Eraldo R. Fernandes

Latent trees for coreference resolution

We describe a structure learning system for unrestricted coreference resolution that explores two key modeling techniques: latent coreference trees and automatic entropy-guided feature induction. The latent tree modeling makes the learning problem computationally feasible because it incorporates a meaningful hidden structure. Additionally, using an automatic feature induction method, we can efficiently build enhanced nonlinear models using linear model learning algorithms. We present empirical results that highlight the contribution of each modeling technique used in the proposed system. Empirical evaluation is performed on the multilingual unrestricted coreference CoNLL-2012 Shared Task datasets, which comprise three languages: Arabic, Chinese and English. We apply the same system to all languages, except for minor adaptations to some language-dependent features such as nested mentions and specific static pronoun lists. A previous version of this system was submitted to the CoNLL-2012 Shared Task closed track, achieving an official score of 58.69, the best among the competitors. The unique enhancement added to the current system version is the inclusion of candidate arcs linking nested mentions for the Chinese language. By including such arcs, the score increases by almost 4.5 points for that language. The current system shows a score of 60.15, which corresponds to a 3.5% error reduction, and is the best performing system for each of the three languages.

A machine learning approach to Portuguese clause identification

In this work, we apply and evaluate a machine-learning-based system to Portuguese clause identification. To the best of our knowledge, this is the first machine-learning-based approach to this task. The proposed system is based on Entropy Guided Transformation Learning. In order to train and evaluate the proposed system, we derive a clause annotated corpus from the Bosque corpus of the Floresta Sinta(c)tica Project – an European and Brazilian Portuguese treebank. We include part-of-speech (POS) tags to the derived corpus by using an automatic state-of-the-art tagger. Additionally, we use a simple heuristic to derive a phrase-chunk-like (PCL) feature from phrases in the Bosque corpus. We train an extractor to this sub-task and use it to automatically include the PCL feature in the derived clause corpus. We use POS and PCL tags as input features in the proposed clause identifier. This system achieves a Fβ=1 of 73.90, when using the golden values of the PCL feature. When the automatic values are used, the system obtains Fβ=1=69.31. These are promising results for a first machine learning approach to Portuguese clause identification. Moreover, these results are achieved using a very simple PCL feature, which is generated by a PCL extractor developed with very little modeling effort.

ETL ensembles for chunking, NER and SRL

We present a new ensemble method that uses Entropy Guided Transformation Learning (ETL) as the base learner. The proposed approach, ETL Committee, combines the main ideas of Bagging and Random Subspaces. We also propose a strategy to include redundancy in transformation-based models. To evaluate the effectiveness of the ensemble method, we apply it to three Natural Language Processing tasks: Text Chunking, Named Entity Recognition and Semantic Role Labeling. Our experimental findings indicate that ETL Committee significantly outperforms single ETL models, achieving state-of-the-art competitive results. Some positive characteristics of the proposed ensemble strategy are worth to mention. First, it improves the ETL effectiveness without any additional human effort. Second, it is particularly useful when dealing with very complex tasks that use large feature sets. And finally, the resulting training and classification processes are very easy to parallelize.

Entropy-Guided feature generation for structured learning of portuguese dependency parsing

Feature generation is a difficult, yet highly necessary, subtask of machine learning modeling. Usually, it is partially solved by a domain expert that generates complex and discriminative feature templates by conjoining the available basic features. This is a limited and expensive way to obtain feature templates and is recognized as a modeling bottleneck. In this work, we propose an automatic method to generate feature templates for structured learning algorithms. The method receives as input the training dataset with basic features and produces a set of feature templates by conjoining basic features that are highly discriminative together. We denote this method entropy guided since it is based on the conditional entropy of local decision variables given the feature values. We illustrate our approach on the Portuguese dependency parsing task and report on experiments with the Bosque corpus. We show that the entropy-guided templates outperform the manually built templates used by MSTParser, which was the best performing system on the Bosque corpus up to now. Furthermore, our approach allows an effortless inclusion of two new basic features that automatically generate additional templates. As a result, our system achieves a per-token accuracy of 92.66%, what represents a reduction by more than 15% on the previous smallest error rate for Portuguese dependency parsing.

RelHunter: a machine learning method for relation extraction from text

We propose RelHunter, a machine learning-based method for the extraction of structured information from text. RelHunter’s key idea is to model the target structures as a relation over entities. Hence, the modeling effort is reduced to the identification of entities and the generation of a candidate relation, which are simpler problems than the original one. RelHunter fits a very broad spectrum of complex computational linguistic problems. We apply it to five tasks: phrase chunking, clause identification, hedge detection, quotation extraction, and dependency parsing. We compare RelHunter to token classification approaches through several computational experiments on seven multilingual corpora. RelHunter outperforms the token classification approaches by 2.14% on average. Moreover, we compare the derived systems against state-of-the-art systems for each corpus. Our systems achieve state-of-the-art performances for three corpora: Portuguese phrase chunking, Portuguese clause identification, and English quotation extraction. Additionally, the derived systems show good quality performance for the other four corpora.

Domain adaptation of POS taggers without handcrafted features

Unsupervised domain adaptation is an attractive option when labeled data is lacking for some domain of interest but is available for other domain. Part-of-speech (POS) tagging is often considered a solved task when enough labeled data is available in the domain of interest. However, when considering a domain adaptation scenario, this is far from true. Several approaches have been proposed for domain adaptation of POS taggers, however as far as we know, all of them are based on handcrafted features. In this work, we employ a machine learning method whose input is exclusively composed of the raw text. This method learns word- and character-level representations (embeddings), and has been successfully applied to intra-domain tasks. We show that this method achieves strong performances on the domain adaptation of English and Portuguese POS taggers.

Portuguese Part-of-Speech Tagging with Large Margin Structure Learning

Part-of-Speech Tagging is a fundamental task on many Natural Language Processing systems. This task consists in identifying the syntactic category, i.e. the part of speech, of each word in a sentence. Despite the fact that the current state-of-the-art accuracy for this task is around 97%, any improvement has an immediate impact on more complex tasks, like Parsing, Semantic Role Labeling and Information Extraction. Thus, it is still relevant to explore this task. In this paper, we introduce a part-of-speech tagger based on the Structure Learning framework that reduces the smallest known error on the Portuguese Mac-Morpho corpus by 7.8%. We also apply our tagger to a recently revised version of Mac-Morpho. Our system accuracy on this latter version is competitive with a semi-supervised Neural Network trained on Mac-Morpho plus a very large non-annotated corpus. Additionally, our system is simpler than previous systems and uses a very limited feature set. Our system employs a Large Margin training criteria to derive a structure predictor that is more robust on unseen data.