General Phrase Debiaser for Masked LMs

• General Phrase Debiaser is a multi-token debiasing paradigm that targets multi-word expressions to reduce stereotypical bias in language models.
• It employs an automatic pipeline that extracts and fine-tunes on stereotypical phrase inventories, mitigating gender discrepancies as measured by Jensen–Shannon divergence.
• The method achieves strong bias reduction on SEAT metrics with minimal utility loss on GLUE, and is applicable to models like BERT, ALBERT, and DistilBERT.

Searching arXiv for recent and foundational papers on phrase-level and representation-space debiasing. [arXiv search] query: "General Phrase Debiaser debiasing masked LLMs multi-token level (Shi et al., 2023)" General Phrase Debiaser denotes a phrase-level debiasing paradigm in which the object of intervention is not an isolated token but a multi-token expression, prompt, or phrase embedding. In its named form, "General Phrase Debiaser: Debiasing Masked LLMs at a Multi-Token Level" introduces an automatic pipeline for masked LLMs that extracts stereotypical phrases from Wikipedia, searches for prompts that maximally expose gendered phrase distributions, and fine-tunes the model so that those distributions become more gender-neutral (Shi et al., 2023). In a broader representation-space sense, related work treats a general phrase debiaser as a reusable operator on text embeddings or hidden states, implemented either as a calibrated linear projection or as a soft projection based on conceptors (Chuang et al., 2023, Yifei et al., 2022).

1. Concept and scope

The central motivation for a general phrase debiaser is that many biased associations are expressed at phrase granularity rather than at the level of single vocabulary items. The motivating examples are occupations such as “software engineer”, “nuclear physicist”, and “primary school teacher”, and disciplinary phrases such as “functional analysis”, “complex differential geometry”, and “dance art”. A word-level method that adjusts the behavior of a model on “engineer” or “analysis” alone may leave substantial asymmetry on the full phrase, especially in technical and disciplinary domains (Shi et al., 2023).

Within masked LLMs, the phrase-level setting is naturally defined through multi-token cloze prediction. Instead of a single target token, the model predicts a sequence of consecutive [MASK] positions corresponding to a candidate phrase. Bias is then measured by the difference between phrase probability distributions induced by alternative demographic attribute phrases, rather than by the probability of one token in isolation. This formulation makes phrase-level debiasing a distributional problem over candidate phrases and phrase lengths, not merely a lexical correction problem (Shi et al., 2023).

The named General Phrase Debiaser is specifically designed for encoder-only masked LLMs such as BERT, ALBERT, and DistilBERT. Related work extends the same general idea to frozen text encoders in vision-LLMs and to contextual hidden states in BERT- and GPT-style LLMs, suggesting that the phrase-debiasing abstraction is not tied to a single architecture (Chuang et al., 2023, Yifei et al., 2022).

2. Multi-token bias as a distributional problem

The formal core of the masked-language-model formulation begins with a phrase set and a set of demographic attribute tuples. Let $\mathcal{M}$ be the MLM, and let a target region consist of $n$ consecutive masks: $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$ For a phrase $ph_i$ and an attribute realization $c_{k,m}^{(n)}$, the model is evaluated on

$$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$

and the phrase-level probability is

$$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$

In practice, the logits of the component tokens of $ph_i$ are summed across the masked positions and normalized with a softmax over candidate phrases of the same length (Shi et al., 2023).

Bias is quantified with Jensen–Shannon divergence. For a gender tuple $k$ with two variants, the phrase-level bias at length $n$ is

$n$0

The model aggregates over phrase lengths and over gender tuples: $n$1 This construction makes phrase-level bias an explicit discrepancy between male- and female-conditioned distributions over stereotypical phrases. The debiasing objective is therefore to make those distributions similar at every relevant phrase length, rather than to suppress a small set of manually listed stereotype words (Shi et al., 2023).

A second formal ingredient appears in the phrase representation used for phrase discovery. For a candidate phrase $n$2, the model computes a [CLS]-based representation by inserting $n$3 into 14 SEAT templates and averaging: $n$4 These template-averaged phrase embeddings support a cosine-similarity retrieval stage that is used to assemble domain-specific phrase inventories before debiasing begins (Shi et al., 2023).

3. Automatic construction of stereotypical phrase inventories

The Phrase Filter Stage is the first half of the named General Phrase Debiaser. Its role is to build a phrase inventory automatically, minimizing manual enumeration of multi-token stereotypes. The pipeline fixes four topics—career, math, art, and science—and starts from a small set of stereotype seeds for each topic. For career, the phrase list is reused from Kaneko and Bollegala; for math, art, and science, the phrases are produced automatically from Wikipedia hyperlinks (Shi et al., 2023).

All hyperlinked phrases extracted from Wikipedia form the raw candidate pool $n$5. Each candidate phrase $n$6 is embedded with the template-averaged [CLS] representation $n$7, and each seed $n$8 is embedded in the same way. Their similarity is measured by

$n$9

For each seed, the method keeps the top $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$0 most similar phrases. The resulting topic-level collections are aggregated into a weighted phrase set $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$1 and a deduplicated set $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$2 (Shi et al., 2023).

The distinction between the two sets is operationally important. $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$3 contains 624 unique phrases and is used during prompt search. $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$4 contains more than 500 phrases and is used during fine-tuning because multiplicity acts as an importance weight: phrases can appear multiple times if they are close to multiple seeds or recur across sublists. This weighting makes the later debiasing objective concentrate on phrases that are repeatedly implicated by the retrieval stage, rather than distributing equal mass over all retrieved candidates (Shi et al., 2023).

This design also clarifies what is meant by “automatic” in General Phrase Debiaser. The method does not eliminate manual supervision entirely; it retains a small seed list per topic and uses fixed SEAT templates. The automation lies in expanding from those seeds to a substantially larger inventory of domain phrases through Wikipedia hyperlinks and model-based similarity search (Shi et al., 2023).

4. Biased-prompt search and debias fine-tuning

The Model Debias Stage has two parts: finding biased prompts and fine-tuning the MLM on them. A prompt is a cloze-style context string that includes a demographic attribute phrase and a multi-mask target region. The search objective is to maximize the JSD-based loss over the phrase inventory, thereby locating contexts in which the model exhibits maximal gender discrepancy over stereotypical phrases (Shi et al., 2023).

Prompt search is implemented with beam search. The maximum prompt length is $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$5, the beam width is $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$6, and the search vocabulary is restricted to the 5,000 most frequent words in Wikipedia. The search uses $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$7 rather than the weighted set, so the ranking is driven by phrase diversity rather than by multiplicity. At each expansion step, candidate prompts are scored by

$[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$8

and the highest-loss prompts are retained as biased prompts (Shi et al., 2023).

Fine-tuning then minimizes the same loss on the selected biased prompts, using $[\mathbf{target}] = [MASK]\ [MASK]\ \dots\ [MASK].$9 and an extended gender attribute set $ph_i$0. The entire MLM is fine-tuned, motivated by the claim that bias is spread throughout the model rather than being confined to a specific embedding layer. The optimization uses AdamW and early stopping, and the reported implementation runs on a single NVIDIA 3090Ti GPU (Shi et al., 2023).

An important design distinction from vocabulary-wide debiasing is that the discrepancy is computed only over $ph_i$1, not over the entire vocabulary $ph_i$2. The stated implication is that gradients are propagated through the phrase inventory associated with the targeted stereotypes rather than through all lexical items. This suggests a more localized intervention in the model’s knowledge than approaches that equalize or regularize the whole output distribution (Shi et al., 2023).

5. Empirical behavior on SEAT and GLUE

The named General Phrase Debiaser is evaluated on BERT, ALBERT, and DistilBERT with SEAT-6, SEAT-6b, SEAT-7, SEAT-7b, SEAT-8, and SEAT-8b, where values closer to $ph_i$3 indicate less bias. On BERT, the average SEAT effect size falls from $ph_i$4 to $ph_i$5; on ALBERT, from $ph_i$6 to $ph_i$7; and on DistilBERT, from $ph_i$8 to $ph_i$9. For BERT specifically, the reported baselines are Context-Debias at $c_{k,m}^{(n)}$0, FairFil at $c_{k,m}^{(n)}$1, Auto-Debias at $c_{k,m}^{(n)}$2, and General Phrase Debiaser at $c_{k,m}^{(n)}$3. On individual tests, BERT’s SEAT-6 moves from $c_{k,m}^{(n)}$4 to $c_{k,m}^{(n)}$5, and SEAT-8 moves from $c_{k,m}^{(n)}$6 to $c_{k,m}^{(n)}$7 (Shi et al., 2023).

The utility evaluation is performed on GLUE. For BERT, the reported changes are small: CoLA $c_{k,m}^{(n)}$8, SST-2 $c_{k,m}^{(n)}$9, MRPC $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$0, STS-B $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$1, QQP $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$2, MNLI $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$3, QNLI $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$4, RTE $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$5, and WNLI $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$6. Comparable patterns are reported for ALBERT and DistilBERT, leading to the conclusion that debiased models show only a little decrease in scores compared to the original models (Shi et al., 2023).

The empirical profile is therefore one of strong SEAT reduction with limited GLUE degradation. The paper characterizes this as state-of-the-art debiasing performance on SEAT across careers and multiple disciplines. The result is especially notable because the disciplinary phrase inventories are multi-token and were built automatically, rather than being restricted to manually curated occupation words (Shi et al., 2023).

6. Representation-space generalizations

A broader line of work treats a general phrase debiaser as a reusable transformation on embeddings rather than as end-to-end MLM fine-tuning. In "Debiasing Vision-LLMs via Biased Prompts", debiasing is performed entirely by a linear operator $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$7 applied to text embeddings $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$8. The base projection removes the span of a spurious prompt matrix $$x'(c_{k,m}^{(n)}) = c_{k,m}^{(n)} \oplus x \oplus [\mathbf{target}],$$9,

$$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$0

and a calibrated projection further down-weights directions induced by positive-pair differences: $$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$1 This operator is used directly on CLIP prompts and Stable Diffusion conditioning vectors, requires no retraining, and is described as a general-purpose phrase debiaser. Reported effects include Waterbirds worst-group accuracy rising from $$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$2 to $$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$3 and the average-minus-worst-group gap falling from $$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$4 to $$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$5 for CLIP ResNet-50; on CelebA, the gap falls from $$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$6 to $$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$7. In Stable Diffusion v2.1, gender discrepancy drops from approximately $$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$8 to approximately $$P_{c_{k,m}^{(n)}} = P([\mathbf{target}] = ph_i \mid \mathcal{M}, x'(c_{k,m}^{(n)})).$$9 on train professions and from approximately $ph_i$0 to approximately $ph_i$1 on test professions, while race discrepancy drops from approximately $ph_i$2 to approximately $ph_i$3 on train professions and from approximately $ph_i$4 to approximately $ph_i$5 on test professions (Chuang et al., 2023).

"Conceptor-Aided Debiasing of LLMs" provides a related but softer operator view. Given contextual embeddings stacked in $ph_i$6, the bias conceptor is

$ph_i$7

and the debiasing operator is the NOT conceptor,

$ph_i$8

Because conceptors support NOT, AND, and OR, the method can combine subspaces for multiple bias sources and construct intersectional debiasers. Post-processing with the NOT conceptor is applied to arbitrary embeddings, including token embeddings, sentence embeddings, and hidden states. For BERT-base, the average absolute SEAT effect size falls from $ph_i$9 to approximately $k$0 under optimized post-processing; for GPT2, it falls from $k$1 to approximately $k$2. On BERT-base, Conceptor P.P. yields an average GLUE score of $k$3, reported as $k$4 relative to the original model, whereas the conceptor-intervened architecture with continued training reduces GLUE to $k$5, reported as $k$6 (Yifei et al., 2022).

These operator-based formulations differ from the named General Phrase Debiaser in mechanics but not in target. In each case, the intervention is phrase-agnostic at inference time: once the operator is computed, any new prompt or phrase embedding in the same space can be transformed by a single matrix application (Chuang et al., 2023, Yifei et al., 2022).

7. Limitations and related debiasing paradigms

The named General Phrase Debiaser has explicit limitations. Its experiments focus on gender bias, not race, religion, age, or nationality. The Phrase Filter Stage depends heavily on Wikipedia hyperlinks, which may under-cover domains with sparse hyperlink structure or different linking conventions. Prompt search with multi-token probabilities is computationally non-trivial even with a 5,000-word search space and beam width $k$7. DistilBERT remains substantially biased after debiasing, with an average SEAT effect size of $k$8. The method includes no explicit utility-regularization term, relying instead on targeted phrase sets and early stopping, and it is designed for encoder-only masked LLMs rather than decoder-only causal LMs (Shi et al., 2023).

Adjacent work expands the design space in three directions. First, "Towards Debiasing NLU Models from Unknown Biases" replaces manually specified bias features with a shallow self-debiasing model trained on a small subset of data and uses its confidence signal in reweighting, product-of-experts, or confidence-regularization objectives. The reported result is that self-debiasing can retain improvement on challenge datasets without specifically targeting certain biases and can improve overall robustness, which suggests a route toward phrase debiasing when bias types are not known a priori (Utama et al., 2020). Second, "DeFrame: Debiasing LLMs Against Framing Effects" introduces framing disparity,

$k$9

and a framing-aware prompting procedure consisting of initial response, framing integration, guideline generation, and self-revision. The paper reports reductions of framing disparity by up to $n$0 and bias score by $n$1 on BBQ, indicating that phrase-level debiasing must account not only for phrase content but also for semantically equivalent alternative framings (Lim et al., 4 Feb 2026). Third, "Bi-directional Bias Attribution: Debiasing LLMs without Modifying Prompts" locates bias at the neuron level of the projection layer through integrated gradients and intervenes directly on selected activations, reducing WinoBias gap on Llama-3.1 from $n$2 to $n$3 under its backward attribution variant. A plausible implication is that phrase debiasing can also be realized as a model-internal control problem rather than as prompt rewriting or phrase-list fine-tuning (Lin et al., 4 Feb 2026).

Taken together, these works define General Phrase Debiaser not as a single algorithm but as a research program: automatic multi-token debiasing in masked LLMs (Shi et al., 2023), closed-form operators on prompt embeddings in vision-language systems (Chuang et al., 2023), soft projection in contextual representation spaces (Yifei et al., 2022), and broader robustness mechanisms for unknown shortcuts, framing variants, and neuron-level stereotype attribution (Utama et al., 2020, Lim et al., 4 Feb 2026, Lin et al., 4 Feb 2026).