Baselines and Bigrams: Simple, Good Sentiment and Topic Classification

article2012ACL1,364 citations

Demonstrates that simple Naive Bayes and Support Vector Machine baselines with word bigrams—particularly the hybrid NBSVM model—can outperform complex, structure-sensitive methods across standard sentiment and topic classification benchmarks.

Listen

Organizations often deploy complex and computationally expensive machine learning models to analyze text sentiment and topic categories, under the assumption that basic baseline algorithms cannot handle nuanced language. However, standard baselines vary widely in performance depending on their specific mathematical formulation, the input features provided, and the length of the processed text. Without well-tuned baseline benchmarks, teams risk investing substantial resources into complex architectures that may deliver inferior performance compared to simpler alternatives.

The main objective of the article is to systematically evaluate simple linear classification methods—specifically variants of Naive Bayes and Support Vector Machines—across multiple text lengths and classification tasks to establish rigorous, lightweight performance baselines.

The researchers conducted an empirical evaluation using standard benchmark datasets covering short sentence snippets, full-length movie reviews up to 50,000 documents, and topic-based newsgroup classifications. The analysis evaluated simple feature representations (single words and two-word pairs) without relying on external linguistic lexicons, rule systems, or complex parsing structures, testing standard algorithms alongside a hybrid model that uses Naive Bayes log-count ratios as features within a Support Vector Machine.

The key findings demonstrate that baseline choice depends heavily on document length and task type. First, for short text snippets, Multinomial Naive Bayes consistently outperformed Support Vector Machines, achieving top accuracies between 79.0% and 93.6% and beating complex rule-based and deep neural models. Second, for long reviews, Support Vector Machines outperformed Naive Bayes, while the hybrid model delivered the strongest results (89.45% and 91.22% accuracy), matching or setting new state-of-the-art benchmarks. Third, adding two-word phrase features consistently improved sentiment classification across all datasets, whereas they offered little to no benefit for topic-based classification. Finally, the hybrid model proved exceptionally robust across all dataset lengths and task types.

These findings imply that engineering teams can achieve state-of-the-art text and sentiment classification without the computational cost, implementation complexity, or runtime overhead of highly specialized architectures. The results challenge the assumption that short text fragments require intricate, hand-crafted linguistic rules or heavy deep learning models, showing instead that simple, well-tuned linear classifiers capture sentiment patterns more effectively at a fraction of the operational cost.

Decision-makers and practitioners should adopt the hybrid model as the standard, default baseline before greenlighting more complex natural language processing models. Teams should ensure that text representations use binary word indicators rather than raw frequency counts, include two-word phrases for sentiment analysis tasks, and rely on Multinomial rather than multivariate Bernoulli Naive Bayes formulations. If deploying the hybrid model, teams should set its interpolation parameter within the recommended range of one-quarter to one-half for stable general performance.

Confidence in these findings is high across the evaluated English benchmark tasks, given the consistent empirical validation across varying sample sizes. However, readers should note that the evaluations were restricted to standard linear models on English text classifications, meaning validation on proprietary or multilingual domain data remains necessary before full operational rollout.

Stanford University computational efficiency deep learning engineering sentiment analysis topic classification word bigram features

sidaw/nbsvm Wang et al (2012).pdf

Paper: Thumbs up? Sentiment Classification using Machine Learning Techniques, Bo Pang et al. (2002). This foundational study established the core benchmark of testing Naive Bayes and Support Vector Machines with n-gram features on movie review sentiment classification.
Paper: A comparison of event models for naive bayes text classification, Andrew McCallum et al. (1998). This paper establishes the formal distinction between multinomial and multivariate Bernoulli event models in Naive Bayes text classification, which directly motivates the source's baseline choices.
Paper: Learning Word Vectors for Sentiment Analysis, Andrew L. Maas et al. (2011). This work introduced the widely used 50,000-document IMDB benchmark dataset and evaluated baseline linear classifiers that the source directly builds upon.
Paper: On Discriminative vs. Generative Classifiers: A comparison of logistic regression and naive Bayes, Andrew Ng et al. (2001). This seminal comparison of generative and discriminative classifiers explains why Naive Bayes and linear models exhibit distinct error regimes across varying data and document scales.
Paper: A re-examination of text categorization methods, Yiming Yang et al. (1999). This empirical study provides foundational comparative analysis on how linear SVMs and Naive Bayes perform across standard text categorization benchmarks.
Paper: Mining the peanut gallery: opinion extraction and semantic classification of product reviews, Kushal Dave et al. (2003). This early work demonstrates the impact of n-gram feature representations and simple statistical scoring methods for product review sentiment classification.

Paper: Convolutional Neural Networks for Sentence Classification, Yoon Kim (2014). This influential paper demonstrates how lightweight convolutional neural networks with pretrained word embeddings offer a next-step neural baseline for sentence and sentiment classification.
Paper: A Simple but Tough-to-Beat Baseline for Sentence Embeddings, Sanjeev Arora et al. (2017). This work extends the source's philosophy of tough-to-beat simple baselines by creating an unsupervised, weighted sentence embedding method for downstream classification.
Paper: EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks, Jason Wei et al. (2019). This research explores lightweight data augmentation techniques to boost performance on text and sentiment classification tasks without complex architectures.
Paper: Recurrent Convolutional Neural Networks for Text Classification, Siwei Lai et al. (2015). This study advances text classification beyond linear n-gram baselines by introducing recurrent convolutional neural networks to capture broader context.
Paper: Document Modeling with Gated Recurrent Neural Network for Sentiment Classification, Duyu Tang et al. (2015). This paper tackles document-level sentiment classification for long reviews using hierarchical gated recurrent neural networks, contrasting with linear n-gram models.
Paper: Deep learning for sentiment analysis: A survey, Lei Zhang et al. (2018). This comprehensive survey provides an overview of the deep learning architectures that succeeded traditional n-gram baselines in sentiment analysis.
Paper: How to Fine-Tune BERT for Text Classification?, Chi Sun et al. (2019). This paper investigates optimal fine-tuning practices for Transformer-based architectures on the standard text and sentiment classification benchmarks evaluated in the source.

Cover for Baselines and Bigrams: Simple, Good Sentiment and Topic Classification

Abstract

Variants of Naive Bayes (NB) and Support Vector Machines (SVM) are often used as baseline methods for text classification, but their performance varies greatly depending on the model variant, features used and task/dataset. We show that: (i) the inclusion of word bigram features gives consistent gains on sentiment analysis tasks; (ii) for short snippet sentiment tasks, NB actually does better than SVMs (while for longer documents the opposite result holds); (iii) a simple but novel SVM variant using NB log-count ratios as feature values consistently performs well across tasks and datasets. Based on these observations, we identify simple NB and SVM variants which outperform most published results on sentiment analysis datasets, sometimes providing a new state-of-the-art performance level.

Topics

computational efficiency deep learning engineering input features linear models naive bayes parsing algorithm sentiment classification support vector text representation

Keywords

sentiment analysis topic classification word bigram features linear SVM logistic regression Bernoulli NB recursive autoencoders NB log-count ratios movie review dataset

1 Introduction
2 The Methods
2.1 Multinomial Naive Bayes (MNB)
2.2 Support Vector Machine (SVM)
2.3 SVM with NB features (NBSVM)
3 Datasets and Task
4 Experiments and Results
4.1 Experimental setup
4.2 MNB is better at snippets
4.3 SVM is better at full-length reviews
4.4 Benefits of bigrams depends on the task
4.5 NBSVM is a robust performer
4.6 Other results
References

Knowls

Knowl 1 — NBSVM: Support Vector Machine with Naive Bayes Feature Weights and Interpolation
model/method
NBSVM is a linear classification model that combines generative Multinomial Naive Bayes (MNB) log-count ratios as feature scalings with discriminative Support Vector Machine (SVM) training and weight interpolation.

Let $\hat{f}^{(i)} \in \{0, 1\}^{|V|}$ denote the binarized feature count vector for training example $i$ with vocabulary $V$ and binary label $y^{(i)} \in \{-1, +1\}$ , where $\hat{f}^{(i)}_j = \mathbf{1}\{f^{(i)}_j > 0\}$ for raw word counts $f^{(i)}$ . The smoothed positive and negative class count vectors are defined as:

$p = \alpha + \sum_{i: y^{(i)}=1} \hat{f}^{(i)}, \quad q = \alpha + \sum_{i: y^{(i)}=-1} \hat{f}^{(i)}$

where $\alpha > 0$ is an additive smoothing parameter (set to $\alpha = 1$ ). The log-count ratio vector $\hat{r} \in \mathbb{R}^{|V|}$ is:

$\hat{r} = \log \left( \frac{p / \|p\|_1}{q / \|q\|_1} \right)$

Each input vector $\hat{f}^{(k)}$ is transformed into $\tilde{f}^{(k)} = \hat{r} \circ \hat{f}^{(k)}$ , where $\circ$ represents elementwise multiplication. The model learns parameters $w \in \mathbb{R}^{|V|}$ and $b \in \mathbb{R}$ by minimizing the $L_2$ -regularized $L_2$ -loss SVM objective:

$w^T w + C \sum_{i=1}^N \max\left(0, 1 - y^{(i)}(w^T \tilde{f}^{(i)} + b)\right)^2$

To regularize towards the Naive Bayes model, the weight vector is interpolated as:

$w' = (1 - \beta)\bar{w} + \beta w$

where $\bar{w} = \|w\|_1 / |V|$ is the mean magnitude of $w$ , and $\beta \in [0, 1]$ is the interpolation parameter. The decision rule for a test instance $k$ is:

$y^{(k)} = \text{sign}\left(w'^T \tilde{f}^{(k)} + b\right)$

Standard default hyperparameters are $\alpha = 1$ , $C = 1$ , and $\beta = 0.25$ .
Knowl 2 — Binarized Multinomial Naive Bayes Linear Formulation
model/method
Multinomial Naive Bayes (MNB) can be parameterized directly as a linear classifier of the form $y^{(k)} = \text{sign}(w^T x^{(k)} + b)$ over binarized feature vectors $\hat{f}^{(k)} = \mathbf{1}\{f^{(k)} > 0\}$ , where $f^{(k)}$ is the raw term count vector for instance $k$ .

Given training cases with labels $y^{(i)} \in \{-1, +1\}$ , let $N_+$ and $N_-$ denote the number of positive and negative training examples, and let $\alpha > 0$ be the smoothing hyperparameter (typically $\alpha = 1$ ). Smoothed count vectors $\hat{p}, \hat{q} \in \mathbb{R}^{|V|}$ are computed as:

$\hat{p} = \alpha + \sum_{i: y^{(i)}=1} \hat{f}^{(i)}, \quad \hat{q} = \alpha + \sum_{i: y^{(i)}=-1} \hat{f}^{(i)}$

The weight vector $w$ corresponds to the log-count ratio vector $\hat{r}$ :

$w = \hat{r} = \log \left( \frac{\hat{p} / \|\hat{p}\|_1}{\hat{q} / \|\hat{q}\|_1} \right)$

and the scalar bias is:

$b = \log\left(\frac{N_+}{N_-}\right)$

The resulting linear prediction for test case $k$ is:

$y^{(k)} = \text{sign}\left(\hat{r}^T \hat{f}^{(k)} + \log\left(\frac{N_+}{N_-}\right)\right)$

Knowl 3 — Sentiment and Subjectivity Classification on Snippet Datasets

data/table

On short snippet sentiment and subjectivity datasets, binarized Multinomial Naive Bayes (MNB) and NBSVM consistently outperform standard linear Support Vector Machines (SVM) and complex models based on parse trees, Recursive Autoencoders, or hand-coded polarity reversal rules.

Method	RT-s	MPQA	CR	Subj.
MNB-uni	77.9	85.3	79.8	92.6
MNB-bi	79.0	86.3	80.0	93.6
SVM-uni	76.2	86.1	79.0	90.8
SVM-bi	77.7	86.7	80.8	91.7
NBSVM-uni	78.1	85.3	80.5	92.4
NBSVM-bi	79.4	86.3	81.8	93.2
RAE	76.8	85.7	–	–
RAE-pretrain	77.7	86.4	–	–
Voting-w/Rev.	63.1	81.7	74.2	–
Rule	62.9	81.8	74.3	–
BoF-noDic.	75.7	81.8	79.3	–
BoF-w/Rev.	76.4	84.1	81.4	–
Tree-CRF	77.3	86.1	81.4	–
BoWSVM	–	–	–	90.0

Values represent 10-fold cross-validation classification accuracy (%). The datasets are: RT-s (Rotten Tomatoes sentence snippet movie reviews, average length $l=21$ ), MPQA (opinion polarity subtask, $l=3$ ), CR (customer review sentences, $l=20$ ), and Subj. (subjectivity vs. objectivity dataset, $l=24$ ). uni denotes unigram features, bi denotes unigram plus bigram features, RAE denotes Recursive Autoencoders, and Tree-CRF denotes dependency tree-based conditional random fields.

Linear SVM is a comparatively weak baseline for short snippet classification. MNB and NBSVM beat or match sophisticated models without relying on parsers, lexicons, or unsupervised pretraining.

Knowl 4 — Sentiment Classification Performance on Full-Length Review Datasets

data/table

On full-length document review tasks, Support Vector Machines (SVM) outperform standard Multinomial Naive Bayes (MNB), while NBSVM achieves superior performance that matches or outperforms complex neural and specialized feature representations.

Method	RT-2k	IMDB	Subj.
MNB-uni	83.45	83.55	92.58
MNB-bi	85.85	86.59	93.56
SVM-uni	86.25	86.95	90.84
SVM-bi	87.40	89.16	91.74
NBSVM-uni	87.80	88.29	92.40
NBSVM-bi	89.45	91.22	93.18
BoW (bnc)	85.45	87.80	87.77
BoW (b $\Delta$ t'c)	85.80	88.23	85.65
LDA	66.70	67.42	66.65
Full+BoW	87.85	88.33	88.45
Full+Unlab'd+BoW	88.90	88.89	88.13
BoWSVM	87.15	–	90.00
Valence Shifter	86.20	–	–
tf. $\Delta$ idf	88.10	–	–
Appr. Taxonomy	90.20	–	–
WRRBM	–	87.42	–
WRRBM + BoW(bnc)	–	89.23	–

Classification accuracy (%) is evaluated on RT-2k (standard 2000 full-length movie reviews, average length $l=787$ , 10-fold CV), IMDB (50k full-length movie reviews, average length $l=231$ , standard train/test split), and the Subj. snippet dataset ( $l=24$ , 10-fold CV) for comparison.

On long documents, the independence assumptions of MNB cause it to lag behind SVM. However, combining NB log-count ratio features with an SVM in NBSVM-bi achieves 89.45% on RT-2k and 91.22% on IMDB, exceeding Word Representation Restricted Boltzmann Machines (WRRBM) and delta-idf baselines.

Knowl 5 — Task-Dependent Utility of Word Bigram Features

data/table

The addition of word bigram features to unigram feature sets consistently improves classification performance on sentiment and subjectivity tasks, but offers negligible or negative utility on topical text classification tasks.

Method	AthR	XGraph	BbCrypt
MNB-uni	85.0	90.0	99.3
MNB-bi	85.1 (+0.1)	91.2 (+1.2)	99.4 (+0.1)
SVM-uni	82.6	85.1	98.3
SVM-bi	83.7 (+1.1)	86.2 (+0.9)	97.7 (-0.5)
NBSVM-uni	87.9	91.2	99.7
NBSVM-bi	87.7 (-0.2)	90.7 (-0.5)	99.5 (-0.2)
ActiveSVM	–	90.0	99.0
DiscLDA	83.0	–	–

Accuracy (%) is reported across three 20-Newsgroups pairwise topic classification tasks: AthR (alt.atheism vs. religion.misc, $l=345$ ), XGraph (comp.windows.x vs. comp.graphics, $l=261$ ), and BbCrypt (rec.sport.baseball vs. sci.crypt, $l=269$ ). uni indicates unigram features; bi indicates unigram + bigram features.

On topical tasks, unigram keywords are sufficiently indicative on their own; adding bigrams to NBSVM yields a slight performance drop of 0.2% to 0.5%. In sentiment tasks, by contrast, bigrams provide substantial gains across all models by capturing local context and modifier structures (e.g., negated adjectives and modified verbs). Adding trigrams, however, slightly degrades performance across tasks.

Knowl 6 — L2-Regularized L2-Loss Support Vector Machine Formulation
model/method
The baseline Support Vector Machine (SVM) classifier is formulated as an $L_2$ -regularized linear model with squared hinge loss ( $L_2$ -loss) over binarized feature vectors $\hat{f}^{(k)} \in \{0, 1\}^{|V|}$ , where $\hat{f}^{(k)} = \mathbf{1}\{f^{(k)} > 0\}$ .

Given training pairs $(\hat{f}^{(i)}, y^{(i)})$ with $y^{(i)} \in \{-1, +1\}$ for $i = 1, \dots, N$ , the weight vector $w \in \mathbb{R}^{|V|}$ and bias scalar $b \in \mathbb{R}$ are obtained by minimizing:

$w^T w + C \sum_{i=1}^N \max\left(0, 1 - y^{(i)}(w^T \hat{f}^{(i)} + b)\right)^2$

where $C > 0$ is the regularization parameter (set to $C = 0.1$ for standard SVM). The $L_2$ -regularized $L_2$ -loss objective demonstrates superior numerical stability over $L_1$ -loss SVM in text classification settings. Predictions for unseen examples $k$ are computed as $y^{(k)} = \text{sign}(w^T \hat{f}^{(k)} + b)$ .
Knowl 7 — Superiority of Binarized Features and Multinomial over Bernoulli Naive Bayes
empirical result
For Multinomial Naive Bayes (MNB) and NBSVM, using binarized feature indicators $\hat{f}^{(k)} = \mathbf{1}\{f^{(k)} > 0\}$ yields approximately 1% higher classification accuracy than utilizing raw term counts $f^{(k)}$ on longer documents, while showing negligible difference on short snippets.

Furthermore, Multivariate Bernoulli Naive Bayes (BNB) performs substantially worse and is less stable than MNB across text classification benchmarks, lagging behind MNB by up to 10% accuracy and only matching MNB on short snippet tasks using unigram features.
Knowl 8 — Regularization and Sensitivity of the NBSVM Interpolation Parameter
empirical result
In the NBSVM model, the weight vector is defined via interpolation between the mean weight magnitude and the discriminatively trained SVM weights:

$w' = (1 - \beta)\bar{w} + \beta w$

where $\bar{w} = \|w\|_1 / |V|$ and $\beta \in [0, 1]$ . This formulation acts as a regularizer: it defaults to generative Naive Bayes weights unless the discriminative SVM is sufficiently confident.

On long document tasks (such as RT-2k and IMDB), classification accuracy is stable within 0.1% for all $\beta \in [0.25, 1.0]$ . On short snippet datasets, setting $\beta = 0.25$ provides an average 0.5% accuracy improvement over $\beta = 1.0$ (pure SVM weights). Selecting $\beta \in [0.25, 0.5]$ provides a robust configuration across both snippet and document classification tasks.

Coverage note — None was omitted; all contributed model formulations (MNB, SVM, NBSVM), experimental benchmarks across snippet, full-length review, and topical datasets, and feature/hyperparameter analyses are included.

References

1.R. Collobert and J. Weston. 2008. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of ICML.
2.George E. Dahl, Ryan P. Adams, and Hugo Larochelle. 2012. Training restricted boltzmann machines on word observations. arXiv:1202.5695v1 [cs.LG].
3.Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-Rui Wang, and Chih-Jen Lin. 2008. LIBLINEAR: A library for large linear classification. Journal of Machine Learning Research, 9:1871–1874, June.
4.Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings ACM SIGKDD, pages 168–177.
5.Alistair Kennedy and Diana Inkpen. 2006. Sentiment classification of movie reviews using contextual valence shifters. Computational Intelligence, 22.
6.Simon Lacoste-Julien, Fei Sha, and Michael I. Jordan. 2008. DiscLDA: Discriminative learning for dimensionality reduction and classification. In Proceedings of NIPS, pages 897–904.
7.Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of ACL.
8.Justin Martineau and Tim Finin. 2009. Delta tfidf: An improved feature space for sentiment analysis. In Proceedings of ICWSM.
9.Andrew McCallum and Kamal Nigam. 1998. A comparison of event models for naive bayes text classification. In AAAI-98 Workshop, pages 41–48.
10.Vangelis Metsis, Ion Androutsopoulos, and Georgios Paliouras. 2006. Spam filtering with naive bayes - which naive bayes? In Proceedings of CEAS.
11.Karo Moilanen and Stephen Pulman. 2007. Sentiment composition. In Proceedings of RANLP, pages 378–382, September 27-29.
12.Tetsuji Nakagawa, Kentaro Inui, and Sadao Kurohashi. 2010. Dependency tree-based sentiment classification using CRFs with hidden variables. In Proceedings of ACL:HLT.
13.Andrew Y Ng and Michael I Jordan. 2002. On discriminative vs. generative classifiers: A comparison of logistic regression and naive bayes. In Proceedings of NIPS, volume 2, pages 841–848.
14.Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of ACL.
15.Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of ACL.
16.Jason D. Rennie, Lawrence Shih, Jaime Teevan, and David R. Karger. 2003. Tackling the poor assumptions of naive bayes text classifiers. In Proceedings of ICML, pages 616–623.
17.Greg Schohn and David Cohn. 2000. Less is more: Active learning with support vector machines. In Proceedings of ICML, pages 839–846.
18.Richard Socher, Jeffrey Pennington, Eric H. Huang, Andrew Y. Ng, and Christopher D. Manning. 2011. Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions. In Proceedings of EMNLP.
19.Casey Whitelaw, Navendu Garg, and Shlomo Argamon. 2005. Using appraisal taxonomies for sentiment analysis. In Proceedings of CIKM-05.
20.Janyce Wiebe, Theresa Wilson, and Claire Cardie. 2005. Annotating expressions of opinions and emotions in language. Language Resources and Evaluation, 39(2-3):165–210.

Citation

MLA

Wang, S. I., and C. D. Manning. “Baselines and Bigrams: Simple, Good Sentiment and Topic Classification”. Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 2012, pp. 90–94, https://aclanthology.org/P12-2018/.

APA

Wang, S. I., & Manning, C. D. (2012). Baselines and Bigrams: Simple, Good Sentiment and Topic Classification. Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 90–94. https://aclanthology.org/P12-2018/

Chicago

Wang, S. I., and C. D. Manning. 2012. “Baselines and Bigrams: Simple, Good Sentiment and Topic Classification”. Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 90–94. https://aclanthology.org/P12-2018/.

Harvard

Wang, S.I. and Manning, C.D. (2012) “Baselines and Bigrams: Simple, Good Sentiment and Topic Classification”, Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). Association for Computational Linguistics, pp. 90–94. Available at: https://aclanthology.org/P12-2018/.

Vancouver

1. Wang SI, Manning CD (2012) Baselines and Bigrams: Simple, Good Sentiment and Topic Classification. In: Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). Association for Computational Linguistics, pp 90–94

BibTeX

@inproceedings{wang-manning-2012-baselines,
    title = "Baselines and Bigrams: Simple, Good Sentiment and Topic Classification",
    author = "Wang, Sida  and
      Manning, Christopher",
    editor = "Li, Haizhou  and
      Lin, Chin-Yew  and
      Osborne, Miles  and
      Lee, Gary Geunbae  and
      Park, Jong C.",
    booktitle = "Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)",
    month = jul,
    year = "2012",
    address = "Jeju Island, Korea",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/P12-2018/",
    pages = "90--94"
}

Metadata:ACL Anthology

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by-nc-sa/4.0/

Baselines and Bigrams: Simple, Good Sentiment and Topic Classification

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — NBSVM: Support Vector Machine with Naive Bayes Feature Weights and Interpolation

Knowl 2 — Binarized Multinomial Naive Bayes Linear Formulation

Knowl 3 — Sentiment and Subjectivity Classification on Snippet Datasets

Knowl 4 — Sentiment Classification Performance on Full-Length Review Datasets

Knowl 5 — Task-Dependent Utility of Word Bigram Features

Knowl 6 — L2-Regularized L2-Loss Support Vector Machine Formulation

Knowl 7 — Superiority of Binarized Features and Multinomial over Bernoulli Naive Bayes

Knowl 8 — Regularization and Sensitivity of the NBSVM Interpolation Parameter

References

Citation

BibTeX

Source Code

Access the Paper