Reconocimiento de Entidades (NER) en Documentos Corporativos con Transformers

Named Entity Recognition (NER) is a fundamental technique in natural language processing that allows identifying and classifying specific elements within a text. This capability is invaluable for companies that need to analyze product mentions, locations or dates in user comments. Through Transformer models, we can implement powerful solutions that automate this process of extracting valuable information.

How to implement named entity recognition with Transformer?

To implement named entity recognition using Transformer, we need to follow a structured process that starts with the proper definition of the pipeline. It is crucial to select the right model according to the language and type of data we will be working with.

The first step is to import the Pipeline library and configure it correctly:

# We import Pipelinefrom transformers import pipeline
 # We configure the pipeline for NER in Spanishner = pipeline("ner",  model="mrm8488/bert-spanish-cased-finetuned-ner",  tokenizer="mrm8488/bert-spanish-cased-finetuned-ner").

It is important to note that we must define the pipeline according to the language (in this case Spanish) and the type of data we will process (product reviews). For NER tasks, we use a Spanish-specific BERT model that has been trained for entity recognition.

When working with Transformer models, it is recommended to use a GPU to optimize performance and processing speed.

How does entity recognition work in practice?

Let's look at a practical example to better understand how this process works:

# Review example review = "Samsung Galaxy S twenty-one product arrived on March twelfth and exceeded my expectations."
 # We apply the NER modelresult = ner(review)print(result).

By executing this code, we will get a dictionary with detailed information about the detected entities. The model identifies different types of entities:

• ORG: Organizations (companies, institutions).
• LOC: Locations (cities, regions, countries)
• MISC: Miscellaneous entities (events, works of art, abstract concepts).

How to interpret the results of the NER model?

The results of the NER model include special labels that help us to understand the structure of the entities:

• B-: Begin - Indicates the beginning of an entity.
• I-: Inside (intermediate) - Indicates the continuation of an entity
• #: Used to join tokens without spaces.

For example, in the above review, the model would detect "Samsung Galaxy S twenty-one" as an organization, with "Sam" marked as B-ORG (organization start) with 99% certainty, followed by tokens marked as I-ORG (continuation).

To reconstruct the complete entity, we must join all related tokens:

# reconstruction of thereconstructed entityreconstructed_entity = "Samsung Galaxy S twenty-one"print(f "Reconstructed entity: {reconstructed_entity}").

What are the practical applications of entity recognition?

Named entity recognition offers multiple benefits for text analysis:

• Monitoring brand and product mentions in reviews and comments.
• Identification of relevant dates in customer communications
• Automating the extraction process for further analysis
• Tracking products and names mentioned by users.

Let's look at some additional examples:

review1 = "I bought the HP laptop in Madrid, and the customer service was excellent."review2 = "The Canon EOS Rebel camera has impressive image quality, ideal for professionals."review3 = "The Casio watch I bought is water resistant and very accurate."
 # We apply the model to each reviewresult1 = ner(review1)result2 = ner(review2)result3 = ner(review3)

In these examples, the model would identify:

• "HP" as MISC (miscellaneous) entity and "Madrid" as LOC (location).
• "Canon EOS Rebel" as MISC entity
• "Casio" as MISC entity

Can the NER model identify new or recent entities?

A common question is whether these models can recognize entities that have recently emerged. The ability of NER models to identify new brands or products is surprisingly good, even with terms that did not exist during their training.

For example, if we analyze a review that mentions "Deep-Seek" (a recent AI model), the system could correctly identify it as a miscellaneous entity, demonstrating its ability to adapt to new terms.

review_new = "The recently released Deep-Seek model has impressive capabilities."new_result = ner(review_new)

This feature makes entity recognition a valuable tool for detecting trending topics, new brands and emerging products in the market.

Named entity recognition represents a powerful tool for extracting structured information from unstructured text. Its implementation with Transformer models facilitates the automatic analysis of large volumes of textual data, allowing companies to obtain valuable insights about their product mentions, competition and user preferences. Have you ever used this technology in your projects? Share your experience in the comments.