Abstract
In this work we present a method for Named Entity Recognition (NER). Our method does not rely on complex linguistic resources, and apart from a hand coded system, we do not use any language-dependent tools. The only information we use is automatically extracted from the documents, without human intervention. Moreover, the method performs well even without the use of the hand coded system. The experimental results are very encouraging. Our approach even outperformed the hand coded system on NER in Spanish, and it achieved high accuracies in Portuguese.
Introduction
Named Entity Recognition (NER)—the task of identifying and categorizing entities like people, locations, and organizations in text—is a cornerstone of Natural Language Processing. Traditionally, these systems have relied heavily on hand-coded rules, regular expressions, and exhaustive dictionaries known as gazetteers. However, this approach faces a significant "portability problem": a system built for one language or domain often performs poorly when applied to another.
Adapting these systems manually is labor-intensive and expensive, and many languages lack the digital linguistic resources required to build them from scratch. This research explores a way to "recycle" existing tools from a source language (Spanish) to assist in a target language (Portuguese) with minimal human intervention, leveraging machine learning to bridge the linguistic gap.
Methodology
The proposed approach utilizes a two-step NER process using supervised machine learning: Named Entity Delimitation (NED), which identifies entity boundaries using the BIO (Beginning, Inside, Outside) scheme, and Named Entity Classification (NEC), which assigns categories to those entities.
Feature Engineering
The model employs two distinct categories of features to inform the classifier:
- Internal Features: These are automatically extracted from the text and include the word itself, its sentence position, and orthographic characteristics (e.g., capitalization, presence of digits, or punctuation).
- External Features: This novel component involves running a hand-coded Spanish NER system over the target text. The resulting Part-of-Speech (POS) tags and predicted BIO tags are used as input features for the machine learning model.
The Algorithm
A modified version of the C4.5 decision tree algorithm (implemented via WEKA) was used for classification. The model was trained using a sliding window of five words to capture essential local context.
Experimental Setup
The study utilized two primary corpora:
- Spanish: The CoNLL-2002 competition dataset, containing over 20,000 training entities.
- Portuguese: The HAREM corpus, consisting of newspaper articles with 648 named entities.
Results and Discussion
The hybrid approach, combining internal and external features, consistently yielded the best performance across both languages.
In Spanish Delimitation (NED), the proposed method achieved an F1-score of 93.7, actually surpassing the standalone hand-coded system (91.7). This suggests that the machine learning model learned to identify and correct patterns where the expert rules typically failed.
In Portuguese, the method reached an F1-score of 87.6 for delimitation. The most challenging task was Portuguese Classification (NEC). While the standalone Spanish system performed poorly on Portuguese classification (11.0 F1), the proposed method successfully integrated features to reach an F1-score of 50.3.
Key Insights
- Cross-Linguistic Synergy: Because Spanish and Portuguese are linguistically related, structural cues like trigger words and POS sequences were similar enough for the model to exploit the "noisy" signal from the foreign tagger.
- Minimalism: High accuracy was achieved using surface-level features (capitalization, suffixes) rather than deep semantic analysis, suggesting NER relies heavily on recognizable orthographic patterns.
Conclusion
While the method showed strong results, limitations remain. The Portuguese HAREM dataset utilized 10 entity classes compared to the 4 recognized by the Spanish system, creating a "class mismatch" that limited classification accuracy. Furthermore, data scarcity in certain categories led to zero recognition for rare classes. Future work will investigate whether this "tagger exploitation" remains effective between less similar language pairs, such as Spanish and English.
Summary updated