Sentence-Level Sentiment Classification

• Sentence-level sentiment classification is the process of assigning polarity labels to individual sentences, enabling fine-grained analysis of opinion and emotion.
• Techniques range from traditional feature engineering and rule-based systems to advanced neural sequence models and sentiment-aware pre-training.
• Practical challenges include handling weak supervision and noisy labels while effectively aggregating sentence-level cues for broader sentiment interpretation.

Sentence-level sentiment classification is the task of assigning a sentiment label to an individual sentence or snippet rather than to an entire document, review, or entity-specific discourse segment. In the literature represented here, the task appears in binary settings such as positive versus negative, in ternary settings that include neutral, and in fine-grained schemes such as five-way polarity or polarity–intensity labels. The area spans directly supervised sentence classification, weakly supervised transfer from document labels, lexicon- and rule-aware neural models, multilingual corpus construction, noisy-label learning, and extensions in which sentence-level sentiment acts as an intermediate representation for document-level or entity-level reasoning rather than the final prediction target (Dong et al., 2014, Qian et al., 2016, Wang et al., 2019, Thakkar et al., 2023, Fan et al., 2022).

1. Task formulation and label structures

Sentence-level sentiment classification is typically defined over isolated textual units such as sentences, clauses, tweets, or short review snippets. In the binary formulation, the label space is usually $\{\text{negative}, \text{positive}\}$, as in several review, Twitter, and noisy-label settings (Nguyen et al., 2017, Wang et al., 2019). Fine-grained formulations also appear. Stanford Sentiment Treebank is used in a five-way setup with $$\{\text{very negative}, \text{negative}, \text{neutral}, \text{positive}, \text{very positive}\}$$ in linguistically regularized LSTMs (Qian et al., 2016), and CroSentiNews 2.0 defines five sentence classes—negative, neutral, positive, mixed, and other / sarcasm—although its main experiments exclude mixed and sarcasm cases and focus on positive, negative, and neutral (Thakkar et al., 2023). Entity-specific sentence annotation can be இன்ன more granular: a Norwegian entity-level sentiment dataset uses the five-way polarity/intensity scheme Negative--Standard, Negative--Slight, Neutral, Positive--Slight, and Positive--Standard, with labels assigned per sentence per entity (Rønningstad et al., 2024).

A recurring distinction in the literature is between sentence-level sentiment as an end task and sentence-level sentiment as an auxiliary signal. Some studies classify standalone sentences directly, such as sentence-level Twitter sentiment, Croatian news sentiment, financial news sentence polarity, or Amazon review sentences (Nguyen et al., 2017, Thakkar et al., 2023, Lutz et al., 2018, Li et al., 2021). Other work uses sentence polarity to improve review-level, document-level, or entity-level prediction. Review-level polarity correction first labels sentences and then filters or reassigns them before final review classification (Orimaye et al., 2015). Document-level sentence-importance modeling estimates sentence contributions within a document rather than predicting standalone sentence labels (Choi et al., 2021). Entity-level sentiment in longer texts treats sentence-level labels as necessary but insufficient evidence for the overall sentiment toward an entity (Rønningstad et al., 2024).

This distinction matters because the same sentence representation machinery can support very different inferential goals. A standalone sentence classifier attempts to infer the sentence’s own polarity. By contrast, a document-level or entity-level model may use sentence-level information only as a latent or intermediate structure whose role is to support a higher-level label. A common misconception is therefore to equate any model that computes sentence representations with a sentence-level sentiment classifier. Several papers explicitly resist that interpretation (Choi et al., 2021, Rønningstad et al., 2024).

2. Early formalisms: parsing, feature engineering, and compositional sentiment

An early line of work treated sentence-level sentiment as a structured compositional problem rather than a flat classification task. A statistical parsing framework formulates a sentiment grammar $$\mathcal{G}_s = \langle V_s, \Sigma_s, S, R_s \rangle$$ with nonterminals for negative polarity, positive polarity, a start symbol, and an empty/OOV category, and uses a modified CYK algorithm to build sentiment parse trees directly from text (Dong et al., 2014). In that framework, the polarity model computes $P(\mathcal{N}\mid w_i^j)$ and $P(\mathcal{P}\mid w_i^j)$ for spans, while logistic-regression composition rules model negation, intensification, and contrast. The system is trained from sentences annotated only with overall polarity labels, without constituent-level sentiment annotations or syntactic parse supervision, and reports accuracies of 85.1 on RT-C, 79.5 on PL05-C, 84.7 on SST, 91.5 on RT-U, 89.3 on IMDB-U, and 86.2 on MPQA (Dong et al., 2014).

Feature-based sentence classification also remained important, particularly in morphologically rich or dialectally diverse settings. A sentence-level Arabic system for Modern Standard Arabic and Egyptian dialect combines an expandable sentiment lexicon, an idioms/sayings lexicon, negation handling, contextual intensifiers, question and wishful-expression features, position features, and syntactic conflict patterns, then feeds these into an SVM-light classifier with a linear kernel (Ibrahim et al., 2015). The lexicon ArSeLEX grows from 400 manually annotated adjectives to 5,244 sentiment words, comprising 2,003 positive, 2,829 negative, and 412 neutral entries, while the idiom resource manually selects 3,296 sentiment phrases from 12,785 sayings/idioms/wisdom expressions (Ibrahim et al., 2015). Reported results include 89.07% accuracy before lexicon expansion and 90.47% after expansion in one setting, and 94.31% before expansion and 95.12% after expansion in a second evaluation setting, with the paper highlighting accuracies of over 95% in the best setup (Ibrahim et al., 2015).

These models made explicit a central property of sentence-level sentiment: local lexical evidence is often insufficient. Negation can invert or damp polarity, intensifiers can strengthen it, and contrastive constructions can make the polarity of one clause dominate the entire sentence. This emphasis on composition persisted in later neural work, although the representational mechanisms changed substantially.

3. Neural sequence models, linguistic regularization, and sentiment-aware interactions

Neural sentence classifiers introduced sequence encoding and distributed representations while retaining a strong interest in sentiment composition. Linguistically regularized LSTMs augment LSTM and Bi-LSTM sentence models with four regularizers: a Non-Sentiment Regularizer for smoothness across adjacent non-opinion words, a Sentiment Regularizer that shifts local distributions according to sentiment-lexicon words, a Negation Regularizer based on learned transformation matrices for negators, and an Intensity Regularizer for intensifiers (Qian et al., 2016). The training objective combines sentence-level cross-entropy with linguistic regularization and $L_2$ regularization:

$$\mathcal{L}(\theta) = -\sum_i \hat{y}_i \log{y_i} + \alpha\sum_i\sum_t L_{t,i} + \beta||\theta||^2.$$

On MR, plain LSTM reaches 77.4, Bi-LSTM 79.3, LR-LSTM 81.5, and LR-Bi-LSTM 82.1, showing clear gains from lexicon-, negation-, and intensity-aware regularization without phrase-level annotations (Qian et al., 2016).

A related direction focused on explicit word interaction. Contextual Factorization Machine and Position-aware Factorization Machine reinterpret factorization machines for snippet and sentence sentiment by restricting interactions to local context windows and, in the position-aware case, making interaction vectors depend on relative distance (Wang et al., 2018). The models learn sentiment-sensitive Word Interaction vectors and report strong snippet-level performance. On movie snippets, FM obtains 0.607 accuracy / 0.540 F1, CFM 0.785 / 0.784, and PFM 0.789 / 0.788; on cellphone snippets, FM 0.745 / 0.731, CFM 0.822 / 0.821, and PFM 0.833 / 0.833; on restaurant snippets, FM 0.698 / 0.674, CFM 0.838 / 0.837, and PFM 0.842 / 0.842 (Wang et al., 2018). The reported pattern is that explicit sentiment-oriented word interaction helps especially at snippet or sentence granularity, where a small number of compositional cues can determine the label.

Twitter sentiment classification highlighted additional problems: noisy orthography, short context, emoticons, hashtags, usernames, and contrastive discourse markers. A hybrid model for Twitter first applies semantic rules involving but, despite, unless, while, and however, then combines character-level DeepCNN representations with word embeddings and a Bi-LSTM sentence encoder (Nguyen et al., 2017). The model keeps emoticons, uses wide convolution at the character level, concatenates character and word features as $v_i = [r_i; e_i]$, and predicts sentiment with a softmax layer. The best reported accuracies are 86.63 on STS with (DeepCNN + Bi-LSTM) + SR + GloVe, 85.14 on Sanders with the same configuration, and 80.9 on HCR with (DeepCNN + Bi-LSTM) + SR + GoogleW2V (Nguyen et al., 2017).

Lexicon integration also remained active in general neural sequence models. A multi-grained lexicon method enriches word representations with fine-grained sentiment embeddings and uses coarse-grained labels to enhance attention, incorporating sentiment lexicons, negation words, and intensifiers into neural sequence models for sentence-level sentiment classification (Zeng et al., 2018). The paper reports gains on SST-5 and MR, and states that an enhanced Bi-LSTM can compare with a Tree-LSTM using expensive phrase-level annotations (Zeng et al., 2018).

4. Contextualized representations, pre-training, and reproducibility

The shift from static embeddings to contextualized representations changed both modeling and evaluation practice. A reanalysis of logic-rule encoding on SST-2 revisits A-but-B sentences and negations, using 100 random seeds rather than the single-run or small-run reporting common in earlier work (Krishna et al., 2018). The central logic rule is that in an A-but-B sentence, the sentiment of the whole sentence should match the sentiment of clause B. After proper averaging, the study finds that projection helps more than distillation and that the distillation model is ineffective. Averaged over 100 seeds, a CNN baseline reaches 85.98 test accuracy and 78.69 on A-but-B sentences; adding projection yields 86.54 and 83.40, while distillation alone gives 86.11 and 79.04 (Krishna et al., 2018). Replacing explicit logic rules with ELMo yields 88.89 test accuracy and 86.51 on A-but-B sentences, with projection adding little further improvement (Krishna et al., 2018). The paper’s methodological claim is equally important: test accuracy for the CNN baseline ranges from 83.47 to 87.20 across 100 seeds, implying that small reported gains may be artifacts of initialization variance rather than robust model improvements (Krishna et al., 2018).

Sentiment-aware pre-training then pushed sentence-level performance further by incorporating both lexical and sentence-level sentiment objectives. SentiWSP combines word-level replaced sentiment word detection with sentence-level contrastive learning (Fan et al., 2022). At the word level, a generator-discriminator framework masks sentiment words, samples replacements, and trains a discriminator to detect whether tokens are original or replaced. At the sentence level, the discriminator acts as the encoder in a contrastive setup: a query sentence is constructed by masking sentiment words, the original sentence serves as the positive example, and similar sentences retrieved through ANN serve as hard negatives (Fan et al., 2022). The best settings reported are 50% sentiment-word masking in word-level pre-training, 70% masking in sentence-level query construction, and hard-negative sampling with $k = 100$ candidate neighbors and $t = 7$ sampled negatives (Fan et al., 2022). On sentence-level benchmarks, the reported accuracies are 96.26 on IMDB, 59.32 on SST-5, 98.25 on Yelp-2, 71.69 on Yelp-5, and 92.41 on MR, with the paper stating new state-of-the-art performance on all five sentence-level sentiment classification benchmarks (Fan et al., 2022).

This progression suggests a notable conceptual shift. Earlier work often injected sentiment knowledge through hand-engineered rules, lexicons, or structural formalisms. Later work increasingly treated sentiment structure as something that contextualized encoders and specialized pre-training objectives can learn implicitly, provided that training objectives expose lexical sentiment corruption and sentence-level polarity contrasts.

5. Weak supervision, noisy labels, and cross-domain transfer

Sentence-level sentiment labels are expensive, and several lines of work therefore ask how far one can go with weak or noisy supervision. In financial news, distributed sentence representations and multi-instance learning are used to transfer document-level labels derived from abnormal stock returns to unlabeled sentences (Lutz et al., 2018). Each announcement is treated as a bag of sentence instances, sentence vectors are obtained with doc2vec, and a multi-instance objective combines sentence-similarity regularization with document-level prediction error:

$$\{\text{very negative}, \text{negative}, \text{neutral}, \text{positive}, \text{very positive}\}$$0

On a manually labeled set of 1,000 financial sentences, the model achieves 69.90% accuracy, 67.80% recall, 70.77% precision, and 69.25% F1, outperforming the best alternative sentence-level baseline by 3.80 percentage points in accuracy (Lutz et al., 2018). A descriptive finding reinforces the motivation: 97.57% of financial documents contain a mix of positive and negative sentences (Lutz et al., 2018).

Noisy sentence labels caused by inheriting review-level labels are addressed directly in NetAb, a dual-CNN architecture for sentence-level sentiment classification under noisy labels (Wang et al., 2019). One CNN, the A-network, predicts clean labels; the other, the Ab-network, models noisy labels through a learned transition matrix $$\{\text{very negative}, \text{negative}, \text{neutral}, \text{positive}, \text{very positive}\}$$1 and is trained with a noisy loss. Training alternates between the two networks in a mutual reinforcement scheme after pre-training the A-network for about 5 epochs (Wang et al., 2019). The experiments consider both synthetic noise rates of 0, 0.1, 0.2, 0.3, 0.4, and 0.5 and real noisy training data obtained by assigning review ratings to sentences. NetAb reports the best overall accuracy on real noisy labels for Movie (0.7047), Laptop (0.7928), and Restaurant (0.8593) (Wang et al., 2019).

A different route to weakly supervised or transfer-based sentence sentiment uses translation as a task-specific intermediate representation. Machine-oriented neural machine translation adapts a Transformer so that the reward is not BLEU or human translation adequacy but the downstream English sentiment classifier’s probability for the correct sentiment label (Tebbifakhr et al., 2019). In German→English sentiment classification on translated tweets, MO-Reinforce reaches F1 80.9 with 5% parallel data and 84.4 with 100% data, compared with 79.7 and 83.2 for a generic MT system; in Italian→English, it reaches 80.3 and 84.5 versus 78.2 and 81.6 (Tebbifakhr et al., 2019). The reported English upper bound is 85.1, so the machine-oriented translation approach comes within 0.7 F1 on German and 0.6 on Italian in the full-data condition (Tebbifakhr et al., 2019).

These results illustrate three distinct supervision strategies: direct sentence annotation, transfer from bag labels through multi-instance learning, and downstream-task feedback through translation adaptation. All three assume that sentence-level polarity can be learned even when clean sentence labels are incomplete, noisy, or entirely absent, but they differ sharply in how they constrain the learning problem.

6. Datasets, annotation regimes, and evaluation practice

The field relies on a heterogeneous dataset landscape, with differences in language, domain, label granularity, and annotation procedure. CroSentiNews 2.0 extends Croatian news sentiment resources from document level to sentence level, yielding 1,988 unique non-empty documents and 14,570 sentence-level sentiment-labeled instances after filtering (Thakkar et al., 2023). Annotation is performed in INCEpTION by native Croatian-speaking undergraduate linguistics students; the five-label scheme includes mixed and sarcasm, but 428 mixed-language cases and 73 sarcasm instances are excluded from experiments, and majority vote determines final labels (Thakkar et al., 2023). Agreement is reported with Fleiss’ $$\{\text{very negative}, \text{negative}, \text{neutral}, \text{positive}, \text{very positive}\}$$2, qualitatively ranging from moderate (0.41–0.60) to substantial (0.61–0.80), and the best Croatian sentence-level baseline reaches macro-F1 79.65 using multitask learning over Croatian document and Croatian sentence supervision with CroSloEngual BERT (Thakkar et al., 2023).

Entity-centric annotation in longer texts complicates the sentence-level picture further. A Norwegian dataset built from the Norwegian Review Corpus contains 412 texts, 11,335 sentences, 2,479 entities, and 9,891 annotations in total, with train/test splits of 368/44 texts and 2,232/247 entities after cleaning (Rønningstad et al., 2024). Annotators assign sentence-level sentiment per entity, sometimes marking a target span and relation type—name_mention, anaphoric, is_member, has_member, or created_by—and sometimes labeling the entire sentence when no target span can be identified (Rønningstad et al., 2024). Document-level mean Cohen’s $$\{\text{very negative}, \text{negative}, \text{neutral}, \text{positive}, \text{very positive}\}$$3 is 0.71 and sentence-level mean Cohen’s $$\{\text{very negative}, \text{negative}, \text{neutral}, \text{positive}, \text{very positive}\}$$4 is 0.72 (Rønningstad et al., 2024). The dataset demonstrates that naive aggregation of sentence-level sentiment over entity mentions fails frequently: only 70% of positive and 55% of negative entities receive the correct overall sentiment label under mention-based aggregation, and 271 of 1,073 non-neutral entities are incorrectly assigned Neutral in the mention-only condition (Rønningstad et al., 2024).

Evaluation metrics also vary with label structure and dataset skew. Accuracy is common in binary or balanced settings, as in many Twitter, financial, and noisy-label studies (Nguyen et al., 2017, Lutz et al., 2018, Wang et al., 2019). Macro-F1 is preferred when class imbalance is substantial, as in CroSentiNews 2.0 (Thakkar et al., 2023). Some work reports classwise F1, confusion matrices, or recovered-label proportions instead of correlation-style measures, especially when comparing aggregation strategies (Rønningstad et al., 2024). The reproducibility study on SST-2 adds another methodological requirement: results should be averaged over far more random seeds than has traditionally been reported, because high variance across initializations can mask or fabricate small gains (Krishna et al., 2018).

7. Sentence-level sentiment beyond isolated sentences: long texts, aggregation, and interpretability

Recent work has emphasized that sentence-level sentiment is often an indispensable intermediate structure even when the final task is broader. In document-level sentiment classification, a model based on ALBERT sentence embeddings, sentence-level class-similarity embeddings, sigmoid gates for sentence importance, GRU encoding, and Luong attention predicts document sentiment while learning a scalar importance $$\{\text{very negative}, \text{negative}, \text{neutral}, \text{positive}, \text{very positive}\}$$5 for each sentence (Choi et al., 2021). Across IMDB, Yelp-hotel, Yelp-rest, and Amazon music reviews, the proposed model outperforms prior baselines and ALBERT-Base, with reported accuracies of 0.548 versus 0.519 on IMDB, 0.833 versus 0.827 on Yelp-hotel, 0.882 versus 0.874 on Yelp-rest, and 0.876 versus 0.858 on Amazon (Choi et al., 2021). The largest ablation drop comes from removing the gated sentence embedding, supporting the claim that sentence importance is the most critical component (Choi et al., 2021).

A related but more explicitly conflict-oriented perspective divides text into constituents and aggregates their predicted sentiments. A 2025 study uses sentence boundaries, clause extraction, or aspect-based sentiment analysis to build an $$\{\text{very negative}, \text{negative}, \text{neutral}, \text{positive}, \text{very positive}\}$$6 matrix of constituent predictions, then reduces it to a three-class decision through averaging, Average WithOut Neutral sub-sentences (AWON), or an MLP over 19 summary features (Kościałkowski et al., 8 May 2025). The strongest results come from the MLP. On SST, Amazon, and Twitter, MLP over sentence splits yields accuracies of 0.71, 0.80, and 0.78, while MLP over ABSA outputs yields 0.70, 0.83, and 0.77; corresponding macro-F1 scores are 0.57, 0.56, and 0.70 for sentence-based MLP and 0.56, 0.59, and 0.68 for ABSA-based MLP (Kościałkowski et al., 8 May 2025). The paper explicitly argues that longer passages exacerbate conflicting-tone problems and that constituent aggregation can be much cheaper than fine-tuning a base transformer, with a stated cost of about $$\{\text{very negative}, \text{negative}, \text{neutral}, \text{positive}, \text{very positive}\}$$7 of fine-tuning the baseline (Kościałkowski et al., 8 May 2025).

Interpretability work on Amazon product reviews also uses sentence-level modeling as a way to inspect what the classifier attends to. A BiLSTM with attention and sentiment lexicon weighting based on WordNet and SentiWordNet reports 96.0% accuracy, 96.0% precision, 99.9% recall, and 97.9% F1 on Amazon Musical Instruments reviews (Li et al., 2021). The study constructs a top-160 aspect-term set using TF-IDF over nouns and observes that aspect terms such as fender, pedal, string, mic, cable, price, and quality can receive the same or even higher attention than overt sentiment words such as great, nice, best, pretty, and fine (Li et al., 2021). This suggests that sentence-level sentiment classification in review domains often depends on aspect anchors as much as on obvious polarity terms.

The broader implication is that sentence-level sentiment classification no longer functions only as a standalone benchmark task. It also serves as a compositional substrate for document-level decision-making, entity-level reasoning, constituent conflict resolution, and interpretability analyses. At the same time, the long-text and entity-centric literature shows that arithmetic aggregation of sentence labels is often insufficient. Sentence-level sentiment is informative, but its contribution depends on discourse structure, reference resolution, bridging relations, indirect sentiment propagation, and learned importance weighting (Rønningstad et al., 2024, Choi et al., 2021).

Sentence-level sentiment classification has therefore evolved from feature-engineered polarity detection toward a family of models and resources that address compositionality, weak supervision, noise robustness, contextual representation learning, multilingual annotation, and structured aggregation. Across these strands, the enduring technical problem is the same: determining how local lexical and clause-level evidence composes into sentence polarity, and when that sentence-level polarity itself is the correct endpoint rather than an intermediate layer in a larger sentiment inference pipeline.