Controllable Text Generation using Transformer based Pre trained Language Models - Songwooseok123/Study_Space GitHub Wiki
-
ctg ์ฐ๊ตฌ๋ฅผ ํด์ผํ๋ ์ด์ :
- "chatgpt๊ฐ ์ํ๋ ๋๋ก ์์ฑ ์ ํด์ฃผ์ง ์๋?"
- chatgpt ์ด์ฉ์ "์ฌ์ด๋จ์ด๋ก ์น์ ํ๊ฒ ํด์ค" ์ด๋ฐ ์์ผ๋ก ๋งค๋ฒ prompt๋ฅผ ์ฃผ๋ฉด ํ ์คํธ ์์ฑ์ ์ ์ดํ ์ ์์. ํ์ง๋ง, ๋ด๊ฐ ์ ํํ๊ฒ ์ํ๋ ์คํ์ผ์ ๊ฒฐ๊ณผ๋ฌผ์ ๋ง๋๋ ค๋ฉด prompt๋ฅผ ์ ์ฐพ์๋ด์ผ ๋๊ณ ์๋๋๊ฒ๋ค๋ ์์. (prompt ์ ๋ ฅ์ ๊น๋ค๋ก์?)
-
ํ์ฒ๋ฆฌ(์กฐ๊ฑด๋ถ ์ธ์ด ๋ชจ๋ธ๋ง)์ ํด์ผํ๋ ์ด์ : ํ์ต์ผ๋ก ํด๊ฒฐํ๊ธฐ์๋ parameter๊ฐ ๋๋ฌด ๋ง์.
๋งํฌ)
CTG survey ๋ ผ๋ฌธ(1. Introduction 2. Concept of CTG & PLM 3. Summarized the approaches to PLM-based CTG 4. Evaluation metrics 5. CHALLENGES AND FUTURE DIRECTIONS
Introduction
CTG(Controllable Text Generation) ์ฐ๊ตฌ ๋๊ธฐ
CTG๋ NLG์ ์ค์ application์์์ ์ค์ฉ์ฑ์ ์ํด ๋ ์ค๋ฅด๊ณ ์๋ ๋ถ์ผ์ ๋๋ค.
-
NLG(Generation) : ์์ฐ์ด ๋ฌธ์ฅ์ ์์ฑํ๋ ๊ธฐ์
- NLG Application : Dialogue systems, Story generation, Data augmentation, Summarization, Question Answering ๋ฑ
-
์ค์ฉ์ฑ : ์ฑ๋ฅ, ์์ฐ์ค๋ฌ์, ์ฌ๋์ด ์ํ๋ ํน์ ์กฐ๊ฑด์ ๋ง์กฑ์ํค๋๊ฐ ๋ฑ
NLP
- NLU(Understanding) : ์์ฐ์ด ํํ์ ๋ฌธ์ฅ์ ์ดํดํ๋ ๊ธฐ์ , ์์ฐ์ด๋ฅผ ์ปดํจํฐ๊ฐ ์ดํดํ ์ ์๋ ๊ฐ์ผ๋ก ๋ฐ๊พธ๋ ๊ณผ์
- NLG(Generation) : ์์ฐ์ด ๋ฌธ์ฅ์ ์์ฑํ๋ ๊ธฐ์ , ์ปดํจํฐ๊ฐ ์ดํดํ ๊ฐ์ ์ฌ๋์ด ์ดํดํ ์ ์๋๋ก ๋ฐ๊พธ๋ ๊ณผ์
NLG using PLM
- ์ต๊ทผ Large-scale pre-trained language models (PLM) , ํนํ Transformer-based PLM์ ์ฌ์ฉํด์ text๋ฅผ ์์ฑํ๋ ๋ฐฉ๋ฒ์ด ๋์ฑ ๋ค์ํ๊ณ ์์ฐ์ค๋ฌ์ด text๋ฅผ ์์ฑํ๊ฒ ํ๋ฉด์ NLG ๋ถ์ผ์์ ๊ฐ๊ด ๋ฐ๊ณ ์์.
PLM์ ์ฌ์ฉํ Method์ ๋ฌธ์ ์
- NLG ์์คํ
์ด ๋ ๋ค์ํ๊ณ ์์ฐ์ค๋ฌ์ด text๋ฅผ ์์ ์ ์ผ๋ก ์์ฑํ๊ธฐ ์ํด์ task๋ณ๋ก control condition๋ค์ ์ถฉ์กฑ์์ผ์ผํจ
- Control condition : task๋ณ๋ก ์ฌ๋์ด ์ํ๋ ํน์ ์กฐ๊ฑด
- ex) story generation(์ผ์ชฝ๊ทธ๋ฆผ) : story ์์ ๋ฑ
- AI Chatbot(์ค๋ฅธ์ชฝ ๊ทธ๋ฆผ) : toxicํ ๋ด์ฉ ํผํ๊ธฐ, ๊ฐ์ , ๊ด๊ณ ๊ณ ๋ ค ๋ฑ
- data augmentation : ensure data distribution balance in different domains
- Application์ ์ค๋ฆฌ์ ์ธ ์ธก๋ฉด? : ์ฑ, ์ธ์ข , ๋ฑ ์ฐจ๋ณ์ ์ธ text ์์ฑ ํผํ๊ธฐ ๋ฑ
- Controllability : ์์ condition์ ๋ง์กฑ์ํค๋ text๋ฅผ ๋ง๋๋ ๋ฅ๋ ฅ
- Control condition : task๋ณ๋ก ์ฌ๋์ด ์ํ๋ ํน์ ์กฐ๊ฑด
๋ฎ์ interpretability์ controllability
- PLM์ neural network (black box ๋ชจ๋ธ)
- latent representation์ผ๋ก๋ถํฐ text ์์ฑ -> ์ฌ๋์ด ์ํ๋๋๋ก text ์์ฑ์ controllํ๊ธฐ ์ด๋ ต๊ณ interpretability๊ฐ ๋จ์ด์ง
- ์ฌ์ ํ ์น ์์ ์กด์ฌํ๋ ๊ฑฐ๋ํ ์์ ํ ์คํธ๋ฅผ ํตํด ํ์ต๋ PLM์ toxicํ ํํ(์์ค, ์ฐจ๋ณ ๋ฑ)์ ์ฌ๊ณผ์์ด ์์ฑํ์ฌ ์ค์ฉ์ ์ด์ง ๋ชปํ ์๋ ์์(์ค๋ฅธ์ชฝ ๊ทธ๋ฆผ)
"real application์์์ ์ค์ฉ์ ์ธ ํ ์คํธ ์์ฑ์ ์ํด plm based ๋ชจ๋ธ์ interpretability์ controllability๊ฐ ๋ณด์ฅ๋์ด์ผ ํ๋ค."
2. Concept of CTG & PLM
2.1 Controllable Text Generation
- Input : Controlled elements
- Process: Generative model (PLM-based model)
- Output: Generated text(satisfying the input control condition)
Goal of CTG
Given a vocabulary $V$ , Generate a target text $Y$ = { $๐ฆ_1, ๐ฆ_2, . . . , ๐ฆ_๐$ } - where $๐ฆ_๐ โ V$, Control element $C$
$$P(๐ |๐ถ) = p(๐ฆ_1, ๐ฆ_2, . . . , ๐ฆ_๐ |C)$$
- ์์ฑ๋ text $Y$ ๋ constraint condition๋ ๋ง์กฑํ๊ณ ๋, general natural languate ํน์ฑ๋ ๋ง์กฑํด์ผํ๋ค.(์ ์ฐฝํจ. ํฉ๋ฆฌ์ , ๊ฐ๋ ์ฑ )
Taxonomy of control condition
- (1) Semantic: content control at the semantic level
- emotion, topic ๋ฑ ์๋ฏธ๋ ๋ด์ฉ์ ์ธ control
- (2)Structural: control over the structure level of the generated text
- data-to text, information extraction ๋ฑ ๊ตฌ์กฐ์ ์ธ level์ control
- (3)Lexical: control at the vocabulary level
- keyword inclusion ๋ฑ such as making the generated text contain certain keywords.
2.1.1 Relevant Tasks Involving CTG
-
Attribute-based Generation
- generate natural language sentences that satisfy specific attributes
- topic, emotion, and keywords ๊ฐ์ ํน์ ํ attribute๋ฅผ ๋ง์กฑ์ํค๋ ๋ฌธ์ฅ์ ์์ฑํ๋ task
- generate natural language sentences that satisfy specific attributes
-
Dialogue Generation
- Input: dialogue content, additional structural information(eg: persona, emotion, dialogue intent, template, etc.)
- Output: Dialogue Response
-
Storytelling
- Input: Story elements (Storylines , story endings)
- Output: Story paragraph
-
Data to Text
- to convert non-linguistic structured data (e.g., a table or a graph) into natural language text
- can be applied in tasks like weather forecast
- to convert non-linguistic structured data (e.g., a table or a graph) into natural language text
-
Data Augmentation
- CTG๋ฅผ ํตํด ์ฃผ์ด์ง attibute์ ๋ง๊ฒ ์๋ก์ด text๋ฅผ ์์ฑํ๋ค.
- ๋์์ด๋ฅผ ํฌํจํ๋๋ก ๋ฑ
- CTG๋ฅผ ํตํด ์ฃผ์ด์ง attibute์ ๋ง๊ฒ ์๋ก์ด text๋ฅผ ์์ฑํ๋ค.
-
Debiasing
- Rewriting biased text to unbiased text or changing the data distribution of the CTG-generated text
- ์ฑ๋ณ,์ธ์ข ๋ฑ ํธ๊ฒฌ์ ๋ฐ์ํ๋ text๋ ๋ชจ๋ธ์ ์ํ์ํค๋ task
- Format Control
- ๋ผ์์ ๋ง์ถฐ์ผํ๋ ์ ํต ์ ๋ฑ
2.2 Transformer-based Pre-trained Language Models(PLM)
- AR : ์์ฐจ์ ์ผ๋ก ๋ค์ ํ ํฐ์ ์์ธกํ๋ LM๋ชจ๋ธ, ์ด์ ์ถ๋ ฅ์ด ๋ค์์ ์ ๋ ฅ์ด ๋๋ ๋ชจ๋ธ
- AE : ๋ ธ์ด์ฆ๊ฐ ์์ธ ์ ๋ ฅ(์๋ฅผ ๋ค์ด MASK)์์ ์ ๋ฐ์ดํฐ๋ฅผ ๋ณต์ํ๋ ๊ณผ์ ์ ๊ฑฐ์น๋ ๋ชจ๋ธ
3. Approaches to PLM-based CTG(into three categories)
PLM-based CTG ์ ํต์ฌ ์์ด๋์ด
- control condition์ ๋ง์กฑํ๋ text๋ฅผ ์์ฑํ๋๋ก ๋ชจ๋ธํํ control signal ์ ์ฃผ๋ ๊ฒ.
Standard language model:
$$P(x_n|X_{1:n-1}) = p(x_n|x_1, x_2, . . . , x_{๐-1})$$
the goal of conditional text generation
$$P(X|C) = \displaystyle \prod_{i=1}^n p(x_i|x_{< i},C)$$
control signal์ ์๋๋ฐฉ์์ ๋ฐ๋ผ ํฌ๊ฒ 3๊ฐ์ง๋ก
3.1. Fine-tuning
- PLM ํ๋ผ๋ฏธํฐ์ ์ผ๋ถ๋ฅผ finetuning ํ๋ ๋ฐฉ๋ฒ
- The most direct way at a lower cost
3.2. Retraining/Refactoring
- PLM์ architecture๋ฅผ ๋ฐ๊พธ๊ฑฐ๋ ์ฌํ์ต์ํค๋ ๋ฐฉ๋ฒ
- fine-tuning๋ณด๋ค๋ ์ข์ ๊ฒฐ๊ณผ๋ฅผ ๋ผ ์ ์์ง๋ง computing resource๊ฐ ๋ง์ด ์๋ชจ๋๊ณ , labeled data ๋ถ์กฑ ๋ฌธ์ ์ ์ง๋ฉดํ ์ ์์.
3.3. Post-Processing
- PLM์ ํ๋ผ๋ฏธํฐ ์ฌ์ด์ฆ๊ฐ ์ฆ๊ฐํ๋ฉด, ์์ ๋ฐฉ๋ฒ๋ค์ resource-intensive(requires significant system resources or time, or requires exclusive access to large amounts of data) ์ด๊ฒ์ ํผํ๊ธฐ ์ํ ๋ฐฉ๋ฒ์ผ๋ก
- PLM์ ํ๋ผ๋ฏธํฐ๋ค์ด fixed ๋๊ณ , control signal์ด ๋์ฝ๋์์ ์๋ํจ.
- require fewer computation resources for training ,can guarantee a better quality of the generated text
- ๋๋ถ๋ถ ์ด ๋ฐฉ๋ฒ์ ์.
3.1 Fine-tuning
3.1.1. Adapted Module
PLM์ 'task-related adapted network module'์ ์ถ๊ฐํ๋ ๋ฐฉ๋ฒ ์ถ๊ฐ๋ adapted network๋ ์ผ๋ฐ fine-tuning ๋ฐฉ๋ฒ์ฒ๋ผ target dataset์ ๋ํด์ PLM๊ณผ ๊ฐ์ด train๋๋ค.
-
- Training ๋ฐฉ๋ฒ : PLM์ logit๊ณผ Auxiliary model์ logits์ ๋ํ๊ณ target task output์ likelihood๋ฅผ ์ต๋ํํ๋ค.
- PLM์ logit : Learning to generate fluent, natural language learn a distribution $p(x_t|x_{< t})$ that assigns high probability to fluent sequences.
- Auxiliary model์ logits : Learning to shift the probability distribution $p(x_t|x_{< t})$ as a function of ฮฑ to obtain $p(x_t|x_{< t} ,ฮฑ)$ we would like the resulting distribution to assign high probability to a subset of fluent sequences that also adhere to the desired attribute.
- Training ๋ฐฉ๋ฒ : PLM์ logit๊ณผ Auxiliary model์ logits์ ๋ํ๊ณ target task output์ likelihood๋ฅผ ์ต๋ํํ๋ค.
- STRUCTADAPT
-
AMR-to-text task AMR:๋ฌธ์ฅ์ ์๋ฏธ ๊ตฌ์กฐ๋ฅผ ๊ทธ๋ํ๋ก ํํํ ๊ฒ
-
Adapter module์ PLM์ encoder์ decoder ๊ฐ๊ฐ์ feed-forward layer ํ์ ๋ํ๋ ๋ฐฉ๋ฒ
-
์ฐธ๊ณ ๋ก layer normalization and residual connections์ ๊ทธ๋ฆผ์์ ์๋ต๋์ด์์.
Adapter
-
- which can encode the graph structure into the PLMs without contaminating its original distributional knowledge
- ํ์ต์, the PLMโs parameters are frozen, and only the injected adapter is trainable
3.1.2. Prompted-based approaches
์ด๋ค task์ Finetuning ๋จ๊ณ์ training ๋ชฉํ๋ฅผ ์๋ PLM์ task์ ๊ฐ๊ฒ ํ๋ค.
Sentiment classification ์์
("I am always happy to see you") ๊ฐ์ฑ ๋ถ์
Traditional approach : encode the sentence into a set of vectors, and then classify their sentiment through a fully connected layer
Prompted-base : ์์ traditional ๋ฐฉ๋ฒ๊ณผ๋ ๋ค๋ฅด๊ฒ, ๋ฐ์ ๊ทธ๋ฆผ์ฒ๋ผ templete์ ๋ง๋ค์ด์ , mask ํ ํฐ์ ์์ธกํ๋๋ก ํ๋ ๊ฒ
semtiment-control text generation ์์
ํ์ฌ ์์ ๊น์ง ์์ฑ๋๊ฑฐ๋ ์ ๋ ฅ๋ ํ ํฐ์ ํตํด ๋ค์ ํ ํฐ์ ์์ธกํ๋ ์์ฑ ๋ชจ๋ธ์ ํน์ง์ ์ด์ฉํ์ฌ ์ฃผ์ด์ง ๋ฌธ์ฅ์ธ prompt๋ก ๋ค์ ์ด์ด์ง ๋ฌธ์ฅ์ ์ฌ์ฉ์์ ์๋์ ๋ง๊ฒ ์์ฑํ๋๋ก ํ๋ ๋ฐฉ๋ฒ
PLM์ input์ ์ด๋ค prompt๋ฅผ ์ถ๊ฐํด์ text ์์ฑ์ control ํ๋ค.
prompt๋ฅผ ์ด๋ป๊ฒ ๊ตฌ์ฑํ๋์ ๋ฐ๋ผ์ ๋ฐฉ๋ฒ์ด ๋๋๋ค.
- Prefix tuning
- PLM์ ํ๋ผ๋ฏธํฐ๋ freeze์ํค๊ณ
- prefix(์ผ์ข ์ ์ค๋ช ์ ๋ํ๋ด๋ ๊ฐ์ํ ํฐ)๋ผ๊ณ ํ๋ ์ฐ์์ task-specific ๋ฒกํฐ๋ฅผ ์ถ๊ฐํ์ฌ ํด๋น prefix์ ๊ด๋ จ๋ ํ๋ผ๋ฏธํฐ๋ง ํ๋ํ๋ ๋ฐฉ๋ฒ
- P-tuning
- PLM์ ์ ๋ ฅํ Prompt๋ฅผ ์์ฑํ๋ ๊ฐ๋จํ ๋ชจ๋ธ (eg. LSTM, MLP, etc.)์ ํ์ตํ์ฌ ์ํ๋ task๋ฅผ ์ํํ๋๋ก ๋ง๋๋ ๋ฐฉ๋ฒ์ ๋๋ค.
PLM์ผ๋ก ๋ค์ด๊ฐ๋ ์ ๋ ฅ ๋ฒกํฐ(prompt)๋ฅผ ํ์ตํ๋ค. p-tuning ์ค๋ช ํ๊ธ ๋ธ๋ก๊ทธ
3.1.3. Reinforcement Learning (RL) inspired Approaches
3.2 Retraining/Refactoring
- CTRL
- ์ ๋ ฅ ์ control code์ธ ํ ํฐ ํ๋๋ฅผ ๋ฌธ์ฅ ๋งจ ์์ ์ถ๊ฐํด์ค์ผ๋ก ์ํ๋ ๋ฌธ์ฅ์ ์์ฑํ ์ ์๋๋ก ๋ชจ๋ธ์ ๋ชฉ์ ํจ์์ ๊ตฌ์กฐ๋ฅผ ๋ณ๊ฒฝํ์ฌ transformer๋ฅผ ์ฒ์๋ถํฐ ํ์ต์ํค๋ ๊ฒ.
- POINTER : lexical constraints๋ฅผ ๋ง์กฑ์ํฌ ์ ์๋ ๋ฐฉ๋ฒ
- ์ด๊ธฐ ์ ๋ ฅ๋ ๋จ์ด๋ค(lexical condition) ์ฌ์ด์ ๋ชจ๋ธ์ด ์ ์ ํ ๋จ์ด๋ฅผ ์์ฑํด์ ๋งค๋๋ฌ์ด ๋ฌธ์ฅ์ด ์์ฑ๋ ์ ์๋๋ก ๋ชจ๋ธ์ ์์ ํ ๋ฐฉ๋ฒ
- CBART
- CoCon
3.3 Post-Processing
๋ค์ ํ ํฐ์ ์์ธกํ ๋, ์์ฑ๋ text๊ฐ condition์ ์ถฉ์กฑ์ํฌ ์ ์๋๋ก ์์ธก ๋ถํฌ๋ฅผ ๋ฐ๊ฟ์ฃผ๋ ๋ฐฉ๋ฒ
-
- ํ์ฒ๋ฆฌ ๋ฐฉ๋ฒ ์ค ๋ํ์ ์ธ ๋ ผ๋ฌธ
- LM ๋ชจ๋ธ ์ํคํ ์ฒ๋ ํ๋ผ๋ฏธํฐ ์์ ์๊ฑด๋๋ฆผ.
- ์ถ๋ก ํ ๋ attribute model ์ด์ฉํจ -> attribute ๋ชจ๋ธ ์ ํ๊ณ lossํจ์ ์ ์ํ๋ฉด ๋๋จ.
- ์ถ๋ก ํ ๋ ๋ค์ ๋จ์ด ๊ณ ๋ฅผ ๋ ํ๋ฅ ๋ถํฌ๋ฅผ attribute ๋ง๊ฒ ๋ฐ๊ฟ์ค.
- pplm ๋ณด์ ๋
ผ๋ฌธ:
- Change or Not: A Simple Approach for Plug and Play Language Models on Sentiment Control
- ์ง๊ธ๊น์ง ์์ฑ๋ ๋ถ๋ถ์ ์ธ ํ ์คํธ๋ก๋ง ๊ฐ์ฑ์ ์์ธกํ๊ธฐ ์ด๋ ต, ๊ณผ๋ํ ์์ ์ผ๋ก ์ธํ fluency๊ฐ ๋จ์ด์ง๋ ๊ฒ์ ๋ณด์
- Valance ์ฐจ์ด ๊ณ์ฐ์ผ๋ก change ํ ์ง ๋ง์ง ์ ํ๊ณ , loss๋ ๋ค์ ์ ์.
- Diffusion-LM Improves Controllable Text Generation
- diffusion lm ๋ชจ๋ธ ์ฌ์ฉํด์ pplm ์ฑ๋ฅ ๊ฐ์
- Attribute Alignment: Controlling Text Generation from Pre-trained Language Models
- Change or Not: A Simple Approach for Plug and Play Language Models on Sentiment Control
-
FAIR
-
MEGATRON-CNTR
-
Adversarial Search
-
GeDi
-
Mix and Match
-
EBM
4. EVALUATION METHODS
General NLG Evaluation Metrics(quality of the generated text) + CTG-specific Evaluation (fulfilling the controlled elements)
4.1 General NLG Evaluation Metrics
4.1.1 Human-Centric Evaluation Metrics.
- ์์ฑ๋ ๋ฌธ์ฅ์ ๋ํด์ ์ฌ๋์ด ์ง์ Score๋ฅผ ์ฃผ๋ ๋ฐฉ๋ฒ.
4.1.2 Automatic Evaluation Metrics
์์ฑ๋ ๋ฌธ์ฅ $G$ ์ Reference(์ ๋ต) text $R$ ์ ๋น๊ตํ๋ ๋ฐฉ๋ฒ
Lexical-based Metrics
- BLEU : Generated Sentence์ ๋จ์ด๊ฐ Reference Sentence์ ํฌํจ๋๋ ์ ๋
- ROUGE : Reference Setence์ ๋จ์ด๊ฐ Generated Sentence์ ํฌํจ๋๋ ์ ๋
- Perplexity (PPL)
-
- Generation probability์ ์ญ์์ ๊ธฐํํ๊ท (๋ฌธ์ฅ ์์ฑ ํ๋ฅ ์ ์ญ์๋ฅผ ๋จ์ด์ ์ N์ผ๋ก ์ ๊ทํ ํจ )
- 'ํท๊ฐ๋ฆฌ๋ ์ ๋'๋ก ๋ฎ์ ์๋ก ์ข์๊ฑฐ
Semantic-based Metrics
Semantic-based metrics aim to handle the evaluation of texts that are lexically different but have a similar semantic meaning
- PLM based ๋ชจ๋ธ๋ก ํ๊ฐ
- BERT์ embedding์ ์ฌ์ฉํด์ capture the semantic similarity between the generated text and its reference.
4.1.3 Semi-automatic Evaluation Metrics.
- Human-Centric Evaluation์ Automatic Evaluation๋ฅผ ํจ๊ป ์ฐ๋ ํ๊ฐ ๋ฐฉ๋ฒ
4.2 CTG-specific Evaluation
4.2.1. Semantic consistency metric(for semantic control conditions)
-
training set์ postive(control conditions์ ๋ง์กฑํ๋ sample๋ค)์ negative(๋ง์กฑํ์ง ์๋) ๋ก ๋๋๋ค.
-
classifier๊ฐ ์์ฑ๋ ๋ฌธ์ฅ์ด condition์ ๋ง์กฑํ๋ ๋ฌธ์ฅ์ธ์ง ์๋์ง (pos/neg classificaiton) ๊ตฌ๋ถํ๋๋ก ํ์ต์ํจ๋ค.
- Accuracy : condition ๋ง์กฑ์ํจ ๋ฌธ์ฅ ๋น์จ
- CTRLEVAL : ์์ฑ๋ ๋ฌธ์ฅ์ coherence, consistency, and attribute relevance๋ฅผ ํ๊ฐํจ
4.2.2. Rule-based metric(for structural and lexical-based controll conditions)
- Coverage : calculate the average percentage of input keywords that are present in the generated text of the CTG model
- Success Rate : measure the matching degree between the generated text and the given structural control elements
4.2.3. Human evaluation metric
5. CHALLENGES AND FUTURE DIRECTIONS
5.1 Challenges
-
- catastrophic forgetting problem in PLM : First, pre-trained language models have learned rich knowledge from large-scale corpus used for pre-training. However, an NLG model needs to learn control constraints on its own training corpus. It is often difficult for the existing PLM-based models to ensure the domain diversity of the generated text while pursuing controllability. This is indeed the well-known catastrophic forgetting problem in PLM. In the field of text generation, it is still a challenge to overcome this problem and make the PLM-based NLG model generate multi-domain text that satisfies specific control conditions with few or zero domain-specific samples.
-
- controlling the generation of text in the decoding stage of a generative mode ์ ๋ฌธ์ ์ :
- distribution gap between the discriminator and the generator, leading to a coarser granularity in the guidance process and decreased quality of the generated text.
- hard to be directly applied to fine-grained control scenarios such as date-to-text or multi-attribute control tasks.
-
- GPT-like models $p(x_n|x_1, x_2, . . . , x_{๐-1})$
- this local normalization format has certain limitations in paragraph/document-level
modeling
- For example, it is hard to keep long-range coherency in both semantic logic and controlled condition. It calls for further research to establish global normalization based on PLMs to ensure that text generation can be controlled locally and globally at the same time.
-
- Fourth, the construction of large-scale pre-trained language models is typically data-driven, which allows the models to learn the primary logic and common sense contained in the training corpus. However, the knowledge captured in those models is rather superficial. The PLMs will lose generalization ability when the training data does not contain relevant common sense and domain-specific knowledge. Therefore, purely relying on PLMs could be difficult to control the generated texts faithfully to common sense and rich knowledge specific to the target domain
-
- Fifth, a reasonable and reliable evaluation has always been a bottleneck restricting the development of more advanced text generation technologies
5.2 Future Directions
- Prompt-based Learning
- Based on the well-designed prompting function, a PLM is able to perform few-shot or even zero-shot learning, adapting to new scenarios with few or no labeled data, thus overcoming the problem of catastrophic forgetting ->CTG ์ ์ฉ
- Fine-grained(์๊ฒ ์ชผ๊ฐ๋ ๊ฒ) Decoding Control
Integration with Classic Generative Theory and Linguistic Knowledge Incorporation of Knowledge Graphs:
Novel Evaluation Metrics and Methods
New CTG tasks [๊ต์๋]
- ๋ ธ์ธ์นํ์ ์ธ ๋ํ ์ ๋
- ์๋ํํ๋ซํผ ์์ : ์ฌ๊ธฐ์ ๊ธฐ ์ํ์ -> ์ด๋ค ์ฝ์ ๋จน์ผ์ธ์ -> ๋ถ๋ฒํ์์ (constraint ์ด์ผ ์ฃผ๋ , ์ด์ผ ํผํ๋ )
- ์์ ์ฐํํ๋ text control ex) ์์๋ชจ์์