Guilherme Fonseca, Gabriel Prenassi, Washington Cunha, Leonardo Chaves Dutra da Rocha, Marcos André Gonçalves. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Guilherme Fonseca, Washington Cunha, Gabriel Prenassi, Marcos André Gonçalves, Leonardo Chaves Dutra Da Rocha. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

We tackle Extreme Multi-Label Text Classification (XMTC), which involves assigning relevant labels to texts from a huge label space. Attempting to optimize the underexplored tail-head trade-off, we address the XMTC task through its core challenges of volume, skewness, and quality by proposing xCoRetriev, a novel two-stage retrieving and fusing ranking pipeline. Our pipeline addresses the volume challenge by dynamically slicing the large label space; it also tackles the skewness challenge by favoring the tail labels while fusing sparse and dense retrievers. Finally, xCoRetriev faces the quality challenge by enhancing the label space with Retrieval-Augmented Generated (RAG)-labels. Our experiments with four XMTC benchmarks with hundreds of thousands of text documents and labels against six state-of-the-art XMTC baselines demonstrate xCoRetriev's strengths in terms of: (i)~scalability for large label spaces, being among the most efficient methods at training and prediction; (ii)~effectiveness in the face of high skewness, with gains of up to 48% in propensity-scored metrics against the best state-of-the-art baselines; and (iii)~capability of handling very noisy datasets by exploiting RAG-labels.

Fine-tuning transformer-based deep-learning models are currently at the forefront of natural language processing (NLP) and information retrieval (IR) tasks. However, fine-tuning these transformers for specific tasks, especially when dealing with ever-expanding volumes of data, constant retraining requirements, and budget constraints, can be computationally and financially costly, requiring substantial energy consumption and contributing to carbon dioxide emissions. This article focuses on advancing the state-of-the-art (SOTA) on instance selection (IS)—a range of document filtering techniques designed to select the most representative documents for the sake of training. The objective is to either maintain or enhance classification effectiveness while reducing the overall training (fine-tuning) total processing time. In our prior research, we introduced the E2SC framework, a redundancy-oriented IS method focused on transformers and large datasets—currently the state-of-the-art in IS. Nonetheless, important research questions remained unanswered in our previous work, mostly due to E2SC’s sole emphasis on redundancy. In this article, we take our research a step further by proposing biO-IS— an extended bi - o bjective i nstance s election solution, a novel IS framework aimed at simultaneously removing redundant and noisy instances from the training. biO-IS estimates redundancy based on scalable, fast, and calibrated weak classifiers and captures noise with the support of a new entropy-based step. We also propose a novel iterative process to estimate near-optimum reduction rates for both steps. Our extended solution is able to reduce the training sets by 41% on average (up to 60%) while maintaining the effectiveness in all tested datasets, with speedup gains of 1.67 on average (up to 2.46x). No other baseline, not even our previous SOTA solution, was capable of achieving results with this level of quality, considering the tradeoff among training reduction, effectiveness, and speedup. To ensure reproducibility, our documentation, code, and datasets can be accessed on GitHub— https://github.com/waashk/bio-is .

Transformer-based deep learning is currently the state-of-the-art in many NLP and IR tasks. However, fine-tuning such Transformers for specific tasks, especially in scenarios of ever-expanding volumes of data with constant re-training requirements and budget constraints, is costly (computationally and financially) and energy-consuming. In this paper, we focus on Instance Selection (IS) - a set of methods focused on selecting the most representative documents for training, aimed at maintaining (or improving) classification effectiveness while reducing total time for training (or fine-tuning). We propose E2SC-IS -- Effective, Efficient, and Scalable Confidence-Based IS -- a two-step framework with a particular focus on Transformers and large datasets. E2SC-IS estimates the probability of each instance being removed from the training set based on scalable, fast, and calibrated weak classifiers. E2SC-IS also exploits iterative heuristics to estimate a near-optimal reduction rate. Our solution can reduce the training sets by 29% on average while maintaining the effectiveness in all datasets, with speedup gains up to 70%, scaling for very large datasets (something that the baselines cannot do).

Hierarchical Topic modeling (HTM) exploits latent topics and relationships among them as a powerful tool for data analysis and exploration.Despite advantages over traditional topic modeling, HTM poses its own challenges, such as (1) topic incoherence, (2) unreasonable (hierarchical) structure, and (3) issues related to the definition of the "ideal" number of topics and depth of the hierarchy.In this paper, we advance the stateof-the-art on HTM by means of the design and evaluation of CluHTM, a novel nonprobabilistic hierarchical matrix factorization aimed at solving the specific issues of HTM.CluHTM's novel contributions include: (i) the exploration of richer text representation that encapsulates both, global (dataset level) and local semantic information -when combined, these pieces of information help to solve the topic incoherence problem as well as issues related to the unreasonable structure; (ii) the exploitation of a stability analysis metric for defining the number of topics and the "shape" the hierarchical structure.In our evaluation, considering twelve datasets and seven stateof-the-art baselines, CluHTM outperformed the baselines in the vast majority of the cases, with gains of around 500% over the strongest state-of-the-art baselines.We also provide qualitative and quantitative statistical analyses of why our solution works so well.

Recent advances in text-related tasks on the Web, such as text (topic) classification and sentiment analysis, have been made possible by exploiting mostly the "rule of more": more data (massive amounts) more computing power, more complex solutions. We propose a shift in the paradigm to do "more with less" by focusing, at maximum extent, just on the task at hand (e.g., classify a single test instance). Accordingly, we propose MetaLazy, a new supervised lazy text classification meta-strategy that greatly extends the scope of lazy solutions. Lazy classifiers postpone the creation of a classification model until a given test instance for decision making is given. MetaLazy exploits new ideas and solutions, which have in common their lazy nature, producing altogether a solution for text classification, which is simpler, more efficient, and less data demanding than new alternatives. It extends and evolves the lazy creation of the model for the test instance by allowing: (i) to dynamically choose the best classifier for the task; (ii) the exploration of distances in the neighborhood of the test document when learning a classification model, thus diminishing the importance of irrelevant training instances; and (iii) a better representational space for training and test documents by augmenting them, in a lazy fashion, with new co-occurrence based features considering just those observed in the specific test instance. In a sizeable experimental evaluation, considering topics and sentiment analysis datasets and nine baselines, we show that our MetaLazy instantiations are among the top performers in most situations, even when compared to state-of-the-art deep learning classifiers such as Deep Network Transformer Architectures.

In this paper, we advance the state-of-the-art in topic modeling by means of a new document representation based on pre-trained word embeddings for non-probabilistic matrix factorization. Specifically, our strategy, called CluWords, exploits the nearest words of a given pre-trained word embedding to generate meta-words capable of enhancing the document representation, in terms of both, syntactic and semantic information. The novel contributions of our solution include: (i)the introduction of a novel data representation for topic modeling based on syntactic and semantic relationships derived from distances calculated within a pre-trained word embedding space and (ii)the proposal of a new TF-IDF-based strategy, particularly developed to weight the CluWords. In our extensive experimentation evaluation, covering 12 datasets and 8 state-of-the-art baselines, we exceed (with a few ties) in almost cases, with gains of more than 50% against the best baselines (achieving up to 80% against some runner-ups). Finally, we show that our method is able to improve document representation for the task of automatic text classification.

In this paper, we advance the state-of-the-art in topic modeling by means of the design and development of a novel (semi-formal) general topic modeling framework. The novel contributions of our solution include: (i) the introduction of new semantically-enhanced data representations for topic modeling based on pooling, and (ii) the proposal of a novel topic extraction strategy - ASToC - that solves the difficulty in representing topics in our semantically-enhanced information space. In our extensive experimentation evaluation, covering 12 datasets and 12 state-of-the-art baselines, totalizing 108 tests, we exceed (with a few ties) in almost 100 cases, with gains of more than 50% against the best baselines (achieving up to 80% against some runner-ups). We provide qualitative and quantitative statistical analyses of why our solutions work so well. Finally, we show that our method is able to improve document representation in automatic text classification.

In this paper, we propose a general framework that allows developers to filter, summarize and analyze user reviews written about applications on App Stores. Our framework extracts automatically relevant features from reviews of apps (e.g., information about functionalities, bugs, requirements, etc) and analyzes the sentiment associated with each of them. Our framework has three main building blocks, namely, (i) topic modeling, (ii) sentiment analysis and (iii) summarization interface. The topic modeling block aims at finding semantic topics from textual comments, extracting the target features based on the most relevant words of each discovered topic. The sentiment analysis block detects the sentiment associated with each discovered feature. The summarization interface provides to developers an intuitive visualization of the features (i.e., topics) and their associated sentiment, providing richer information than a 'star rating' strategy. Our evaluation shows that the topic modeling block is able to organize information provided by users in subcategories that facilitate the understanding of which features more positively/negatively impact the overall evaluation of the application. Regarding user satisfaction, we can observe that, in spite of the star rating being a good measure of evaluation, the Sentiment Analysis technique is more accurate in capturing the sentiment transmitted by the user by means of a comment.