Train Entity Recognition Model (GeoAI)

Summary

Trains a named entity recognition model to extract a predefined set of entities from raw text.

Learn more about how Entity Recognition works

Usage

• This tool requires deep learning frameworks be installed. To set up your machine to use deep learning frameworks in ArcGIS AllSource, see Install deep learning frameworks for ArcGIS.

• This tool can also be used to fine-tune an existing trained model.

• To run this tool using GPU, set the Processor Type environment to GPU. If you have more than one GPU, specify the GPU ID environment instead.

• The input can be a feature class or table with a text field and named entity labels, or a folder with training data in .json or .csv files.

• This tool uses transformer-based backbones for training NER models and also supports in-context learning with prompts using the Mistral LLM. To install the Mistral backbone, see ArcGIS Mistral Backbone.

• For information about requirements for running this tool and issues you may encounter, see Deep Learning frequently asked questions.

Parameters

Dialog

Label	Explanation	Data Type
Input Folder or Table	The input can be either of the following: A feature class or table containing a text field with the input text for the model and the labelled entities where the selected text field will be used as input text for the model and the remaining fields will be treated as named entities labels. A folder containing training data in the form of standard datasets for NER tasks. The training data must be in .json or .csv files. The file format determines the dataset type of the input. When the input is a folder, the following dataset types are supported: ner_json—The training data folder should contain a .json file with text and the labelled entities formatted using the spaCy JSON training format. IOB—The IOB (I - inside, O - outside, B - beginning tags) format proposed by Ramshaw and Marcus in the paper Text Chunking using Transformation-Based Learning. The training data folder should contain the following two .csv files: tokens.csv—Contains text as input chunks tags.csv—Contains IOB tags for the text chunks BILUO—An extension of the IOB format that additionally contains L - last and U - unit tags. The training data folder should contain the following two .csv files: tokens.csv—Contains text as input chunks tags.csv—Contains BILUO tags for the text chunks	Folder; Feature Layer; Table View; Feature Class
Output Model	The output folder location where the trained model will be stored.	Folder
Pretrained Model File (Optional)	A pretrained model that will be used to fine-tune the new model. The input can be an Esri model definition file (.emd) or a deep learning package file (.dlpk). A pretrained model with similar entities can be fine-tuned to fit the new model. The pretrained model must have been trained with the same model type and backbone model that will be used to train the new model.	File
Address Entity (Optional)	An address entity that will be treated as a location. During inference, such entities will be geocoded using the specified locator, and a feature class will be produced as a result of the entity extraction process. If no locator is provided or the trained model does not extract address entities, a table containing the extracted entities will be produced instead.	String
Max Epochs (Optional)	The maximum number of epochs for which the model will be trained. A maximum epoch value of 1 means the dataset will be passed through the neural network one time. The default value is 5.	Long
Model Backbone (Optional)	Specifies the preconfigured neural network that will be used as the architecture for training the new model. bert-base-cased—The model will be trained using the BERT neural network. BERT is pretrained using the masked language modeling objective and next sentence prediction. roberta-base— The model will be trained using the RoBERTa neural network. RoBERTa modifies the key hyperparameters of BERT, eliminating the pretraining objective and training of the next sentence with small batches and higher learning rates. albert-base-v1—The model will be trained using the ALBERT neural network. ALBERT uses a self-supervised loss that focuses on modeling intersentence coherence, resulting in better scalability than BERT. xlnet-base-cased—The model will be trained using the XLNet neural network. XLNet is a generalized autoregressive pretraining method. It allows learning bidirectional contexts by maximizing the expected probability on all permutations of the factorization order, which overcomes the drawbacks of BERT. xlm-roberta-base—The model will be trained using the XLM-RoBERTa neural network. XLM-RoBERTa is a multilingual model trained on 100 different languages. Unlike some XLM multilingual models, it does not require language tensors to understand which language is used and identifies the correct language from the input IDs. distilroberta-base—The model will be trained using the DistilRoBERTa neural network. DistilRoBERTa is an English language model pretrained with the supervision of roberta-base neural network based solely on OpenWebTextCorpus, a reproduction of OpenAI's WebText dataset. distilbert-base-cased—The model will be trained using the DistilBERT neural network. DistilBERT is a smaller general-purpose language representation model. mistral—The model will be trained using the Mistral large language model (LLM). Mistral is a decoder-only transformer that uses Sliding Window Attention, Grouped Query Attention, and the Byte-fallback BPE tokenizer. To install the Mistral backbone, see ArcGIS Mistral Backbone.	String
Batch Size (Optional)	The number of training samples that will be processed at one time. The default value is 2. Increasing the batch size can improve tool performance; however, as the batch size increases, more memory is used. If an out of memory error occurs, use a smaller batch size.	Double
Model Arguments (Optional)	Additional arguments that will be used for initializing the model. The supported model argument is sequence_length, which is used to set the maximum sequence length of the training data that will be considered for training the model.	Value Table
Learning Rate (Optional)	The step size indicating how much the model weights will be adjusted during the training process. If no value is specified, an optimal learning rate will be derived automatically.	Double
Validation Percentage (Optional)	The percentage of training samples that will be used for validating the model. The default value is 10 for transformer-based model backbones and 50 for the Mistral backbone.	Double
Stop when model stops improving (Optional)	Specifies whether model training will stop when the model is no longer improving or continue until the Max Epochs parameter value is reached. Checked—The model training will stop when the model is no longer improving, regardless of the Max Epochs parameter value. This is the default. Unchecked—The model training will continue until the Max Epochs parameter value is reached.	Boolean
Make model backbone trainable (Optional)	Specifies whether the backbone layers in the pretrained model will be frozen, so that the weights and biases remain as originally designed. Checked—The backbone layers will not be frozen, and the weights and biases of the Model Backbone parameter value can be altered to fit the training samples. This takes more time to process but typically produces better results. This is the default. Unchecked—The backbone layers will be frozen, and the predefined weights and biases of the Model Backbone parameter value will not be altered during training.	Boolean
Text Field	A text field in the input feature class or table that contains the text that will be used by the model as input. This parameter is required when the Input Folder or Table parameter value is a feature class or table.	Field
Prompt (Optional)	A specific input or instruction given to a large language model (LLM) to generate an expected output. The default value is Extract named entities belonging to the specified classes within the provided text. Do not tag entities belonging to any other class.	String

Python

arcpy.geoai.TrainEntityRecognitionModel(in_folder, out_model, {pretrained_model_file}, {address_entity}, {max_epochs}, {model_backbone}, {batch_size}, {model_arguments}, {learning_rate}, {validation_percentage}, {stop_training}, {make_trainable}, text_field, {prompt})

Name	Explanation	Data Type
in_folder	The input can be either of the following: A feature class or table containing a text field with the input text for the model and the labelled entities where the selected text field will be used as input text for the model and the remaining fields will be treated as named entities labels. A folder containing training data in the form of standard datasets for NER tasks. The training data must be in .json or .csv files. The file format determines the dataset type of the input. When the input is a folder, the following dataset types are supported: ner_json—The training data folder should contain a .json file with text and the labelled entities formatted using the spaCy JSON training format. IOB—The IOB (I - inside, O - outside, B - beginning tags) format proposed by Ramshaw and Marcus in the paper Text Chunking using Transformation-Based Learning. The training data folder should contain the following two .csv files: tokens.csv—Contains text as input chunks tags.csv—Contains IOB tags for the text chunks BILUO—An extension of the IOB format that additionally contains L - last and U - unit tags. The training data folder should contain the following two .csv files: tokens.csv—Contains text as input chunks tags.csv—Contains BILUO tags for the text chunks	Folder; Feature Layer; Table View; Feature Class
out_model	The output folder location where the trained model will be stored.	Folder
pretrained_model_file (Optional)	A pretrained model that will be used to fine-tune the new model. The input can be an Esri model definition file (.emd) or a deep learning package file (.dlpk). A pretrained model with similar entities can be fine-tuned to fit the new model. The pretrained model must have been trained with the same model type and backbone model that will be used to train the new model.	File
address_entity (Optional)	An address entity that will be treated as a location. During inference, such entities will be geocoded using the specified locator, and a feature class will be produced as a result of the entity extraction process. If no locator is provided or the trained model does not extract address entities, a table containing the extracted entities will be produced instead.	String
max_epochs (Optional)	The maximum number of epochs for which the model will be trained. A maximum epoch value of 1 means the dataset will be passed through the neural network one time. The default value is 5.	Long
model_backbone (Optional)	Specifies the preconfigured neural network that will be used as the architecture for training the new model. bert-base-cased—The model will be trained using the BERT neural network. BERT is pretrained using the masked language modeling objective and next sentence prediction. roberta-base— The model will be trained using the RoBERTa neural network. RoBERTa modifies the key hyperparameters of BERT, eliminating the pretraining objective and training of the next sentence with small batches and higher learning rates. albert-base-v1—The model will be trained using the ALBERT neural network. ALBERT uses a self-supervised loss that focuses on modeling intersentence coherence, resulting in better scalability than BERT. xlnet-base-cased—The model will be trained using the XLNet neural network. XLNet is a generalized autoregressive pretraining method. It allows learning bidirectional contexts by maximizing the expected probability on all permutations of the factorization order, which overcomes the drawbacks of BERT. xlm-roberta-base—The model will be trained using the XLM-RoBERTa neural network. XLM-RoBERTa is a multilingual model trained on 100 different languages. Unlike some XLM multilingual models, it does not require language tensors to understand which language is used and identifies the correct language from the input IDs. distilroberta-base—The model will be trained using the DistilRoBERTa neural network. DistilRoBERTa is an English language model pretrained with the supervision of roberta-base neural network based solely on OpenWebTextCorpus, a reproduction of OpenAI's WebText dataset. distilbert-base-cased—The model will be trained using the DistilBERT neural network. DistilBERT is a smaller general-purpose language representation model. mistral—The model will be trained using the Mistral large language model (LLM). Mistral is a decoder-only transformer that uses Sliding Window Attention, Grouped Query Attention, and the Byte-fallback BPE tokenizer. To install the Mistral backbone, see ArcGIS Mistral Backbone.	String
batch_size (Optional)	The number of training samples that will be processed at one time. The default value is 2. Increasing the batch size can improve tool performance; however, as the batch size increases, more memory is used. If an out of memory error occurs, use a smaller batch size.	Double
model_arguments [model_arguments,...] (Optional)	Additional arguments that will be used for initializing the model. The supported model argument is sequence_length, which is used to set the maximum sequence length of the training data that will be considered for training the model.	Value Table
learning_rate (Optional)	The step size indicating how much the model weights will be adjusted during the training process. If no value is specified, an optimal learning rate will be derived automatically.	Double
validation_percentage (Optional)	The percentage of training samples that will be used for validating the model. The default value is 10 for transformer-based model backbones and 50 for the Mistral backbone.	Double
stop_training (Optional)	Specifies whether model training will stop when the model is no longer improving or continue until the max_epochs parameter value is reached. STOP_TRAINING—The model training will stop when the model is no longer improving, regardless of the max_epochs parameter value. This is the default. CONTINUE_TRAINING—The model training will continue until the max_epochs parameter value is reached.	Boolean
make_trainable (Optional)	Specifies whether the backbone layers in the pretrained model will be frozen, so that the weights and biases remain as originally designed. TRAIN_MODEL_BACKBONE—The backbone layers will not be frozen, and the weights and biases of the model_backbone parameter value can be altered to fit the training samples. This takes more time to process but typically produces better results. This is the default. FREEZE_MODEL_BACKBONE—The backbone layers will be frozen, and the predefined weights and biases of the model_backbone parameter value will not be altered during training.	Boolean
text_field	A text field in the input feature class or table that contains the text that will be used by the model as input. This parameter is required when the in_folder parameter value is a feature class or table.	Field
prompt (Optional)	A specific input or instruction given to a large language model (LLM) to generate an expected output. The default value is Extract named entities belonging to the specified classes within the provided text. Do not tag entities belonging to any other class.	String

Code sample

TrainEntityRecognitionModel (stand-alone script)

The following example demonstrates how to use the TrainEntityRecognitionModel function.

# Name: TrainEntityRecognizer.py
# Description: Train an Entity Recognition model to extract useful entities such as "Address", "Date" from text.  

# Import system modules
import arcpy

arcpy.env.workspace = "C:/textanalysisexamples/data"

# Set local variables
in_folder = "train_data"
out_folder = "test_bio_format"

# Run Train Entity Recognition Model
arcpy.geoai.TrainEntityRecognitionModel(in_folder, out_folder)

Environments

Processor Type, GPU ID