Ensemble AI Nearly Perfectly Spots Cause and Effect in Sentences

When a sentence says that one event made another happen, most human readers grasp the connection instantly. For machines, however, detecting that a phrase such as “the storm caused the power outage” expresses genuine causation, rather than mere correlation or coincidence, remains one of the stubbornly difficult problems in natural language processing. A new study published in the Journal of Big Data by researchers at Amrita Vishwa Vidyapeetham in Coimbatore, India, together with a collaborator at Universitat Pompeu Fabra in Barcelona, reports that carefully assembled teams of machine learning models can push this task to near-perfect accuracy, and that explainable artificial intelligence can reveal exactly why those models reach their decisions.

The research, led by Rohith Chemmancheri Madathil and corresponding author Sachin Kumar Sukumaran, tackles what computer scientists call causality detection in sentences: the automatic identification of causal relationships expressed in natural language text. The stakes are higher than they might first appear. Financial analysts need to trace which market events triggered which outcomes, healthcare systems need to connect treatments with effects, safety engineers need to understand what led to accidents, and policymakers need evidence chains that link actions to consequences. In each of these domains, text is the primary carrier of causal claims, and misreading them can translate directly into bad decisions.

What makes the new work distinctive is its breadth of comparison. Rather than championing a single method, the team systematically pitted ensemble approaches, which combine several models into a voting committee, against non-ensemble baselines across three fundamentally different ways of representing language. The first category relied on traditional embeddings: TF-IDF and Bag-of-Words, which count words and phrases statistically, plus Word2Vec, FastText, and GloVe, which represent words as dense numerical vectors learned from large text corpora. The second category used deep learning architectures, including convolutional neural networks (CNN), recurrent networks such as LSTM, GRU, and BiLSTM, and hybrid CNN-GRU and CNN-LSTM designs that combine local pattern detection with sequential memory. The third and most modern category employed transformer-based embeddings from models including BERT, RoBERTa, BART, MiniLM, DistilBERT, and DeBERTa, the family of architectures that underpins most contemporary language technology.

On the ensemble side, the researchers built a framework integrating Random Forest, Naive Bayes, XGBoost, and Linear Support Vector Machines, combining their votes through soft voting and weighted-soft voting schemes. In soft voting, each model contributes its probability estimate that a sentence is causal, and the framework aggregates those probabilities; in weighted-soft voting, more reliable models receive greater influence over the final verdict. The non-ensemble baselines included Support Vector Machines, Random Forest, XGBoost, and a Neural Tangent Kernel approach for the traditional embeddings, while an SVM with a radial basis function kernel (SVM-RBF) served as the comparator for the deep learning and transformer representations.

The evaluation was deliberately rigorous. Experiments ran on four publicly available datasets, with performance measured using five standard metrics: F1-score, which balances precision and recall; precision itself, which measures how often flagged sentences are truly causal; recall, which measures how many true causal sentences are caught; accuracy; and ROC-AUC, which captures the trade-off between true and false positive rates across decision thresholds. The team applied hyperparameter tuning for the non-ensemble approaches and used 10-fold cross-validation, a technique in which the data is split into ten parts and the model is trained and tested ten separate times so that results do not depend on one lucky or unlucky split.

The headline result is striking: the ensemble-based approaches consistently outperformed their non-ensemble counterparts, achieving F1-scores of 99.98 percent, 98.38 percent, 99.72 percent, and 99.99 percent across the four datasets. Scores at that level mean the models made almost no errors in distinguishing causal from non-causal sentences on the test data. The consistency of the advantage across all three embedding paradigms suggests that the benefit comes from the ensemble principle itself, in which the diverse errors of individual classifiers cancel each other out, rather than from any particular choice of text representation. It also indicates that even relatively simple, interpretable classifiers such as Naive Bayes and Linear SVMs, when combined intelligently, can rival the raw power of much larger transformer-based systems on this specific task.

High benchmark scores alone, however, are no longer enough for researchers who want their models trusted in high-stakes settings. A model that flags a sentence as causal in a medical report or a financial disclosure needs to show its reasoning. To that end, the study employed three explainable AI techniques: LIME, Anchor, and Counterfactuals. LIME, which stands for Local Interpretable Model-agnostic Explanations, works by perturbing an input sentence and observing how the model’s prediction changes, thereby building a simple local approximation of which words drove the decision. Anchor methods go a step further by identifying the minimal set of words that, if preserved, guarantee the model’s output with high confidence. Counterfactual explanations ask the inverse question: what minimal change to the sentence would flip the model’s judgment from causal to non-causal? Together, these tools provide complementary windows into the internal logic of causality detectors.

The implications reach well beyond the leaderboard. Because the ensemble framework combines lightweight classifiers, it can plausibly be deployed in settings where the computational cost of running large transformer models on every incoming document would be prohibitive, such as continuous monitoring of news feeds for risk management or screening of clinical narratives. Meanwhile, the explainability layer addresses a persistent obstacle to adoption: domain experts in medicine, finance, and safety engineering are typically unwilling to act on a machine’s judgment without understanding the evidence behind it. By surfacing which specific words and constructions trigger a causal classification, the study’s interpretability methods make it possible for a human analyst to verify, challenge, or override the machine’s conclusion.

The study also contributes a methodological map for future researchers. By organizing the field into traditional embeddings, deep learning architectures, and transformer embeddings, and by testing both ensemble and non-ensemble classifiers within each, the authors have produced one of the more complete comparative pictures of causality detection to date. The finding that ensembles dominate across all three paradigms gives practitioners a clear default strategy, while the open publication of the work, released as open access under a Creative Commons license, means other teams can reproduce and extend the experiments. The article was received in September 2025, accepted in September 2026, and published on 1 October 2026, with the authors declaring no competing interests.

Causality, philosophers have argued for centuries, is the invisible thread that binds events into explanations. Teaching machines to see that thread in ordinary language has long been a benchmark of genuine language understanding, and the Amrita-led team’s results suggest that the combination of ensemble learning, rich text representations, and transparent explanations has brought that goal measurably closer. If the near-perfect scores hold up as the models encounter messier real-world text, the technology could soon become a quiet but essential layer in the systems that professionals use to read, triage, and act upon the flood of documents in which cause and effect are claimed every day.

Subject of Research: Machine learning methods for detecting causal relationships in sentences using ensemble and non-ensemble classifiers across embedding paradigms with explainable AI

Article Title: Detection of causality in sentences: a comparative study of ensemble and non-ensemble approaches across embedding paradigms with explainable AI

Article References: Chemmancheri Madathil, R., Kolangaraparambil Biju, S., Mohan, N., Kar, M. K., Okkath Krishnanunni, S., & Sukumaran, S. K. (2026). Detection of causality in sentences: a comparative study of ensemble and non-ensemble approaches across embedding paradigms with explainable AI. Journal of Big Data. https://doi.org/10.1186/s40537-026-01553-y

Image Credits: AI Generated

DOI: 10.1186/s40537-026-01553-y

Keywords: causality detection, natural language processing, ensemble learning, word embeddings, transformers, deep learning, explainable AI, LIME, machine learning, text classification, XGBoost, Journal of Big Data