---
title: General-to-Particular (G2P) Principle
url: https://www.emergentmind.com/topics/general-to-particular-g2p
type: topic
---

# General-to-Particular (G2P) Principle

Searching arXiv for the primary "General-to-Particular" paper and adjacent usages of the acronym G2P across domains.
General-to-Particular (G2P) is a modeling principle in which a system first learns a representation of information shared across conditions and then transforms that representation into a specialized form required by a specific domain, task, or context. The clearest formalization in the provided arXiv literature appears in multi-domain neural machine translation (NMT), where a shared encoder captures general knowledge and additional transformation modules convert it into domain-specific representations before decoding [1911.09912]. In the same literature, however, the acronym **G2P** is also used for grapheme-to-phoneme conversion, gloss-to-pose generation in sign language production, and Jefferson Lab’s \(g_2^p\) proton-spin program, so terminological disambiguation is essential [2004.03136], [2208.09141], [1704.02308].

## 1. Terminology and scope

The phrase **General-to-Particular** is explicitly operationalized in the NMT paper "Go From the General to the Particular: Multi-Domain Translation with Domain Transformation Networks" [1911.09912]. There, the contrast is between **general knowledge**, shared across domains, and **particular knowledge**, distinctive to each domain. The problem is not simply multi-tasking; it is the controlled conversion of one representational type into another.

Acronym ambiguity is substantial in adjacent literatures.

| Usage of G2P | Meaning | Representative source |
|---|---|---|
| General-to-Particular | Transforming shared representations into domain-specific ones | [1911.09912] |
| Grapheme-to-phoneme | Converting orthographic text into phonemic or pinyin output | [2004.03136] |
| G2P in SLP | Generating sign pose sequence from gloss sequence | [2208.09141] |
| \(g_2^p\) / g2p | Jefferson Lab proton-spin experiment and target system | [1305.3295] |

This ambiguity matters because the conceptual content differs sharply across fields. In multi-domain NMT, G2P is a representation-learning doctrine. In speech technology, it usually denotes orthography-to-pronunciation conversion. In sign language production, it denotes gloss-to-pose mapping. In nuclear physics, \(g_2^p\) refers to measurements of the proton spin structure function \(g_2\).

## 2. Formalization in multi-domain neural machine translation

The G2P formulation in multi-domain NMT begins from a diagnosis of standard mixed-domain training. For domains \(D_1,\dots,D_N\), a conventional unified model optimizes

$$
\mathcal{J}(\theta) = \argmax_{\theta} \frac{1}{N} \sum_{n=1}^N \mathcal{L}_n(\theta),
$$

but this objective provides no explicit signal about domains [1911.09912]. The reported consequence is that mixed-domain models usually learn the **general knowledge** shared across domains while tending to miss **particular/domain-specific knowledge**, because the encoder must encode both in the same hidden states. Fine-tuning can preserve domain-specific knowledge, but it requires multiple separate models.

The proposed architecture inserts **Domain Transformation Networks (Dtns)** between a shared encoder and a shared decoder. The encoder maps a source sentence \(\mathbf{x}\) to general representations

$$
\mathbf{H} = (\mathbf{H}_1,\dots,\mathbf{H}_I), \qquad \mathbf{H}\in\mathbb{R}^{I\times d},
$$

and a domain-specific module transforms those representations into domain-specific ones by

$$
\mathbf{H'}=\mathcal{F}(\mathbf{H}, W_n)+\mathbf{H}.
$$

Here, \(W_n\) denotes the parameters for domain \(n\), and \(\mathcal{F}\) may be implemented by an FNN, CNN, or self-attention network; the paper reports that **multi-head attention** works best in practice [1911.09912]. The residual connection is not incidental: it preserves the general representation \(\mathbf{H}\) while learning a domain-specific difference term.

The decoder then conditions on \(\mathbf{H'}\), rather than directly on \(\mathbf{H}\). The baseline NMT distribution remains the standard autoregressive factorization,

$$
P(\mathbf{y}|\mathbf{x}; \theta) = \prod_{j=1}^{J} P(y_j \mid \mathbf{y}_{<j}, \mathbf{x}; \theta),
$$

but the conditioning representation has been explicitly specialized. The paper also adopts **domain-aware batch learning**, in which each mini-batch contains examples from only one domain. For unbalanced data, the batch-sampling probability is

$$
q_i = \frac{p_i^\alpha}{\sum_{j=1}^N p_j^\alpha}, \qquad
p_i = \frac{n_i}{\sum_{k=1}^N n_k},
$$

with \(\alpha=0.7\), to reduce bias toward high-resource domains [1911.09912].

## 3. Supervision mechanisms for the general-to-particular transformation

The architectural separation alone is not treated as sufficient. The NMT paper adds two complementary supervision signals—**domain distillation** and **domain discrimination**—to ensure that the transformation from general to particular actually occurs [1911.09912].

**Domain distillation** is a teacher-student mechanism. A fine-tuned domain model serves as teacher, and the unified model is trained against both hard labels and the teacher’s soft output distribution:

$$
\mathcal{L}(\theta)=\argmax_{\theta}\sum_{(\mathbf{x},\mathbf{y})\in D}\bigg\{(1-\lambda)\log P(\mathbf{y}|\mathbf{x};\theta)
+\lambda\sum_{j=1}^J\sum_{k=1}^{|V|}\hat{P}(y_j=k|\mathbf{y}_{<j},\mathbf{x};\hat{\theta}) \log P(y_j=k|\mathbf{y}_{<j},\mathbf{x};\theta)\bigg\}.
$$

The stated intuition is that the teacher carries domain-specific knowledge, so distillation transfers that knowledge into the unified model without requiring separate deployed models.

**Domain discrimination** is used to make the transformed representations domain-aware in the intended location. The model first computes a weighted summary of encoder states,

$$
\widetilde{\mathbf{H}}=\sum_{i=1}^I \alpha_i \mathbf{H}_i,
$$

and applies domain classifiers to both the encoder summary and the transformed summary:

$$
P(d|\mathbf{x};\psi)=\textup{softmax}(W_D^\top \widetilde{\mathbf{H}}),
$$

$$
P(d|\mathbf{x};\gamma)= \textup{softmax}(W_D'^\top \widetilde{\mathbf{H}'}).
$$

The resulting objective is

$$
\mathcal{L}(\theta, \gamma, \psi) =\argmax_{\theta, \gamma, \psi}
\sum_{(\mathbf{x},\mathbf{y},d)\in D}
\bigg\{ \log P({\bf y}|{\bf x}; \theta)
+ \log P(d|\mathbf{x};\gamma)
+ \log P(d|\mathbf{x};\psi)
+ \delta \times H(P(d|\mathbf{x};\psi)) \bigg\}.
$$

Training alternates between optimizing \(\{\theta,\gamma\}\) and \(\psi\); when optimizing \(\{\theta,\gamma\}\), the adversarial term \(\log P(d|\mathbf{x};\psi)\) is dropped [1911.09912]. In the paper’s interpretation, the two signals are complementary: distillation is a rich weak-supervision channel, while discrimination is a strong explicit domain signal. Together they are intended to keep the encoder general, force the transformation network to learn the domain-specific residual, and make the decoder use the transformed representation effectively.

## 4. Experimental evidence in translation

The NMT experiments cover balanced and unbalanced multi-domain settings across several language pairs. The balanced settings include a small-scale Zh\(\Rightarrow\)En corpus with the domains **Law, Oral, Thesis, News**, and a large-scale Zh\(\Rightarrow\)En corpus with **Law, News, Patent, Subtitle**. The unbalanced settings include De\(\Rightarrow\)En with **Law, Medical, IT, Koran**, and En\(\Rightarrow\)Fr with **Medical, Parliament** [1911.09912]. The implementation uses a Transformer base architecture with 6 encoder and 6 decoder layers, hidden size 512, BPE with 30K merge operations, 4-gram BLEU, Fairseq, Adam, and \(\lambda=0.1\), \(\delta=0.1\).

The main BLEU results are as follows.

| Setting | Transformer baseline | Best unified G2P variant |
|---|---:|---:|
| Balanced Zh\(\Rightarrow\)En, small-scale | 30.90 | 34.25 |
| Balanced Zh\(\Rightarrow\)En, large-scale | 41.45 | 42.93 |
| Unbalanced De\(\Rightarrow\)En | 48.68 | 49.70 |
| Unbalanced En\(\Rightarrow\)Fr | 50.75 | 51.83 |

In the strongest balanced small-scale condition, adding domain transformation raises BLEU from **30.90** to **32.34**, and adding domain supervision further raises it to **34.25**, a gain of **+3.35 BLEU** over the baseline [1911.09912]. The same system is reported as close to fine-tuning: **34.92** for fine-tuning versus **34.25** for the proposed unified model. The paper further states that the method consistently beats domain control, domain discrimination, and domain context, and gains roughly **1–2 BLEU** over prior unified multi-domain methods.

The analytical evidence is intended to show that the model is not merely larger, but more structurally specialized. In the reported t-SNE visualizations, encoder representations are relatively concentrated, indicating shared/general information, whereas transformed representations are more spread out by domain. When decoding with different transformation modules, each module performs best on its own domain. Ablation studies show that distillation alone helps, discrimination alone helps, and both together help most. A case study in the thesis domain reports that the supervised transformation model yields more fluent and domain-accurate phrasing, including correctly translating “yuan wai” as “outside hospital” [1911.09912].

## 5. Adjacent meanings of the acronym and related formalisms

A common misconception is that **G2P** uniformly refers to **General-to-Particular**. In the supplied arXiv literature, the acronym is more frequently used in other senses.

In speech technology, G2P most often means **grapheme-to-phoneme conversion**. For Mandarin Chinese, "g2pM: A Neural Grapheme-to-Phoneme Conversion Package for Mandarin Chinese Based on a New Open Benchmark Dataset" defines G2P as conversion of Chinese text into **pinyin**, with **polyphone disambiguation** as the central challenge; among **17,720** unique Chinese characters, only **762** are polyphones (**4.30%**), but they account for **67M out of 363M** character instances (**18.49%**). The paper introduces the **CPP** benchmark with **99,264 sentences**, trains a compact Bi-LSTM, and reports **97.31%** test accuracy, compared with **97.85%** for Chinese BERT [2004.03136]. For Persian, "Multi-Module G2P Converter for Persian Focusing on Relations between Words" treats G2P as orthography-to-pronunciation conversion under missing vowels, homographs, OOVs, and ezafe; its multi-module sequence-level system reaches **94.48% word-level accuracy** [2208.01371]. A multilingual TTS front-end extends G2P to a unified pronunciation-learning task across **24 locales**, reporting **2.09%** average word-level PER and **0.97%** average sentence-level PER, while also showing that multilingual sharing is competitive but not uniformly dominant for context-sensitive phenomena such as Mandarin polyphones and French post-lexical rules [2307.16709]. "IPA-CHILDES & G2P+" uses **G2P+** as a normalization wrapper around multiple backends, aligned to **Phoible**, to create **IPA CHILDES**, a phonemic corpus of **45 million words** across **31 languages** [2504.03036].

The same acronym also appears in work that reduces or bypasses conventional grapheme-to-phoneme conversion. "MixedG2P-T5" replaces the standard G2P stage in Japanese TTS with a learned mapping from mixed-script text to SSL-derived discrete units, using **ContentVec**, a **T5** predictor, and **FastSpeech 2**; the proposed system reports **7.47%** UER and synthesis quality described as comparable to a conventional G2P-based baseline [2509.01391]. In robust phonetic transcription, G2P is used as a source of weak supervision rather than as the target task: "Scaling Human and G2P Supervision for Robust Phonetic Transcription" finds that G2P-generated labels help only when fewer than **20–30 hours** of human annotation are available, while the best curriculum—**XLSR pretraining + ASR finetuning + 40 hours of human phonetic supervision** and notably excluding G2P—achieves a **2.3× reduction** in weighted phone feature error rate over prior systems [2606.16019].

Outside speech technology, "G2P-DDM" uses **G2P** to mean **gloss sequence to sign pose sequence** in sign language production. Its method converts continuous pose generation into discrete latent-code generation via **Pose-VQVAE**, a discrete denoising diffusion model, **CodeUnet**, and heuristic **Sequential-KNN clustering** for length prediction [2208.09141]. In nuclear physics, \(g_2^p\) denotes the proton spin-structure program at Jefferson Lab rather than any representation-learning framework; the target-system paper describes a dynamically polarized ammonia target for the \(g_2^p\) and \(G_E^p\) experiments, and the SANE analysis paper reports a measurement of the proton spin structure function \(g_2\) over a four-momentum transfer range of \(2\) to \(6\ \mathrm{GeV}^2\) [1305.3295], [1704.02308].

## 6. Methodological implications, limits, and broader significance

Within machine learning, the strongest evidence for **General-to-Particular** as a substantive design principle comes from cases where the shared-to-specific transition is made explicit. The NMT formulation does this by separating encoder generality from domain-specific residual transformation [1911.09912]. A related, though more feature-local, example appears in "Selective Augmentation: Improving Universal Automatic Phonetic Transcription via G2P Bootstrapping", which explicitly describes a **General-to-Particular bootstrapping** logic: a shared IPA-like representation is used to align helper-language and baseline outputs, **phonation type** is transferred from Hindi, and **place of articulation** is retained from the original target-language transcription. In the reported German plosive evaluation, **VoicingAcc** rises from **71.3** to **83.8**, **Asp%** rises from **0.0** to **61.2**, and **Ten%** falls from **73.8** to **50.0** [2604.27204]. This is a concrete instance of preserving a general shared space while injecting a more particular distinction where it is needed.

The limits of the principle are equally instructive. A unified multilingual pronunciation model can outperform monolingual systems on English homograph disambiguation and some Arabic settings, yet monolingual models still do better on Mandarin polyphone disambiguation and French post-lexical rules [2307.16709]. Likewise, large-scale G2P-generated supervision is useful as a bootstrap when human annotation is extremely scarce, but after roughly **20–30 hours** of human labels it provides no statistically significant benefit and can reduce cross-dialect robustness [2606.16019]. This suggests that a move from general to particular is not guaranteed by parameter sharing, multilinguality, or data quantity alone. The transformation must be structurally identified, and the supervision must be matched to the kind of particularity that matters.

In that sense, General-to-Particular is best understood as an explicit specialization strategy rather than a synonym for unified modeling. The central claim of the explicit G2P literature is not that one representation can absorb all variation, but that generality and particularity should be assigned different representational roles. Where that division is architecturally clear and supported by targeted supervision, the resulting models can approach the performance of separately specialized systems while remaining unified [1911.09912].

Source: https://www.emergentmind.com/topics/general-to-particular-g2p