 'x'. The dummy value is for formatting purposes and does not require user modifications. The file headers are a string pair called document and summary. The file path includes the dataset name containing the file, like the following: \dataset_name\input_file.csv. Ensure that the dataset is in text format. The values can contain article text, cleaned of any special characters. When this Batch Predict task is used the wikipeda_connector, ensure the input path contains the file location of the Wikipedia Connector outputs, like /input/wikipedia_connector/wiki_output.csv
• --modelpath is the model that generates the summaries. By default, it is the cnvrg model trained on wiki_lingua dataset. /model/model/ is the model’s default path. For a custom model, the path is /input/train-task-name/model-file where train-task-name is the library used to run the code. More information can be found here.
• --tokenizer is the tokenizer used while performing the summarization. There is no option to train the tokenizer.
• --min_percent is the lowest ratio of the paragraph, to which the summary lengths can be reduced.
• --encoder_max_length is the maximum length of the encoder used while inputting text to summarize.

Text Summarization Outputs

• --output.csv is the name of the file containing the custom text and article summaries.

Instructions

NOTE

The minimum resource recommendations to run this blueprint are 3.5 CPU and 8 GB RAM.

NOTE

This blueprint’s performance can benefit from using GPU as its compute.

NOTE

NOTE: The PubMed Connector Batch and Wikipedia Connector Batch Predict use different models.

Complete the following steps to run the text-summarizer model in batch mode:

1. Click the Use Blueprint button. The cnvrg Blueprint Flow page displays.

2. For the PubMed Summarization Batch blueprint, complete the following steps:

   • Click the PubMed Connector task to display its dialog, and within the Parameters tab, provide the following Key-Value pair information:
     • Key: email − Value: provide the user’s email address
     • Key: field − Value: leave the field to be extracted from the papers as full for the full articles
     • Key: max_results − Value: provide the maximum number of results to return
     • Key: query − Value: provide the text to query the PubMed database
   • Click the Advanced tab to change resources to run the blueprint, as required.
   • Click the Batch task, and within the Parameters tab, provide the following Key-Value pair information:
     • Key: page_wise − Value: leave the value as false for the full article summaries
     • Key: dir − Value: provide the path to the data file including the PubMed prefix. Ensure the path adheres to the following format: /input/pubmed_connector/pdfs
   • Click the Advanced tab to change resources to run the blueprint, as required.

3. For the Text Summarization Batch blueprint, complete the following steps:

   • Click the S3 Connector task to display its dialog, and within the Parameters tab, provide the following Key-Value pair information:

     • Key: bucketname − Value: provide the data bucket name
     • Key: prefix − Value: provide the main path to the articles folder

   • Click the Advanced tab to change resources to run the blueprint, as required.

   • Click the Wikipedia Connector task to display its dialog, and within the Parameters tab, provide the following Key-Value pair information:

     • Key: topics − Value: provide the topics to extract from Wikipedia
     • Format − the flexible format can be in either comma-separated text (shown), tabular CSV, or URL link

   • Click the Advanced tab to change resources to run the blueprint, as required.

   • Click the Batch-Predict task to display its dialog, and within the Parameters tab, provide the following Key-Value pair information:

     • Key: input_path − Value: provide the path to the training file including the Wikipedia prefix in the following format: /input/wikipedia_connector/wiki_output.csv
     • Key: modelpath − Value: provide the path to the model including the S3 prefix in the following format: /input/s3_connector/model_files/summarization/bart_large_cnn/
     • Key: tokenizer − Value: provide the path to the tokenizer files including the S3 prefix in the following format: /input/s3_connector/model_files/summarization/tokenizer_files/

     NOTE

     You can use the prebuilt example data paths provided.

   • Click the Advanced tab to change resources to run the blueprint, as required.

4. Click the Run button.

   Depending on the blueprint run, the cnvrg software deploys a text-summarization model that summarizes PubMed papers or custom text and Wikipedia articles and downloads CSV files containing the summaries.

5. Track the blueprint’s real-time progress in its Experiments page, which displays artifacts such as logs, metrics, hyperparameters, and algorithms. The Text Summarization Batch Experiments page is shown.

6. Select Batch Predict > Experiments > Artifacts and locate the output CSV files. The Text Summarization Batch Artifacts are shown below.

7. Select the desired CSV File Name, click the right Menu icon, and click Open File to view the output CSV file. An output CSV file from the Text Summarization Batch blueprint is shown below.

Depending on the blueprint run, a custom text-summarizer model that summarizes PubMed papers or custom text and Wikipedia articles has now been deployed in batch mode. For information on a blueprint's software version and release details, click here or click here.

Connected Libraries

Refer to the following libraries connected to this blueprint:

• PubMed Connector
• S3 Connector
• Wikipedia Connector
• Text Summarization Batch

Related Blueprints

Refer to the following blueprints related to this batch blueprint:

• Text Summarization Inference
• Text Summarization Train
• Topic Modeling Batch
• Text Extraction Inference
• Text Detection Batch

Inference

Text summarization is a process of creating a fluent, concise, and continuous summary from large textual paragraphs and passages while preserving main ideas, key points, and important information.

Purpose

Use this inference blueprint to immediately summarize English text to short sentences. To use this pretrained text-summarization model, create a ready-to-use API-endpoint that can be quickly integrated with your data and application.

This inference blueprint’s model was trained using the WikiLingua Multilingual Abstractive Summarization dataset. To use custom data according to your specific business, run this counterpart’s training blueprint, which trains the model and establishes an endpoint based on the newly trained model.

Instructions

NOTE

The minimum resource recommendations to run this blueprint are 3.5 CPU and 8 GB RAM.

NOTE

This blueprint’s performance can benefit from using GPU as its compute.

Complete the following steps to deploy a text-summarization API endpoint:

1. Click the Use Blueprint button.
2. In the dialog, select the relevant compute to deploy the API endpoint and click the Start button.
3. The cnvrg software redirects to your endpoint. Complete one or both of the following options:
   • Use the Try it Live section with any text to check your model.
   • Use the bottom integration panel to integrate your API with your code by copying in the code snippet.

An API endpoint that summarizes English text has now been deployed. For information on this blueprint's software version and release details, click here.

Related Blueprints

Refer to the following blueprints related to this inference blueprint:

• Text Summarization Train
• Text Summarization Batch
• Topic Modeling Train
• Topic Modeling Inference

Training

Text summarization is a process of creating a fluent, concise, and continuous summary from large textual paragraphs and passages while preserving main ideas, key points, and important information.

Overview

The following diagram provides an overview of this blueprint’s inputs and outputs.

Purpose

Use this training blueprint to train a custom model that can summarize Wikipedia articles and custom textual paragraphs to short sentences using the BERT model. This blueprint also establishes an endpoint that can be used to summarize paragraphs based on the newly trained model.

To train this model with your data, create a summarization folder in the S3 Connector that comprises the training file in CSV format containing text to be summarized (optional). Also, include two subdirectories that contain the model and tokenizer files, namely:

• default_model − A folder with the base model to fine-tune to obtain the custom model
• tokenizer − A folder with the tokenizer files to use to assist in text summarization

NOTE

This documentation uses the Wikipedia connection and subsequent summarization as an example. The users of this blueprint can select any source of text and input it to the Batch Predict task.

Deep Dive

The following flow diagram illustrates this blueprint’s pipeline:

Flow

The following list provides a high-level flow of this blueprint’s run:

• In the S3 Connector, provide the data bucket name and the path to the training data file and two subfolders, with one folder containing the base model and another containing the tokenizers.
• In the Train task, provide the training_file path and two subfolder default_model and tokenizer paths, all with the same previous S3 prefix.
• In the Wikipedia task, provide the Wikipedia topics to extract, in the format of text, tabular, or URL link.
• In the Batch Predict task, provide the input_path, modelpath, and tokenizer paths to the Wikipedia Connector’s output CSV file, the Train task’s custom model (or user-specified custom model location), and the S3 Connector’s tokenizer files, respectively.

Arguments/Artifacts

For more information and examples of this blueprint's tasks, its inputs, and outputs, click here.

Train Inputs

• --training_file is the name of the file and its path which contains the user’s model-training data or the custom wiki_lingua dataset. It looks like this: training_data/wiki_lingua_file.csv, where training_data is the name of the dataset. This CSV file has two columns:
  • the first with a document header and cell(s) below containing the article (titled document)
  • the second with a summary header and cell(s) below containing the summary
• --train_rows is the number of rows of either the wiki_lingua dataset or the custom-user dataset on which the model is trained. NOTE: The cnvrg team recommends <500 rows for faster results.
• --tokenizer are the tokenizer files to be used to generate the summary, which can be trained alongside model, but isn't trained in this blueprint.
• --default_model is the model bart_large_cnn on which the training is performed.
• --encoder_max_length is the maximum length into which the data is encoded.
• --decoder_max_length is the maximum length into which the model decodes the encoded sequences.
• --label_smooth_factor is the factor that smooths the labels. Zero means no label smoothing. Otherwise, the underlying one-hot-encoded labels are changed from 0s and 1s, label_smoothing_factor/num_labels and label_smoothing_factor + label_smoothing_factor/num_labels, respectively.
• --weight_decay_factor is the factor that regularizes and shrinks the weights during backpropagation to help prevent overfitting the training data and exploding the gradient problem.

Train Outputs

• my_custom_model is the name of the folder that contains the files for the mode.

Wikipedia Connector Inputs

• topics is a list of topics to extract from Wikipedia. Its input format is flexible, which can be either comma-separated text, a tabular CSV file format, or a URL link.

Wikipedia Connector Outputs

• --wiki_output.csv is the two-column output CSV file that contains the text in a text-title format.

Instructions

NOTE

The minimum resource recommendations to run this blueprint are 3.5 CPU and 8 GB RAM.

NOTE

This blueprint’s performance can benefit from using GPU as its compute.

Complete the following steps to train the text-summarizer model:

1. Click the Use Blueprint button. The cnvrg Blueprint Flow page displays.

2. In the flow, click the S3 Connector task to display its dialog.

   • Within the Parameters tab, provide the following Key-Value pair information:
     • Key: bucketname − Value: enter the data bucket name
     • Key: prefix − Value: provide the main path to the images folder
   • Click the Advanced tab to change resources to run the blueprint, as required.

3. Return to the flow and click the Train task to display its dialog.

   • Within the Parameters tab, provide the following Key-Value pair information:

     • Key: training_file − Value: provide the path to the CSV file including the S3 prefix in the following format: /input/s3_connector/<prefix>/wiki_lingua_file.csv
     • Key: default_model − Value: provide the path to the base model including the S3 prefix in the following format: /input/s3_connector/<prefix>/bart_large_cnn
     • Key: tokenizer − Value: provide the path to the tokenizer files including the S3 prefix in the following format: /input/s3_connector/<prefix>/tokenizer_files

     NOTE

     You can use the prebuilt example data paths provided.

   • Click the Advanced tab to change resources to run the blueprint, as required.

4. Click the Wikipedia Connector task to display its dialog.

   • Within the Parameters tab, provide the following Key-Value pair information:
     • Key: topics − Value: provide the topics to extract from Wikipedia
     • Format − provide the topics value in one of the following three flexible formats: comma-separated text (shown), tabular CSV, or URL link
   • Click the Advanced tab to change resources to run the blueprint, as required.

5. Click the Batch Predict task to display its dialog.

   • Within the Parameters tab, provide the following Key-Value pair information:

     • Key: input_path − Value: provide the path to the Wikipedia Connector’s output CSV file in the following format: /input/wikipedia_connector/wiki_ output.csv
     • Key: modelpath − Value^: provide the path to the Train task’s custom model in the following format: /input/train/my_custom_model
     • Key: tokenizer − Value: provide the path to the S3 Connector’s tokenizer files including in the following format: /input/s3_connector/model_files

     ^NOTE

     The Value for the modelpath Key can also point to a user-defined model, not just the one trained in the Train task.

     NOTE

     You can use the prebuilt example data paths provided.

   • Click the Advanced tab to change resources to run the blueprint, as required.

6. Click the Run button. The cnvrg software launches the training blueprint as set of experiments, generating a trained text-summarizer model and deploying it as a new API endpoint.

   NOTE

   The time required for model training and endpoint deployment depends on the size of the training data, the compute resources, and the training parameters.

   For more information on cnvrg endpoint deployment capability, see cnvrg Serving.

7. Track the blueprint's real-time progress in its Experiments page, which displays artifacts such as logs, metrics, hyperparameters, and algorithms.

8. Click the Serving tab in the project and locate your endpoint.

9. Complete one or both of the following options:

   • Use the Try it Live section with any text to check the model’s ability to summarize.
   • Use the bottom integration panel to integrate your API with your code by copying in your code snippet.

A custom model and API endpoint, which can summarize text, have now been trained and deployed. If using the Batch Predict task, a custom text-summarizer model has now been deployed in batch mode. For information on this blueprint's software version and release details, click here.

Connected Libraries

Refer to the following libraries connected to this blueprint:

• S3 Connector
• Wikipedia Connector
• Text Summarization Train
• Text Summarization Batch
• Text Summarization Inference

Related Blueprints

Refer to the following blueprints related to this training blueprint:

• Text Summarization Inference
• Text Summarization Batch
• Topic Modeling Train
• Topic Modeling Inference