The dominant approaches for controlling language models are based on
fine-tuning large language models or prompt engineering. However, these methods
often require condition-specific data or considerable hand-crafting. We propose
a new simple guided decoding method, Gamma Sampling, which does not require
complex engineering and any extra data. Gamma Sampling introduces
attribute-related information (provided by humans or language models
themselves) into the sampling process to guide language models to generate
texts with desired attributes. Experiments on controlling topics and sentiments
of generated text show Gamma Sampling to be superior in diversity, attribute
relevance and overall quality of generated samples while maintaining a fast
generation speed. In addition, we successfully applied Gamma Sampling to
control other attributes of language such as relatedness and repetition, which
further demonstrates the versatility and effectiveness of this method. Gamma
Sampling is now available in the python package samplings via import gamma
sampling from samplings.
SAMPLING WITH ATTRIBUTE-RELATED INFORMATION FOR CONTROLLING LANGUAGE MODELS
言語モデル制御のための属性関連情報付きサンプリング
0.47
Shangda Wu†, Maosong Sun†‡ † Department of Music AI and Information Technology, Central Conservatory of Music ‡ Department of Computer Science and Technology, Tsinghua University shangda@mail.ccom.ed u.cn, sms@tsinghua.edu.cn
Gamma Sampling introduces attributerelated information (provided by humans or language models themselves) into the sampling process to guide language models to generate texts with desired attributes.
Experiments on controlling topics and sentiments of generated text show Gamma Sampling to be superior in diversity, attribute relevance and overall quality of generated samples while maintaining a fast generation speed.
In addition, we successfully applied Gamma Sampling to control other attributes of language such as relatedness and repetition, which further demonstrates the versatility and effectiveness of this method.
1 INTRODUCTION Benefiting from large scale text data crawled from the web, the state-of-the-art large language models (LMs) achieve great success in the task of language generation.
However, although existing models can generate high-quality texts, we have little control over the attributes of the generated outputs (e g topic and sentiment).
How to steer unconditional LMs, i.e. controllable text generation, becomes a topic of real-world significance.
非条件lms、すなわち制御可能なテキスト生成をどのように操るかは、現実世界の重要事項となる。
0.61
Despite significant advances in controllable text generation (Weng 2021), it remains an open question what the ideal method for controlling attributes of the generated language is.
No matter whether training a conditional language model from scratch (Keskar et al 2019) or fine-tuning an LM (Ziegler et al 2019, Xu et al 2021), the need for condition-specific data makes these approaches cannot easily apply to unconditional language models when there is no data for training.
条件言語モデルをゼロからトレーニングする(Keskar et al 2019)か、LMを微調整する(Ziegler et al 2019, Xu et al 2021)かに関わらず、条件固有データの必要性により、トレーニング用データが存在しない場合には、これらのアプローチが非条件言語モデルに容易に適用できない。
0.78
On the other hand, although some prompting (Shin et al 2020, Zou et al 2021) and guided decoding (Ghazvininejad et al 2017, Pascual et al 2020, Lu et al 2021) are data-free methods, they have very limited steerability or controllable attributes and often require considerable hand-crafting.
一方、いくつかのプロンプト (Shin et al 2020, Zou et al 2021) やガイド付きデコード (Ghazvininejad et al 2017 Pascual et al 2020, Lu et al 2021) はデータフリーの手法であるが、ステアビリティや制御可能な属性が非常に限られており、しばしば手作りを必要とする。
0.80
To overcome these issues we propose Gamma Sampling1, which does not require complex engineering and any training data.
This method is inspired by gamma correction (Applebaum 1952, Kubinger et al 1998), a nonlinear operation used to encode and decode luminance or tristimulus values in video or still image systems by applying to every pixel value.
この方法はガンマ補正(Applebaum 1952, Kubinger et al 1998)にインスパイアされ、すべてのピクセル値を適用してビデオや静止画像の輝度や三刺激値を符号化・復号する非線形演算である。
0.78
The basic assumption of our approach is that some attributes of the generated text are closely related to the number of occurrences of some tokens.
提案手法の基本的前提は,生成したテキストの属性がトークンの発生数と密接に関連していることである。
0.66
For example, sentence length is highly correlated with some punctuation marks (e g full stops, question marks, exclamation marks), while the topic of a text is highly correlated with the topic word and words related to it (based on word similarity or word semantic distance).
To properly examine Gamma Sampling and current methods for controllable text generation, we compare generations from them based on GPT-2 (Radford et al 2019) along several axes, i.e. fluency, diversity, attribute relevance and overall quality.
GPT-2(Radford et al 2019)に基づいて,ガンマサンプリングと現行のテキスト生成方法について,複数の軸,すなわち流速,多様性,属性関連性,全体的な品質を比較検討した。
0.73
Our results show Gamma Sampling generates texts that are not only the most diverse and relevant to the topic, but are of the highest overall quality compared to baselines while maintaining a fast generation speed.
• We propose Gamma Sampling for controllable generation, which can easily introduce attribute-related information into the sampling process by selecting attribute-related tokens manually or automatically, enabling any LM to support controllable generation.
Based on Gamma Sampling, we implement several attributes that can be controlled at the same time without compromising the quality of the generated texts while increasing their diversity.
• We compared texts generated at different control strengths and compared Gamma Sampling with CTRL (Keskar et al 2019), PPLM (Dathathri et al 2020), Fine-tuned GPT2 and vanilla GPT2 (Radford et al 2019) through both automatic and human evaluations.
(<i>Gamma Smpling</i>) と CTRL (Keskar et al 2019), PPLM (Dathathri et al 2020), Fine-tuned GPT2 と vanilla GPT2 (Radford et al 2019) を比較した。 訳抜け防止モード: •異なる制御強度で生成されたテキストを比較し、Gamma SmplingとCTRLを比較した(Keskar et al 2019)。 PPLM (Dathathri et al 2020 ), Fine - tuned GPT2 そして、Vanilla GPT2 (Radford et al 2019 ) は、自動評価と人的評価の両方を通じて。
0.68
The results show that our method generally outperformed all the baselines in terms of diversity, attribute relevance and overall quality while maintaining a fast generation speed, even though we only introduced attribute-related information in the sampling phase without any further training.
2 BACKGROUND The typical controllable generation is about modelling a probabilistic model p(x|a), that is, generating text x based on an attribute a to obtain a satisfactory result.
2 BACKGROUND 典型的な制御可能な生成は確率モデル p(x|a) をモデル化することであり、すなわち、属性 a に基づいてテキスト x を生成して満足な結果を得る。
0.79
In contrast, for unconditional LMs only p(x) can be obtained directly.
対照的に、無条件の LM に対して p(x) のみが直接得られる。
0.74
However, it is still possible to make the generated text x with an attribute a by introducing an additional layer (small attribute model) or by introducing attributerelated information into the input text (prompt) or the decoding process (guided decoding).
Furthermore, conditional LMs such as CTRL (Keskar et al 2019) fall short to control what not to generate (e g detoxification and anti-degeneration), and using control code may decrease the diversity of generated texts since they usually resemble texts in training data with that code.
さらに、ctrl(keskar et al 2019)のような条件付きlmmは、生成しないもの(デトキシフィケーションやアンチデジェネレーションなど)を制御するために不足しており、制御コードを使用すると、そのコードを持つトレーニングデータのテキストに通常似ているため、生成されたテキストの多様性が低下する可能性がある。 訳抜け防止モード: さらに、CTRL(Keskar et al 2019 )のような条件付きLMも不足している。 生成しないものを制御する(例えば、解毒と抗退化) 制御コードを使って 生成されたテキストの多様性を減らし 通常はテキストに似ています
0.72
Fine-tuned Language Model Compared to conditional LMs, fine-tuned models (Ziegler et al 2019, Xu et al 2021) usually strike a good balance between training cost and generation quality.
微調整言語モデル 条件付きLMと比較すると、微調整モデル(Ziegler et al 2019, Xu et al 2021)は通常、トレーニングコストと生成品質のバランスが良い。
0.83
These models are based on existing large models with all the weights in them fine-tuned, limiting the fine-tuning to the top or additional layers only, or introducing discriminators (Krause et al 2021, Liu et al 2021a).
これらのモデルは、すべての重量が微調整された既存の大型モデルに基づいており、微調整を上層または追加層のみに制限したり、識別器を導入する(Krause et al 2021, Liu et al 2021a)。
0.73
However, fine-tuning models still require training on condition-specific data (usually a small amount).
しかし、微調整モデルは条件固有のデータ(通常は少量)のトレーニングを必要とする。
0.75
Furthermore, models such as PPLM (Dathathri et al 2020), which combines one or multiple small attribute models with a large LM, could cause computational efficiency to become unacceptable as multiple passes at every decoding step.
さらに、PPLM(Dathathri et al 2020)のような、1つまたは複数の小さな属性モデルと大きなLMを組み合わせたモデルでは、デコードステップ毎に複数のパスが許されるので、計算効率が受け入れられなくなる可能性がある。
0.63
Prompting As large LMs, e g GPT-2 (Radford et al 2019) and GPT-3 (Brown et al 2020), are trained on huge amounts of data, they are very powerful on many NLP tasks.
プロンプティング 大きなLM e g GPT-2 (Radford et al 2019) と GPT-3 (Brown et al 2020) は大量のデータに基づいて訓練されているため、多くのNLPタスクにおいて非常に強力である。
0.82
By selecting an appropriate prompt (Shin et al 2020, Zou et al 2021), we can control the model output so that an unconditional language model can be used to solve a wide range of downstream tasks.
適切なプロンプト(Shin et al 2020, Zou et al 2021)を選択することで、モデル出力を制御し、非条件言語モデルを使用して幅広い下流タスクを解くことができる。
0.77
Although prompt engineering as a data-free approach has become a recent research hotspot (Liu et al 2021b), it has very limited steerability and requires prompt design manually or automatically (Li & Liang 2021), which has a significant impact on the performance of downstream tasks.
データフリーなアプローチとしてのプロンプトエンジニアリングは、最近の研究ホットスポット(liu et al 2021b)となっているが、操作性は非常に限られており、手動または自動でプロンプトデザインを必要とする(li & liang 2021)。
0.72
2
2
0.42
英語（論文から抽出）
日本語訳
スコア
Figure 1: Common methods of controllable generation.
図1: 制御可能な生成の一般的な方法。
0.69
Modules that introduce attribute-related information are marked in red.
属性関連の情報を導入するモジュールは、赤でマークされる。
0.65
Guided Decoding Although decoding does not affect any trainable parameters, it is a critical component of text generation.
Common decoding methods include temperature sampling (Ackley et al 1985), top-k sampling (Fan et al 2018) and nucleus sampling (Holtzman et al 2020).
一般的な復号法には、温度サンプリング(ackley et al 1985)、トップkサンプリング(fan et al 2018)、核サンプリング(holtzman et al 2020)がある。
0.60
They are based on the assumption that sampled tokens with a higher probability make the generated text more fluent but prone to degeneration, and vice versa, making the generated text more novel but potentially lacking incoherence.
Penalized sampling2 (Keskar et al 2019), a variant of temperature sampling, prevents degeneration by discounting the scores of previously generated tokens.
温度サンプリングの変種であるpenalized sampling2(keskar et al 2019)は、以前に生成されたトークンのスコアを割引することで、デジェネレーションを防止する。
0.59
However, these decoding methods are mainly focused on how to balance fluency and novelty and thus cannot be applied to controllable generation in a broader sense.
In contrast, some decoding methods introduce additional information (Ghazvininejad et al 2017, Pascual et al 2020, Lu et al 2021) that can be applied to control other attributes of the generated text.
対照的に、いくつかの復号法では、生成されたテキストの他の属性を制御するために適用可能な追加情報(Ghazvininejad et al 2017 Pascual et al 2020, Lu et al 2021)を導入している。
0.67
However, these methods often require manually setting scoring functions for different tasks or only for a specific task, which makes them somewhat less flexible.
As shown in Fig 1, regardless of what approaches are, LMs need to be provided with attributerelated information, and they differ only in where this information is introduced.
The key intuition behind the first two approaches is that, after considerable training, LMs can learn attribute-related information from large amounts of data and then use it to achieve controllable generation.
However, on the other hand, the last two assume that large LMs already learned enough information from data and only need to introduce attribute-related information into the input or decoding to be applied to various downstream tasks.
3 METHODOLOGY Some attributes of the generated text are closely related to the number of occurrences of some tokens (attribute-related tokens), so adjusting the proportion of their occurrences in the text can significantly alter the relevant attributes of the text.
A naive approach is to directly multiply the probability of these attribute-related tokens by a constant greater than 0 to adjust their numbers of occurrences in the generated text.
2A more generalised version of it is implemented as temperature sampling, which is included in the
2A より一般化されたバージョンは温度サンプリングとして実装され、これに含まれる。
0.64
python package samplings.
pythonパッケージサンプリング。
0.74
3 LargeLanguage ModelSoftmaxInputwit h ConditionOutput Text𝑡with Attribute 𝑎Conditional Language ModelLargeLanguage ModelSmallAttribute ModelSoftmaxInputOut put Text𝑡with Attribute 𝑎Fine-tunedLanguage ModelGuided DecodingLargeLanguag e ModelSoftmaxInputwit h PromptOutput Text𝑡with Attribute 𝑎Language Modelwith PromptingLargeLangua ge ModelSoftmaxInputOut put Text𝑡with Attribute 𝑎Language Modelwith Guided Decoding
3 LargeLanguage ModelSoftmaxInput with Condition Output Texttwith Attribute aConditional Language ModelLargeLanguage ModelSmallAttribute ModelSoftmaxInputOut put Texttwith Attribute aFine-tunedLanguage ModelGuided DecodingLargeLanguag e ModelSoftmaxInput with PromptOutput Texttwith Attribute aLanguage Model with PromptingLargeLangua ge ModelSoftmaxInputOut put Texttwith Attribute aLanguage Model with Guided Decoding
0.45
英語（論文から抽出）
日本語訳
スコア
Specifically, due to its linear characteristic, it may cause the probability of attribute-related tokens to increase too rapidly or to decay too drastically, leading to degeneration or, equivalently, to banning these selected tokens.
Furthermore, as it does not map an interval to the same one, there is no guarantee that a token with an original probability of 1, i.e. a token that models have absolute confidence to generate, will still keep the same probability after processing.
The goal of Gamma Sampling is to raise or lower the probability p of attribute-related tokens based on the parameter γ.
ガンマサンプリングの目標は、パラメータ γ に基づいて属性関連のトークンの確率 p を上昇または低下させることである。 訳抜け防止モード: ガンマサンプリングの目標は 属性 - パラメータ γ に基づいた関連するトークンの確率 p を上昇または低下させる。
0.82
For clarity, we assume that only one token is selected.
明確にするために、1つのトークンのみが選択されると仮定する。
0.56
Obviously, the backbone function f of Gamma Sampling is monotonically increasing with two inputs, i.e. γ and p.
明らかに、ガンマサンプリングのバックボーン関数 f は、2つの入力、すなわち γ と p で単調に増加する。
0.76
The basic expression can be formalised as:
基本的な表現は次のように定式化できる。
0.56
p∗ = f (γ, p),
p∗ = f (γ, p)
0.41
(1) where p∗ is the processed probability of the selected token.
1) p∗ が選択されたトークンの処理確率である場合。
0.79
As a probability as well as a dependent variable, p∗ determines an important characteristic of the backbone function f: if we consider γ as a constant (regardless of its value), the domain and the range of f are the same (from 0 to 1).
確率として、従属変数と同様に、p∗ はバックボーン関数 f の重要な特性を決定する: γ を定数(その値によらず)として考えると、その領域と f の範囲は同じ(0 から 1 まで)。
0.80
In contrast, simple functions such as linear multiplication f (γ, p) = γp do not possess this property, which is why they cannot be properly applied to controllable generation.
対照的に、線型乗法 f (γ, p) = γp のような単純な函数はこの性質を持っていないので、制御可能な生成には正しく適用できない。
0.78
Furthermore, f is expected to keep p unchanged when it is at the endpoints of its interval (i.e. 0 and 1), that is, to ensure that the following two equations always hold:
さらに、f はその区間の終点(すなわち 0 と 1 )にあるとき p を不変に保つことが期待され、次の2つの方程式が常に成り立つことを保証する。
0.75
f (γ, 0) = 0, f (γ,
f(γ) である。 0) = 0, f (γ, )
0.53
1) = 1, (2)
1) = 1, (2)
0.43
where γ can be any number in its range.
γはその範囲内の任意の数である。
0.80
Without this characteristic, when the model has absolute confidence in generating (or not) a token, it becomes very unreasonable to change the probability of that token.
As shown in Fig 23, Gamma correction is responsible for performing nonlinear processes on the pixels of the input image based on γ ∈ [0, +∞) and thereby remodelling the saturation of the image.
In the common case of A = 1, inputs and outputs are typically in the range 0 to 1.
a = 1 の一般的な場合、入力と出力は典型的には 0 から 1 の範囲である。
0.88
It is noticeable that the function of Eq 4 satisfies exactly all the characteristics needed for the backbone function f for Gamma Sampling.
eq 4 の関数は、ガンマサンプリングのバックボーン関数 f に必要な全ての特性を満たすことは明らかである。
0.76
Unlike gamma correction, Gamma Sampling introduces attribute-related information by only applying f to attribute-related tokens, and therefore is not a global modification.
Based on the expression of the gamma correction, we formalise f as follows:
ガンマ補正の表現に基づいて、f を次のように定式化する。
0.68
f (a(γ), p) = pa(γ),
f(a(γ), p) = pa(γ)
0.39
a(γ) = tan(
a(γ) = tan()
0.46
πγ 2 ), (5)
πγ 2 ), (5)
0.38
where the activation function a is used to map the γ from [0, 1] to [0,∞). Consequently, the range of γ in Gamma Sampling is set from 0 to 1. Unlike gamma correction, since Gamma Sampling is aimed at probabilities rather than pixels, the selection of powers can be more aggressive. For example, if γ = ∞, all the p that are not equal to 1 are mapped to 0 (that is the γ0). After mapping by a, the range of γ for increasing probability p of the selected token is γ ∈ [0, 0.5), while the range for decreasing p is γ ∈ (0.5, 1], which can make the tuning of γ more intuitive.
where the activation function a is used to map the γ from [0, 1] to [0,∞). Consequently, the range of γ in Gamma Sampling is set from 0 to 1. Unlike gamma correction, since Gamma Sampling is aimed at probabilities rather than pixels, the selection of powers can be more aggressive. For example, if γ = ∞, all the p that are not equal to 1 are mapped to 0 (that is the γ0). After mapping by a, the range of γ for increasing probability p of the selected token is γ ∈ [0, 0.5), while the range for decreasing p is γ ∈ (0.5, 1], which can make the tuning of γ more intuitive. 訳抜け防止モード: ここではアクティベーション関数aが使われます γ を [ 0, 1 ] から [ 0,∞ ) に写す。 その結果、ガンマサンプリングにおけるγの範囲は0から1に設定される。 ガンマ補正とは違い ガンマサンプリングはピクセルではなく確率を目標としている。 力の選び方は より攻撃的です 例えば γ = ∞ の場合、1 に等しいでないすべての p は 0 に写像される(これは γ0 である)。 選択されたトークンの確率pを上げるためのγの範囲は γ ∈ [0, 0.5 ) p を減少させる範囲は γ ∈ (0.5, 1 ] である。 これにより γ のチューニングをより直感的にできる。
0.84
Fig 3 visualises a simple example of language generation with Gamma Sampling.
Fig 3はGamma Samplingで言語生成の簡単な例を視覚化する。
0.78
There is not only one f and a that can be used for Gamma Sampling.
ガンマサンプリングに使えるfとaは1つだけではない。
0.56
As other f expressions are not as concise as the gamma correction-based one, while they are functionally the same, we do not discuss them further.
他のf式はガンマ補正ベースの式ほど簡潔ではないので、機能的には同じであるが、それ以上は議論しない。
0.67
In contrast, different activation functions a have different features and can be applied to various scenarios.
対照的に、異なるアクティベーション関数 a は異なる特徴を持ち、様々なシナリオに適用できる。
0.83
For example, gatekeeper activation aG(γ) = − ln2 lnγ always holds f (aG(γ), γ) = 0.5, which means for all p < γ there is p∗ < 0.5 after processing, and vice versa.
After processing by f, the sum of the probability distributions will no longer be equal to 1, which violates the definition of probability distributions.
f による処理の後、確率分布の和はもはや 1 に等しいものではなくなり、確率分布の定義に違反する。
0.78
Also, for simplicity, we previously assumed that the probability of only one token would be adjusted.
また、単純さのため、1つのトークンのみの確率を調整できると仮定した。
0.71
However, most attributes need to manipulate the probabilities of multiple attribute-related tokens at the same time.
しかし、ほとんどの属性は複数の属性関連トークンの確率を同時に操作する必要がある。
0.69
Furthermore, the need to adjust multiple attributes of texts is not considered in the previous discussion.
さらに、以前の議論では、テキストの複数の属性を調整する必要性は考慮されていない。
0.62
To solve these problems, we introduce further computational processes alongside f:
これらの問題を解決するために、f:と共にさらなる計算プロセスを導入する。
0.57
A p∗ A = ptan( πγ 2 ) t = pt · p∗ p∗ A , pA pA − p∗ p∗ A i = pi + p(cid:123)(A∪F )
A p∗ A = ptan( πγ 2 ) t = pt · p∗ p∗ A , pA pA − p∗ p∗ A i = pi + p(cid:123)(A?
0.44
· (1 − pF )1−tan( πγ 2 ), t ∈ A,
· (1 − pf )1−tan( πγ 2 ) t ∈ a,
0.45
, i /∈ A ∪ F, (6)
, は、F。 (6)
0.42
where A is the set of attribute-related tokens of the current turn, F is the set of frozen tokens (whose probabilities have been changed in previous turns), pA is the sum of the probabilities of tokens in A, while pF is the sum of the probabilities of tokens in F, and p(cid:123)(A∪F ) is the sum of the probabilities of tokens that are neither in A nor F. It is important to note that there are some risks when using Gamma Sampling on its own.
A が現在のターンの属性関連トークンの集合である場合、F は凍結トークンの集合であり(前回のターンで確率が変化した)、pA は A のトークンの確率の和であり、pF は F のトークンの確率の和であり、p(cid:123)(A) は A と F のどちらにも属さないトークンの確率の和である。 訳抜け防止モード: A は属性の集合である 現在のターンの関連トークン F は凍結トークンの集合である(その確率は以前のターンで変化している)。 pA は A におけるトークンの確率の和であり、pF は F におけるトークンの確率の和である。 と p(cid:123)(A)F は A にも F にもならないトークンの確率の和である。 Gamma Smplingを単独で使用する場合には,いくつかのリスクがある点に注意が必要だ。
0.79
When γ is too small, attribute-related tokens with small probabilities would dramatically increase to values that are not acceptable (nearly equal to 1), and vice versa.
To get around this problem, we can take top-k sampling (Fan et al 2018) or nucleus sampling (Holtzman et al 2020) before Gamma Sampling to truncate the unreliable tail of the probability distribution.
この問題を回避するために、ガンマサンプリングの前にトップkサンプリング(Fan et al 2018)または核サンプリング(Holtzman et al 2020)を用いて確率分布の信頼性の低い尾を切断する。
0.76
In our experiments, we used and investigated the effect of using nucleus sampling with Gamma Sampling.
実験では,ガンマサンプリングを用いた核サンプリングの効果について検討した。
0.73
3.3 CONTROLLABLE ATTRIBUTES
3.3 制御可能なアトリビュート
0.47
To achieve controllable text generation, we design the following 6 controllable attributes.
制御可能なテキスト生成を実現するために,以下の6つの制御可能な属性を設計する。
0.59
Sentence Length This is an objective attribute of text which highly correlated with some punctuation marks (e g full stops, question marks, exclamation marks).
Repetition creasing the probabilities of tokens that have been generated recently.
最近生成されたトークンの確率を繰り返すこと。
0.58
Relatedness Generating more coherent text by increasing the probabilities of words that are related to recently generated nouns, which are labelled by NLTK (Bird & Loper 2004).
Diversity DIST-1, DIST-2 and DIST-3 scores (Li et al 2016) are used to evaluate the diversity of generated samples.
DIST-1, DIST-2およびDIST-3スコア(Li et al 2016)を用いて, 生成標本の多様性を評価する。
0.70
A higher value of DIST-N means a higher proportion of distinct 1-2-3-grams in the generated text.
DIST-Nの値が高いと、生成されたテキストの1-2-3-gramの比率が高くなる。
0.52
Attribute Relevance The experiments focused on evaluating three types of attributes: sentence length, topic and sentiment.
属性関連 実験は、文の長さ、話題、感情の3種類の属性を評価することに焦点を当てた。
0.60
The average sentence length (ASL) counts the average number of words per sentence in the generated text.
平均文長(ASL)は、生成されたテキスト中の文あたりの平均単語数をカウントする。
0.77
The latter two attributes are evaluated using two metrics from external classifiers (Lewis et al 2020, Loureiro et al 2022), including external classifier accuracy (ECA) and external classifier confidence (ECC).
後者の2つの属性は、外部分類器(Lewis et al 2020, Loureiro et al 2022)と外部分類器の精度(ECA)と外部分類器の信頼性(ECC)の2つの指標を用いて評価される。
0.76
There are some weaknesses in all of the above evaluation metrics: the PPL of generated text with degeneration can be very low and is not necessarily reflect its quality, and n-gram blocking (Paulus et al 2018, Klein et al 2017, Kulikov et al 2019) can easily achieve a very high DIST-N but the generated text may not be coherent, and repetitive generating topic words or words with a clear sentiment can easily cheat external classifiers.
生成したテキストのPPLは非常に低く、必ずしもその品質を反映するわけではないし、n-gramのブロッキング(Paulus et al 2018, Klein et al 2017, Kulikov et al 2019)は極めて高いDIST-Nを容易に達成できるが、生成されたテキストは一貫性がない可能性がある。 訳抜け防止モード: 上記の評価指標にはいくつかの弱点がある。 変性テキストのPPLは非常に低い 必ずしもその品質を反映していないし、n-gramブロック(Paulus et al 2018)もある。 Klein et al 2017 Kulikov et al 2019 )は非常に高い DIST - N を実現できます。 生成されたテキストは一貫性がないかもしれません そして、明確な感情を持つ話題の単語や単語を繰り返し生成することで、外部の分類器を簡単に騙すことができます。
0.63
Therefore, we need a more comprehensive metric to evaluate the generated text.
したがって、生成したテキストを評価するために、より包括的なメトリクスが必要である。
0.56
Overall Quality Based on TOEFL Independent Writing Rubrics5, we designed scoring criteria called T4MT (Toefl writing rubrics for Machine-generated Text) to assess the quality of the generated text more comprehensively.
総合品質はTOEFL独立書記規則5に基づき,T4MT (Toefl writing rubrics for Machine- generated Text) と呼ばれる評価基準を設計し,テキストの質をより総合的に評価した。
0.76
MTurkers need to rate each generated text on a scale of 0 to 5 (nonsense to advanced) depending on its quality and whether they fit the given topic or sentiment.
Given the characteristics of machine-generated text, we set whether the text itself is finished (stop mid-sentence) or contains factual errors does not affect the scoring, but extensive repetition, complete off-topic or obvious common sense errors will result in a low score.
Figure 4: Average Sentence Length (ASL) in various settings of γ (Gamma).
図4: γ(ガンマ)の様々な設定における平均文長(ASL)。
0.84
4.2 CONTROLLABLE SENTENCE LENGTH
4.2 制御可能な文長
0.52
To achieve controllable sentence length, we selected all sentence enders (such as full stops, question marks and exclamation marks) as attribute-related tokens.
All generated samples start with “The issue focused” and GPT-2 is asked to continue generating the following 100 tokens.
から始まり、GPT-2は以下の100個のトークンを引き続き生成するよう求められている。
0.58
To ensure that each sample has at least one sentence ender, we enabled Perfect Ending, so that γ decreases linearly to 0 (when generating the 100th token) as the number of tokens generated increases, starting at the 80th token.
Some typical generation samples of controllable sentence length can be found in Appendix B. We first investigated the steerability of Gamma Sampling, generating 100 samples (set random seeds from 0 to 99) at each different value of γ, for a total of 1000 samples.
As shown in Fig 4, there is a significant increase in ASL with increasing values of γ, from an average of 12.98 words per sentence (γ = 0.1) to an average of 95.68 words per sentence (γ = 1.0).
In contrast, the average number of words per sentence generated by GPT-2 generally averaged 38.39 (γ = 0.5).
対照的に、GPT-2で生成された文あたりの平均単語数は38.39(γ = 0.5)である。
0.65
Note that if Perfect Ending is not turned on, a sentence will never end when γ = 1.0.
Perfect Ending が有効でない場合、文は γ = 1.0 で終わらないことに注意。
0.78
To examine the effect of nucleus sampling on Gamma Sampling, we then selected three representative γ (0.1, 0.5 and 0.9) and used nucleus sampling with top-p from 0.8 to 1.0.
In one extreme case (γ = 0.1, top-p = 1.0), the PPL of the generated samples even reaches 6844.10, which means that it is almost impossible to generate such text for the original GPT-2.
Table 1: Results of generated samples in various settings of γ (Gamma) and top-p
表1:γ(ガンマ)およびtop-pの様々な設定における生成サンプルの結果
0.85
In summary, by changing the probabilities of attribute-related tokens, Gamma Sampling can significantly change the corresponding attributes of generated samples.
PPLM and Gamma Sampling do not require prompts as they do not introduce attribute-related information at the input.
PPLMとガンマサンプリングは、入力に属性関連情報を導入しないため、プロンプトを必要としない。
0.70
The former is based on GPT-2 with 345M parameters and the latter is based on the one with only 117M parameters, close to that of the original GPT (Radford et al 2018).
前者は345Mパラメータを持つGPT-2に基づいており、後者はオリジナルのGPTに近い117Mパラメータを持つものに基づいている(Radford et al 2018)。
0.76
4.3 CONTROLLABLE TOPIC AND SENTIMENT
4.3 制御可能な話題と感情
0.51
As shown in Table 2, for controllable topic and sentiment, we include the following baselines: Vanilla GPT2 (Radford et al 2019), Fine-tuned GPT26, CTRL (Keskar et al 2019) and PPLM (Dathathri et al 2020).
コントロール可能なトピックと感情について表2に示すように、Vanilla GPT2 (Radford et al 2019), Fine-tuned GPT26, CTRL (Keskar et al 2019), PPLM (Dathathri et al 2020) といったベースラインが含まれている。
0.80
PPLM has two different schemes (PPLM-BoW and PPLM-Discrim) for controllable topic and controllable sentiment respectively.
Since both of them generate r samples (default r = 10) and choose the best sample based on the log-likelihood, we refer to them together as PPLM-BCR in the overall evaluation.
両者は r サンプル (デフォルト r = 10) を生成し,ログ類似度に基づいて最適なサンプルを選択するため,総合評価では PPLM-BCR と呼ぶ。
0.75
We provide a baseline Gamma Sampling (GS) that only increases the probability of attribute-related tokens (γ = 0.1), and another one GS-M that uses Gamma Sampling to control multiple attributes at once.
Although the number of parameters for all these methods seems to be very small, it does not affect to demonstrate the effectiveness of Gamma Sampling, as GS and GS-M are both based on GPT-2 with only 117M parameters, less than a tenth of CTRL.
All models are asked to generate 100 essays for each task.
すべてのモデルは、各タスクに対して100のエッセイを生成するように求められます。
0.56
Given the prefix (“The issue focused” for topics and “The movie” for sentiments), models need to continue to write the following 100 words.
プレフィックス(トピックの “the issue focus” と感情の “the movie” )を考えると、モデルは次の100ワードを書き続ける必要がある。
0.72
In addition, we applied nucleus sampling with top-p = 0.8 for all methods.
さらに,全手法でトップp=0.8で核サンプリングを行った。
0.70
For automatic evaluation, we get 6 methods × 5 tasks × 100 generations = 3000 samples in total, and for human evaluation, we selected 30 generations for each model per task, in total 6 methods × 5 tasks × 30 generations = 900 samples.
PPLM-BCR, in particular, is second only to GS-M in terms of diversity (DIST-N).
特にPPLM-BCRは多様性(DIST-N)においてGS-Mに次いで第2位である。
0.60
However, as shown in Table 2, the number of parameters in CTRL is 10 times larger than GS-M which is based on Vanilla GPT2, and the generation time of PPLM-BCR is at least 100 times longer than in GS-M.
Overall, our evaluations demonstrate that simply using Gamma Sampling to increase the probability of attribute-related tokens can improve the attribute relevance of the generated text, but does not necessarily lead to high quality.
However, together with lowering repetition and raising relatedness can greatly improve the quality of the generated text while maintaining a high attribute relevance.
5 CONCLUSIONS In this work, we presented Gamma Sampling, which introduces attribute-related information into sampling for controlling LMs by changing the probabilities of attribute-related tokens.
For controllable topic and sentiment tasks, by steering GPT-2, Gamma Sampling significantly outperforms conditional LMs and fine-tuned LMs in terms of diversity, attribute relevance and overall quality while maintaining a fast generation speed.
The two strong baselines CTRL and PPLM that approach our method on the human evaluation metric T4MT are either 10 times larger models or take at least 100 times longer for generations.
1) We only present 6 controllable attributes of texts, whereas there are many more attributes of text that can be defined from token level (e g controllable tense).
2) This paper is only focused on text, but Gamma Sampling can also be applied to other sequence generation tasks, such as controlling the attributes of generated music or changing the behaviour of an agent in reinforcement learning.
3) Cross-modal generation can be easily achieved by extracting γ sequences from other modalities (e g video) to control the attributes of sequences generated in the target modality, or using the same γ sequence to control sequences from different modalities, allowing the attributes of multiple modalities to change synchronously.
Future research on Gamma Sampling includes two directions: how to automatically find the corresponding attribute-related tokens based on a given attribute, and how to automatically adjust γ to balance the attribute relevance and quality of the model generated texts.
REFERENCES David H. Ackley, Geoffrey E. Hinton, and Terrence J. Sejnowski.
David H. Ackley, Geoffrey E. Hinton, Terrence J. Sejnowski を参照。
0.90
A learning algorithm for boltzmann machines.
ボルツマンマシンの学習アルゴリズム
0.46
Cogn. Sci. , 9(1):147–169, 1985.
コーグ Sci , 9(1):147–169, 1985.
0.34
doi: 10.1207/s15516709cog 0901\ 7.
doi: 10.1207/s15516709cog 0901\ 7。
0.49
URL 10
URL 10
0.64
英語（論文から抽出）
日本語訳
スコア
https://doi.org/10.1 207/s15516709cog0901 _7.
https://doi.org/10.1 207/s15516709cog0901 _7。
0.14
Sidney Applebaum.
シドニー・アップルバウム。
0.50
Gamma correction in constant luminance color television systems.
定輝度カラーテレビシステムにおけるガンマ補正
0.81
Proceedings of the IRE, 40(10):1185–1195, 1952.
経過 ireのうち40(10):1185–1195、1952。
0.58
Steven Bird and Edward Loper.
スティーブン・バードとエドワード・ローパー
0.66
NLTK: the natural language toolkit.
NLTK: 自然言語ツールキット。
0.71
In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, Barcelona, Spain, July 2126, 2004 - Poster and Demonstration.
Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al Language models are few-shot learners.
tom brown, benjamin mann, nick ryder, melanie subbiah, jared d kaplan, prafulla dhariwal, arvind neelakantan, pranav shyam, girish sastry, amanda askell, et al language modelsは、わずかなショット学習モデルである。
0.71
Advances in neural information processing systems, 33:1877–1901, 2020.
ニューラル情報処理システムの進歩, 33:1877–1901, 2020。
0.75
Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu.
Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, Rosanne Liu
0.38
Plug and play language models: A simple approach to controlled text In 8th International Conference on Learning Representations, ICLR 2020, Addis generation.
Plug and Play言語モデル: コントロールされたテキストに対するシンプルなアプローチ 第8回Learning Representationsに関する国際会議、ICLR 2020、Addis生成。
0.84
Ababa, Ethiopia, April 26-30, 2020.
アベバ、エチオピア、2020年4月26-30日。
0.58
OpenReview.net, 2020.
OpenReview.net、2020年。
0.65
URL https://openreview.
URL https://openreview.c om
0.73
net/forum?
ネット/フォーラム?
0.50
id=H1edEyBKDS.
id=H1edEyBKDS
0.27
Angela Fan, Mike Lewis, and Yann N. Dauphin.
アンジェラ・ファン、マイク・ルイス、ヤン・N・ダウフィン。
0.53
Hierarchical neural story generation.
階層型ニューラルストーリー生成。
0.74
In Iryna Gurevych and Yusuke Miyao (eds.), Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pp. 889–898.
iryna gurevych and yusuke miyao (eds.)では、第56回計算言語学会年次総会(acl 2018, australia, july 15-20, 2018, volume 1: long papers, pp. 889–898)が開催された。
0.66
Association for Computational Linguistics, 2018.
計算言語学協会、2018年。
0.65
doi: 10.18653/v1/ P18-1082.
doi: 10.18653/v1/p18-1082 。
0.45
URL https://aclanthology .org/P18-1082/.
URL https://aclanthology .org/P18-1082/
0.23
Marjan Ghazvininejad, Xing Shi, Jay Priyadarshi, and Kevin Knight.
Marjan Ghazvininejad、Xing Shi、Jay Priyadarshi、Kevin Knight。
0.35
Hafez: an interactive poetry generation system.
Hafez:対話型詩生成システム。
0.69
In Mohit Bansal and Heng Ji (eds.), Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, System Demonstrations, pp. 43–48.
Mohit Bansal and Heng Ji (eds.), Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, August 4, System Demonstrations, pp. 43–48。 訳抜け防止モード: モヒト・バンサルとヘンジ (eds .) では,計算言語学会第55回大会が開催されている。 ACL 2017, Vancouver, Canada, July 30 - 8月4日 システムデモ , pp. 43-48。
0.63
Association for Computational Linguistics, 2017.
計算言語学協会、2017年。
0.66
doi: 10.18653/v1/P17-4008 .
10.18653/v1/P17-4008 。
0.34
URL https://doi.org/10.1 8653/v1/P17-4008.
URL https://doi.org/10.1 8653/v1/P17-4008
0.17
Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi.
アリ・ホルツマン、ヤン・バイス、リ・デュ、マクスウェル・フォーブス、エジン・チョイ。
0.42
The curious case of neural text degeneration.
神経テキスト変性の興味深い例。
0.59
In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020.
Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong, and Richard Socher.
Nitish Shirish Keskar、Bryan McCann、Lav R. Varshney、Caiming Xiong、Richard Socher。
0.38
CoRR, transformer language model for controllable generation.
コル 制御可能な生成のためのトランスフォーマー言語モデル
0.50
CTRL: A conditional abs/1909.05858, 2019.
ctrl: 条件付きabs/1909.05858, 2019。
0.66
URL http://arxiv.org/abs /1909.05858.
url http://arxiv.org/abs /1909.05858。
0.40
Guillaume Klein, Yoon Kim, Yuntian Deng, Jean Senellart, and Alexander M. Rush.
ギヨーム・クライン、ユンティアン・デン、ジャン・セネラート、アレクサンダー・m・ラッシュ。
0.38
Opennmt: Open-source toolkit for neural machine translation.
Opennmt: ニューラルネットワーク翻訳のためのオープンソースツールキット。
0.67
In Mohit Bansal and Heng Ji (eds.), Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, System Demonstrations, pp. 67–72.
mohit bansal and heng ji (eds.)において、第55回計算言語学会年次総会(acl 2017)の議事録、バンクーバー、カナダ、7月30日-8月4日、システムデモ、p.67-72。
0.60
Association for Computational Linguistics, 2017.
計算言語学協会、2017年。
0.66
doi: 10.18653/v1/P17-4012 .
10.18653/v1/P17-4012 。
0.35
URL https://doi.org/10.
URL https://doi.org/10。
0.29
18653/v1/P17-4012.
18653/v1/P17-4012。
0.15
Ben Krause, Akhilesh Deepak Gotmare, Bryan McCann, Nitish Shirish Keskar, Shafiq R. Joty, Richard Socher, and Nazneen Fatema Rajani.
Ben Krause, Akhilesh Deepak Gotmare, Bryan McCann, Nitish Shirish Keskar, Shafiq R. Joty, Richard Socher, Nazneen Fatema Rajani
In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Findings of the Association for Computational Linguistics: EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 16-20 November, 2021, pp. 4929–4952.
Wilfried Kubinger, Markus Vincze, and Minu Ayromlou.
Wilfried Kubinger、Markus Vincze、Minu Ayromlou。
0.29
The role of gamma correction in colour In 9th European Signal Processing Conference, EUSIPCO 1998, Island of image processing.
第9回欧州信号処理会議 (EUSIPCO 1998, Island of Image Processing) におけるガンマ補正の役割
0.70
Rhodes, Greece, 8-11 September, 1998, pp. 1–4.
ギリシャのローデス、1998年9月8-11日、p.1-4。
0.53
IEEE, 1998.
1998年、IEEE。
0.70
URL https://ieeexplore.
URL https://ieeexplore.c om
0.36
ieee.org/document/70 90032/.
ieee.org/ドキュメント/7090032/。
0.32
11
11
0.43
英語（論文から抽出）
日本語訳
スコア
Ilia Kulikov, Alexander H. Miller, Kyunghyun Cho, and Jason Weston.
Ilia Kulikov、Alexander H. Miller、Yunghyun Cho、Jason Weston。
0.39
Importance of search and evaluation strategies in neural dialogue modeling.
ニューラルダイアログモデリングにおける探索と評価戦略の重要性
0.85
In Kees van Deemter, Chenghua Lin, and Hiroya Takamura (eds.), Proceedings of the 12th International Conference on Natural Language Generation, INLG 2019, Tokyo, Japan, October 29 - November 1, 2019, pp. 76–87.
Kees van Deemter, Chenghua Lin, and Hiroya Takamura (eds.), Proceedings of the 12th International Conference on Natural Language Generation, INLG 2019, Tokyo, Japan, October 29–11月1日, pp. 76–87。 訳抜け防止モード: Kees van Deemter, Chenghua Lin, Hiroya Takamura (eds ) 第12回自然言語生成国際会議(INLG 2019)の開催報告 東京、日本、2019年10月29日～11月1日 76-87頁。
0.68
Association for Computational Linguistics, 2019.
計算言語学協会、2019年。
0.65
doi: 10.18653/v1/W19-8609 .
doi: 10.18653/v1/w19-8609 。
0.38
URL https: //aclanthology.org/W 19-8609/.
URL https: //aclanthology.org/W 19-8609/
0.23
Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer.
Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, Luke Zettlemoyer。
0.39
BART: denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension.
BART: 自然言語の生成、翻訳、理解のためのシーケンス・ツー・シーケンスの事前学習。
0.57
In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel R. Tetreault (eds.), Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pp. 7871–7880.
dan jurafsky, joyce chai, natalie schluter, joel r. tetreault (eds.) では、第58回計算言語学会年次総会(acl 2020, online, july 5-10, 2020, pp. 7871-7880)が開催された。
A diversity-promoting objective function for neural conversation models.
ニューラル会話モデルのための多様性促進目的関数
0.82
In Kevin Knight, Ani Nenkova, and Owen Rambow (eds.), NAACL HLT 2016, The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego California, USA, June 12-17, 2016, pp. 110–119.
Kevin Knight, Ani Nenkova, Owen Rambow (eds.), NAACL HLT 2016 The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego California, USA, June 12-17, 2016, pp. 110–119. 訳抜け防止モード: ケビン・ナイト、Ani Nenkova、Owen Rambow (eds .)、NAACL HLT 2016に登場。 2016年北米計算言語学会会議 : ヒューマン・ランゲージ・テクノロジー サンディエゴ・カリフォルニア - 2016年6月12日 - 17日。 110-119頁。
0.57
The Association for Computational Linguistics, 2016.
計算言語学協会、2016年。
0.62
doi: 10.18653/v1/n16-1014 .
doi: 10.18653/v1/n16-1014 。
0.38
URL https://doi.org/10.1 8653/v1/n16-1014.
URL https://doi.org/10.1 8653/v1/n16-1014
0.16
Xiang Lisa Li and Percy Liang.
xiang lisa li と percy liang。
0.61
Prefix-tuning: Optimizing continuous prompts for generation.
Prefix-tuning: 生成のための継続的プロンプトの最適化。
0.57
In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli (eds.), Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pp. 4582–4597.
第59回計算言語学会年次大会および第11回自然言語処理国際会議 ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pp. 4582–4597 訳抜け防止モード: 清清宗では、Fei Xia、Wenjie Li、Roberto Navigli(eds )。 )計算言語学会第59回大会の報告 第11回自然言語処理国際会議, ACL / IJCNLP 2021 (巻1:長編)仮想行事、2021年8月1日～6日 pp.4582-4597。
Dexperts: Decoding-time controlled text generation with experts and antiIn Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli (eds.), Proceedings of experts.
the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pp. 6691–6706.
第59回計算言語学会年次総会と第11回自然言語処理国際合同会議(acl/ijcnlp 2021, ( volume 1: long papers), virtual event, august 1-6, 2021, pp. 6691-6706)
Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing.
pretrain, prompt, and predict:自然言語処理におけるプロンプト手法の体系的な調査。
0.89
CoRR, abs/2107.13586, 2021b.
CoRR, abs/2107.13586, 2021b。
0.33
URL https://arxiv.org/ab s/2107.13586.
URL https://arxiv.org/ab s/2107.13586
0.46
Daniel Loureiro, Francesco Barbieri, Leonardo Neves, Luis Espinosa Anke, and Jos´e CamachoCollados.
Daniel Loureiro、Francesco Barbieri、Leonardo Neves、Luis Espinosa Anke、Jos ́e CamachoCollados。
0.38
Timelms: Diachronic language models from twitter.
timelms: twitterのdiachronic language model。
0.76
CoRR, abs/2202.03829, 2022.
コラー、abs/2202.03829、2022。
0.46
URL https://arxiv.org/ab s/2202.03829.
url https://arxiv.org/ab s/2202.03829。
0.41
Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, Noah A. Smith, and Yejin Choi.
Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, Noah A. Smith, Yejin Choi 訳抜け防止モード: Ximing Lu, Sean Welleck, Peter West, Liwei Jiang 加西順吾、ダニエル・ハシャビ、ロナン・ル・ブラス、リアンワイ・チン Youngjae Yu, Rowan Zellers, Noah A. Smith, Yejin Choi
0.73
Neurologic a*esque decoding: Constrained text generation with lookahead heuristics.
Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh.
テイラー・シン、ヤサマン・ラゼギ、ロバート・l・ローガン4世、エリック・ウォレス、サマー・シン。
0.52
Autoprompt: Eliciting knowledge from language models with automatically generated prompts.
Autoprompt: 自動生成プロンプトによる言語モデルからの知識の排除。
0.87
In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu (eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pp. 4222–4235.
Bonnie Webber, Trevor Cohn, Yulan He, Yang Liu (eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pp. 4222-4235 訳抜け防止モード: Bonnie Webber, Trevor Cohn, Yulan He, Yang Liu (eds )。 ) 自然言語処理における実証的手法に関する2020年会議の成果 EMNLP 2020, Online, November 16–20, 2020 pp.4222-4235。
Raise a child in large language model: Towards effective and generalizable fine-tuning.
大きな言語モデルで子供を育てる: 効果的で一般化可能な微調整を目指して。
0.63
In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pp. 9514–9528.
Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pp. 9514-9528 訳抜け防止モード: In Marie - Francine Moens, Xuanjing Huang, Lucia Specia そしてScott Wen - tau Yih (eds .) - 2021年の自然言語処理における経験的手法に関する会議の成果 EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic 7 - 11 November , 2021 , pp . 9514–9528 .
Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul F. Christiano, and Geoffrey Irving.
Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul F. Christiano, Geoffrey Irving 訳抜け防止モード: Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown アレック・ラドフォード、ダリオ・アモディ、ポール・F・クリスティアーノ、ジェフリー・アーヴィング。
0.80
Fine-tuning language models from human preferences.
人間の好みによる微調整言語モデル。
0.69
CoRR, abs/1909.08593, 2019.
corr、abs/1909.08593、2019年。
0.47
URL http://arxiv.org/abs /1909.08593.
URL http://arxiv.org/abs /1909.08593
0.46
Xu Zou, Da Yin, Qingyang Zhong, Hongxia Yang, Zhilin Yang, and Jie Tang.
Xu Zou, Da Yin, Qingyang Zhong, Hongxia Yang, Zhilin Yang, Jie Tang
0.34
Controllable generation from pre-trained language models via inverse prompting.
逆プロンプトによる事前学習言語モデルからの制御可能生成
0.77
In Feida Zhu, Beng Chin Ooi, and Chunyan Miao (eds.), KDD ’21: The 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, August 14-18, 2021, pp. 2450–2460.
Feida Zhu, Beng Chin Ooi, and Chunyan Miao (eds.), KDD ’21: The 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, August 14-18, 2021, pp. 2450–2460。 訳抜け防止モード: Feida Zhu, Beng Chin Ooi, Chunyan Miao (eds ) KDD'21 : 第27回知識発見とデータマイニングに関するACM SIGKDD会議 シンガポールバーチャルイベント 2021年8月14日-18日 pp.2450-2460。
0.75
ACM, 2021.
ACM、2021年。
0.81
doi: 10.1145/3447548.3467 418.
doi 10.1145/3447548.3467 418
0.29
URL https://doi.org/10.1 145/3447548.3467418.
URL https://doi.org/10.1 145/3447548.3467418
0.20
13
13
0.85
英語（論文から抽出）
日本語訳
スコア
APPENDIX A
appendIX A
0.24
Please read the “General Rules” below carefully to ensure that you understand the scoring criteria.
評価基準を理解するために、下記の “General Rules” を慎重に読んでください。
0.78
More details are given in the “Writing Rubrics” (in full instructions).
詳細は“Writing Rubrics”(完全なインストラクションで)に記載されている。
0.77
If your score is too far away from the average of other annotators (difference in absolute value greater than 1), it will be rejected.
given topic • Most of the essays are around 100 words in length, and most of them stop mid-sentence (essays are
トピックや ※エッセイのほとんどが長さ約100語程度で、ほとんどが中途半端である(エッセイ)
0.62
not finished). It is common and does not affect the scoring
未完成) 一般的であり、得点に影響しない
0.63
• If an essay is completely off-topic, or included something that clearly contradicts common sense,
※エッセイが完全に話題外である場合、あるいは常識と明確に矛盾するものを含む場合。
0.59
only a maximum of 1 mark can be scored
最大1点しか得点できない
0.50
• If an essay contains factual errors (e g South African anti-Apartheid leader Nelson Mandela dying in prison in the 1980s) but not common sense errors (e g the sun rises in the west), it does not affect the scoring
Writing Rubrics: 14 Score Level TASK DESCRIPTION 5 Advanced An essay at this level largely accomplishes all of the following: ✓ Effectively addresses the topic and task ✓ Is well organized and well developed, using clearly appropriate explanations, exemplifications and/or details ✓ Displays unity, progression and coherence ✓ Displays consistent facility in the use of language, demonstrating syntactic variety, appropriate word choice and idiomaticity, though it may have minor lexical or grammatical errors ✓ Does not contain any obvious errors of common sense 4 High- Intermediate An essay at this level largely accomplishes all of the following: ✓ Addresses the topic and task well, though some points may not be fully elaborated ✓ Is generally well organized and well developed, using appropriate and sufficient explanations, exemplifications and/or details ✓ Displays unity, progression and coherence, though it may contain occasional redundancy, digression, or unclear connections ✓ Displays facility in the use of language, demonstrating syntactic variety and range of vocabulary, though it will probably have occasional noticeable minor errors in structure, word form or use of idiomatic language that do not interfere with meaning ✓ May occasionally contain common sense errors but do not affect the presentation 3 Low- Intermediate An essay at this level is marked by one or more of the following: ✓ Addresses the topic and task using somewhat developed explanations, exemplifications and/or details ✓ Displays unity, progression and coherence, though connection of ideas may be occasionally obscured ✓ May demonstrate inconsistent facility in sentence formation and word choice that may result in lack of clarity and occasionally obscure meaning ✓ May display accurate but limited range of syntactic structures and vocabulary ✓ May contain common sense errors rendering the presentation dubious 2 Basic An essay at this level may reveal one or more of the following weaknesses: ✓ Limited development in response to the topic and task ✓ Inadequate organization or connection of ideas ✓ Inappropriate or insufficient exemplifications, explanations or details to support or illustrate generalizations in response to the task ✓ A noticeably inappropriate choice of words or word forms ✓ An accumulation of errors in sentence structure and/or usage ✓ May contain obvious common sense errors that make the presentation obscured 1 Below Basic An essay at this level is seriously flawed by one or more of the following weaknesses: ✓ Serious disorganization or underdevelopment ✓ Little or no detail, or irrelevant specifics, or questionable responsiveness to the task ✓ Serious and frequent errors in sentence structure or usage ✓ May contain a lot of sentence-level repetition ✓ May contain serious common sense errors rendering the presentation difficult to understand 0 Nonsense An essay at this level is a complete mess consisting of random characters and/or a great amount of word-level or phrase-level repetition.
英語表記。 14 Score Level TASK DESCRIPTION 5 Advanced An essay at this level largely accomplishes all of the following: ✓ Effectively addresses the topic and task ✓ Is well organized and well developed, using clearly appropriate explanations, exemplifications and/or details ✓ Displays unity, progression and coherence ✓ Displays consistent facility in the use of language, demonstrating syntactic variety, appropriate word choice and idiomaticity, though it may have minor lexical or grammatical errors ✓ Does not contain any obvious errors of common sense 4 High- Intermediate An essay at this level largely accomplishes all of the following: ✓ Addresses the topic and task well, though some points may not be fully elaborated ✓ Is generally well organized and well developed, using appropriate and sufficient explanations, exemplifications and/or details ✓ Displays unity, progression and coherence, though it may contain occasional redundancy, digression, or unclear connections ✓ Displays facility in the use of language, demonstrating syntactic variety and range of vocabulary, though it will probably have occasional noticeable minor errors in structure, word form or use of idiomatic language that do not interfere with meaning ✓ May occasionally contain common sense errors but do not affect the presentation 3 Low- Intermediate An essay at this level is marked by one or more of the following: ✓ Addresses the topic and task using somewhat developed explanations, exemplifications and/or details ✓ Displays unity, progression and coherence, though connection of ideas may be occasionally obscured ✓ May demonstrate inconsistent facility in sentence formation and word choice that may result in lack of clarity and occasionally obscure meaning ✓ May display accurate but limited range of syntactic structures and vocabulary ✓ May contain common sense errors rendering the presentation dubious 2 Basic An essay at this level may reveal one or more of the following weaknesses: ✓ Limited development in response to the topic and task ✓ Inadequate organization or connection of ideas ✓ Inappropriate or insufficient exemplifications, explanations or details to support or illustrate generalizations in response to the task ✓ A noticeably inappropriate choice of words or word forms ✓ An accumulation of errors in sentence structure and/or usage ✓ May contain obvious common sense errors that make the presentation obscured 1 Below Basic An essay at this level is seriously flawed by one or more of the following weaknesses: ✓ Serious disorganization or underdevelopment ✓ Little or no detail, or irrelevant specifics, or questionable responsiveness to the task ✓ Serious and frequent errors in sentence structure or usage ✓ May contain a lot of sentence-level repetition ✓ May contain serious common sense errors rendering the presentation difficult to understand 0 Nonsense An essay at this level is a complete mess consisting of random characters and/or a great amount of word-level or phrase-level repetition. 訳抜け防止モード: 英語表記。 14 スコアレベル TASK DESCRIPTION 5 Advanced このレベルでのエッセイは、以下のすべてを達成する。 明確な適切な説明、例示及び/又は詳細を使用する。 ... 統一性、進歩性、一貫性を示す。 言語の使用において一貫した施設を表示する。 構文的多様性、適切な単語選択、慣用性を示すが、小さな語彙的または文法的な誤りがあるかもしれない ^ 共通感覚の明確な誤りは含まない ^ 中間的 ^ このレベルのエッセイは、主に以下のすべてのことを達成している。 適切な説明と十分な説明を使って 例示及び/又は詳細: 時折冗長性を含むかもしれないが、統一性、進歩性、一貫性を示す。 digression (複数形 digressions) 構文的多様性と語彙の範囲を示すが、構造上の小さな誤りが時々起こるだろう。 意味に干渉しない慣用的な言語の語形または用法 しかし、プレゼンテーション3の低中間性 このレベルのエッセイには、以下の1つ以上の特徴がある。 例示と/または詳細 ^ 統一性、進歩性、一貫性を示すが、アイデアのつながりは時々曖昧になることがある。 単語の選択は、明確さの欠如と時折不明瞭な意味を生じる可能性がある。 ... 正確で限定的な構文構造と語彙を表示できる。 ... 疑わしい表現を描画する常識的誤りを含むかもしれない。 2 基本 このレベルのエッセイは、以下の弱点の1つまたは2つを明らかにすることができる。 タスクに反応して一般化を支援したり説明したりするための説明や詳細 ^ 単語や単語の形式を著しく不適切な選択 ^ 文構造や用法における誤りの蓄積 ^ プレゼンテーションを曖昧にする明らかな常識的誤り ^ 下記の基礎 このレベルのエッセイは、以下の弱点の1つまたはそれ以上によって深刻な欠陥がある。 または、無関係な特質、またはタスクに対する疑わしい応答性 ... 文構造や使用法における重大かつ頻繁なエラー ... 文がたくさん含まれているか? レベル繰り返し? ... プレゼンテーションを理解しにくくする深刻な常識的エラーを含むか 0 ナンセンス このレベルのエッセイは、ランダムな文字で構成された完全な混乱である。 そして/または大量の単語 - レベルまたはフレーズ - レベルを繰り返します。
0.42
英語（論文から抽出）
日本語訳
スコア
APPENDIX B
appendIX B
0.29
Table 4: Examples of GPT2-generated texts in various settings of γ (Gamma)
表4:γ(ガンマ)の様々な設定におけるGPT2生成テキストの例
0.82
15 GammaText0.1The issue focused on today's visualizing of America.
15 GammaText0.1 今日のアメリカの可視化に焦点を当てた問題。
0.51
At one point? Will our massive commercialisationint erest allow us to avoid such digital environments?
ある時点で? 我々の大規模な商業化はそのようなデジタル環境を避けることができるだろうか?
0.61
For much of that space? Is there any hope?
そのスペースは? 希望はあるのか?
0.66
I really thought we'd?" he told your Daily Mail. "For most other places in the world?
Don't put people. "Addressing another episode of Call Me By Your Name.0.2The issue focused on practice and implementation of criminal responsibility for Caplindale'splanned smoking ban. It came after court orders told officials that Smokers Protection Society could always appeal the ban. CITES held about 60 meetings with department officials in April. An employee said there were several meetings where seven states had ordered Smokers Protection Society to copy booklets from Smokers Protection Society. These conflict with Smokers Protection Society's views of enforcement.0.3The issue focused on student contracts and debt at Wal-Mart. As such, members have essentially zero input into whether government agencies should collect student contracts or stop collecting them. This led to big gains for Wal-Mart by participating in workers' collective bargaining (WCT). But now we're seeing evidence that Wal-Mart only receives federal revenue from student contracts and won't pick up some of it from store employees. It also produced bad results for minimum wage increases.0.4The issue focused on Muslim rape laws and forced marriages in 1999. MrKhaled told local magistrates that after five years of preparation for his trial in 2004, he began feeling uncomfortable about going to court at all. He received several letters from male guests telling him he could not enter his date's house under certain circumstances: his partner would only return when they return from work; his partner would return when they leave as well; or if they tried to enter his house with him.0.5The issue focused on economy and Social Security was raised when MacCallum proposed 2010 increases in taxes on college tuition and connected wages to reduce benefits for Americans who received college degrees to increase income among workers who received higher education. For example, President Obama proposed finding $32 billion in zero-income tax credits for individuals under age 65 with college degrees (who currently receive home ownership systems from low-income households). These taxes would continue to fall under current tax laws and leave everyone without enough money to pay for."0.6The issue focused on trustful providers for Canadian consumers and helped shape how Canada named its internet policies and regulations in recent years.Disclosure: Canadian Federation of Independent Business Inc. policy gives university perspective on internet privacy from Keening There have been roundups in various meetings over the past month with Canadian universities that offer options for cloud hosting but those solutions don't require books or accessible internet access.
人を配置するな。 "Addressing another episode of Call Me By Your Name.0.2The issue focused on practice and implementation of criminal responsibility for Caplindale'splanned smoking ban. It came after court orders told officials that Smokers Protection Society could always appeal the ban. CITES held about 60 meetings with department officials in April. An employee said there were several meetings where seven states had ordered Smokers Protection Society to copy booklets from Smokers Protection Society. These conflict with Smokers Protection Society's views of enforcement.0.3The issue focused on student contracts and debt at Wal-Mart. As such, members have essentially zero input into whether government agencies should collect student contracts or stop collecting them. This led to big gains for Wal-Mart by participating in workers' collective bargaining (WCT). But now we're seeing evidence that Wal-Mart only receives federal revenue from student contracts and won't pick up some of it from store employees. It also produced bad results for minimum wage increases.0.4The issue focused on Muslim rape laws and forced marriages in 1999. MrKhaled told local magistrates that after five years of preparation for his trial in 2004, he began feeling uncomfortable about going to court at all. He received several letters from male guests telling him he could not enter his date's house under certain circumstances: his partner would only return when they return from work; his partner would return when they leave as well; or if they tried to enter his house with him.0.5The issue focused on economy and Social Security was raised when MacCallum proposed 2010 increases in taxes on college tuition and connected wages to reduce benefits for Americans who received college degrees to increase income among workers who received higher education. For example, President Obama proposed finding $32 billion in zero-income tax credits for individuals under age 65 with college degrees (who currently receive home ownership systems from low-income households). These taxes would continue to fall under current tax laws and leave everyone without enough money to pay for."0.6The issue focused on trustful providers for Canadian consumers and helped shape how Canada named its internet policies and regulations in recent years.Disclosure: Canadian Federation of Independent Business Inc. policy gives university perspective on internet privacy from Keening There have been roundups in various meetings over the past month with Canadian universities that offer options for cloud hosting but those solutions don't require books or accessible internet access. 訳抜け防止モード: 人を配置するな。 『Call Me By Your Name.0.2』では、キャプリンデールの禁煙に関する犯罪責任の実践と実施に焦点を当てている。 CITESは4月に部局と約60回会合を開き、7カ国が喫煙者保護協会に「喫煙者保護協会」の冊子を複製するよう命令したと明らかにした。 喫煙者保護協会の実施観との対立0.3 学生契約に焦点をあてて ウォル・マートにおける債務について。そうであるように、会員は基本的に、政府機関が学生契約の収集をすべきかどうかについてのインプットをゼロにしている。 それとも、集めるのをやめるか。これは、労働者の集団交渉(WCT)に参加することで、Wal-Martにとって大きな利益をもたらした。しかし今、Wal-Martが学生契約から連邦政府の収入しか受け取らないという証拠を見ている。 店員から受け取りません。 最低賃金引き上げの悪い結果も出ました。0.4 イスラム教徒のレイプ法に焦点をあてた問題。 1999年に結婚を強制したカレド氏は、2004年の裁判の準備に5年間を費やした後、地元の判事に話した。 裁判に出るのが不愉快で 男性客から 一定条件下では 出入りできないという 手紙が数通届きました 相手が帰ってくるのは 彼らは仕事から戻って 経済と社会保障に焦点をあてた問題は、2010年に大学授業に対する税増税を提案したときに提起された。 高等教育を受けた労働者の収入を増やすために、大学学位を受けたアメリカ人の利益を下げるために、賃金をつないだ。 オバマ大統領は、65歳未満(現在は低所得世帯から住宅所有制度を受け取っている)の個人に対して、ゼロで32億ドルの所得税額控除を課すことを提案した。 これらの税は、現行の税法に該当し続け、すべての人に十分な費用を支払わなければならなくなるだろう。 カナダ独立企業連盟(Canadian Federation of Independent Business, Inc.)の政策は、Keening(ケニング)のインターネットのプライバシーに関する大学の視点を与えている。 しかし、これらのソリューションは本やインターネットアクセスは必要ありません。
0.69
Part of that includes experts able to blog about product launch events and deliver smaller updated versions of works.0.7The issue focused on trust between banks and consumers that became hot under President Donald Trump's administration.
This included policies such as limits on Wall Street bailouts and restrictions on immigration from various African countries.
これにはウォール街救済の制限やアフリカ諸国からの移民制限といった政策も含まれていた。
0.65
It also focused on climate change and stated that financial institutions should consider reducing costs as well as requiring investment planning for increased risks and duration of risks to businesses in order to reduce stock prices.
The strategy involved focusing on environmental concerns as well as assets such as servers and loan guarantees.0.8The issue focused on trust between banks and consumers for increased transparency and transparency from consumer agents and rated agencies that sell digital currency which require banks to label different consumers "qualified trust agents or rated agencies," according to NIST officials.Speculatio n over whether Bitcoin could fall through this winter now continues to haunt Bitcoin investors as policy makers fix issues through legislation that require banks to label customers "qualified trust agents or rated agencies," according to.
The strategy involved focusing on environmental concerns as well as assets such as servers and loan guarantees.0.8The issue focused on trust between banks and consumers for increased transparency and transparency from consumer agents and rated agencies that sell digital currency which require banks to label different consumers "qualified trust agents or rated agencies," according to NIST officials.Speculatio n over whether Bitcoin could fall through this winter now continues to haunt Bitcoin investors as policy makers fix issues through legislation that require banks to label customers "qualified trust agents or rated agencies," according to. 訳抜け防止モード: 環境問題とサーバ・ローン保証等の資産に重点を置き、消費者エージェントからの透明性と透明性を高めるための銀行と消費者の信頼に焦点を当てた戦略。 評価された機関は デジタル通貨を売る 銀行は異なる消費者に「適格な信託業者または格付機関」のラベルを付ける必要がある。 NISTの関係者によると、ビットコインが今冬に下落するかどうかの憶測が続いている。 Bitcoinの投資家を悩ませる 政策立案者が 法律で問題を解決し 顧客を「適格な信託業者または格付け機関」に格付けする銀行 によると
0.67
And around the world a ban on Bitcoin units like trading accounts or dealing in.0.9The issue focused on trust between banks and consumers with regard to EU controls on retail banking while animal welfare theme took center stage at an EU summit on animal welfare this week in Brussels.Under EU financial controls on retail banking the animal welfare theme had little impact on climate change global warming policy," according to recently published comments from Food say systems using river stocks have lower chances of sustained desertification than systems using sustainable streams with sufficient nutrients for mild drought and desertification.1.0T he issue focused on trust between banks and consumers with regard to EU controls on retail banking while animal welfare theme took center stage at an EU summit on animal welfare this week in Brussels (March 25-26), according to French expert Reynaud NaisongeDengeauof Merck & Co., which supports EU consumer protection measures (pdf), Google Cell or Android OS X version 7 or higher (encryption security another factor there), net neutrality taking center stage at an EU summit on animal welfare this week in Brussels.
And around the world a ban on Bitcoin units like trading accounts or dealing in.0.9The issue focused on trust between banks and consumers with regard to EU controls on retail banking while animal welfare theme took center stage at an EU summit on animal welfare this week in Brussels.Under EU financial controls on retail banking the animal welfare theme had little impact on climate change global warming policy," according to recently published comments from Food say systems using river stocks have lower chances of sustained desertification than systems using sustainable streams with sufficient nutrients for mild drought and desertification.1.0T he issue focused on trust between banks and consumers with regard to EU controls on retail banking while animal welfare theme took center stage at an EU summit on animal welfare this week in Brussels (March 25-26), according to French expert Reynaud NaisongeDengeauof Merck & Co., which supports EU consumer protection measures (pdf), Google Cell or Android OS X version 7 or higher (encryption security another factor there), net neutrality taking center stage at an EU summit on animal welfare this week in Brussels. 訳抜け防止モード: And around the world a ban on Bitcoin units like trading accounts or dealing in.0.9The issue focused on trust between banks and consumers with regard to EU controls on retail banking while animal welfare theme took center stage at an EU summit on animal welfare this week in Brussels . Under EU financial controls on retail banking the animal welfare theme had little impact on climate change global warming policy, " according to recently published comments from Food say systems using river stocks have lower chances of sustained desertification than systems using sustainable streams with sufficient nutrients for mild drought 動物福祉のテーマは、今週ブリュッセルで開かれたEUの動物福祉サミット(3月25日～26日)で中心となった。 欧州連合(EU)の消費者保護対策(pdf)を支援するフランスの専門家、レイノー・ナイソンジュ・デントーフ・メルク(Reynaud NaisongeDengeauof Merck & Co.)によると、 Google CellかAndroid OS Xバージョン7以上 (暗号化セキュリティの別の要因) ブリュッセルで今週開かれた動物福祉に関する欧州連合(EU)首脳会議では、ネット中立性(ネット中立性)が中心となる。
0.89
英語（論文から抽出）
日本語訳
スコア
APPENDIX C
appendIX C
0.30
Table 5: Examples of texts generated by various methods.
表5: 様々なメソッドによって生成されるテキストの例。
0.72
The topic here is COMPUTERS.
ここでのトピックはComputersです。
0.61
16 MethodTextVanilla GPT2The issue focused on natural selection on physical processes and physical processes are related to computer evolution: Part 2 A theory of evolution The theory of evolution claims that nature has evolved from man and man has evolved from plant and animal life.
16 methodtextvanilla gpt2 物理的プロセスと物理的プロセスに関する自然選択に焦点を当てた問題は、コンピュータの進化に関係している: part 2 a theory of evolution 進化の理論は、自然は人間から進化し、人間は植物や動物から進化してきたと主張している。 訳抜け防止モード: 16 MethodTextVanilla GPT2 物理過程と物理過程の自然選択に焦点を当てた問題はコンピュータの進化に関連している 進化論 その2 進化論は進化論を主張する 自然は人間から進化し 人間は植物と動物の生活から進化しました
0.62
These processes can be understood as selection for natural selection and predation on animals.
これらの過程は、動物に対する自然選択と捕食の選択として理解することができる。
0.66
Over time, natural selection will eventually produce human beings that have higher intelligence and higher survival rates.
時が経つにつれ、自然選択は最終的には高い知性と高い生存率を持つ人間を生み出す。
0.75
Some scientists believe that humans will eventually surpass animals and become intelligent as human beings.
Some scientists believe that humans will eventuallyFine-tuned GPT2The issue focused on both ways in which technology can make use of already existing assets in order to reach new markets.
Google said it would provide access to two Google Apps services that help businesses save money on cloud computing andCTRLThe issue focused on making sure that government agents could get access to text messages if they wanted to.
Second off, there are different levels of interception (example: voice, video, etc.).
第2に、さまざまなレベルのインターセプション(例:音声、ビデオなど)があります。
0.64
Most messages will need to be monitored by some sort of monitoring tool (example: PGP)PPLM-BoWThe issue focused on the way police handled and recorded audio recordings, according to the ACLU's report, which also found that police officers sometimes use cell-site technology to record and analyze data on the internet.TheACLU report, titled "The Rise of Cell Stealing by Mobile Data Devices," was released today at the annual meeting of the American Civil Liberties Union International.
ACLUのレポートによると、警察はインターネット上のデータを記録・分析するために細胞サイト技術を使うことがあるという。ACLUの報告書「The Rise of Cell Stealing by Mobile Data Devices」は、アメリカ市民自由連合(American Civil Liberties Union International)の年次会合で発表された。 訳抜け防止モード: ほとんどのメッセージは、ある種の監視ツール(例えば: PGP)PPLM - BoW この問題は、警察がオーディオ録音を扱い記録する方法に焦点を当てている。 ACLUの報告書によると 警察官は時々 細胞 - サイト技術を使って インターネット上のデータを 記録し分析することです TheACLU Report (英語) 「携帯電話によるセルステアリングの台頭」 アメリカ自由人権連合(American Civil Liberties Union International)の年次総会で発表された。
0.78
The report, which examines the privacy implications of mobile surveillance technologies, also looks at other privacy-related issues, including privacy of communications data inGSThe issue focused on Windows 10 hardware devices and devices running Windows 10 Home Edition.
モバイル監視技術のプライバシーへの影響を調査中のこのレポートは、プライバシーに関する他の問題にも目を向けている。 訳抜け防止モード: このレポートは、モバイル監視技術のプライバシーへの影響を調査している。 また、他のプライバシ - 通信データのプライバシを含む関連問題にも注目する。 およびWindows 10 Home Editionが動作するデバイス。
0.67
Windows 10 Home Edition devices also included Windows 10 Mobile devices.
Windows 10 Home Editionデバイスには、Windows 10 Mobileデバイスも含まれている。
0.71
Windows 10 Mobile devices include devices running Windows 10 Mobile devices that use Windows 10 Mobile devices in Windows 10 Mobile devices.
Windows 10 Mobileデバイスには、Windows 10 MobileデバイスでWindows 10 Mobileデバイスを使用するWindows 10 Mobileデバイスを実行するデバイスが含まれる。
0.76
Windows 10 Mobile devices also include devices running Windows 10 Mobile devices that use Windows 10 Mobile devices in Windows 10 Mobile devices.
Windows 10 Mobileデバイスには、Windows 10 MobileデバイスをWindows 10 Mobileデバイスで使用するWindows 10 Mobileデバイスも搭載されている。
0.80
Windows 10 Mobile devices also include devices running Windows 10 Mobile devices that use Windows 10 Mobile devices in Windows 10 Mobile devices.
Windows 10 Mobileデバイスには、Windows 10 MobileデバイスをWindows 10 Mobileデバイスで使用するWindows 10 Mobileデバイスも搭載されている。
0.80
Windows 10 Mobile devicesGS-MThe issue focused on Windows 10 hardware devices and Internet Explorer software.
Windows 10 Mobile DeviceGS-M Windows 10ハードウェアデバイスとInternet Explorerソフトウェアに焦点を当てた問題。
0.83
Microsoft engineers created the framework that monitors applications from desktop PCs to tablets, smartphones and laptops using Wi-Fi networks.
The goals are similar to those for websites or databases in browsers: servers take care of communication between computers; sensors measure computer activity and traffic volume; and algorithms automatically identify machines as belonging to one particular machine group.Even though Windows 10 technology has yet to become widely adopted by developers, hackers who believe they can steal systems will often
サーバはコンピュータ間の通信を処理し、センサーはコンピュータのアクティビティとトラフィック量を計測し、アルゴリズムは自動的に特定のマシングループに属するマシンを識別する。 訳抜け防止モード: 目標はブラウザのWebサイトやデータベースと似ている サーバーはコンピュータ間の通信を処理します センサーはコンピュータの活動量と交通量を測定する アルゴリズムは特定のマシングループに属するマシンを自動的に識別する。 Windows 10の技術はまだ開発者に広く採用されていない。 システムを盗むことができると信じているハッカーは
0.78
英語（論文から抽出）
日本語訳
スコア
Table 6: Examples of texts generated by various methods.
表6: 様々なメソッドによって生成されるテキストの例。
0.74
The topic here is LEGAL.
ここでのトピックはLEGALです。
0.71
17 MethodTextVanilla GPT2The issue focused on issues related to gender transition and gender expression in science and technology; following this led to these articles which appeared in daily newspapers and educational outlets; found support for information about gender transition and gender expression in science and technology; created article which describes research into gender transition and gender expression in science and technology; discovered recent scientific literature which supports male and female transgender individuals in science and technology; stated that science and technology should create evidence that gender transition is better than gender expression; suggested that gender transition can be indicatedFine-tuned GPT2The issue focused on by police after the deaths of three young men was how they came to end up on the streets of Birmingham.Theywere shot dead by two police officers on 16 October after they stopped and searched them for drugs in West Kirkham Street.Someofficers may have known about the drugs but did not stop them and made no arrests.Oneofficer has been suspended while an internal investigation is carried out.
17 MethodTextVanilla GPT2The issue focused on issues related to gender transition and gender expression in science and technology; following this led to these articles which appeared in daily newspapers and educational outlets; found support for information about gender transition and gender expression in science and technology; created article which describes research into gender transition and gender expression in science and technology; discovered recent scientific literature which supports male and female transgender individuals in science and technology; stated that science and technology should create evidence that gender transition is better than gender expression; suggested that gender transition can be indicatedFine-tuned GPT2The issue focused on by police after the deaths of three young men was how they came to end up on the streets of Birmingham.Theywere shot dead by two police officers on 16 October after they stopped and searched them for drugs in West Kirkham Street.Someofficers may have known about the drugs but did not stop them and made no arrests.Oneofficer has been suspended while an internal investigation is carried out. 訳抜け防止モード: 17 methodtextvanilla gpt2 : 科学技術における性転換と性表現に関する問題を中心に これに続き、日刊紙や教育機関に掲載された。 科学技術におけるジェンダー・トランジションとジェンダー表現の情報支援 科学技術におけるジェンダー移行とジェンダー表現の研究について述べる記事;科学技術における男女トランスジェンダーの個人を支援する最近の科学文献の発見 科学とテクノロジーは性転換が性表現より優れているという証拠を生み出すべきだ、と述べた。 gpt2チューニング 3人の若者が死んだ後、警察が焦点を絞った問題は、バーミンガムの街路にたどり着く方法だった。 撃たれた 10月16日、ウェストカークハム・ストリートで2人の警官が立ち止まり、麻薬を捜索した。 一部の職員は薬について知っていたかもしれないが、止めなかった 内部調査中にoneofficerは停止されました。
0.60
Police said they had received some allegations that police officers had stopped them for drugs on at least two occasions.CTRLThe issue focused around my ex telling me I should try to open up a credit card at her store when it is open at another location.
She even told me not to open up a store right nextPPLM-BoWThe issue focused on the way police handled and captured suspects who had been charged with murder in cases, according to the suit.Onedefendant, identified in court papers as Jerald J. Williams, died of a heart attack, and the other defendant, identified in court papers as David R. Williams, died of a stroke.
訴状によると、被告人はJerald J. Williams(ジェラルド・J・ウィリアムズ)氏が心臓発作で死亡し、もう一人の被告人はDavid R. Williams(デイビッド・ウィリアムズ)氏が脳卒中で死亡した。 訳抜け防止モード: 彼女は、nextpplmのすぐ隣に店を開くなとさえ言った。 容疑者を逮捕し 事件で殺人罪で起訴されました ジェラルド・j・ウィリアムズ(jerald j. williams)は、心臓発作で死亡した。 そしてもう1人の被告は、デイヴィッド・r・ウィリアムズ(david r. williams)として裁判所の書類で特定され、脳卒中で死亡した。
0.67
Prosecutors charged the two of them with first-degree murder.Prosecutorssa id that Williams had been convicted of murder after a police officer killed the man's girlfriend and then killed himselfGSThe issue focused on potential legislation that would legalize medical marijuana and medical marijuana businesses.
The legislation would legalize medical marijuana businesses and medical marijuana businesses would have to prove they would get legal medical marijuana coverage from federal law.These factors could include medical marijuana businesses filing lawsuits or filing lawsuits against medical marijuana businesses that don't provide coverage for medical marijuana business tax credits.However, medical marijuana businesses would need to prove they would meet federal rules for coverage that include medical marijuana businesses filing lawsuits or filing lawsuits against medical marijuanaGS-MThe issue focused on China's financial regulatory enforcement and management of criminal enterprises.Accordin g to court documents, prosecutors alleged that federal authorities prosecuted illegal Chinese firm Law Criminal Legal Enterprise Group (CCLGE) for engaging in illicit activities such as trading unlawful currency or goods at legal tender.
The legislation would legalize medical marijuana businesses and medical marijuana businesses would have to prove they would get legal medical marijuana coverage from federal law.These factors could include medical marijuana businesses filing lawsuits or filing lawsuits against medical marijuana businesses that don't provide coverage for medical marijuana business tax credits.However, medical marijuana businesses would need to prove they would meet federal rules for coverage that include medical marijuana businesses filing lawsuits or filing lawsuits against medical marijuanaGS-MThe issue focused on China's financial regulatory enforcement and management of criminal enterprises.Accordin g to court documents, prosecutors alleged that federal authorities prosecuted illegal Chinese firm Law Criminal Legal Enterprise Group (CCLGE) for engaging in illicit activities such as trading unlawful currency or goods at legal tender. 訳抜け防止モード: この法律は、医療大麻事業と医療大麻事業を合法化する。 連邦法から 合法的な マリファナの報道を受ける 医療大麻事業を対象としない医療大麻事業を訴えたり、医療大麻事業を対象としない医療大麻事業を提訴する場合もある。 しかし、医療マリファナビジネスは、医療マリファナビジネスが訴訟を起こしたり、医療マリファナGSに対して訴訟を起こしたりすることを含む、カバー範囲に関する連邦規則を満たすことを証明する必要がある。 刑事法人の経営・管理 裁判所文書によると、連邦当局は違法中国法人刑事法商集団(CCLGE)を起訴した。 違法な通貨や商品を合法的に取引するなどの違法な活動に従事すること
0.67
Law Criminal Legal Enterprise Group could have faced prosecution under international law if it would have engaged in economic violations against national laws.China also reportedly seized tax assets worth $18 billion from lawyers involved in filing lawsuits challenging government corruption charges
However, no doubt there are still common themes that come up with specific references to microalgae (see item 8).
しかし、マイクロ藻類への特定の参照を導き出す一般的なテーマがまだある(第8号参照)。
0.70
For example, two species of bacteria were identified as microalgae (Page 24).
例えば、2種類の細菌がマイクロ藻として同定された(Page 24)。
0.76
These bacteria were identified by comparison with non-microalgae samples (Page 22).
これらの細菌は非ミクロ藻類(Page 22)と比較された。
0.77
These bacteria were identified by comparison with microaerobes(Page 23).
これらの細菌をマイクロエローブと比較した(Page 23)。
0.79
These bacteria were identified by comparison with microalgae samples (PageFine-tuned GPT2The issue focused on public transport and transport links between south Wales and north Wales is being explored by an expert panel to improve transport links between the two regions.Anexpert panel will review public transport projects and ideas for improvement in south Wales and north Wales.Therewill also be an exercise on transport links between south Wales and south west Wales.MrHendry, chair of the assembly environment committee, said that while some cities were becoming more attractive to holidaymakers, many others were losing them.Hesaid that while people couldCTRLThe issue focused on how people view health issues on social media.
These bacteria were identified by comparison with microalgae samples (PageFine-tuned GPT2The issue focused on public transport and transport links between south Wales and north Wales is being explored by an expert panel to improve transport links between the two regions.Anexpert panel will review public transport projects and ideas for improvement in south Wales and north Wales.Therewill also be an exercise on transport links between south Wales and south west Wales.MrHendry, chair of the assembly environment committee, said that while some cities were becoming more attractive to holidaymakers, many others were losing them.Hesaid that while people couldCTRLThe issue focused on how people view health issues on social media. 訳抜け防止モード: これらの細菌は、小藻類サンプル(pagefine - tuned gpt2)と比較して同定された。 2つの領域間の輸送リンクを改善する。 anexpert panelは、サウスウェールズとノースウェールズの公共交通機関プロジェクトと改善のためのアイデアをレビューする。 また、サウス・ウェールズとサウス・ウェスト・ウェールズを結ぶ交通の訓練も行う。 議会環境委員会の議長は 一部の都市はホリデーメーカーにとって魅力的になっていた。 他にも多くの者が失っていた。 この問題は、ソーシャルメディア上で健康問題をどう見ているかに焦点が当てられている。
0.62
Can anyone explain what people mean when they say 'fat' or'skinny', or how they measure things like Body Mass Index (BMI)?
人が"太った"とか"スキニー"とか、BMI(Body Mass Index)などと言えば何を意味するのか、誰か説明して頂けますか?
0.63
People often ask what people mean when they say 'healthy', but what does it mean for someone who has never done any exercise or diet or diet pills?
Just curious what people mean when they say 'healthy'.
人は「健康」と言うときに何を意味するのかを知りたがるだけです。
0.49
Score: 11 Title: Woman receives hip replacement after accident causes hipPPLM-BoWThe issue focused on the way police handled and recorded interactions with citizens in a series of videos posted by the group "CitizenWatch," which tracks mass surveillance by police across the country, and the video released last year of an incident in the Bronx that shows two people being tased for not displaying a "reasonable amount" of blood.
Score: 11 Title: Woman received hip replacement after accident cause hipPPLM-BoWこの問題は、全国の警察による大量監視を追跡するグループ「CitizenWatch」が投稿した一連のビデオで、警察が市民とのやりとりを扱い記録することに焦点を当てている。 訳抜け防止モード: スコア:11のタイトル: ヒップPPLMの事故で股関節置換術を受けた女性 問題の焦点は、「市民監視団」が投稿した一連のビデオの中で、警察が市民とのやりとりを扱い記録することであった。 これは全国の警察による大量監視を追跡しており、昨年ブロンクス事件の映像が公開された。 血液の「妥当な量」を示さないために2人が味わわれているのがわかります。
0.67
The video was posted by the ACLU.
ビデオはACLUによって投稿された。
0.75
The New York State Police declined to provide copies of the videos to The Intercept.TheNew York City Council passed a law earlier inGSThe issue focused on climate change and climate change scientists have noted climate scientists have turned their attention to climate change as the answer to climate change.
Scientists have suggested climate scientists will consider climate change more strongly if climate scientists research climate change as the answer to climate change.
Scientists have suggested climate scientists will consider climate change more strongly if climate scientists research climate change as the answer to climate change.
Scientists have suggested climate scientists will consider climate change more strongly if climate scientists research climate change as the answer to climate change.
Scientists have suggestedGS-MThe issue focused on climate change and scientists had concerns about scientific literature that researchers used to research science.
科学者たちは、科学者が科学の研究に用いた科学文献に懸念を抱いていた。
0.68
Scientists such as physicist Carl Sagan would often argue that studies could not tell us whether there is an inherent biological difference between humans and other animals.
Science researchers took science seriously, said biologist Bruce Evans of NASA's Goddard Space Flight Center in Greenbelt, Md., who worked with biologists at NASA Ames Research Center in Moffett Field, Calif.
と、カリフォルニア州モフェットフィールドにあるnasaエイムズ研究センター(nasa ames research center)で生物学者と共に働いていたnasaゴダード宇宙飛行センターの生物学者bruce evans氏は語った。
0.77
"There are things scientists know," he said.
「科学者は知っている」と彼は言った。
0.78
Scientists should research scientifically if they
科学者は科学的に研究すべきで
0.65
英語（論文から抽出）
日本語訳
スコア
Table 8: Examples of texts generated by various methods.
表8: 様々なメソッドによって生成されるテキストの例。
0.73
The sentiment here is NEGATIVE.
ここの感情は否定的だ。
0.55
19 MethodTextVanilla GPT2The movie set up at not very far off ground (from 10 feet down to around 20 feet), they'll give you your limited options and really give you to each player you'll control.
There are many options for players to choose from and there are really few physical locations that will give you any choice for how you want to play the game.
There are lots of examples of films which do this: Her, George Clooney's competition with drama No Country for Old Men, Philomena, Good Will Hunting.
ジョージ・クルーニーのドラマ『no country for old men』、フィロメナ、グッド・ウィル・ハンティング。 訳抜け防止モード: これを行う映画の例はたくさんあります:彼女、 ジョージ・クルーニーとドラマ『no country for old men』のコンペティション。 フィロメナ グッド・ウィル・ハンティング
0.66
There are lots of examples of films which don't: Home and Mindy Kaling's Rosario Dawson –which could have been –Butch Cassidy and the Sundance Kid.
There are lots of examples of films which don't feel like one particular movie: Skyfall, Lost River,CTRLThe movie is worth buying if you like horror movies, but there are many scenes that will leave you bored with its plot, even though it is something new in its genre.
One that stands out isPPLM-DiscrimThe movie, which is in development and the movie is now on.Iwas, and still am, in a very bad mood, as well as the like, as well as the a and the theare the and the and the are.Thisis a long post, but I am, you know, a very verystrong, very, not a strong, not a very weak, but just very verystrong, and not at all.Avery, not at all,GSThe movie doesn't exist in its entirety in spite of its fairly decent gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross inGS-MThe movie doesn't exist in its entirety.
One that stands out isPPLM-DiscrimThe movie, which is in development and the movie is now on.Iwas, and still am, in a very bad mood, as well as the like, as well as the a and the theare the and the and the are.Thisis a long post, but I am, you know, a very verystrong, very, not a strong, not a very weak, but just very verystrong, and not at all.Avery, not at all,GSThe movie doesn't exist in its entirety in spite of its fairly decent gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross in gross inGS-MThe movie doesn't exist in its entirety.
0.45
But don't count on it not being dead for no reason.
しかし、理由もなく死んでいないことを考慮してはいけない。
0.70
If you want to see more material from them, then read over their reports and feel free to tell us what you think of the film.
And if you do find something missing or need help with any aspect of this material, contact one of our films library officers at your local cinema today.
They can get you nothing but content that isn't broken up into multiple pictures.
それらは、複数の写真に分割されていないコンテンツしか得られません。
0.61
We won't be sorry.
申し訳ありません。
0.35
英語（論文から抽出）
日本語訳
スコア
Table 9: Examples of texts generated by various methods.
表9: 様々なメソッドによって生成されるテキストの例。
0.72
The sentiment here is POSITIVE.
ここでの感情はPOSITIVEです。
0.71
20 MethodTextVanilla GPT2The movie should have been offered as a separate movie for others to watch in order to avoid criticism and keeping their feelings about it secret.
As one reviewer described it: "If you don't want to watch this movie in public I won't watch it."
あるレビュアーは「あなたがこの映画を公の場で見たくないなら、私はそれを見ない」と評した。
0.78
How then do you explain this secret secretbehind this movie?
ではどうやってこの秘密の映画を 説明しますか?
0.71
What if you think it's really bad?
本当に悪いと思ったらどうしますか。
0.78
What if you think it's really important?
もしそれが本当に重要だと思うなら?
0.84
What if you think it's really cool?
本当にカッコイイと思ったらどうですか?
0.63
What if you think it's really weird?
もし本当に変だと思ったら?
0.55
What if youFine-tuned GPT2The movie –just –can't get enough of Edmondson and Musgrove.
映画『Edmondson』と『Musgrove』の2作目。
0.42
How did this kid who wants to "make people laugh" turn into an accomplished actor?
人を笑わせたい」というこの子供は、どうやって優秀な俳優になったのか?
0.70
One must wonder what makes him keep coming back for more.
なぜ彼がさらに戻ってこられるのか不思議に思うだろう。
0.65
23 January. 23) Selma Selmais about the 1965 US civil-rights marches led by Martin Luther King that set off from Selma, Alabama, to the state capital of Montgomery to protest against the insidious obstruction of voter registration for black Americans.CTRLThe movie is simply excellent.
1月23日。 23) セルマ・セルメイズ(selma selmais) アラバマ州セルマを出発したマーティン・ルーサー・キング(martin luther king)率いる1965年の米国民権運動について、黒人の有権者登録を不当に妨害する抗議のため、州都モンゴメリーに抗議する。 訳抜け防止モード: 1月23日。 23) selma selmais about the 1965 us civil - rights marches led by martin luther king, launched out from selma, to the state capital of montgomery 黒人の投票者登録の不正な妨害に抗議するため、ctrlは単に素晴らしい映画だ。
0.50
If you love "Planet of the Apes" you will love this movie.
もしあなたが『Apes"Planet of the Apes"』が大好きなら、この映画は気に入るだろう。
0.76
Great characters.
素晴らしいキャラクター。
0.63
Lots of action. Rating: 4.0 While not an Oscar winner, this movie is pretty good...especially if you like scifiaction.
Its story line is somewhat predictable, but its pretty fun to watch.
ストーリーラインはある程度予測可能であるが、見るのはとても楽しい。
0.80
Rating: 4.0 Great movie, however, it could have been done without all the cussing.
評価:4.0 素晴らしい映画だが、全てを使わずに実現できた。
0.65
Great action, love the interaction between the characters, just wish they could have kept it clean.
素晴らしいアクション、キャラクター間の相互作用を愛し、きれいに保てればよかったのに。
0.69
Rat@@PPLM-DiscrimThe movie, which is in the early stage and is not in a "I want to have the sex of the person I love," says it's time timewe got we got done and are on the on the to the "I was my I had my life," the time," and a lot of times, you know my mylife!
rat@@pplm-discrimthe movieは、初期段階にあり、"i want to have the sex of the person i love", it's time time we got done, are the to the to the "i was my had my life, the time" というタイトルで、mylife! 訳抜け防止モード: Rat@@PPLM - 映画。 初期段階にあり は含まれていない。 好きな人のセックスをしたい」 時間ができた時がやってきて、"私は私の人生を過ごした"、"その時"の途中にある、と言います。 何度も 俺のマイライフを知ってるだろ!
0.69
I've been here. I can tell.
私はここにいた。 わかるよ。
0.59
I have had a life.
私は人生を過ごした。
0.81
I have been here and my family's here, I've done it.
私はここにいて、家族はここにいて、それをやった。
0.79
You are inGSThe movie just gets even sweeter as time goes on.
映画は時間が経つにつれ、さらに甘くなっていく。
0.58
No longer do we want to fight over who gets to run for president of the United States.