Sparse L1 Unigram (`--use_sparse_pruning`): Vocabulary Selection via Continuous L1 Regularization¶

SentencePiece 0.2.3 adds Sparse Pruning (--use_sparse_pruning=true) as an optional vocabulary selection algorithm for the Unigram model (default = false; in 0.3.0, this flag is planned to be removed and Sparse Pruning will become the only behavior). Unlike the discrete heuristic pruning of the current default Unigram training (Discrete Pruning), Sparse Pruning uses a continuous L1-regularized EM (Expectation-Maximization) procedure. This mode is required when Auto-character Coverage is used with the Unigram model.

1. Background: Limitations of the Current Default Method (Discrete Pruning)¶

1.1 The Unigram Model and the Vocabulary Selection Problem¶

The Unigram model (Kudo, 2018) assigns an occurrence probability \(P(x)\) to each subword \(x \in V\) in a vocabulary \(V\), and defines the probability of a segmentation path \(\pi = (x_1, \ldots, x_n)\) of a sentence \(d\) as \(P(\pi) = \prod_k P(x_k)\). The training objective is the negative log marginal likelihood over the corpus \(\mathcal{D}\):

\[ \mathcal{L}(V) = \sum_{d \in \mathcal{D}} -\log \sum_{\pi \in \Omega(d)} P(\pi) \]

Vocabulary selection is the problem of solving \(\min_{|V| \le K} \mathcal{L}(V)\) under a vocabulary size budget \(K\). Once \(V\) is fixed, \(P(x)\) can be estimated with EM; the difficulty lies in the discrete choice of \(V\). Because searching over all substrings in the corpus is not practical, SentencePiece uses a Suffix Array to extract substrings with frequency at least 2 as the initial candidate (seed) vocabulary, and selects \(V\) of size \(K\) from it by pruning.

1.2 The Current Default Method: Discrete Pruning¶

The current default implementation (--use_sparse_pruning=false, TrainDiscretePruning / PruneSentencePieces) starts from the seed vocabulary and repeats the following two steps until the vocabulary reaches \(K\):

Fix the vocabulary \(V\) and run EM for a few iterations (--num_sub_iterations) to estimate \(P(x)\).
Estimate a heuristic "removal cost" for each subword \(x_i\) and remove the bottom 25% (--shrinking_factor=0.75).

Because computing the exact leave-one-out loss \(\Delta\mathcal{L}_i = \mathcal{L}(V \setminus \{x_i\}) - \mathcal{L}(V)\) for every candidate is expensive, step 2 relies on the approximation

\[ \Delta_i \approx f_i \cdot \bigl(\log P(x_i) - \textstyle\sum_k \log P(x'_{i,k})\bigr) \]

where \(f_i\) is the number of occurrences of \(x_i\) when the corpus is segmented with Viterbi (1-best) decoding, and \((x'_{i,1}, x'_{i,2}, \ldots)\) is the second-best segmentation of the string \(x_i\) without using \(x_i\) itself.

1.3 Limitations¶

(a) Theoretical grounding

The removal cost depends on the 1-best frequency \(f_i\). A subword whose probability mass is spread over several competing paths in the lattice may be useful, yet receive \(f_i = 0\) and be removed (1-best bias).
The approximation assumes that every occurrence of \(x_i\) in the corpus is replaced by the same local second-best sequence \(x'_i\), which does not match how probability mass is redistributed across the sentence lattice.
Although the target problem is formulated as minimizing negative log-likelihood under a discrete budget (\(\min_{|V| \le K} \mathcal{L}(V)\)), the vocabulary size constraint \(|V| \le K\) is handled by a separate heuristic pruning step rather than continuous optimization of a single objective. In particular, sequence length (token count) does not appear in \(\mathcal{L}(V)\).
The initial character set selected by --character_coverage is excluded from pruning, so the budget allocation between characters and subwords is not optimized.

(b) Implementation complexity

A 2-best lattice search (Lattice::NBest(2)) must be run for every subword string to obtain alternatives[i].
An additional pass over the whole corpus with Viterbi decoding is needed to count \(f_i\), separate from the Forward-Backward pass of EM.
Special cases such as "no second-best path exists" or "the 1-best path is already split" require dedicated branches such as always_keep.

2. Formulation of Sparse L1 Unigram¶

2.1 Change of Variables to Unnormalized Frequencies¶

Applying L1 regularization directly to normalized probabilities \(P(x)\) has no effect, because the probability constraint \(\sum_x P(x) = 1\) makes the L1 norm \(\sum_x P(x) = 1\) constant. We therefore represent each subword \(x_i\) by its non-negative unnormalized expected corpus frequency \(w_i \ge 0\) (measured on the corpus token-count scale so that \(w_i = \mathbb{E}[c_i]\) when \(\lambda = 0\)) and define \(P_{\mathbf{w}}(x_i) = w_i / \|\mathbf{w}\|_1\) with \(\|\mathbf{w}\|_1 = \sum_j w_j\).

2.2 Objective¶

The discrete size constraint \(|V| \le K\) is relaxed into the following continuous L1-regularized objective:

\[ \min_{\mathbf{w} \ge 0} \; \mathcal{J}(\mathbf{w}) = \mathcal{L}(\mathbf{w}) + \lambda \|\mathbf{w}\|_1 \]

where \(\lambda > 0\) is the regularization parameter. A subword with \(w_i = 0\) drops out of the vocabulary, so \(\|\mathbf{w}\|_0 = |V|\), and the L1 term is the convex (Lasso-type) relaxation of the L0 constraint \(|V| \le K\).

2.3 L1-Regularized EM and the Proximal Update¶

The objective is optimized with EM.

E-step: Run Forward-Backward under the current \(P_{\mathbf{w}}\) and compute the expected count \(\mathbb{E}[c_i] = \sum_d \mathbb{E}[c_i \mid d]\) of each subword (identical to standard Unigram).
M-step: Apply the proximal operator of the L1 term (soft-thresholding):

\[ w_i^{(t+1)} = \mathrm{prox}_{\lambda\|\cdot\|_1}\bigl(\mathbb{E}[c_i]\bigr) = \max\bigl(0,\; \mathbb{E}[c_i] - \lambda\bigr) \]

While standard maximum-likelihood estimation sets \(w_i \leftarrow \mathbb{E}[c_i]\), the L1 version subtracts the same \(\lambda\) from every subword and sets negative values to 0 (= removal). The E-step is unchanged, and the M-step adds this subtraction.

2.4 Choosing \(\lambda\): Adjust \(\lambda\) so that \(K\) weights remain non-zero¶

Rather than fixing \(\lambda\), the implementation (TrainSparsePruning) iteratively adjusts \(\lambda\) so that the vocabulary size reaches \(K\):

Run the E-step to obtain \(\mathbb{E}[c_i]\).
While the number of active subwords is sufficiently larger than \(K\), set \(\lambda^{(t)}\) to the threshold that removes 25% of the active subwords, i.e., the lower 25% quantile of \(\mathbb{E}[c_i]\). \(\lambda\) is updated monotonically (\(\lambda^{(t)} \leftarrow \max(\lambda^{(t-1)}, \text{quantile})\)).
Apply the M-step \(w_i \leftarrow \max(0, \mathbb{E}[c_i] - \lambda^{(t)})\) and normalize to probabilities.
Repeat until the number of active subwords reaches \(K\).
Finally, run one ordinary EM step with \(\lambda = 0\) on the surviving vocabulary to remove the shrinkage bias introduced by L1 (--post_l1_debias=true, i.e., a post-Lasso / debiased refit).

This is the standard Lagrangian-relaxation / dual-ascent view of the constrained problem \(\min \mathcal{L}(\mathbf{w})\) s.t. \(\|\mathbf{w}\|_0 \le K\): replace the constraint by a multiplier \(\lambda\) and adjust the multiplier until the constraint is met. Raising \(\lambda\) gradually in 25% steps rather than all at once follows the same schedule as shrinking_factor=0.75 in standard Unigram, and can be seen as a discretized regularization path along which the solution is tracked while EM re-estimates the parameters.

Related known techniques

Variable selection by L1 regularization (Lasso): Tibshirani (1996)
Soft-thresholding as the proximal operator of L1: Parikh & Boyd (2014)
Re-estimation without penalty after selection (post-Lasso): Belloni & Chernozhukov (2013)

Sparse Pruning combines these ingredients with Unigram EM, whose latent variables are segmentation paths.

3. Theoretical Properties¶

Introducing L1 regularization gives the pruning step the following four interpretations.

3.1 \(\|\mathbf{w}\|_1\) is the expected token count¶

On the expected-count scale \(w_i = \mathbb{E}[c_i]\) (before subtracting \(\lambda\)),

\[ \|\mathbf{w}\|_1 = \sum_i \mathbb{E}[c_i] = \mathbb{E}[\text{total number of tokens in the corpus}] \]

The objective \(\mathcal{J} = \mathcal{L} + \lambda \|\mathbf{w}\|_1\) is therefore a weighted sum of the negative log-likelihood and the expected total number of tokens needed to encode the corpus, including sequence length (token count) directly in the objective.

3.2 \(\mathbb{E}[c_i]\) is a lower bound on the leave-one-out loss¶

Splitting \(Z(d) = \sum_\pi P(\pi)\) into paths that contain \(x_i\) and paths that do not, the increase in loss on sentence \(d\) when \(x_i\) is removed from the vocabulary (with other weights held fixed) is

\[ \Delta\mathcal{L}_i(d) = \log Z(d) - \log Z_{-i}(d) = -\log\bigl(1 - P(x_i \mid d)\bigr) \]

Since \(-\log(1-p) \ge p\) for \(0 \le p < 1\), summing over the corpus gives

\[ \Delta\mathcal{L}_i = \sum_d -\log\bigl(1 - P(x_i \mid d)\bigr) \;\ge\; \sum_d P(x_i \mid d) = \mathbb{E}[c_i] \]

That is, the expected count \(\mathbb{E}[c_i]\), already available from the E-step, is a lower bound on the loss increase caused by removing \(x_i\).

Unlike \(\Delta_i\), which relies on 1-best frequencies and a second-best alternative, \(\mathbb{E}[c_i]\) accounts for all paths and has no 1-best bias.
Both \(-\log(1-p)\) and \(p\) are monotonically increasing, so the ranking within each sentence is preserved. In the low-frequency region where pruning decisions are made (\(P(x_i \mid d) \ll 1\)), \(-\log(1-p) \approx p\) and the bound is tight.
When \(P(x_i \mid d) \to 1\) (there is no segmentation of \(d\) that avoids \(x_i\)), the loss diverges while the bound saturates at 1. This property is used when pruning characters (Auto-char, Section 4).
No 2-best search or extra Viterbi pass is required; the pruning criterion is already computed in the E-step, which simplifies the implementation.

3.3 Subtracting \(\lambda\) acts as a rich-get-richer regularizer¶

Subtracting the same \(\lambda\) from every subword in the M-step is a penalty that is non-uniform in relative terms:

High-frequency subwords (\(\mathbb{E}[c_i] \gg \lambda\)): the relative reduction is small.
Low-frequency subwords (\(\mathbb{E}[c_i] \approx \lambda\)): a large part of the mass is removed, or the weight becomes 0.

In the next E-step, paths containing the shrunk low-frequency subwords receive lower posterior probability, and the mass moves to paths containing higher-frequency subwords. Through this positive feedback (rich-get-richer), competing redundant fragments are eliminated across iterations. This is the same mechanism by which Lasso shrinkage drives solutions toward sparsity, operating inside EM.

The same mechanism can be read as a penalty on the number of splits. Representing a string with a single token \(x\) subtracts \(\lambda\) once; splitting it into two tokens \((y, z)\) subtracts \(\lambda\) once for each token, i.e., twice in total. Each additional split costs an extra \(\lambda\) on the expected-count scale, so among segmentations with similar probabilities, the one with fewer tokens is preferred.

Note: This shrinkage is a regularizer for vocabulary selection and need not remain in the final probability estimates. After the vocabulary is fixed, --post_l1_debias=true (default) re-estimates with \(\lambda = 0\), restoring ordinary maximum-likelihood estimates for the surviving vocabulary.

3.4 \(\lambda\) is a shadow price (Lagrange multiplier)¶

In terms of units, \(\mathcal{L}\) is measured in nats and \(\|\mathbf{w}\|_1\) in tokens, so

\[ \lambda = \frac{\Delta \mathcal{L}}{\Delta(\text{number of tokens})} \quad [\text{nats/token}] \]

In optimization terminology, \(\lambda\) is the Lagrange multiplier of the constraint \(\|\mathbf{w}\|_0 \le K\) (relaxed to \(\|\mathbf{w}\|_1\)); in economics the same quantity is called a shadow price: the log-likelihood cost of saving one token by using one vocabulary slot.

A subword with \(\mathbb{E}[c_i] \ge \lambda\) is judged to be worth its vocabulary slot (in terms of tokens saved) and survives.
The iteration in Section 2.4 can be interpreted as gradually raising this price until demand (the number of surviving subwords) equals supply (the budget \(K\)), i.e., finding the equilibrium price \(\lambda^*\).

4. Relation to Auto-character Coverage¶

Auto-character Coverage (--auto_character_coverage=true) does not pre-select characters with --character_coverage, puts every character in the corpus into the seed vocabulary, prunes unnecessary characters under the same criterion as subwords, and leaves the pruned characters to byte fallback at inference time (--byte_fallback=true). For Unigram, this mode requires --use_sparse_pruning=true.

In both Discrete Pruning's 2-best comparison (\(\Delta_i = f_i \cdot \delta\mathrm{loss}_i\)) and the exact leave-one-out loss (\(\Delta\mathcal{L}_i = -\sum_d \log(1 - P(x_i \mid d))\)), evaluating the removal of a single character \(c\) is problematic when no shorter alternative path exists in the training lattice: the alternative path probability is \(0\) (\(P(c \mid d) = 1\)), so the loss increase diverges to \(+\infty\) (which is why Discrete Pruning unconditionally excludes single characters from pruning).

Using the lower bound \(\mathbb{E}[c_i]\) from Section 3.2 resolves this naturally because \(\mathbb{E}[c_i]\) handles both cases with a single formula:

When alternative paths exist (\(P(x_i \mid d)\) is small, as for typical subwords): Since \(-\log(1 - u) \approx u\), the expected count \(\mathbb{E}[c_i]\) behaves directly as an approximation to the leave-one-out log-likelihood loss.
When no alternative path exists (\(P(c \mid d) \to 1\), as for single-character tokens): Instead of diverging like \(-\log(1 - u) \to \infty\), each occurrence saturates at \(u = 1\), and the corpus sum \(\mathbb{E}[c]\) reduces to the expected number of times character \(c\) is used as a single-character token—which is proportional to the number of extra byte fallback tokens that would be emitted at inference time if \(c\) were removed.

Thus, without introducing any special branch or ad-hoc byte probabilities in the training lattice, the single rule \(\mathbb{E}[c_i] < \lambda\) simultaneously prunes subwords with small leave-one-out loss and leaves rare characters (used fewer than \(\lambda\) times as single-character tokens) to byte fallback.

5. Benchmark Summary (13 Languages, 390 MB Wikipedia Corpus)¶

To isolate the effect of Sparse Pruning (--use_sparse_pruning=true) from character selection, we evaluate Unigram Discrete, Unigram Sparse, and BPE under the same fixed character coverage (--character_coverage=0.9995, 5,940 single characters) on a 390 MB Wikipedia corpus (wikimedia/wikipedia, 30 MB per language across 13 languages: en, de, fr, vi, ru, ar, he, hi, th, ja, zh, ko, el) and a separate 13 MB evaluation set (1 MB per language). All models use normalization_rule_name=identity, --split_by_whitespace=true, --split_by_unicode_script=false, --split_by_number=false, and --byte_fallback=true (see auto_character_coverage.md Appendix A for full training flags).

For benchmark results combining Sparse Pruning with automatic character selection (--auto_character_coverage=true --use_sparse_pruning=true, including comparisons against Byte-Level BPE), see Auto-character Coverage §4.

5.1 Overall Performance by Vocabulary Size¶

All models share the same 5,940 single characters (--character_coverage=0.9995), 256 byte tokens, and 3 control tokens; the remaining \(K - 6,199\) slots are multi-character subwords (1,801 / 9,801 / 25,801 / 57,801 / 121,801 at \(K = 8\text{k} \dots 128\text{k}\)).

Vocab Size	Model	Total Tokens (↓)	Overall BPT (↑)	Corpus Passes (E-step + Prune)
8,000	Unigram Discrete	4,911,865	2.7762	45 (30 + 15)
	Unigram Sparse	4,854,101	2.8092	20 (20 + 0)
	BPE	4,757,263	2.8664	—
16,000	Unigram Discrete	3,701,860	3.6836	39 (26 + 13)
	Unigram Sparse	3,642,149	3.7440	17 (17 + 0)
	BPE	3,650,479	3.7355	—
32,000	Unigram Discrete	3,136,294	4.3479	33 (22 + 11)
	Unigram Sparse	3,078,441	4.4296	15 (15 + 0)
	BPE	3,105,444	4.3911	—
64,000	Unigram Discrete	2,731,505	4.9922	24 (16 + 8)
	Unigram Sparse	2,693,741	5.0622	13 (13 + 0)
	BPE	2,717,130	5.0186	—
128,000	Unigram Discrete	2,431,350	5.6085	18 (12 + 6)
	Unigram Sparse	2,399,733	5.6824	10 (10 + 0)
	BPE	2,415,118	5.6462	—

Key observations: - Direct comparison (Unigram Discrete vs Unigram Sparse): Holding the character vocabulary fixed at 5,940 single characters and changing only --use_sparse_pruning=false \(\to\) true improves overall BPT across all vocabulary sizes by +1.19% to +1.88% (reducing token count on the 13 MB evaluation set by 31k–60k tokens). - Comparison against BPE: Under Discrete Pruning, Unigram Discrete produces lower BPT than BPE at every vocabulary size. With Sparse Pruning, Unigram Sparse exceeds BPE from 16,000 upward (3.7440 vs 3.7355 at 16k; 4.4296 vs 4.3911 at 32k; 5.0622 vs 5.0186 at 64k; 5.6824 vs 5.6462 at 128k). - Computational cost: Discrete Pruning executes two E-steps (--num_sub_iterations=2) plus one Viterbi corpus segmentation pass (PruneSentencePieces) and 2-best lattice searches per outer stage. Sparse Pruning executes one E-step per outer stage (plus one refit E-step at the end) with no Viterbi pruning pass or 2-best search, reducing total corpus passes by roughly half (e.g., 33 \(\to\) 15 passes at \(K = 32,000\)).

5.2 Per-Language BPT at Vocabulary Size 32,000¶

Language	Script Family	Unigram Discrete	Unigram Sparse	BPE
th (Thai)	Abugida (Thai)	7.6930	7.6209	7.6615
hi (Hindi)	Abugida (Devanagari)	6.9042	7.0455	6.9782
ru (Russian)	Cyrillic	4.9693	5.0806	5.0566
el (Greek)	Greek	4.9912	5.0895	5.0605
ar (Arabic)	Abjad (Arabic)	4.7515	4.8679	4.7355
he (Hebrew)	Abjad (Hebrew)	4.5965	4.7431	4.6214
vi (Vietnamese)	Latin + diacritics	4.3439	4.4109	4.4208
ja (Japanese)	Ideographic + Syllabic	4.0335	4.0456	4.0078
ko (Korean)	Syllabic (Hangul)	3.8206	4.0252	4.0054
de (German)	Latin (compounds)	3.6314	3.6577	3.6053
fr (French)	Latin	3.4634	3.5081	3.4720
zh (Chinese)	Ideographic (CJK)	3.4903	3.5108	3.5215
en (English)	Basic Latin	3.3877	3.4889	3.4577

At \(K = 32,000\), Unigram Sparse improves BPT over Unigram Discrete in 12 of the 13 languages, with the largest gains in Korean (+5.35%), Hebrew (+3.19%), English (+2.99%), Arabic (+2.45%), and Russian (+2.24%), and achieves the highest BPT among the three models in 10 of the 13 languages.

6. Usage¶

spm_train \
  --input=corpus.txt \
  --model_prefix=spm_unigram_sparse \
  --model_type=unigram \
  --vocab_size=32000 \
  --use_sparse_pruning=true \
  --post_l1_debias=true      # default: true

Combined with Auto-char:

spm_train \
  --input=corpus.txt \
  --model_prefix=spm_unigram_autochar \
  --model_type=unigram \
  --vocab_size=32000 \
  --use_sparse_pruning=true \
  --auto_character_coverage=true \
  --byte_fallback=true

Flag	Default	Description
`--use_sparse_pruning`	`false`	Perform Unigram vocabulary selection with L1-regularized EM (Sparse Pruning). Currently an experimental `spm_train` CLI-only flag; in 0.3.0 this flag will be removed and Sparse Pruning will become the only behavior.
`--post_l1_debias`	`true`	After the vocabulary is fixed, re-estimate with \(\lambda = 0\) to remove the L1 shrinkage bias. Currently a `spm_train` CLI-only flag.

The training log reports, for each epoch, active (number of active subwords), lambda, obj (per-sentence negative log-likelihood), |w| (Viterbi 1-best token count across the corpus), and bytes/tok (corpus bytes divided by |w|).

7. Summary of Differences¶

Aspect	Discrete Pruning (current default)	Sparse Pruning (L1)
Pruning criterion	\(\Delta_i \approx f_i \cdot \delta\mathrm{loss}_i\) (1-best frequency × 2-best alternative)	\(\mathbb{E}[c_i]\) (expected count over all paths; lower bound on the leave-one-out loss)
Objective formulation	Discrete \(\min_{\lvert V \rvert \le K} \mathcal{L}(V)\) approximated by alternating EM and heuristic pruning	Continuous \(\mathcal{L} + \lambda\\|\mathbf{w}\\|_1\) optimized consistently
Token count	Absent from the objective	Explicit as \(\\|\mathbf{w}\\|_1\) = expected token count
Probability re-estimation	\(w_i \leftarrow \mathbb{E}[c_i]\) (with digamma update)	\(w_i \leftarrow \max(0, \mathbb{E}[c_i] - \lambda)\) (debiased at the end)
Extra computation	2-best search + Viterbi re-segmentation pass	None (already computed in the E-step)
Characters	Initial character set excluded from pruning	Same criterion as subwords (Auto-char)

Sparse L1 Unigram (--use_sparse_pruning): Vocabulary Selection via Continuous L1 Regularization¶