cs.CLOct 29, 2025

TOPol: Reconstructing and Interpreting Multidimensional Semantic Polarity Fields

Authors: Lucia Gomez, Gabin Taibi, Julius Sipke Kooistra

Organizations: University of Geneva · Bern Unversity of Applied Sciences · University of Twente · LGT Bank

Abstract

Polarity in natural language processing has long been operationalized as a single axis: the sentiment construct. We challenge this one-dimensional view by asking what a topic's documents are bipolar about, without specifying an answer in advance. TOPol inverts the usual concept-direction paradigm: rather than mapping a named construct to one direction, it maps a topic to the full set of statistically reproducible bipolar semantic axes its documents support in a transformer's residual stream, and asks only afterwards which of them correspond to a known construct. These axes need not be evaluative. The method is a recursive, orthogonal-by-construction bipartition of each topic's document cloud, with a candidate axis promoted to a finding only if it survives one pre-specified split-half stability gate. We calibrate against a known answer first: on 9,986 Amazon reviews (14 discovered topics) in Gemma 3 4B, the star-rating difference-in-means vector serves as a ruler known by construction to be sentiment. Discovery blindly re-derives the sentiment axis in every topic (cosine 0.71-0.97) without ever seeing the rating label, while the same gate certifies 543 axes, of which 106 are near-orthogonal to sentiment yet equally reproducible; each topic supports 33-45 mutually orthogonal axes, and sentiment axes resolve in later layers than the rest. Differential vocabulary, an internals-grounded natural-language decoder and vocabulary-based clustering agree on what the non-sentiment axes contrast; clustering by shared vocabulary surfaces 99 families, 9 recurring across topics. Direction similarity to the ruler predicts an independent sentiment estimate only weakly (r=0.243r=0.243). Polarity is better treated as a property of a topic's representational geometry, with a measurable rank and a per-axis depth, than as one universal scalar.

Figures & tables

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Apr 8, 2026

SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA)

We present the SemEval-2026 shared task on Dimensional Aspect-Based Sentiment Analysis (DimABSA), which improves traditional ABSA by modeling sentiment along valence-arousal (VA) dimensions rather than using categorical polarity labels. To extend ABSA beyond consumer reviews to public-issue discourse (e.g., political, energy, and climate issues), we introduce an additional task, Dimensional Stance Analysis (DimStance), which treats stance targets as aspects and reformulates stance detection as regression in the VA space. The task consists of two tracks: Track A (DimABSA) and Track B (DimStance). Track A includes three subtasks: (1) dimensional aspect sentiment regression, (2) dimensional aspect sentiment triplet extraction, and (3) dimensional aspect sentiment quadruplet extraction, while Track B includes only the regression subtask for stance targets. We also introduce a continuous F1 (cF1) metric to jointly evaluate structured extraction and VA regression. The task attracted more than 400 participants, resulting in 112 final submissions and 42 system description papers. We report baseline results, discuss top-performing systems, and analyze key design choices to provide insights into dimensional sentiment analysis at the aspect and stance-target levels. All resources are available on our GitHub repository.
Apr 8, 2026

SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization

We present SemEval-2026 Task 9, a shared task on online polarization detection, covering 22 languages and comprising over 110K annotated instances. Each data instance is multi-labeled with the presence of polarization, polarization type, and polarization manifestation. Participants were asked to predict labels in three sub-tasks: (1) detecting the presence of polarization, (2) identifying the type of polarization, and (3) recognizing the polarization manifestation. The three tasks attracted over 1,000 participants worldwide and more than 10k submission on Codabench. We received final submissions from 67 teams and 73 system description papers. We report the baseline results and analyze the performance of the best-performing systems, highlighting the most common approaches and the most effective methods across different subtasks and languages. The dataset of this task is publicly available.
Oct 4, 2026

Usage-Modulated Sentiment Representations in Large Language Models

Prior work suggests that sentiment can often be captured by approximately linear directions in LLM activation spaces, but a single direction may not fully capture sentiment representations. In natural communication, sentiment is shaped not only by polarity but also by usage factors, such as tone and audience adaptation. We test whether these factors systematically modulate sentiment representations beyond a shared sentiment direction. We construct a controlled paired dataset that holds event content fixed while varying sentiment polarity and usage factors, and analyze Llama, Mistral, and Gemma. We identify a shared sentiment direction, remove it, and test the residual structure through erasure and generation-time tone steering. Across models, the shared direction is robust (median cosine 0.953-0.975), yet removing it leaves 0.833-0.909 of the original positive-negative representation-difference norm. The residuals contain compact, reproducible usage-conditioned structure. Targeted erasure weakens held-out usage metrics more than random and label-shuffled controls. On Llama, outputs steered along residualized tone components are preferred in 92.8% of blind target-tone comparisons while preserving the requested sentiment polarity in 98.7% of evaluated outputs.