Icon

Example 3 - Building Sentiment Predictor - LSTM

Building a Sentiment Analysis Predictive Model - Deep Learning using an RNN

This workflow uses a Kaggle Dataset including 14K customer tweets towards six US airlines (https://www.kaggle.com/crowdflower/twitter-airline-sentiment). Contributors annotated the valence of the tweets as positive, negative and neutral. Once users are satisfied with the model evaluation, they should export (1) the Dictionary, (2) the Category to Number Model, and (3) the Trained Network for deployment in non-annotated data.

This workflow is tailored for Windows. If you run it on another system, you may have to (1) adapt the environment of the Conda Environment Propagation node and (2) make sure that the Keras Embedding Layer node has the right number of units, which depends on the native encoding of the system and is indicated in the CSV Reader node.

If you use this workflow, please cite:
F. Villaroel Ordenes & R. Silipo, “Machine learning for marketing on the KNIME Hub: The development of a live repository for marketing applications”, Journal of Business Research 137(1):393-410, DOI: 10.1016/j.jbusres.2021.08.036.

Building a Sentiment Analysis Predictive Model - LSTM model This workflow uses a Kaggle Dataset including 14K customer tweets towards six US airlines (https://www.kaggle.com/crowdflower/twitter-airline-sentiment). Contributors annotated the valence of the tweets as positive, negative andneutral. Once users are satisfied with the model evaluation, they should export (1) the Dictionary, (2) the Category to Number Model, and (3) the Trained Network for deployment in non-annotated data. This workflow is tailored for Windows. If you run it on another system, you may have to (1) adapt the environment of the Conda Environment Propagation node and (2) make sure that the Keras Embedding Layer node has the rightnumber of units, which depends on the native encoding of the system and is indicated in the CSV Reader node. 2. Read annotatedtwitter dataset. 1. Define the Network Architecture The Keras Layer nodes define an LSTM based recurrent neuralnetwork. The network structure can be extended by adding more Keras Layer nodes. 3. Manipulate and Encode DataThe metanode performs an index encoding to encode each word with an index.This blog post describes different encoding options https://www.knime.com/blog/text-encoding-a-reviewIn general recurrent neural networks can handle sequences with different lengths.During training though, all sequences must have the same length. Therefore, themetanode adds zeros to the end of the sequences, so that all sequences have thesame length. This approach is known as zero padding. 4. Train and Apply Network The Keras Network Learner node trains thedefined network. In the configuration window you can define the inputcolumn(s), target column(s), the loss function, and the training parameters,e.g. number of epochs, batch size, and optimizer. The Keras NetworkExecutor node applies the trained network to the input data. In theconfiguration window you can select the input column(s) and define theoutput by clicking on the "add output" button. In this worklflow the softmaxoutput layer is defined as output. This means that the output are theprobabilities for the three classes. The Conda Environment Propagation node ensures the existence of aConda environment with all packages. Another option is to setup yourPython integration to use a Conda environment with all packages asdescribed here: https://docs.knime.com/2019-06/deep_learning_installation_guide/index.html#dl_python_setup 5. Evaluate Trained Network The metanodeExtract Predictions uses the probabilitiesproduced by the Keras Network Executor nodeand extracts the class with the highestprobability. Workflow source: https://hub.knime.com/knime/spaces/Machine%20Learning%20and%20Marketing/latest/Sentiment%20Analysis~-z5F8gFJ9Fm1756a/F. Villaroel Ordenes & R. Silipo, “Machine learning for marketing on the KNIME Hub: The development of a live repository for marketingapplications”, Journal of Business Research 137(1):393-410, DOI: 10.1016/j.jbusres.2021.08.036. Kaggle DatasetN=14640Tweets fromconsumers toairlinesLoss function:Categorical CrossEntropyEpochs: 5080% training20% testingEncode eachclass with an index Output: Softmax layer=> Probability for the different classesInput: # word in dictionary + 1128 units Softmax with 3 unitsNote: An appropriate output layer for amulticlass classification task is a softmax layer with as many unitsas classes.Units for cellstate: 256Shape: ?Note: Using ? as inputshape allows to handle different sequence lengths Class withhighest probabilitySet up a conda environment: dl_sentiment_kerasCSV Reader Create CollectionColumn Keras NetworkLearner Partitioning Category To Number Keras NetworkExecutor Scorer Keras EmbeddingLayer Keras Dense Layer Keras LSTM Layer Keras Input Layer Index encodingand zero padding Extract Prediction Conda EnvironmentPropagation Building a Sentiment Analysis Predictive Model - LSTM model This workflow uses a Kaggle Dataset including 14K customer tweets towards six US airlines (https://www.kaggle.com/crowdflower/twitter-airline-sentiment). Contributors annotated the valence of the tweets as positive, negative andneutral. Once users are satisfied with the model evaluation, they should export (1) the Dictionary, (2) the Category to Number Model, and (3) the Trained Network for deployment in non-annotated data. This workflow is tailored for Windows. If you run it on another system, you may have to (1) adapt the environment of the Conda Environment Propagation node and (2) make sure that the Keras Embedding Layer node has the rightnumber of units, which depends on the native encoding of the system and is indicated in the CSV Reader node. 2. Read annotatedtwitter dataset. 1. Define the Network Architecture The Keras Layer nodes define an LSTM based recurrent neuralnetwork. The network structure can be extended by adding more Keras Layer nodes. 3. Manipulate and Encode DataThe metanode performs an index encoding to encode each word with an index.This blog post describes different encoding options https://www.knime.com/blog/text-encoding-a-reviewIn general recurrent neural networks can handle sequences with different lengths.During training though, all sequences must have the same length. Therefore, themetanode adds zeros to the end of the sequences, so that all sequences have thesame length. This approach is known as zero padding. 4. Train and Apply Network The Keras Network Learner node trains thedefined network. In the configuration window you can define the inputcolumn(s), target column(s), the loss function, and the training parameters,e.g. number of epochs, batch size, and optimizer. The Keras NetworkExecutor node applies the trained network to the input data. In theconfiguration window you can select the input column(s) and define theoutput by clicking on the "add output" button. In this worklflow the softmaxoutput layer is defined as output. This means that the output are theprobabilities for the three classes. The Conda Environment Propagation node ensures the existence of aConda environment with all packages. Another option is to setup yourPython integration to use a Conda environment with all packages asdescribed here: https://docs.knime.com/2019-06/deep_learning_installation_guide/index.html#dl_python_setup 5. Evaluate Trained Network The metanodeExtract Predictions uses the probabilitiesproduced by the Keras Network Executor nodeand extracts the class with the highestprobability. Workflow source: https://hub.knime.com/knime/spaces/Machine%20Learning%20and%20Marketing/latest/Sentiment%20Analysis~-z5F8gFJ9Fm1756a/F. Villaroel Ordenes & R. Silipo, “Machine learning for marketing on the KNIME Hub: The development of a live repository for marketingapplications”, Journal of Business Research 137(1):393-410, DOI: 10.1016/j.jbusres.2021.08.036. Kaggle DatasetN=14640Tweets fromconsumers toairlinesLoss function:Categorical CrossEntropyEpochs: 5080% training20% testingEncode eachclass with an index Output: Softmax layer=> Probability for the different classesInput: # word in dictionary + 1128 units Softmax with 3 unitsNote: An appropriate output layer for amulticlass classification task is a softmax layer with as many unitsas classes.Units for cellstate: 256Shape: ?Note: Using ? as inputshape allows to handle different sequence lengths Class withhighest probabilitySet up a conda environment: dl_sentiment_kerasCSV Reader Create CollectionColumn Keras NetworkLearner Partitioning Category To Number Keras NetworkExecutor Scorer Keras EmbeddingLayer Keras Dense Layer Keras LSTM Layer Keras Input Layer Index encodingand zero padding Extract Prediction Conda EnvironmentPropagation

Nodes

Extensions

Links

Download

To use this workflow in KNIME, download it from the below URL and open it in KNIME:

Download Workflow
Created by: fvill
Created at: 2020-11-10
On NodePit since: 2023-12-06
Last update: 2024-04-27
Created with KNIME version: v4.6.3
Tags: sentiment analysissentimentmachine learningsupervised learningRNNKerasDeep Learning

Deploy, schedule, execute, and monitor your KNIME workflows locally, in the cloud or on-premises – with our brand new NodePit Runner.

Try NodePit Runner!