Emotional Textual Analysis

• Emotional textual analysis is a field that computationally studies affect in text by differentiating between coarse sentiment polarity and refined emotion categories with contextual roles.
• Methodologies include lexicon-based, rule-based, deep learning, and structured extraction techniques that capture polarity, appraisal, and causal structures in narratives.
• Applications span social media monitoring, literary analysis, and clinical assessments, emphasizing the need for context-sensitive modeling and robust annotation strategies.

Emotional textual analysis is the computational study of affect in written language, spanning coarse sentiment polarity, discrete emotion detection, dimensional affect modeling, and richer representations that encode who feels what, toward whom or what, under which conditions, and for what reason. Across natural language processing, computational literary studies, affective computing, and cognitively oriented annotation work, the topic has expanded from sentence-level positive/negative labeling to context-sensitive analyses of narrative progression, reader appraisal, interactional dynamics, and causal structure (Mohammad, 2013, Kim et al., 2018, Cortal et al., 2022).

1. Conceptual scope and theoretical orientation

A basic distinction in the field is between sentiment and emotion. In the general-purpose NLP formulation surveyed by Kim and Klinger, sentiment is often reduced to binary or ternary polarity, whereas emotion analysis seeks finer distinctions such as anger, fear, joy, sadness, disgust, and surprise (Kim et al., 2018). The engineering-oriented framework of Bostan and Klinger likewise treats sentiment analysis as assigning binary or ternary polarity and emotion detection as assigning one label from a broader affective inventory, often close to Ekman’s emotions, while also noting that practical systems frequently collapse complex affective structure into a reduced output representation for interoperability (Mohammad, 2013).

This reduction has been repeatedly criticized as insufficient for psychologically meaningful interpretation. In “Natural Language Processing for Cognitive Analysis of Emotions,” the central objection is that sentence- or document-level categorization obscures distinctions such as who feels the emotion, toward what or whom it is directed, why it arose, and which valued “territories” or objects are implicated (Cortal et al., 2022). Related critiques appear in reader-centric and interactional work: Persona-E$^2$ argues that emotion is not only a property of the writer or of the text, but also of the reader’s appraisal of an event, and Moltbook-based PSR modeling argues that emotion in agent interaction is conditioned by persona and stimulus rather than being a stand-alone property of a text snippet (Yang et al., 10 Apr 2026, Hasan et al., 19 May 2026).

A further conceptual divide concerns the validity target. “Manipulating emotions for ground truth emotion analysis” argues that many standard datasets provide third-person labels on text with unknown validity for the writer’s actual internal state, whereas experimentally manipulated author emotions offer a stronger criterion for evaluating text-based emotion inference (Kleinberg, 2020). This suggests that emotional textual analysis is not a single task but a family of related tasks whose object may be perceived emotion, expressed emotion, elicited reader emotion, or latent author state.

2. Representational schemes

The simplest representations are scalar polarity and subjectivity. In the analysis of blog and forum discussions, each comment is assigned a positive probability $P_{pos}$ and a subjective probability $P_{sub}$, both in $[0,1]$, and emotional dependence is then studied statistically over chains of comments rather than via hard labels (Weroński et al., 2011). This probabilistic representation preserves uncertainty and supports sequential analyses such as conditional probability ratios, PMI, and mutual information.

Categorical and dimensional schemes remain standard. The survey on computational literary studies reviews polarity, Ekman-style categories, Plutchik’s wheel, and dimensional models such as valence–arousal and valence–arousal–dominance (VAD) (Kim et al., 2018). In Moltbook, textual interactions are mapped to 28 GoEmotions-style categories and then projected into VAD coordinates through a manually specified lookup table, allowing emotions to be represented as points or sets of points in $\mathbb{R}^3$ (Hasan et al., 19 May 2026). AffectSpeech extends the representational agenda further by treating emotion in speech as a textualizable object described through six dimensions: sentiment polarity, open-vocabulary emotion captions, intensity level, prosodic attributes, prominent segments, and semantic content (Qi et al., 5 Apr 2026).

Structured role-based schemes push beyond flat labels. The CAE-based annotation framework defines the base roles cue, experiencer, target, and cause, then extends them with territory, object, attack, attacker, modifier, and negation (Cortal et al., 2022). In this design, emotion causation is not a free-form span only; it can be decomposed into what was threatened or changed, by whom, and through what kind of action. CauseMotion similarly formalizes emotional events as sextuplets $(h_j, t_j, a_j, o_j, p_j, r_j)$, encoding holder, target, aspect, opinion, sentiment, and rationale, and then reasons over causal links among these structures in long conversations (Zhang et al., 1 Jan 2025).

These representational choices are not merely technical. They determine whether the system models polarity, named emotion, appraisal, role structure, temporal prominence, or causal explanation. A plausible implication is that the development of emotional textual analysis has been driven less by a single dominant ontology than by repeated attempts to match representation to use case.

3. Corpora, annotation, and the problem of disagreement

A recurring constraint is the quality and design of annotated corpora. The CAE paper explicitly states that available annotated corpora are often small and homogeneous, motivating a new French dataset of autobiographical accounts of emotional scenes in a coaching setting (Cortal et al., 2022). Each account is organized into Facts, Emotions, Reasons, and Actions, which embeds emotion naming inside a broader reflective pipeline rather than isolating it as a sentence-level label. At the same time, the paper is explicit that annotation had not yet been completed, so it reports no corpus size, splits, label distributions, or inter-annotator agreement.

Other recent datasets deliberately move away from the single-gold-label assumption. Persona-E$^2$ contains 3,111 events and 111,996 annotations, with 36 labels per event, and grounds personality-conditioned emotional appraisal in measured MBTI and Big Five traits rather than role-played personas (Yang et al., 10 Apr 2026). Noblet’s evaluative French corpus reports strong disagreement, but argues that the disagreement follows stable statistical trends and can be modeled by LLMs, suggesting that annotation variability is driven by underlying linguistic features rather than by pure noise (Noblet, 1 Sep 2025). AffectSpeech, although speech-centered, is directly relevant because each of its 253,799 authentic English utterances is aligned with 1,522,794 textual descriptions created through a human–LLM collaborative pipeline, and those descriptions are further reformulated into six stylistic variants to reduce stylistic bias (Qi et al., 5 Apr 2026).

Long-context and multimodal corpora expand the scope of annotation. CauseMotion introduces ATLAS-6 as a benchmark for long-sequence emotional causal reasoning, with 20,000 synthetic long dialogues and 2,745 real-world long-sequence dialogues, all in the range of 70 to 300 turns, and annotates holder, target, aspect, opinion, sentiment, and rationale (Zhang et al., 1 Jan 2025). DreamNet uses 1,500 dream narratives annotated with 8 emotion categories and 12 semantic themes, framing emotion prediction as multilabel classification rather than single-label categorization (Panchagnula, 26 Feb 2025).

The status of disagreement is therefore a major methodological question. The experimental induction study shows that even when emotion is manipulated under random assignment, text-based measures recover only part of the true variance (Kleinberg, 2020). The evaluative corpus study argues that disagreement should not automatically be treated as annotation failure (Noblet, 1 Sep 2025). Together, these results weaken the assumption that emotional textual analysis can always be reduced to recovering one hidden correct label.

4. Methodological families

Lexicon-based and rule-based methods remain foundational because they are interpretable and portable. The general-purpose toolkit of Bostan and Klinger combines a lexicon-based emotion detector using WordNet-Affect, a symbolic dependency-based sentiment analyzer using the Liu lexicon and Stanford CoreNLP, and a Random Forest classifier, all normalized through EmotionML (Mohammad, 2013). Empath generalizes the lexicon paradigm by learning a neural embedding across more than 1.8 billion words of modern fiction, generating lexical categories from a small set of seed terms, and validating them through crowdsourcing; its categories are reported to be highly correlated with similar LIWC categories, with $r=0.906$ after validation (Fast et al., 2016). Social-media emotion work has also used hybrid lexicon-and-rule systems with emoticons, degree words, negation, grammatical dependencies, and person-aware weighting to classify tweets into Happiness, Sadness, Fear, Anger, Surprise, and Disgust (Gaind et al., 2019).

Structured extraction methods add syntax, co-reference, and semantic roles. The CAE prototype uses dependency parsing, part-of-speech tagging, co-reference resolution, and regular expression filtering, implemented with SpaCy and an external crosslingual coreference system, plus lexical resources such as WOLF, SentiWordNet, and the NRC Emotion Lexicon (Cortal et al., 2022). Its operational heuristics include regular-expression detection of first-person experiencers and passive-voice rules for identifying territories and attackers. CauseMotion extends the extraction agenda to long dialogues with RAG, sliding windows, and multimodal fusion, retrieving prior windows by cosine similarity and linking extracted emotional events through semantic, temporal, and rationale scores (Zhang et al., 1 Jan 2025).

Supervised and deep architectures dominate where task-specific labeled data are available. DreamNet uses RoBERTa-base, a bidirectional LSTM, and joint multilabel prediction heads for emotions and dream themes, reporting 92.1% accuracy and 88.4% F1-score in text-only mode (Panchagnula, 26 Feb 2025). In multimodal sentiment analysis, DEVA converts raw audio and visual emotion-relevant attributes into textual emotional descriptions and then uses text-guided progressive fusion, improving over strong baselines on MOSI, MOSEI, and CH-SIMS (Wu et al., 2024). In contact-center emotion recognition, DistilBERT is used as the text branch of a hybrid acoustic-textual system and is reported at 93.6% accuracy on a custom tweet-based dataset, although the text-specific methodology is under-specified (Wewelwala et al., 27 Mar 2025).

A common recent pattern is textualization of non-text signals. DEVA verbalizes prosodic and facial features into textual descriptions (Wu et al., 2024). AffectSpeech uses open-vocabulary natural-language descriptions as the primary representation of vocal affect (Qi et al., 5 Apr 2026). This suggests a representational convergence in which natural language itself becomes the medium for integrating heterogeneous affective evidence.

5. Context, interaction, narrative, and temporal structure

A central result of the literature is that emotional content is rarely independent across textual units. The analysis of blogs and BBC forums demonstrates that discussions “cannot be treated as random insertions of comments,” since consecutive comments exhibit measurable emotional correlations (Weroński et al., 2011). In the Blogs dataset, subjective clusters are reported as 26–36% larger than under global shuffling and about 7% larger than under thread shuffling, and mutual information for positivity falls from 4.53 in the original Blogs data to 0.05 under global shuffle (Weroński et al., 2011). This established an early statistical case for sequence-aware and thread-aware sentiment modeling.

Narrative research reaches a related conclusion through different methods. Mohammad’s work on novels and fairy tales introduces emotion word density and relative salience as ways to compare texts and trace emotional change, finding that fairy tales have a much wider range of emotion word densities than novels (Denis et al., 2013). “Pattern Recognition in Narrative” goes further by arguing that emotion is revealed not as a quality in its own right but through interaction and context, and models emotional relations in Casablanca and Madame Bovary through correspondence analysis and sequence-constrained hierarchical clustering rather than lexicon counts alone (Murtagh et al., 2014). The literary survey summarizes this broader tradition as one concerned with emotional arcs, character relations, and structural changes of sentiment over narrative time (Kim et al., 2018).

Interaction-centered frameworks generalize these ideas to contemporary AI and social settings. Persona–Stimulus–Reaction represents persona, stimulus, and reaction as affective objects in VAD space and classifies reactions as aligned, persona-consistent, stimulus-driven, transformative, or conflict-resolving (Hasan et al., 19 May 2026). Persona-E$^2$ shows that people with similar personality profiles agree more with one another than with out-group annotators, and reports positive Personality Agreement Gaps across Big Five and MBTI groupings (Yang et al., 10 Apr 2026). In long-form conversations, CauseMotion treats emotional analysis as causal-chain construction over dialogue history, explicitly targeting nonlocal triggers and delayed emotional consequences (Zhang et al., 1 Jan 2025). Across these studies, the isolated document is no longer the sole unit of analysis; thread, scene, segment, event, persona, and dialogue history all become relevant.

6. Applications, limitations, and open research questions

The application range is broad. CAE-oriented work targets coaching, self-understanding, and emotion management (Cortal et al., 2022). Literary studies use emotional textual analysis for affect-based search, genre comparison, historical analysis, and character-network interpretation (Denis et al., 2013, Kim et al., 2018). Social-media and platform studies apply it to moderation, stance-rich political discussion, and user-level emotional-state inference enriched by communication trees and profile signals (Moradbeiki et al., 11 Apr 2025, Weroński et al., 2011). Clinical and psychologically oriented applications include induced-emotion validation, depression-sensitive language analysis, and dream narrative analysis (Kleinberg, 2020, Vos et al., 12 Aug 2025, Panchagnula, 26 Feb 2025). Multimodal speech work uses textual emotion descriptions for captioning and synthesis control (Qi et al., 5 Apr 2026).

The limitations are equally prominent. Domain dependence remains a central obstacle: supervised systems often degrade sharply across domains, and symbolic systems require manually crafted rules whose maintenance cost rises with coverage (Mohammad, 2013). Several influential proposals are explicitly preliminary. The CAE dataset lacked completed annotation and therefore no quantitative evaluation of the extractor (Cortal et al., 2022). Moltbook PSR provides descriptive analyses but no in-domain annotation or formal validation of the emotional classifier (Hasan et al., 19 May 2026). AffectSpeech reports large-scale verification statistics but not conventional inter-annotator agreement coefficients (Qi et al., 5 Apr 2026). Even where ground truth is experimentally strengthened, text captures only part of actual emotion, with at most about one-third of the variance in induced happiness and about one-quarter in sadness explained by lexicon measures (Kleinberg, 2020).

A major open question concerns what exactly should count as the target. Writer-side emotion, reader-side appraisal, perceived emotion, and interaction-conditioned response are not equivalent (Yang et al., 10 Apr 2026, Kleinberg, 2020). Another concerns the level of structure required for useful analysis. Coarse polarity can be statistically informative, as in thread-level clustering (Weroński et al., 2011), but many use cases require semantic roles, causes, trajectories, or open-vocabulary descriptions (Cortal et al., 2022, Qi et al., 5 Apr 2026). A final recurring issue is that disagreement may be intrinsic to emotional interpretation rather than a defect to be eliminated (Noblet, 1 Sep 2025).

The trajectory visible across the literature is therefore not a simple replacement of lexicons by deep models. It is a progressive broadening of what counts as emotional evidence and what counts as an emotional unit: from polarity scores to roles, from isolated utterances to causal chains, from writer expression to reader appraisal, and from fixed labels to descriptive language. Emotional textual analysis, in this sense, is increasingly defined by its attempt to model emotionally meaningful textual phenomena as structured, contextual, and interpretable rather than merely classifiable.