TTiti的学习笔记
首页 / 专业知识 / 40-References/Papers/gpt - GPT/01_original.md

Improving Language Understanding by Generative Pre-Training Alec Radford Karthik Narasimhan Tim Salimans Ilya Sutskever OpenAI OpenAI OpenAI OpenAI

专业知识 · 40-References/Papers/gpt - GPT/01_original.md

--- title: "Improving Language Understanding by Generative Pre-Training Alec Radford Karthik Narasimhan Tim Salimans Ilya Sutskever OpenAI OpenAI OpenAI OpenAI" aliases: - "GPT" source: "https://openai.com/research/language-unsupervised" arxiv: "" created: 2026-07-16 type: paper-translation status: extraction-complete_translation-pending tags: - paper - ml - deep-learning - nlp


Improving Language Understanding by Generative Pre-Training Alec Radford Karthik Narasimhan Tim Salimans Ilya Sutskever OpenAI OpenAI OpenAI OpenAI

原文全文

Page 1

<a id="S0001"></a> Source: p.1 S0001

Improving Language Understanding by Generative Pre-Training Alec Radford Karthik Narasimhan Tim Salimans Ilya Sutskever OpenAI OpenAI OpenAI OpenAI alec@openai.com karthikn@openai.com tim@openai.com ilyasu@openai.com Abstract Natural language understanding comprises a wide range of diverse tasks such as textual entailment, question answering, semantic similarity assessment, and document classification.

<a id="S0002"></a> Source: p.1 S0002

Although large unlabeled text corpora are abundant, labeled data for learning these specific tasks is scarce, making it challenging for discriminatively trained models to perform adequately.

<a id="S0003"></a> Source: p.1 S0003

We demonstrate that large gains on these tasks can be realized by generative pre-training of a language model on a diverse corpus of unlabeled text, followed by discriminative fine-tuning on each specific task.

<a id="S0004"></a> Source: p.1 S0004

In contrast to previous approaches, we make use of task-aware input transformations during fine-tuning to achieve effective transfer while requiring minimal changes to the model architecture.

<a id="S0005"></a> Source: p.1 S0005

We demonstrate the effectiveness of our approach on a wide range of benchmarks for natural language understanding.

<a id="S0006"></a> Source: p.1 S0006

Our general task-agnostic model outperforms discriminatively trained models that use architectures specifically crafted for each task, significantly improving upon the state of the art in 9 out of the 12 tasks studied.

<a id="S0007"></a> Source: p.1 S0007

For instance, we achieve absolute improvements of 8.9% on commonsense reasoning (Stories Cloze Test), 5.7% on question answering (RACE), and 1.5% on textual entailment (MultiNLI). 1 Introduction The ability to learn effectively from raw text is crucial to alleviating the dependence on supervised learning in natural language processing (NLP).

<a id="S0008"></a> Source: p.1 S0008

Most deep learning methods require substantial amounts of manually labeled data, which restricts their applicability in many domains that suffer from a dearth of annotated resources [61].

<a id="S0009"></a> Source: p.1 S0009

In these situations, models that can leverage linguistic information from unlabeled data provide a valuable alternative to gathering more annotation, which can be time-consuming and expensive.

<a id="S0010"></a> Source: p.1 S0010

Further, even in cases where considerable supervision is available, learning good representations in an unsupervised fashion can provide a significant performance boost.

<a id="S0011"></a> Source: p.1 S0011

The most compelling evidence for this so far has been the extensive use of pretrained word embeddings [10, 39, 42] to improve performance on a range of NLP tasks [8, 11, 26, 45].

<a id="S0012"></a> Source: p.1 S0012

Leveraging more than word-level information from unlabeled text, however, is challenging for two main reasons.

<a id="S0013"></a> Source: p.1 S0013

First, it is unclear what type of optimization objectives are most effective at learning text representations that are useful for transfer.

<a id="S0014"></a> Source: p.1 S0014

Recent research has looked at various objectives such as language modeling [44], machine translation [38], and discourse coherence [22], with each method outperforming the others on different tasks.1 Second, there is no consensus on the most effective way to transfer these learned representations to the target task.

<a id="S0015"></a> Source: p.1 S0015

Existing techniques involve a combination of making task-specific changes to the model architecture [43, 44], using intricate learning schemes [21] and adding auxiliary learning objectives [50].

<a id="S0016"></a> Source: p.1 S0016

These uncertainties have made it difficult to develop effective semi-supervised learning approaches for language processing. 1https://gluebenchmark.com/leaderboard Preprint.

Page 2

<a id="S0017"></a> Source: p.2 S0017

In this paper, we explore a semi-supervised approach for language understanding tasks using a combination of unsupervised pre-training and supervised fine-tuning.

<a id="S0018"></a> Source: p.2 S0018

Our goal is to learn a universal representation that transfers with little adaptation to a wide range of tasks.

<a id="S0019"></a> Source: p.2 S0019

We assume access to a large corpus of unlabeled text and several datasets with manually annotated training examples (target tasks).

<a id="S0020"></a> Source: p.2 S0020

Our setup does not require these target tasks to be in the same domain as the unlabeled corpus.

<a id="S0021"></a> Source: p.2 S0021

We employ a two-stage training procedure.

<a id="S0022"></a> Source: p.2 S0022

First, we use a language modeling objective on the unlabeled data to learn the initial parameters of a neural network model.

<a id="S0023"></a> Source: p.2 S0023

Subsequently, we adapt these parameters to a target task using the corresponding supervised objective.

<a id="S0024"></a> Source: p.2 S0024

For our model architecture, we use the Transformer [62], which has been shown to perform strongly on various tasks such as machine translation [62], document generation [34], and syntactic parsing [29].

<a id="S0025"></a> Source: p.2 S0025

This model choice provides us with a more structured memory for handling long-term dependencies in text, compared to alternatives like recurrent networks, resulting in robust transfer performance across diverse tasks.

<a id="S0026"></a> Source: p.2 S0026

During transfer, we utilize task-specific input adaptations derived from traversal-style approaches [52], which process structured text input as a single contiguous sequence of tokens.

<a id="S0027"></a> Source: p.2 S0027

As we demonstrate in our experiments, these adaptations enable us to fine-tune effectively with minimal changes to the architecture of the pre-trained model.

<a id="S0028"></a> Source: p.2 S0028

We evaluate our approach on four types of language understanding tasks – natural language inference, question answering, semantic similarity, and text classification.

<a id="S0029"></a> Source: p.2 S0029

Our general task-agnostic model outperforms discriminatively trained models that employ architectures specifically crafted for each task, significantly improving upon the state of the art in 9 out of the 12 tasks studied.

<a id="S0030"></a> Source: p.2 S0030

For instance, we achieve absolute improvements of 8.9% on commonsense reasoning (Stories Cloze Test) [40], 5.7% on question answering (RACE) [30], 1.5% on textual entailment (MultiNLI) [66] and 5.5% on the recently introduced GLUE multi-task benchmark [64].

<a id="S0031"></a> Source: p.2 S0031

We also analyzed zero-shot behaviors of the pre-trained model on four different settings and demonstrate that it acquires useful linguistic knowledge for downstream tasks. 2 Related Work Semi-supervised learning for NLP Our work broadly falls under the category of semi-supervised learning for natural language.

<a id="S0032"></a> Source: p.2 S0032

This paradigm has attracted significant interest, with applications to tasks like sequence labeling [24, 33, 57] or text classification [41, 70].

<a id="S0033"></a> Source: p.2 S0033

The earliest approaches used unlabeled data to compute word-level or phrase-level statistics, which were then used as features in a supervised model [33].

<a id="S0034"></a> Source: p.2 S0034

Over the last few years, researchers have demonstrated the benefits of using word embeddings [11, 39, 42], which are trained on unlabeled corpora, to improve performance on a variety of tasks [8, 11, 26, 45].

<a id="S0035"></a> Source: p.2 S0035

These approaches, however, mainly transfer word-level information, whereas we aim to capture higher-level semantics.

<a id="S0036"></a> Source: p.2 S0036

Recent approaches have investigated learning and utilizing more than word-level semantics from unlabeled data.

<a id="S0037"></a> Source: p.2 S0037

Phrase-level or sentence-level embeddings, which can be trained using an unlabeled corpus, have been used to encode text into suitable vector representations for various target tasks [28, 32, 1, 36, 22, 12, 56, 31].

<a id="S0038"></a> Source: p.2 S0038

Unsupervised pre-training Unsupervised pre-training is a special case of semi-supervised learning where the goal is to find a good initialization point instead of modifying the supervised learning objective.

<a id="S0039"></a> Source: p.2 S0039

Early works explored the use of the technique in image classification [20, 49, 63] and regression tasks [3].

<a id="S0040"></a> Source: p.2 S0040

Subsequent research [15] demonstrated that pre-training acts as a regularization scheme, enabling better generalization in deep neural networks.

<a id="S0041"></a> Source: p.2 S0041

In recent work, the method has been used to help train deep neural networks on various tasks like image classification [69], speech recognition [68], entity disambiguation [17] and machine translation [48].

<a id="S0042"></a> Source: p.2 S0042

The closest line of work to ours involves pre-training a neural network using a language modeling objective and then fine-tuning it on a target task with supervision.

<a id="S0043"></a> Source: p.2 S0043

Dai et al. [13] and Howard and Ruder [21] follow this method to improve text classification.

<a id="S0044"></a> Source: p.2 S0044

However, although the pre-training phase helps capture some linguistic information, their usage of LSTM models restricts their prediction ability to a short range.

<a id="S0045"></a> Source: p.2 S0045

In contrast, our choice of transformer networks allows us to capture longerrange linguistic structure, as demonstrated in our experiments.

<a id="S0046"></a> Source: p.2 S0046

Further, we also demonstrate the effectiveness of our model on a wider range of tasks including natural language inference, paraphrase detection and story completion.

<a id="S0047"></a> Source: p.2 S0047

Other approaches [43, 44, 38] use hidden representations from a 2

Page 3

<a id="S0048"></a> Source: p.3 S0048

pre-trained language or machine translation model as auxiliary features while training a supervised model on the target task.

<a id="S0049"></a> Source: p.3 S0049

This involves a substantial amount of new parameters for each separate target task, whereas we require minimal changes to our model architecture during transfer.

<a id="S0050"></a> Source: p.3 S0050

Auxiliary training objectives Adding auxiliary unsupervised training objectives is an alternative form of semi-supervised learning.

<a id="S0051"></a> Source: p.3 S0051

Early work by Collobert and Weston [10] used a wide variety of auxiliary NLP tasks such as POS tagging, chunking, named entity recognition, and language modeling to improve semantic role labeling.

<a id="S0052"></a> Source: p.3 S0052

More recently, Rei [50] added an auxiliary language modeling objective to their target task objective and demonstrated performance gains on sequence labeling tasks.

<a id="S0053"></a> Source: p.3 S0053

Our experiments also use an auxiliary objective, but as we show, unsupervised pre-training already learns several linguistic aspects relevant to target tasks. 3 Framework Our training procedure consists of two stages.

<a id="S0054"></a> Source: p.3 S0054

The first stage is learning a high-capacity language model on a large corpus of text.

<a id="S0055"></a> Source: p.3 S0055

This is followed by a fine-tuning stage, where we adapt the model to a discriminative task with labeled data. 3.1 Unsupervised pre-training Given an unsupervised corpus of tokens U = {u , . . . , u }, we use a standard language modeling 1 n objective to maximize the following likelihood: (cid:88) L (U) = log P (u |u , . . . , u ; Θ) (1) 1 i i−k i−1 i where k is the size of the context window, and the conditional probability P is modeled using a neural network with parameters Θ.

<a id="S0056"></a> Source: p.3 S0056

These parameters are trained using stochastic gradient descent [51].

<a id="S0057"></a> Source: p.3 S0057

In our experiments, we use a multi-layer Transformer decoder [34] for the language model, which is a variant of the transformer [62].

<a id="S0058"></a> Source: p.3 S0058

This model applies a multi-headed self-attention operation over the input context tokens followed by position-wise feedforward layers to produce an output distribution over target tokens: h = U W + W 0 e p h l = transformer_block(h l−1 )∀i ∈ [1, n] (2) P (u) = softmax(h W T ) n e where U = (u , . . . , u ) is the context vector of tokens, n is the number of layers, W is the token −k −1 e embedding matrix, and W is the position embedding matrix. p 3.2 Supervised fine-tuning After training the model with the objective in Eq. 1, we adapt the parameters to the supervised target task.

<a id="S0059"></a> Source: p.3 S0059

We assume a labeled dataset C, where each instance consists of a sequence of input tokens, x1, . . . , xm, along with a label y.

<a id="S0060"></a> Source: p.3 S0060

The inputs are passed through our pre-trained model to obtain the final transformer block’s activation hm, which is then fed into an added linear output layer with l parameters W to predict y: y P (y|x1, . . . , xm) = softmax(hmW ). (3) l y This gives us the following objective to maximize: (cid:88) L (C) = log P (y|x1, . . . , xm). (4) 2 (x,y) We additionally found that including language modeling as an auxiliary objective to the fine-tuning helped learning by (a) improving generalization of the supervised model, and (b) accelerating convergence.

<a id="S0061"></a> Source: p.3 S0061

This is in line with prior work [50, 43], who also observed improved performance with such an auxiliary objective.

<a id="S0062"></a> Source: p.3 S0062

Specifically, we optimize the following objective (with weight λ): L (C) = L (C) + λ ∗ L (C) (5) 3 2 1 Overall, the only extra parameters we require during fine-tuning are W , and embeddings for delimiter y tokens (described below in Section 3.3). 3

Page 4

<a id="S0063"></a> Source: p.4 S0063

Figure 1: (left) Transformer architecture and training objectives used in this work. (right) Input transformations for fine-tuning on different tasks.

<a id="S0064"></a> Source: p.4 S0064

We convert all structured inputs into token sequences to be processed by our pre-trained model, followed by a linear+softmax layer. 3.3 Task-specific input transformations For some tasks, like text classification, we can directly fine-tune our model as described above.

<a id="S0065"></a> Source: p.4 S0065

Certain other tasks, like question answering or textual entailment, have structured inputs such as ordered sentence pairs, or triplets of document, question, and answers.

<a id="S0066"></a> Source: p.4 S0066

Since our pre-trained model was trained on contiguous sequences of text, we require some modifications to apply it to these tasks.

<a id="S0067"></a> Source: p.4 S0067

Previous work proposed learning task specific architectures on top of transferred representations [44].

<a id="S0068"></a> Source: p.4 S0068

Such an approach re-introduces a significant amount of task-specific customization and does not use transfer learning for these additional architectural components.

<a id="S0069"></a> Source: p.4 S0069

Instead, we use a traversal-style approach [52], where we convert structured inputs into an ordered sequence that our pre-trained model can process.

<a id="S0070"></a> Source: p.4 S0070

These input transformations allow us to avoid making extensive changes to the architecture across tasks.

<a id="S0071"></a> Source: p.4 S0071

We provide a brief description of these input transformations below and Figure 1 provides a visual illustration.

<a id="S0072"></a> Source: p.4 S0072

All transformations include adding randomly initialized start and end tokens ((cid:104)s(cid:105), (cid:104)e(cid:105)).

<a id="S0073"></a> Source: p.4 S0073

Textual entailment For entailment tasks, we concatenate the premise p and hypothesis h token sequences, with a delimiter token ($) in between.

<a id="S0074"></a> Source: p.4 S0074

Similarity For similarity tasks, there is no inherent ordering of the two sentences being compared.

<a id="S0075"></a> Source: p.4 S0075

To reflect this, we modify the input sequence to contain both possible sentence orderings (with a delimiter in between) and process each independently to produce two sequence representations hm l which are added element-wise before being fed into the linear output layer.

<a id="S0076"></a> Source: p.4 S0076

Question Answering and Commonsense Reasoning For these tasks, we are given a context document z, a question q, and a set of possible answers {a }.

<a id="S0077"></a> Source: p.4 S0077

We concatenate the document context k and question with each possible answer, adding a delimiter token in between to get [z; q; $; a ].

<a id="S0078"></a> Source: p.4 S0078

Each k of these sequences are processed independently with our model and then normalized via a softmax layer to produce an output distribution over possible answers. 4 Experiments 4.1 Setup Unsupervised pre-training We use the BooksCorpus dataset [71] for training the language model.

<a id="S0079"></a> Source: p.4 S0079

It contains over 7,000 unique unpublished books from a variety of genres including Adventure, Fantasy, and Romance.

<a id="S0080"></a> Source: p.4 S0080

Crucially, it contains long stretches of contiguous text, which allows the generative model to learn to condition on long-range information.

<a id="S0081"></a> Source: p.4 S0081

An alternative dataset, the 1B Word Benchmark, which is used by a similar approach, ELMo [44], is approximately the same size 4

Page 5

<a id="S0082"></a> Source: p.5 S0082

Table 1: A list of the different tasks and datasets used in our experiments.

<a id="S0083"></a> Source: p.5 S0083

Task Datasets Natural language inference SNLI [5], MultiNLI [66], Question NLI [64], RTE [4], SciTail [25] Question Answering RACE [30], Story Cloze [40] Sentence similarity MSR Paraphrase Corpus [14], Quora Question Pairs [9], STS Benchmark [6] Classification Stanford Sentiment Treebank-2 [54], CoLA [65] but is shuffled at a sentence level - destroying long-range structure.

<a id="S0084"></a> Source: p.5 S0084

Our language model achieves a very low token level perplexity of 18.4 on this corpus.

<a id="S0085"></a> Source: p.5 S0085

Model specifications Our model largely follows the original transformer work [62].

<a id="S0086"></a> Source: p.5 S0086

We trained a 12-layer decoder-only transformer with masked self-attention heads (768 dimensional states and 12 attention heads).

<a id="S0087"></a> Source: p.5 S0087

For the position-wise feed-forward networks, we used 3072 dimensional inner states.

<a id="S0088"></a> Source: p.5 S0088

We used the Adam optimization scheme [27] with a max learning rate of 2.5e-4.

<a id="S0089"></a> Source: p.5 S0089

The learning rate was increased linearly from zero over the first 2000 updates and annealed to 0 using a cosine schedule.

<a id="S0090"></a> Source: p.5 S0090

We train for 100 epochs on minibatches of 64 randomly sampled, contiguous sequences of 512 tokens.

<a id="S0091"></a> Source: p.5 S0091

Since layernorm [2] is used extensively throughout the model, a simple weight initialization of N (0, 0.02) was sufficient.

<a id="S0092"></a> Source: p.5 S0092

We used a bytepair encoding (BPE) vocabulary with 40,000 merges [53] and residual, embedding, and attention dropouts with a rate of 0.1 for regularization.

<a id="S0093"></a> Source: p.5 S0093

We also employed a modified version of L2 regularization proposed in [37], with w = 0.01 on all non bias or gain weights.

<a id="S0094"></a> Source: p.5 S0094

For the activation function, we used the Gaussian Error Linear Unit (GELU) [18].

<a id="S0095"></a> Source: p.5 S0095

We used learned position embeddings instead of the sinusoidal version proposed in the original work.

<a id="S0096"></a> Source: p.5 S0096

We use the ftfy library2 to clean the raw text in BooksCorpus, standardize some punctuation and whitespace, and use the spaCy tokenizer.3 Fine-tuning details Unless specified, we reuse the hyperparameter settings from unsupervised pre-training.

<a id="S0097"></a> Source: p.5 S0097

We add dropout to the classifier with a rate of 0.1.

<a id="S0098"></a> Source: p.5 S0098

For most tasks, we use a learning rate of 6.25e-5 and a batchsize of 32.

<a id="S0099"></a> Source: p.5 S0099

Our model finetunes quickly and 3 epochs of training was sufficient for most cases.

<a id="S0100"></a> Source: p.5 S0100

We use a linear learning rate decay schedule with warmup over 0.2% of training. λ was set to 0.5. 4.2 Supervised fine-tuning We perform experiments on a variety of supervised tasks including natural language inference, question answering, semantic similarity, and text classification.

<a id="S0101"></a> Source: p.5 S0101

Some of these tasks are available as part of the recently released GLUE multi-task benchmark [64], which we make use of.

<a id="S0102"></a> Source: p.5 S0102

Figure 1 provides an overview of all the tasks and datasets.

<a id="S0103"></a> Source: p.5 S0103

Natural Language Inference The task of natural language inference (NLI), also known as recognizing textual entailment, involves reading a pair of sentences and judging the relationship between them from one of entailment, contradiction or neutral.

<a id="S0104"></a> Source: p.5 S0104

Although there has been a lot of recent interest [58, 35, 44], the task remains challenging due to the presence of a wide variety of phenomena like lexical entailment, coreference, and lexical and syntactic ambiguity.

<a id="S0105"></a> Source: p.5 S0105

We evaluate on five datasets with diverse sources, including image captions (SNLI), transcribed speech, popular fiction, and government reports (MNLI), Wikipedia articles (QNLI), science exams (SciTail) or news articles (RTE).

<a id="S0106"></a> Source: p.5 S0106

Table 2 details various results on the different NLI tasks for our model and previous state-of-the-art approaches.

<a id="S0107"></a> Source: p.5 S0107

Our method significantly outperforms the baselines on four of the five datasets, achieving absolute improvements of upto 1.5% on MNLI, 5% on SciTail, 5.8% on QNLI and 0.6% on SNLI over the previous best results.

<a id="S0108"></a> Source: p.5 S0108

This demonstrates our model’s ability to better reason over multiple sentences, and handle aspects of linguistic ambiguity.

<a id="S0109"></a> Source: p.5 S0109

On RTE, one of the smaller datasets we evaluate on (2490 examples), we achieve an accuracy of 56%, which is below the 61.7% reported by a multi-task biLSTM model.

<a id="S0110"></a> Source: p.5 S0110

Given the strong performance of our approach on larger NLI datasets, it is likely our model will benefit from multi-task training as well but we have not explored this currently. 2https://ftfy.readthedocs.io/en/latest/ 3https://spacy.io/ 5

Page 6

<a id="S0111"></a> Source: p.6 S0111

Table 2: Experimental results on natural language inference tasks, comparing our model with current state-of-the-art methods. 5x indicates an ensemble of 5 models.

<a id="S0112"></a> Source: p.6 S0112

All datasets use accuracy as the evaluation metric.

<a id="S0113"></a> Source: p.6 S0113

Method MNLI-m MNLI-mm SNLI SciTail QNLI RTE ESIM + ELMo [44] (5x) - - 89.3 - - - CAFE [58] (5x) 80.2 79.0 89.3 - - - Stochastic Answer Network [35] (3x) 80.6 80.1 - - - - CAFE [58] 78.7 77.9 88.5 83.3 GenSen [64] 71.4 71.3 - - 82.3 59.2 Multi-task BiLSTM + Attn [64] 72.2 72.1 - - 82.1 61.7 Finetuned Transformer LM (ours) 82.1 81.4 89.9 88.3 88.1 56.0 Table 3: Results on question answering and commonsense reasoning, comparing our model with current state-of-the-art methods.. 9x means an ensemble of 9 models.

<a id="S0114"></a> Source: p.6 S0114

Method Story Cloze RACE-m RACE-h RACE val-LS-skip [55] 76.5 - - - Hidden Coherence Model [7] 77.6 - - - Dynamic Fusion Net [67] (9x) - 55.6 49.4 51.2 BiAttention MRU [59] (9x) - 60.2 50.3 53.3 Finetuned Transformer LM (ours) 86.5 62.9 57.4 59.0 Question answering and commonsense reasoning Another task that requires aspects of single and multi-sentence reasoning is question answering.

<a id="S0115"></a> Source: p.6 S0115

We use the recently released RACE dataset [30], consisting of English passages with associated questions from middle and high school exams.

<a id="S0116"></a> Source: p.6 S0116

This corpus has been shown to contain more reasoning type questions that other datasets like CNN [19] or SQuaD [47], providing the perfect evaluation for our model which is trained to handle long-range contexts.

<a id="S0117"></a> Source: p.6 S0117

In addition, we evaluate on the Story Cloze Test [40], which involves selecting the correct ending to multi-sentence stories from two options.

<a id="S0118"></a> Source: p.6 S0118

On these tasks, our model again outperforms the previous best results by significant margins - up to 8.9% on Story Cloze, and 5.7% overall on RACE.

<a id="S0119"></a> Source: p.6 S0119

This demonstrates the ability of our model to handle long-range contexts effectively.

<a id="S0120"></a> Source: p.6 S0120

Semantic Similarity Semantic similarity (or paraphrase detection) tasks involve predicting whether two sentences are semantically equivalent or not.

<a id="S0121"></a> Source: p.6 S0121

The challenges lie in recognizing rephrasing of concepts, understanding negation, and handling syntactic ambiguity.

<a id="S0122"></a> Source: p.6 S0122

We use three datasets for this task – the Microsoft Paraphrase corpus (MRPC) [14] (collected from news sources), the Quora Question Pairs (QQP) dataset [9], and the Semantic Textual Similarity benchmark (STS-B) [6].

<a id="S0123"></a> Source: p.6 S0123

We obtain state-of-the-art results on two of the three semantic similarity tasks (Table 4) with a 1 point absolute gain on STS-B.

<a id="S0124"></a> Source: p.6 S0124

The performance delta on QQP is significant, with a 4.2% absolute improvement over Single-task BiLSTM + ELMo + Attn.

<a id="S0125"></a> Source: p.6 S0125

Classification Finally, we also evaluate on two different text classification tasks.

<a id="S0126"></a> Source: p.6 S0126

The Corpus of Linguistic Acceptability (CoLA) [65] contains expert judgements on whether a sentence is grammatical or not, and tests the innate linguistic bias of trained models.

<a id="S0127"></a> Source: p.6 S0127

The Stanford Sentiment Treebank (SST-2) [54], on the other hand, is a standard binary classification task.

<a id="S0128"></a> Source: p.6 S0128

Our model obtains an score of 45.4 on CoLA, which is an especially big jump over the previous best result of 35.0, showcasing the innate linguistic bias learned by our model.

<a id="S0129"></a> Source: p.6 S0129

The model also achieves 91.3% accuracy on SST-2, which is competitive with the state-of-the-art results.

<a id="S0130"></a> Source: p.6 S0130

We also achieve an overall score of 72.8 on the GLUE benchmark, which is significantly better than the previous best of 68.9. 6

Page 7

<a id="S0131"></a> Source: p.7 S0131

Table 4: Semantic similarity and classification results, comparing our model with current state-of-theart methods.

<a id="S0132"></a> Source: p.7 S0132

All task evaluations in this table were done using the GLUE benchmark. (mc= Mathews correlation, acc=Accuracy, pc=Pearson correlation) Method Classification Semantic Similarity GLUE CoLA SST2 MRPC STSB QQP (mc) (acc) (F1) (pc) (F1) Sparse byte mLSTM [16] - 93.2 - - - - TF-KLD [23] - - 86.0 - - - ECNU (mixed ensemble) [60] - - - 81.0 - - Single-task BiLSTM + ELMo + Attn [64] 35.0 90.2 80.2 55.5 66.1 64.8 Multi-task BiLSTM + ELMo + Attn [64] 18.9 91.6 83.5 72.8 63.3 68.9 Finetuned Transformer LM (ours) 45.4 91.3 82.3 82.0 70.3 72.8 Overall, our approach achieves new state-of-the-art results in 9 out of the 12 datasets we evaluate on, outperforming ensembles in many cases.

<a id="S0133"></a> Source: p.7 S0133

Our results also indicate that our approach works well across datasets of different sizes, from smaller datasets such as STS-B (≈5.7k training examples) – to the largest one – SNLI (≈550k training examples). 5 Analysis Impact of number of layers transferred We observed the impact of transferring a variable number of layers from unsupervised pre-training to the supervised target task.

<a id="S0134"></a> Source: p.7 S0134

Figure 2(left) illustrates the performance of our approach on MultiNLI and RACE as a function of the number of layers transferred.

<a id="S0135"></a> Source: p.7 S0135

We observe the standard result that transferring embeddings improves performance and that each transformer layer provides further benefits up to 9% for full transfer on MultiNLI.

<a id="S0136"></a> Source: p.7 S0136

This indicates that each layer in the pre-trained model contains useful functionality for solving target tasks.

<a id="S0137"></a> Source: p.7 S0137

Figure 2: (left) Effect of transferring increasing number of layers from the pre-trained language model on RACE and MultiNLI. (right) Plot showing the evolution of zero-shot performance on different tasks as a function of LM pre-training updates.

<a id="S0138"></a> Source: p.7 S0138

Performance per task is normalized between a random guess baseline and the current state-of-the-art with a single model.

<a id="S0139"></a> Source: p.7 S0139

Zero-shot Behaviors We’d like to better understand why language model pre-training of transformers is effective. A hypothesis is that the underlying generative model learns to perform many of the tasks we evaluate on in order to improve its language modeling capability and that the more structured 7

Page 8

<a id="S0140"></a> Source: p.8 S0140

Table 5: Analysis of various model ablations on different tasks.

<a id="S0141"></a> Source: p.8 S0141

Avg. score is a unweighted average of all the results. (mc= Mathews correlation, acc=Accuracy, pc=Pearson correlation) Method Avg.

<a id="S0142"></a> Source: p.8 S0142

Score CoLA SST2 MRPC STSB QQP MNLI QNLI RTE (mc) (acc) (F1) (pc) (F1) (acc) (acc) (acc) Transformer w/ aux LM (full) 74.7 45.4 91.3 82.3 82.0 70.3 81.8 88.1 56.0 Transformer w/o pre-training 59.9 18.9 84.0 79.4 30.9 65.5 75.7 71.2 53.8 Transformer w/o aux LM 75.0 47.9 92.0 84.9 83.2 69.8 81.1 86.9 54.4 LSTM w/ aux LM 69.1 30.3 90.5 83.2 71.8 68.1 73.7 81.1 54.6 attentional memory of the transformer assists in transfer compared to LSTMs.

<a id="S0143"></a> Source: p.8 S0143

We designed a series of heuristic solutions that use the underlying generative model to perform tasks without supervised finetuning.

<a id="S0144"></a> Source: p.8 S0144

We visualize the effectiveness of these heuristic solutions over the course of generative pre-training in Fig 2(right).

<a id="S0145"></a> Source: p.8 S0145

We observe the performance of these heuristics is stable and steadily increases over training suggesting that generative pretraining supports the learning of a wide variety of task relevant functionality.

<a id="S0146"></a> Source: p.8 S0146

We also observe the LSTM exhibits higher variance in its zero-shot performance suggesting that the inductive bias of the Transformer architecture assists in transfer.

<a id="S0147"></a> Source: p.8 S0147

For CoLA (linguistic acceptability), examples are scored as the average token log-probability the generative model assigns and predictions are made by thresholding.

<a id="S0148"></a> Source: p.8 S0148

For SST-2 (sentiment analysis), we append the token very to each example and restrict the language model’s output distribution to only the words positive and negative and guess the token it assigns higher probability to as the prediction.

<a id="S0149"></a> Source: p.8 S0149

For RACE (question answering), we pick the answer the generative model assigns the highest average token log-probability when conditioned on the document and question.

<a id="S0150"></a> Source: p.8 S0150

For DPRD [46] (winograd schemas), we replace the definite pronoun with the two possible referrents and predict the resolution that the generative model assigns higher average token log-probability to the rest of the sequence after the substitution.

<a id="S0151"></a> Source: p.8 S0151

Ablation studies We perform three different ablation studies (Table 5).

<a id="S0152"></a> Source: p.8 S0152

First, we examine the performance of our method without the auxiliary LM objective during fine-tuning.

<a id="S0153"></a> Source: p.8 S0153

We observe that the auxiliary objective helps on the NLI tasks and QQP.

<a id="S0154"></a> Source: p.8 S0154

Overall, the trend suggests that larger datasets benefit from the auxiliary objective but smaller datasets do not.

<a id="S0155"></a> Source: p.8 S0155

Second, we analyze the effect of the Transformer by comparing it with a single layer 2048 unit LSTM using the same framework.

<a id="S0156"></a> Source: p.8 S0156

We observe a 5.6 average score drop when using the LSTM instead of the Transformer.

<a id="S0157"></a> Source: p.8 S0157

The LSTM only outperforms the Transformer on one dataset – MRPC.

<a id="S0158"></a> Source: p.8 S0158

Finally, we also compare with our transformer architecture directly trained on supervised target tasks, without pre-training.

<a id="S0159"></a> Source: p.8 S0159

We observe that the lack of pre-training hurts performance across all the tasks, resulting in a 14.8% decrease compared to our full model. 6 Conclusion We introduced a framework for achieving strong natural language understanding with a single task-agnostic model through generative pre-training and discriminative fine-tuning.

<a id="S0160"></a> Source: p.8 S0160

By pre-training on a diverse corpus with long stretches of contiguous text our model acquires significant world knowledge and ability to process long-range dependencies which are then successfully transferred to solving discriminative tasks such as question answering, semantic similarity assessment, entailment determination, and text classification, improving the state of the art on 9 of the 12 datasets we study.

<a id="S0161"></a> Source: p.8 S0161

Using unsupervised (pre-)training to boost performance on discriminative tasks has long been an important goal of Machine Learning research.

<a id="S0162"></a> Source: p.8 S0162

Our work suggests that achieving significant performance gains is indeed possible, and offers hints as to what models (Transformers) and data sets (text with long range dependencies) work best with this approach.

<a id="S0163"></a> Source: p.8 S0163

We hope that this will help enable new research into unsupervised learning, for both natural language understanding and other domains, further improving our understanding of how and when unsupervised learning works.

<a id="S0164"></a> Source: p.8 S0164

Ma. A simple but tough-to-beat baseline for sentence embeddings. 2016. 8

Page 9

<a id="S0165"></a> Source: p.9 S0165

Layer normalization. arXiv preprint arXiv:1607.06450, 2016. [3] Y.

<a id="S0166"></a> Source: p.9 S0166

Greedy layer-wise training of deep networks.

<a id="S0167"></a> Source: p.9 S0167

In Advances in neural information processing systems, pages 153–160, 2007. [4] L.

<a id="S0168"></a> Source: p.9 S0168

The fifth pascal recognizing textual entailment challenge.

<a id="S0169"></a> Source: p.9 S0169

Manning. A large annotated corpus for learning natural language inference.

<a id="S0170"></a> Source: p.9 S0170

Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055, 2017. [7] S.

<a id="S0171"></a> Source: p.9 S0171

Story comprehension for predicting what happens next.

<a id="S0172"></a> Source: p.9 S0172

In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1603–1614, 2017. [8] D.

<a id="S0173"></a> Source: p.9 S0173

Manning. A fast and accurate dependency parser using neural networks.

<a id="S0174"></a> Source: p.9 S0174

In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 740–750, 2014. [9] Z.

<a id="S0175"></a> Source: p.9 S0175

Quora question pairs. https://data.quora.com/First-Quora- Dataset-Release-Question-Pairs, 2018. [10] R.

<a id="S0176"></a> Source: p.9 S0176

Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning.

<a id="S0177"></a> Source: p.9 S0177

In Proceedings of the 25th international conference on Machine learning, pages 160–167.

<a id="S0178"></a> Source: p.9 S0178

Natural language processing (almost) from scratch.

<a id="S0179"></a> Source: p.9 S0179

Journal of Machine Learning Research, 12(Aug):2493–2537, 2011. [12] A.

<a id="S0180"></a> Source: p.9 S0180

Supervised learning of universal sentence representations from natural language inference data.

<a id="S0181"></a> Source: p.9 S0181

In Advances in Neural Information Processing Systems, pages 3079–3087, 2015. [14] W. B.

<a id="S0182"></a> Source: p.9 S0182

Automatically constructing a corpus of sentential paraphrases.

<a id="S0183"></a> Source: p.9 S0183

In Proceedings of the Third International Workshop on Paraphrasing (IWP2005), 2005. [15] D.

<a id="S0184"></a> Source: p.9 S0184

Why does unsupervised pre-training help deep learning?

<a id="S0185"></a> Source: p.9 S0185

Journal of Machine Learning Research, 11(Feb):625–660, 2010. [16] S.

<a id="S0186"></a> Source: p.9 S0186

Gpu kernels for block-sparse weights. 2017. [17] Z.

<a id="S0187"></a> Source: p.9 S0187

Learning entity representation for entity disambiguation.

<a id="S0188"></a> Source: p.9 S0188

In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), volume 2, pages 30–34, 2013. [18] D.

<a id="S0189"></a> Source: p.9 S0189

Bridging nonlinearities and stochastic regularizers with gaussian error linear units. arXiv preprint arXiv:1606.08415, 2016. [19] K. M.

<a id="S0190"></a> Source: p.9 S0190

Teaching machines to read and comprehend.

<a id="S0191"></a> Source: p.9 S0191

In Advances in Neural Information Processing Systems, pages 1693– 1701, 2015. [20] G. E.

<a id="S0192"></a> Source: p.9 S0192

Teh. A fast learning algorithm for deep belief nets.

<a id="S0193"></a> Source: p.9 S0193

Neural computation, 18(7):1527–1554, 2006. [21] J.

<a id="S0194"></a> Source: p.9 S0194

Universal language model fine-tuning for text classification.

<a id="S0195"></a> Source: p.9 S0195

Association for Computational Linguistics (ACL), 2018. [22] Y.

<a id="S0196"></a> Source: p.9 S0196

Discourse-based objectives for fast unsupervised sentence representation learning. arXiv preprint arXiv:1705.00557, 2017. [23] Y.

<a id="S0197"></a> Source: p.9 S0197

Discriminative improvements to distributional sentence similarity.

<a id="S0198"></a> Source: p.9 S0198

In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 891–896, 2013. 9

Page 10

<a id="S0199"></a> Source: p.10 S0199

Semi-supervised conditional random fields for improved sequence segmentation and labeling.

<a id="S0200"></a> Source: p.10 S0200

In Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pages 209–216.

<a id="S0201"></a> Source: p.10 S0201

Association for Computational Linguistics, 2006. [25] T.

<a id="S0202"></a> Source: p.10 S0202

Scitail: A textual entailment dataset from science question answering.

<a id="S0203"></a> Source: p.10 S0203

Convolutional neural networks for sentence classification.

<a id="S0204"></a> Source: p.10 S0204

Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. [28] R.

<a id="S0205"></a> Source: p.10 S0205

In Advances in neural information processing systems, pages 3294–3302, 2015. [29] N.

<a id="S0206"></a> Source: p.10 S0206

Constituency parsing with a self-attentive encoder.

<a id="S0207"></a> Source: p.10 S0207

Race: Large-scale reading comprehension dataset from examinations.

<a id="S0208"></a> Source: p.10 S0208

Unsupervised machine translation using monolingual corpora only.

<a id="S0209"></a> Source: p.10 S0209

Distributed representations of sentences and documents.

<a id="S0210"></a> Source: p.10 S0210

In International Conference on Machine Learning, pages 1188–1196, 2014. [33] P.

<a id="S0211"></a> Source: p.10 S0211

Semi-supervised learning for natural language.

<a id="S0212"></a> Source: p.10 S0212

PhD thesis, Massachusetts Institute of Technology, 2005. [34] P. J.

<a id="S0213"></a> Source: p.10 S0213

Generating wikipedia by summarizing long sequences.

<a id="S0214"></a> Source: p.10 S0214

Stochastic answer networks for natural language inference. arXiv preprint arXiv:1804.07888, 2018. [36] L.

<a id="S0215"></a> Source: p.10 S0215

An efficient framework for learning sentence representations.

<a id="S0216"></a> Source: p.10 S0216

Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2017. [38] B.

<a id="S0217"></a> Source: p.10 S0217

Learned in translation: Contextualized word vectors.

<a id="S0218"></a> Source: p.10 S0218

In Advances in Neural Information Processing Systems, pages 6297–6308, 2017. [39] T.

<a id="S0219"></a> Source: p.10 S0219

Distributed representations of words and phrases and their compositionality.

<a id="S0220"></a> Source: p.10 S0220

In Advances in neural information processing systems, pages 3111–3119, 2013. [40] N.

<a id="S0221"></a> Source: p.10 S0221

Lsdsem 2017 shared task: The story cloze test.

<a id="S0222"></a> Source: p.10 S0222

In Proceedings of the 2nd Workshop on Linking Models of Lexical, Sentential and Discourse-level Semantics, pages 46–51, 2017. [41] K.

<a id="S0223"></a> Source: p.10 S0223

Semi-supervised text classification using em.

<a id="S0224"></a> Source: p.10 S0224

Semi-Supervised Learning, pages 33–56, 2006. [42] J.

<a id="S0225"></a> Source: p.10 S0225

Glove: Global vectors for word representation.

<a id="S0226"></a> Source: p.10 S0226

In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543, 2014. [43] M. E.

<a id="S0227"></a> Source: p.10 S0227

Semi-supervised sequence tagging with bidirectional language models.

<a id="S0228"></a> Source: p.10 S0228

Deep contextualized word representations.

<a id="S0229"></a> Source: p.10 S0229

When and why are pre-trained word embeddings useful for neural machine translation?

Page 11

<a id="S0230"></a> Source: p.11 S0230

Resolving complex cases of definite pronouns: the winograd schema challenge.

<a id="S0231"></a> Source: p.11 S0231

In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 777–789.

<a id="S0232"></a> Source: p.11 S0232

Association for Computational Linguistics, 2012. [47] P.

<a id="S0233"></a> Source: p.11 S0233

Squad: 100,000+ questions for machine comprehension of text.

<a id="S0234"></a> Source: p.11 S0234

Unsupervised pretraining for sequence to sequence learning. arXiv preprint arXiv:1611.02683, 2016. [49] M.

<a id="S0235"></a> Source: p.11 S0235

Efficient learning of sparse representations with an energy-based model.

<a id="S0236"></a> Source: p.11 S0236

In Advances in neural information processing systems, pages 1137–1144, 2007. [50] M.

<a id="S0237"></a> Source: p.11 S0237

Semi-supervised multitask learning for sequence labeling.

<a id="S0238"></a> Source: p.11 S0238

Monro. A stochastic approximation method.

<a id="S0239"></a> Source: p.11 S0239

The annals of mathematical statistics, pages 400–407, 1951. [52] T.

<a id="S0240"></a> Source: p.11 S0240

Reasoning about entailment with neural attention. arXiv preprint arXiv:1509.06664, 2015. [53] R.

<a id="S0241"></a> Source: p.11 S0241

Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. [54] R.

<a id="S0242"></a> Source: p.11 S0242

Recursive deep models for semantic compositionality over a sentiment treebank.

<a id="S0243"></a> Source: p.11 S0243

In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631–1642, 2013. [55] S.

<a id="S0244"></a> Source: p.11 S0244

Riedl. A simple and effective approach to the story cloze test. arXiv preprint arXiv:1803.05547, 2018. [56] S.

<a id="S0245"></a> Source: p.11 S0245

Learning general purpose distributed sentence representations via large scale multi-task learning. arXiv preprint arXiv:1804.00079, 2018. [57] J.

<a id="S0246"></a> Source: p.11 S0246

Semi-supervised sequential labeling and segmentation using giga-word scale unlabeled data.

<a id="S0247"></a> Source: p.11 S0247

Proceedings of ACL-08: HLT, pages 665–673, 2008. [58] Y.

<a id="S0248"></a> Source: p.11 S0248

Hui. A compare-propagate architecture with alignment factorization for natural language inference. arXiv preprint arXiv:1801.00102, 2017. [59] Y.

<a id="S0249"></a> Source: p.11 S0249

Multi-range reasoning for machine comprehension. arXiv preprint arXiv:1803.09074, 2018. [60] J.

<a id="S0250"></a> Source: p.11 S0250

Ecnu at semeval-2017 task 1: Leverage kernel-based traditional nlp features and neural networks to build a universal model for multilingual and cross-lingual semantic textual similarity.

<a id="S0251"></a> Source: p.11 S0251

In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 191–197, 2017. [61] Y.

<a id="S0252"></a> Source: p.11 S0252

Opportunities and challenges in working with low-resource languages.

<a id="S0253"></a> Source: p.11 S0253

In Advances in Neural Information Processing Systems, pages 6000–6010, 2017. [63] P.

<a id="S0254"></a> Source: p.11 S0254

Extracting and composing robust features with denoising autoencoders.

<a id="S0255"></a> Source: p.11 S0255

In Proceedings of the 25th international conference on Machine learning, pages 1096–1103.

<a id="S0256"></a> Source: p.11 S0256

Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. [65] A.

<a id="S0257"></a> Source: p.11 S0257

Corpus of linguistic acceptability. http://nyu-mll.github.io/cola, 2018. [66] A.

<a id="S0258"></a> Source: p.11 S0258

Bowman. A broad-coverage challenge corpus for sentence understanding through inference.

<a id="S0259"></a> Source: p.11 S0259

Towards human-level machine reading comprehension: Reasoning and inference with multiple strategies. arXiv preprint arXiv:1711.04964, 2017. 11

Page 12

<a id="S0260"></a> Source: p.12 S0260

Roles of pre-training and fine-tuning in context-dependent dbn-hmms for real-world speech recognition.

<a id="S0261"></a> Source: p.12 S0261

NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2010. [69] R.

<a id="S0262"></a> Source: p.12 S0262

Split-brain autoencoders: Unsupervised learning by cross-channel prediction.

<a id="S0263"></a> Source: p.12 S0263

In CVPR, volume 1, page 6, 2017. [70] X.

<a id="S0264"></a> Source: p.12 S0264

Semi-supervised learning literature survey. 2005. [71] Y.

<a id="S0265"></a> Source: p.12 S0265

Aligning books and movies: Towards story-like visual explanations by watching movies and reading books.

<a id="S0266"></a> Source: p.12 S0266

In Proceedings of the IEEE international conference on computer vision, pages 19–27, 2015. 12