FastText.zip: Compressing text classification models

Armand JoulinEdouard GravePiotr BojanowskiMatthijs DouzeHervé JégouTomas Mikolov

article2016arXiv1,365 citations

Proposes a product quantization approach to compress text classification models by two orders of magnitude with minimal accuracy loss, enabling fastText deployment on memory-constrained devices.

Listen

Text classification is vital for applications such as search ranking, email sorting, and spam filtering. However, standard high-performing models often require gigabytes of memory to store large vocabularies and embedding matrices, making on-device deployment difficult on resource-constrained platforms like smartphones.

The article demonstrates an effective compression pipeline that shrinks text classification models by multiple orders of magnitude while preserving competitive accuracy and fast processing speeds.

The researchers evaluated their approach on eight standard benchmark datasets and a large-scale hashtag prediction dataset with over 300,000 output labels. They combined several techniques: product quantization (a method that splits vectors into smaller pieces to store them compactly), separate encoding of vector magnitude and angle, vocabulary pruning based on feature importance and training-set coverage, and dictionary hashing.

The key findings show that standard product quantization compresses models by a factor of 10 without meaningful accuracy loss. When paired with structured vocabulary pruning and hashing, the pipeline achieves extreme compression—reducing model footprints by a factor of 1,000 to 4,000 (often shrinking models from over 100 megabytes down to less than 64 kilobytes) with an average accuracy drop of less than one percent. Retraining the output layer after compressing input matrices compensates for quantization distortions, and max-coverage pruning maintains strong predictive performance on massive datasets where naive pruning methods fail.

These results demonstrate that engineering teams can deploy lightweight, accurate language classifiers locally on edge hardware, substantially reducing server infrastructure costs, cloud latency, and memory overhead compared to heavier neural network alternatives.

Decision-makers should consider adopting this compression pipeline for embedded or mobile text processing tasks. If deployment demands extreme size reduction, teams should test model-specific cut-offs to ensure required feature coverage. Future work could explore scaling vector dimensions based on feature frequencies and splitting low-value words into character sub-units to further optimize performance on short inputs.

Meta classification classification models hashing fastText library text classification models quantization methods

arXiv: 1612.03651

Paper: Bag of Tricks for Efficient Text Classification, Armand Joulin et al. (2017). It introduces the baseline fastText linear text classification architecture and n-gram embedding framework that FastText.zip directly compresses using product quantization.
Paper: Deep Compression: Compressing Deep Neural Network with Pruning, Trained Quantization and Huffman Coding, Song Han et al. (2015). It introduces fundamental model compression techniques combining pruning, quantization, and coding that form the conceptual foundation for compressing embedding models.
Paper: Feature hashing for large scale multitask learning, Kilian Q. Weinberger et al. (2009). It provides foundational principles for feature hashing and dimensionality reduction in large-scale text classification that motivate FastText.zip's hashing-based compression approaches.
Paper: Distributed Representations of Words and Phrases and their Compositionality, Tomas Mikolov et al. (2013). It details foundational techniques for learning dense distributed representations and word embeddings that serve as the primary memory bottleneck addressed in FastText.zip.

Paper: Accelerating Large-Scale Inference with Anisotropic Vector Quantization, Ruiqi Guo et al. (2020). It builds upon product quantization for large-scale embeddings by introducing an anisotropic, score-aware quantization loss to accelerate inference and retrieval.
Paper: A Survey of Quantization Methods for Efficient Neural Network Inference, Amir Gholami et al. (2021). It provides an extensive survey synthesizing modern quantization paradigms, theoretical trade-offs, and inference acceleration strategies across deep learning architectures.
Paper: To prune, or not to prune: exploring the efficacy of pruning for model compression, Michael Zhu et al. (2017). It systematically investigates the trade-offs between pruning and architectural downsizing under strict memory budgets across language and vision tasks.
Paper: Stop Indexing at Full Precision: Revisiting Clustering for Vector Embeddings, Leonardo Kuffo et al. (2026). It explores aggressive quantization and clustering techniques directly on vector embeddings to optimize memory footprint and search efficiency.

Cover for FastText.zip: Compressing text classification models

Abstract

We consider the problem of producing compact architectures for text classification, such that the full model fits in a limited amount of memory. After considering different solutions inspired by the hashing literature, we propose a method built upon product quantization to store word embeddings. While the original technique leads to a loss in accuracy, we adapt this method to circumvent quantization artefacts. Our experiments carried out on several benchmarks show that our approach typically requires two orders of magnitude less memory than fastText while being only slightly inferior with respect to accuracy. As a result, it outperforms the state of the art by a good margin in terms of the compromise between memory usage and accuracy.

Topics

classification classification models hashing text classification word embedding quantization schemes quantizers random access memory

Keywords

fastText library text classification models quantization methods softmax approximation locality-sensitive hashing Bloom filters compact codes linear classifiers

1 Introduction
2 Related work
3 Proposed approach
3.1 Text classification
3.2 Bottom-up product quantization
3.3 Further text specific tricks
4 Experiments
4.1 Small datasets
4.2 Large dataset: FlickrTag
5 Future Work
6 Conclusion
References

Knowls

Knowl 1 — Normalized Product Quantization for Word and N-Gram Embeddings
model/method
In linear text classification models, storing dense embedding matrices for large vocabularies and n-grams requires substantial memory. Standard Product Quantization (PQ) splits a $d$ -dimensional vector $\mathbf{x} \in \mathbb{R}^d$ into $k$ subvectors $\mathbf{x}^1, \dots, \mathbf{x}^k \in \mathbb{R}^{d/k}$ and quantizes each subvector to the nearest centroid among $2^b$ codebook centroids ( $b=8$ bits per subquantizer) using $k$ -means:

$\hat{\mathbf{x}} = \sum_{i=1}^k q_i(\mathbf{x}^i)$

However, text embedding norms vary widely (by ratios up to $1000:1$ between maximum and minimum norms). Because standard $k$ -means minimizes squared $L_2$ reconstruction error, it tends to map low-norm vectors to zero, severely hurting classification accuracy.

Normalized Product Quantization (NPQ) circumvents this artifact by decomposing each embedding vector $\mathbf{x}$ into its scalar $L_2$ norm $\|\mathbf{x}\|_2$ and its unit-direction vector $\mathbf{u} = \mathbf{x} / \|\mathbf{x}\|_2$ :
1. The direction vector $\mathbf{u}$ is quantized into $k$ subquantizers ( $k \in [2, d/2]$ , typically $k = d/2$ ).
2. The scalar norm $\|\mathbf{x}\|_2$ is quantized separately into a 1-dimensional codebook of $2^b$ scalar values using an additional $b$ bits (typically 1 extra byte per vector).
This magnitude/direction decoupling allows faithful reconstruction of low-norm discriminative embeddings with no significant drop in classification performance.
Knowl 2 — Bottom-Up Retraining for Quantized Factorized Classifiers
model/method
For any vector quantizer satisfying the Lloyd optimality conditions, the reconstructed vector magnitude is systematically underestimated because the expected squared magnitude is reduced by the mean quantization error:

$\mathbb{E}[\|\hat{\mathbf{x}}\|^2] = \mathbb{E}[\|\mathbf{x}\|^2] - \mathbb{E}[\|\mathbf{x} - \hat{\mathbf{x}}\|^2]$

In a low-rank linear text classifier parameterized as $f(\mathbf{x}_n) = B A \mathbf{x}_n$ , where $A \in \mathbb{R}^{d \times V}$ is the input feature embedding matrix and $B \in \mathbb{R}^{C \times d}$ is the linear output classification matrix over $C$ classes, direct off-the-shelf quantization of $A$ and $B$ degrades classification accuracy by $0.1\%$ to $0.5\%$ .

To correct for this distortion, a bottom-up retraining procedure is applied:
1. The input embedding matrix $A$ is quantized into $\hat{A}$ using product quantization.
2. The input matrix is frozen at $\hat{A}$ , and the uncompressed output matrix $B$ is retrained on the training data to adapt its weights to the quantized input representations.
3. The retrained output matrix $B$ is quantized into $\hat{B}$ using product quantization.
This bottom-up retraining recovers the loss in classification accuracy across diverse datasets.
Knowl 3 — Online Max-Cover Vocabulary Pruning
algorithm
When reducing vocabulary size by keeping only a budget of $K$ features (words and n-grams), purely ranking embeddings by norm can leave certain training documents with zero retained features. To guarantee document coverage while maximizing retained representation weight, feature selection is posed as a constrained set-cover problem:

$\max_{S \subseteq V} \sum_{s \in S} w_s \quad \text{s.t.} \quad |S| \le K, \quad P \mathbf{1}_S \ge \mathbf{1}_{|\mathcal{D}|}$

where $V$ is the full set of vocabulary features, $\mathcal{D}$ is the training document collection, $w_s = \|\mathbf{a}_s\|_2$ is the $L_2$ norm of the embedding vector $\mathbf{a}_s$ corresponding to feature $s$ , and $P \in \{0, 1\}^{|\mathcal{D}| \times |V|}$ is the binary document-feature occurrence matrix ( $P_{d,s} = 1$ if feature $s$ occurs in document $d$ ).

The optimization is solved using an online greedy algorithm:
Input: Training document collection $\mathcal{D}$ , feature set $V$ , feature weights $w_s = \|\mathbf{a}_s\|_2$ , budget $K$
Output: Selected feature subset $S$
$S \leftarrow \emptyset$
for each document $d \in \mathcal{D}$ do
    if no feature in document $d$ belongs to $S$ then
         $s^* \leftarrow \arg\max_{s \in d} w_s$
         $S \leftarrow S \cup \{s^*\}$
    end if
end for
if $|S| < K$ then
     $R \leftarrow V \setminus S$
    sort $R$ in descending order of $w_s$
     $S \leftarrow S \cup \text{first } (K - |S|) \text{ elements of } R$
end if
return $S$

Knowl 4 — Norm-Based vs. Entropy-Based Vocabulary Pruning

empirical result

Selecting features based on the $L_2$ norm of their trained embedding vectors $\|\mathbf{a}_s\|_2$ outperforms unsupervised frequency- or entropy-based selection for text classifier compression. High-entropy/frequency words are dominated by non-discriminative stopwords and punctuation, whereas high-norm embeddings correspond to strongly class-discriminative tokens.

Word	Entropy Rank	Norm Rank	Word	Entropy Rank	Norm Rank
.	1	354	mediocre	1399	1
,	2	176	disappointing	454	2
the	3	179	so-so	2809	3
and	4	1639	lacks	1244	4
i	5	2374	worthless	1757	5
a	6	970	dreadful	4358	6
to	7	1775	drm	6395	7
it	8	1956	poorly	716	8
of	9	2815	uninspired	4245	9
this	10	3275	worst	402	10

On the Amazon full review dataset, the top 10 words ranked by entropy consist entirely of punctuation and stopwords (e.g., '.', ',', 'the', 'and'), all having low norm ranks. Conversely, the top 10 words ranked by embedding norm are sentiment-laden terms ('mediocre', 'disappointing', 'worst', 'worthless') that provide high discriminative power despite moderate corpus frequency.

Knowl 5 — Dictionary Elimination via Feature Hashing and Index Pruning
model/method
Storing raw character strings for words and n-grams in the dictionary accounts for 1 to 2 MB of memory in standard models. This overhead is eliminated by mapping both words and n-grams directly into a fixed set of $M$ hash buckets using a hash function.

When vocabulary pruning is combined with feature hashing to retain only $K$ active bucket embeddings:
1. Sorted Index List + Binary Search: An array of the $K$ surviving bucket indices is stored. At test time, verifying whether a hashed feature belongs to the pruned vocabulary requires a binary search with time complexity $O(\log K)$ and a memory footprint of several hundred kilobytes.
2. Bloom Filter Alternative: A Bloom filter achieves $O(1)$ membership testing at test time with lower memory overhead, but false positives introduce corrupted embedding lookups that degrade test accuracy compared to exact index tables.

Knowl 6 — Extreme Compression Performance on Standard Text Classification Benchmarks

data/table

Combining Normalized Product Quantization with $k = 1$ subquantizer, feature hashing, and aggressive norm-based feature pruning compresses fastText models to under $64\,\text{KiB}$ , $32\,\text{KiB}$ , and $16\,\text{KiB}$ on standard sentiment and topic classification datasets ( $d=8$ , 2M initial buckets).

Dataset	Full Size	Full Acc. (%)	64KiB Acc. (%)	32KiB Acc. (%)	16KiB Acc. (%)
AG	65M	92.1	91.4	90.6	89.1
Amazon full	108M	60.0	58.8	56.0	52.9
Amazon pol.	113M	94.5	93.3	92.1	89.3
DBPedia	87M	98.4	98.2	98.1	97.4
Sogou	73M	96.4	96.4	96.3	95.5
Yahoo	122M	72.1	70.0	69.0	69.2
Yelp full	78M	63.8	63.2	62.4	58.7
Yelp pol.	77M	95.7	95.3	94.9	93.2
Average diff.		0.0	-0.8	-1.7	-3.5

Compared to uncompressed models of size 65–122 MB, the compressed models achieve compression factors between $\times 1000$ and $\times 4000$ with an average accuracy drop of only $0.8\%$ at $64\,\text{KiB}$ and $1.7\%$ at $32\,\text{KiB}$ .

Knowl 7 — Model Size and Accuracy Comparison with Character-Level CNNs

data/table

FastText combined with Normalized Product Quantization ( $k = d/2$ ) and vocabulary pruning matches or outperforms character-level Convolutional Neural Networks (CNNs) in classification accuracy while requiring two to three orders of magnitude less memory at inference time.

Dataset	Zhang et al. (2015)		Xiao Cho (2016)		fastText+PQ ( $k=d/2$ )
	Acc. (%)	Size	Acc. (%)	Size	Acc. (%)	Size (RAM)
AG	90.2	108M	91.4	80M	91.9	889K
Amazon full	59.5	10.8M	59.2	1.6M	59.6	449K
Amazon pol.	94.5	10.8M	94.1	1.6M	94.3	449K
DBPedia	98.3	108M	98.6	1.2M	98.5	98K
Sogou	95.1	108M	95.2	1.6M	96.5	98K
Yahoo	70.5	108M	71.4	80M	71.7	889K
Yelp full	61.6	108M	61.8	1.4M	63.3	98K
Yelp pol.	94.8	108M	94.5	1.2M	95.5	449K

CNN sizes assume float32 storage, while fastText+PQ reports active memory consumed in RAM during test time. Across all 8 benchmarks, fastText+PQ operates in under 1 MB of RAM while achieving equal or higher accuracy than character CNN architectures.

Knowl 8 — Large-Scale Multi-Class Quantization and Pruning on FlickrTag
empirical result
On the FlickrTag hashtag prediction task (312,116 classes, 1,427,667 words, 10M n-gram buckets, hierarchical softmax, embedding dimension $d=256$ ), the uncompressed baseline occupies $12\,\text{GiB}$ and achieves $45.4\%$ accuracy.
1. Input and Output Matrix Quantization: Applying product quantization ( $k=128$ ) to both input matrix $A$ and output matrix $B$ , combined with separate norm encoding and bottom-up retraining, reduces model size from $12\,\text{GiB}$ to $1.5\,\text{GiB}$ ( $8\times$ reduction) with zero loss in test accuracy ( $45.4\%$ ).
2. Pruning Method Comparison: When pruning the quantized model down to 2M and 1M embeddings, Max-Cover pruning preserves $88.4\%$ test set document coverage (identical to the unpruned full model) and maintains $45.5\%$ and $43.9\%$ accuracy at $305\,\text{MiB}$ and $179\,\text{MiB}$ respectively.
3. Failure of Naive Pruning: Naive norm pruning drops test coverage to $73.2\%$ (2M) and $61.9\%$ (1M), yielding lower accuracy ( $41.6\%$ and $35.8\%$ ). Entropy pruning drops coverage to $70.5\%$ , yielding $32.1\%$ (2M) and $30.5\%$ (1M) accuracy.

Coverage note — None was omitted; all primary contributions—including Normalized Product Quantization (NPQ), separate norm encoding, bottom-up retraining, online max-cover pruning, hashing/Bloom filter dictionary reduction, and empirical evaluations on standard benchmarks and the large-scale FlickrTag dataset—are fully covered.

References

1.Alekh Agarwal, Olivier Chapelle, Miroslav Dud<unk>ık, and John Langford. A reliable effective terascale linear learning system. Journal of Machine Learning Research, 15(1):1111–1133, 2014.
2.Francis Bach, Rodolphe Jenatton, Julien Mairal, and Guillaume Obozinski. Optimization with sparsity-inducing penalties. Foundations and Trends<unk><sup>®</sup> in Machine Learning, 4(1):1–106, 2012.
3.Ashwinkumar Badanidiyuru, Baharan Mirzasoleiman, Amin Karbasi, and Andreas Krause. Streaming submodular maximization: Massive data summarization on the fly. In SIGKDD, pp. 671–680. ACM, 2014.
4.Mohammad Hossein Bateni, Mohammad Taghi Hajiaghayi, and Morteza Zadimoghaddam. Submodular secretary problem and extensions. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, pp. 39–52. Springer, 2010.
5.Moses S. Charikar. Similarity estimation techniques from rounding algorithms. In STOC, pp. 380–388, May 2002.
6.Welin Chen, David Grangier, and Michael Auli. Strategies for training large vocabulary neural language models. arXiv preprint arXiv:1512.04906, 2015.
7.Flavio Chierichetti, Ravi Kumar, and Andrew Tomkins. Max-cover in map-reduce. In International Conference on World Wide Web, 2010.
8.Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized neural networks: Training neural networks with weights and activations constrained to +1 or -1. arXiv preprint arXiv:1602.02830, 2016.
9.M. Datar, N. Immorlica, P. Indyk, and V.S. Mirrokni. Locality-sensitive hashing scheme based on p-stable distributions. In Proceedings of the Symposium on Computational Geometry, pp. 253–262, 2004.
10.Scott Deerwester, Susan T Dumais, George W Furnas, Thomas K Landauer, and Richard Harshman. Indexing by latent semantic analysis. Journal of the American society for information science, 1990.
11.Misha Denil, Babak Shakibi, Laurent Dinh, Marc-Aurelio Ranzato, and Nando et all de Freitas. Predicting parameters in deep learning. In NIPS, pp. 2148–2156, 2013.
12.Uriel Feige. A threshold of ln n for approximating set cover. JACM, 45(4):634–652, 1998.
13.Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. Optimized product quantization for approximate nearest neighbor search. In CVPR, June 2013.
14.Yunchao Gong and Svetlana Lazebnik. Iterative quantization: A procrustean approach to learning binary codes. In CVPR, June 2011.
15.Yunchao Gong, Liu Liu, Ming Yang, and Lubomir Bourdev. Compressing deep convolutional networks using vector quantization. arXiv preprint arXiv:1412.6115, 2014.
16.Edouard Grave, Armand Joulin, Moustapha Ciss<unk>é, David Grangier, and Herv<unk>é J<unk>égou. Efficient softmax approximation for gpus. arXiv preprint arXiv:1609.04309, 2016.
17.Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016.
18.Herv<unk>é J<unk>égou, Matthijs Douze, and Cordelia Schmid. Hamming embedding and weak geometric consistency for large scale image search. In ECCV, October 2008.
19.Herv<unk>é J<unk>égou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE Trans. PAMI, January 2011.
20.Thorsten Joachims. Text categorization with support vector machines: Learning with many relevant features. Springer, 1998.
21.Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016.
22.Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. NIPS, 2:598–605, 1990.
23.Zhouhan Lin, Matthieu Courbariaux, Roland Memisevic, and Yoshua Bengio. Neural networks with few multiplications. arXiv preprint arXiv:1510.03009, 2015.
24.Andrew McCallum and Kamal Nigam. A comparison of event models for naive bayes text classification. In AAAI workshop on learning for text categorization, 1998.
25.Lukas Meier, Sara Van De Geer, and Peter B<unk>ùhlmann. The group lasso for logistic regression. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 70(1):53–71, 2008.
26.Tomas Mikolov. Statistical language models based on neural networks. In PhD thesis. VUT Brno, 2012.
27.Tomas Mikolov, Ilya Sutskever, Anoop Deoras, Hai-Son Le, Stefan Kombrink, and J Cernocky. Subword language modeling with neural networks. preprint, 2012.
28.Behnam Neyshabur and Nathan Srebro. On symmetric and asymmetric lshs for inner product search. In ICML, pp. 1926–1934, 2015.
29.Mohammad Norouzi and David Fleet. Cartesian k-means. In CVPR, June 2013.
30.Bo Pang and Lillian Lee. Opinion mining and sentiment analysis. Foundations and trends in information retrieval, 2008.
31.Alexandre Sablayrolles, Matthijs Douze, Herv<unk>é J<unk>égou, and Nicolas Usunier. How should we evaluate supervised hashing? arXiv preprint arXiv:1609.06753, 2016.
32.Jorge S<unk>ánchez and Florent Perronnin. High-dimensional signature compression for large-scale image classification. In CVPR, 2011.
33.Anshumali Shrivastava and Ping Li. Asymmetric LSH for sublinear time maximum inner product search. In NIPS, pp. 2321–2329, 2014.
34.Andreas Stolcke. Entropy-based pruning of backoff language models. arXiv preprint cs/0006025, 2000.
35.David Talbot and Thorsten Brants. Randomized language models via perfect hash functions. In ACL, 2008.
36.Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. In Communications of the ACM, 2016.
37.Jingdong Wang, Heng Tao Shen, Jingkuan Song, and Jianqiu Ji. Hashing for similarity search: A survey. arXiv preprint arXiv:1408.2927, 2014.
38.Jun Wang, Wei Liu, Sanjiv Kumar, and Shih-Fu Chang. Learning to hash for indexing big data - A survey. CoRR, abs/1509.05472, 2015.
39.Sida Wang and Christopher D Manning. Baselines and bigrams: Simple, good sentiment and topic classification. In ACL, 2012.
40.Kilian Q Weinberger, Anirban Dasgupta, John Langford, Alex Smola, and Josh Attenberg. Feature hashing for large scale multitask learning. In ICML, 2009.
41.Yair Weiss, Antonio Torralba, and Rob Fergus. Spectral hashing. In NIPS, December 2009.
42.Yijun Xiao and Kyunghyun Cho. Efficient character-level document classification by combining convolution and recurrent layers. arXiv preprint arXiv:1602.00367, 2016.
43.Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In NIPS, 2015.

Citation

MLA

Joulin, A., et al. “FastText.zip: Compressing Text Classification Models”. arXiv, 2016, http://arxiv.org/abs/1612.03651v1.

APA

Joulin, A., Grave, E., Bojanowski, P., Douze, M., Jégou, H., & Mikolov, T. (2016). FastText.zip: Compressing text classification models. arXiv. http://arxiv.org/abs/1612.03651v1

Chicago

Joulin, A., E. Grave, P. Bojanowski, M. Douze, H. Jégou, and T. Mikolov. 2016. “FastText.zip: Compressing Text Classification Models”. arXiv. http://arxiv.org/abs/1612.03651v1.

Harvard

Joulin, A. et al. (2016) “FastText.zip: Compressing text classification models”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1612.03651v1.

Vancouver

1. Joulin A, Grave E, Bojanowski P, Douze M, Jégou H, Mikolov T (2016) FastText.zip: Compressing text classification models. arXiv

BibTeX

@article{joulin2016fasttext,
  title = {FastText.zip: Compressing text classification models},
  author = {Joulin, Armand and Grave, Edouard and Bojanowski, Piotr and Douze, Matthijs and Jégou, Hérve and Mikolov, Tomas},
  year = {2016},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1612.03651v1},
  eprint = {1612.03651}
}

Metadata:arXiv

Access the Paper

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

Open PDF

License: Published with permission

FastText.zip: Compressing text classification models

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Normalized Product Quantization for Word and N-Gram Embeddings

Knowl 2 — Bottom-Up Retraining for Quantized Factorized Classifiers

Knowl 3 — Online Max-Cover Vocabulary Pruning