In this work, we investigate the problems of semantic parsing in a few-shot
learning setting. In this setting, we are provided with utterance-logical form
pairs per new predicate. The state-of-the-art neural semantic parsers achieve
less than 25% accuracy on benchmark datasets when k= 1. To tackle this problem,
we proposed to i) apply a designated meta-learning method to train the model;
ii) regularize attention scores with alignment statistics; iii) apply a
smoothing technique in pre-training. As a result, our method consistently
outperforms all the baselines in both one and two-shot settings.
To this problem, k metadesignated i) proposed we a apply to regii) the to method learning train model; ularize attention scores with alignment statistics; iii) apply a smoothing technique in pretraining.
One representations, such as logical of seapplication the wide preventing key obstacle task-specific training mantic parsing is the lack of new predicates of data.
Due capable of to new business requirement it needs to book ground transport as well.
新しいビジネス要件に対応できるため、地上輸送も予約する必要があります。
0.75
”How much does it A user could ask the assistant cost to go from Atlanta downtown to airport?”.
アトランタのダウンタウンから空港に行くのに、アシスタントの費用はいくらかかりますか?
0.54
The as corresponding LF is follows: $1 ) transport ground (exists $1 (and ( e atlanta:ci $1 atlanta:ci) )(from airport $0 fare $1 ) )))) ( =(ground fare where both ground transport and ground are are used other predicates new predicates while the airport.
As to city, from in flight booking, such as training data large parallel construction of manual consider the time-consuming, we and is expensive few-shot which reformulation of the problem, quires only a handful of utterance-LF training pairs ∗corresponding author
The cost of preparing fewfor each correspondlow, is examples the thus shot training permit significantly faster prototyping techniques development than supervised approaches ing and expansions.
Moreover, the SOTA parsers achieve less than 32% of accuracy on five widely used corpora, when the LFs in the share LF templates training in the sets do not test al., 2018).
An LF template et sets (Finegan-Dollak normalizing is derived by entities and attribute the values of an LF into typed variable names (Finegansetting imposes few-shot The al., 2018).
new LF templates, which are mixtures of known and new predicates.
新しいLFテンプレートは、既知の述語と新しい述語の混合物です。
0.80
In contrast, imtasks (e g the work by studied classification) age on the prior (Snell few-shot learning al., 2017; Finn et al., et 2017) considers an instance exclusively belonging to either a known class or a new class.
対照的に、前者(Snell few-shot learning al., 2017; Finn et al., et 2017)のイタスク(例えば、研究された分類による仕事)は、既知のクラスまたは新しいクラスにのみ属すインスタンスを検討している。
0.86
Thus, it is non-trivial learning apply to few-shot conventional LFs with mixed algorithms to generate types of predicates.
To address present we challenges, above ProtoParser, a transition-based neural semantic parser, which a sequence applies of parse actions transduce to utterance an into an LF template is parser The slots.
corresponding the and fills pretrained on a foltraining set with known predicates, support set lowed by fine-tuning on a that contains few-shot examples extends of new predicates.
It architecsequence-to- sequence attention-based the 2014) ture et al., the following (Sutskever
it architecsequence-to- sequence attention- based the 2014) ture et al., the following (sutskever)
0.77
with
と
0.60
英語（論文から抽出）
日本語訳
スコア
the is • • a metric-based
はあ? は • • metric (複数形 metrics)
0.63
few-shot specific problems
few‐shot 特定の問題
0.67
novel techniques to alleviate in the few-shot setting: Predicate-droput.
数ショット設定で緩和する斬新なテクニック:Predicate-droput。
0.67
Predicate-droput a meta• to improve representation learning technique known and new predicates.
Predicate-droput a meta• to improve representation learning technique known and new predicates。
0.89
learning both for We predicates known that found empirically represented with supervisely learned are better embeddings, while new predicates are better initialized learnby et ing algorithm (Snell In order al., 2017).
両方の述語を学ぶ 経験的に学習された述語がより良い埋め込みであるのに対して、新しい述語はより良い初期化学習とingアルゴリズム(snell in order al., 2017)である。
0.70
to embeddings work together let the two types of in a single model, we devised a training procepredicate-dropo ut called dure the simulate to scenario in pre-training.
ひとつのモデルに2つのタイプを組み込むことで,事前学習時のシナリオをシミュレーションするdure the simulationと呼ばれるトレーニング用プロセプティケートドロップアウトを考案した。
0.72
testing Attention regularization.
テスト 注意の正規化。
0.59
In this work, new twice predicates appear approximately once or to learn is Thus, during training.
この作品では、新しい2つの述語がおよそ1回現れるか、あるいは学習するために現れる。
0.62
it insufficient attention reliable the Seq2Seq scores in architecture those predicates.
これらの述語によるアーキテクチャにおけるSeq2Seqスコアの信頼性は不十分である。
0.50
the for In spirit of al., 2016), we prosupervised attention (Liu et them with regularize to scores alignment pose estimated statistics by using co-occurrence and string similarity between words and predThe prior work on supervised attention icates.
The for In spirit of al., 2016, we prosupervised attention (Liu et them with regularize to scores alignments pose estimated statistics by using co-occurrence and string similarity between words and pred The prior work on supervised attention icates。
0.82
requires it because applicable, is not either al., large parallel signifdata (Liu et 2016), Rabi(Bao effort icant manual 2018; et al., designed is novich et al., 2017), or it only for applications other than semantic parsing (Liu 2017; Kamigaito et et al., al., 2017).
適用可能なのは,al., large parallel signifdata (liu et 2016), rabi (bao effort icant manual 2018; et al., designed is novich et al., 2017) か,あるいは意味構文解析以外のアプリケーションのみである (liu 2017, kamigaito et al., al., 2017)。 訳抜け防止モード: 適用可能なのは、al ., large parallel signifdata (Liu et 2016 ) のどちらかではないためである。 Rabi (Bao effort icant manual 2018 ; et al ., designed is novich et al ., 2017) あるいは意味解析以外のアプリケーションにのみ適用できます (Liu 2017 ; Kamigaito et al )。 al . , 2017 ) .
0.83
smoothing. Pre-training of vocabulary The in fine-tuning predicates is higher than that in pre-training, which leads to a distribution disstages.
スムース。 語彙の事前学習 微調整述語における述語は事前学習よりも高いため、分布が崩壊する。
0.58
two training between the crepancy Insmoothing spired by Laplace (Manning et al., 2008), we achieve significant performance gain by applying a smoothing technique during the discrepancy.
The crepancy Insmoothing spied by Laplace (Manning et al., 2008) の2つの訓練により, 相違点に平滑化技術を適用することにより, 大幅な性能向上を実現した。
0.72
experiments on three benchmark corextensive Our outperforms show that pora the ProtoParser significant a competitive margin.
baselines effectiveness ablation study demonstrates the The technique.
ベースライン効果アブレーション研究は、そのテクニックを実証する。
0.60
of each The results individual statistically significant with p≤0.05 according are the Wilcoxon to (Wilcoxon, test 1992).
各々の統計的に重要な結果 p≤0.05 は、Wilcoxon to (Wilcoxon, test 1992) である。
0.77
2 Related Work parsing Semantic machine learning
2 セマンティック機械学習を解析する関連作業
0.74
with proposed signed-rank pre-training to alleviate
署名されたランクで 緩和するための事前訓練
0.44
is for ample of semantic
それは 意味が豊富です
0.49
work on parsing.
パースに取り組みます。
0.49
There models (Kamath and Das, 2018; Zhu surveys recent The a wide range this of work in cover al., et 2019) formalism semantic The of repmeaning area.
モデルがあります (Kamath and Das, 2018; Zhu surveys recent The a wide range this of work in cover al., et 2019) formalism semantic The of repducing area。
0.72
(Monfrom lambda calculas range resentations abstract meaning representato 1973), SQL, tague, the core of most tion (Banarescu et al , 2013).
(Monfrom lambda calculas range resentations abstract meaning representationato 1973), SQL, tague, the core of most tion (Banarescu et al , 2013)
0.73
At real., (Chen cent models et 2018; Cheng et al., 2019; et Lin al., 2019b; Yin and Neual., 2019; Zhang et big, 2018) is SEQ2SEQ with attention (Bahdanau task as al., 2014) by formulating the et a machine problem.
実際のところ(Chen cent model et 2018; Cheng et al., 2019; et Lin al., 2019b; Yin and Neual., 2019; Zhang et big, 2018)は、マシン問題を定式化することにより、注目のSEQ2SEQ(Bahdanau task as al., 2014)である。
0.95
translation and COARSE2FINE (Dong Lapata, 2018) the highest accuracy on GEOreports QUERY (Zelle and Mooney, 1996) and ATIS (Price, in et (Guo IRNET setting.
翻訳とCOARSE2FINE (Dong Lapata, 2018)はGEOreports QUERY (Zelle and Mooney, 1996)とATIS (Price, in et (Guo IRNET setting)の最高精度である。
0.82
1990) al., supervised a are al., et 2019) 2019) two and RATSQL (Wang best performing models on the Text-to-SQL benchalso deal., 2018).
1990) al., supervised a are al., et 2019) 2019) two and ratsql (wang best performing models on the text-to-sql bench also deal., 2018)。
0.89
They are (Yu et mark, SPIDER to unseen database to generalize able signed to be schemas.
それらは (Yu et mark, SPIDER) で、スキーマとして署名可能なデータベースを一般化する。
0.71
However, supervised models perform well only when there sufficient training data.
しかし、監督モデルは十分なトレーニングデータがある場合にのみうまく機能します。
0.61
is Sparsity Data datasets parsing semantic Most are small in size.
Sparsity Dataデータセット セマンティックを解析する ほとんどの場合、サイズは小さい。
0.67
To address this issue, one line of research is augment existing datasets with to data generated automatically 2017; and Yan, (Su and Liang, 2016; Cai Jia and Yates, 2013).
この問題に対処するために、既存のデータセットを2017年に自動生成したデータに拡張し、yan, (su and liang, 2016; cai jia and yates, 2013)。
0.71
Another line research is to exploit available resources, of such as bases (Krishnamurthy al., knowledge et et 2018; Chang 2019; al., and Berant, 2017; Herzig et 2019; Zhang Lee, al., 2019a; Guo al., 2019; et Wang al., 2019), features in different et semantic (Dadashkarimi et al , 2018; Li et al , 2020), domains al., 2018; Koˇcisk`y et (Yin et or unlabeled data al., 2016; Sun et orthogal., 2019).
もう1つのライン研究は、利用可能なリソース(Krishnamurthy al., knowledge et 2018; Chang 2019; al., and Berant, 2017; Herzig et 2019; Zhang Lee, al., 2019a; Guo al., 2019; et Wang al., 2019)、異なるetセマンティクスの機能(Dadashkarimi et al , 2018; Li et al , 2020)、ドメインal., 2018; Koscisk`y et (Yin et or unlabeled data al., 2016; Sun et orthogal., 2019)の活用である。
0.84
Those works are onal setting because aims our to to our approach of data labeled of handful exploit efficiently new a predicates, which are not limited to the ones in knowledge bases.
Our setting also does not require involvement of humans in the loop such as active 2018; Ni et (Duong 2019) learning and al., et al., 2015; Herzig and Becrowd-sourcing (Wang et al., rant, 2019).
私たちの設定はまた、アクティブな2018年、Ni et (Duong 2019)学習とal., et al., 2015; Herzig and Becrowd-sourcing (Wang et al., rant, 2019)などのループに人間が関与する必要はありません。
0.88
We assume availability of resources than the prior work and focus on the probdifferent caused lems develop We new predicates.
ful of training examples labeled a et The comprehengives 2019) al., (Zhu survey and algorithms the data, models, sive overview of
ful of training examples labeled a et the comprehengives 2019) al. (zhu survey and algorithms the data, models, sive overview of
0.84
英語（論文から抽出）
日本語訳
スコア
t t1 t2 t3 t4 t5 t6 t7
t1 t2 t3 t4 t5 t6 t7
0.69
Actions GEN [(ground va)] transport GEN [(to ve)] city va GEN [(from airport ve)] va GEN [(= (ground fare va) va)] REDUCE [and :- NT NT NT NT] REDUCE [exists :va NT] e NT] :REDUCE [lambda va
行動 GEN [(ground va)] transport GEN [(to ve)] city va GEN [(from airport ve)] va GEN [(= (ground fare va) va)] REDUCE [and :- NT NT NT NT] REDUCE [exists :va NT] e NT] :REDUCE [lambda va] 訳抜け防止モード: 行動 GEN [ ( Ground va ) ] transport GEN [ ( to ve ) ] city va GEN [ ( from airport ve ) ] va GEN [ (= ( Ground fare va ) ) ) ] REDUCE [ (= ( Ground fare va ) ) ] and : - NT NT NT NT ] REDUCE [ exists : va NT ] e NT ] : REDUCE [ lambda va ]
0.84
Table 1: An example action sequence.
テーブル 1:例 アクションシーケンス。
0.70
et (Snell type of problems.
など (スネル) 問題の種類。
0.60
al., It categorizes this proposed for al., 2018), into multitask learning (Hu et the models 2017; learning Vinyals embedding learning with external memory (Lee et al., 2016), and Choi, 2018; Sukhbaatar et al , 2015), and gener2017) (Reed terms in al., et ative modeling of what al., (Lee used.
アル... これはこの提案を、マルチタスク学習(Hu et the model 2017; Learning Vinyals embedding learning with external memory (Lee et al., 2016), and Choi, 2018; Sukhbaatar et al , 2015), and gener2017 (Reed term in al., et ative model of what al., (Lee et al., 2015)に分類する。
0.57
prior tackknowledge is et 2019) problem of poor les across SQL generalization the templates for SQL query generation in the one-shot learning setting.
before tackknowledge is et 2019) SQLジェネライゼーションの貧弱なlesの問題は、ワンショット学習設定でSQLクエリ生成用のテンプレートである。
0.74
In their setting, the they assume all temSQL templates on test set are the shared with assume only plates on support set.
support the in 3 Semantic Parser semanthe SOTA neural ProtoParser follows et al., Lapata, Guo 2018; parsers tic and into an LF in two steps: to map an utterance 2019) template generation and slot filling1.
in 3セマンティクスパーサであるsemanthe sota neural protoparserをサポートする。 et al., lapata, guo 2018; parsers tic と lf への2つのステップ: 発話2019)テンプレート生成とスロットフィリング1。 訳抜け防止モード: 3Semantic Parser Semanthe SOTA Neural ProtoParserは以下のとおりである。 Lapata, Guo 2018 ; Parsers tic and into a LF in two steps : to map an utterance 2019 ) template generation and slot fill1.
0.81
implements It designated system transition temgenerate a to plates, slot variables with followed by filling the address values To from utterances.
extracted the few-shot in the challenges setting, we proposed 4. training methods, detailed in Sec.
課題設定のいくつかのショットを抽出し,4つのトレーニング手法を提案した。
0.68
three differ only Many LFs in mentioned atoms, such and attribute values.
3 は、前述の原子、そのような属性値の多くの LF だけが異なる。
0.66
An LF template as entities is the in LFs with atoms created replacing by typed slot variables.
エンティティとしてのLFテンプレートは、型付きスロット変数で置き換えられた原子を持つLFです。
0.76
As the LF template of an example, our example in Sec.
例のLFテンプレートとして、Secの私たちの例です。
0.81
1 is created by substituting i) typed a “atlanta:ci”; entity the for atom variable ve shared variable ii) a name “$0“ for all variables va and “$1“.
1 は i) "atlanta:ci"; entity the for atom variable ve shared variable ii) すべての変数 va と "$1" に対して "$0" という名前をタイプすることで作られる。
0.82
(Dong (lambda va city va
(東) (lambda va city va)
0.79
(exists va )(from airport
(現存va)(空港から)
0.77
ve e (and ( ve) va
ve E (と(ve)va
0.74
ground ( =(ground
ground (複数形 grounds)
0.81
transport va fare ) va
交通機関 va 運賃 ) va
0.76
) va )))) (to
) va )))) (to)
0.85
1Code be itory: few-shot-semantic-pa rsing
1Code be itory:Small-semantic -parsing
0.60
datasets reposhttps://github. com/zhuang-li/
datasets reposhttps://github. com/zhuang-li/
0.42
found this and
見つかった これ そして
0.71
can in let x = {x1, ..., xn} Formally, an NL utterdenote ance, and its LF is represented semantic tree as a = {v1, = (V,E), where V ..., vm} the denotes y ∈ V, E ⊆ V × V set with node and is its edge vi set V = Vp ∪ Vv set.
できる で x = {x1, ..., xn} を形式的に NL utterdenote ance とし、その LF を a = {v1, = (V,E) として表現する。 訳抜け防止モード: できる で x = { x1, ..., xn } を形式的に言う。 NL utterdenote ance とその LF は a = { v1, として意味木で表される V , E ) ここで V ..., vm } は y ∈ V, ノードで設定された E × V × V は、そのエッジ vi 集合 V = Vp × Vv 集合である。
0.76
The node divided further is set Vp set slot and predicate template into value a a Vv.
さらに分割されたノードは、Vpセットスロットと述語テンプレートを値 a Vv に設定する。
0.82
predicate A template node represents a predicate symbol or a term, while a slot value node represents an atom mentioned in utterances.
Thus, an of composed tree y a tree abstract is semantic representing and a set slot value a template of τy nodes Vv,y attaching to the abstract tree.
したがって、構成されたツリー y のツリー抽象は意味的表現であり、セットスロットは抽象ツリーにアタッチする sy ノード Vv,y のテンプレートを値とする。
0.71
In the few-shot setting, we are provided with a set Ds, set Dtrain, set test a and support a train Dtest.
数ショット設定では、セットDs、セットDtrain、セットテストaとサポート列車Dtestが提供されます。 訳抜け防止モード: 少数のショット設定では、セットDsが提供されます。 set Dtrain, set test a, support a train Dtest 。
0.75
Each either of those is an example in sets utterance-LF pair (xi, yi).
これらはいずれも集合 utterance-lf pair (xi, yi) の例である。
0.74
The new predicates apin Dtrain.
新しい述語はapin dtrainです。
0.76
and Dtest in Ds not For pear only but each new are K (xi, yi) per K-shot learning, there predicate p in Ds.
Ds における Dtest は真珠のみではなく、K-shot 学習毎に K (xi, yi) となり、Ds において p を述語する。
0.75
Each new predicate appears also in the test set.
それぞれの新しい述語はテストセットにも現れる。
0.76
The goal is to maximize the accuin Dtest given estimating LFs of racy by utterances on Dtrain ∪ Ds.
目標は、Dtrainの発話による希少性のLFを推定し、Dtestを最大化することである。
0.71
a parser using trained 3.1 Transition System apply transition We the system (Cheng et al., 2019) to perform a sequence of transition actions to generate the template of a semantic tree.
トレーニングされた3.1トランジションシステムを使用したパーサ トランジションを適用する システム(Cheng et al., 2019)は、セマンティックツリーのテンプレートを生成するためのトランジションアクションのシーケンスを実行する。 訳抜け防止モード: トレーニングされた3.1トランジションシステムを使用したパーサ 移行を適用する システム( Cheng et al , 2019) セマンティックツリーのテンプレートを生成するために、遷移アクションのシーケンスを実行する。
0.77
The transition usoutputs partially-constructe d system maintains stack.
遷移出力は部分的に構成されたシステムがスタックを維持する。
0.48
starts with an empty stack.
空のスタックから始めます。
0.55
The parser ing a the of At each it performs following transtep, one parsing sition actions to update the generstate and process node.
パーサー ing a the of それぞれが次のトランテプを実行し、1つのパーシング sition アクションでジェネラステートとプロセスノードを更新します。
0.74
tree ate stack the until repeats The a contains a complete tree.
tree ate the until repeats a には完全なツリーが含まれている。
0.80
• GEN [y] creates a new leaf node y stack.
• GEN[y] は新しい葉ノード y スタックを生成します。
0.82
it on the REDUCE [r].
It on the REDUCE [r].
0.70
reduce The action identifies : −body.
reduce アクションは:-bodyを識別する。
0.75
head rule an implication The rule A new stack.
ヘッドルール 新しいスタックのルールを含意する。
0.68
from the popped first body is attaching by formed subtree the is rule head as a new parent node to the rule body .
(1) particular Template Generation utterance, an Given the task generate a is to actions of sequence a = τy.
(1) 特に テンプレート生成の発話、与えられたタスクは、シーケンス a = τy のアクションに a を生成する。
0.77
tree build an abstract to ..., ak a1, We found out LFs often contain idioms, which are frequent subtrees shared across LF templates.
tree build a abstract to ..., ak a1, 私たちはLFがしばしば、LFテンプレート間で共有される頻繁なサブツリーであるイディオムを含むことを発見しました。 訳抜け防止モード: tree build an abstract to ..., ak a1, we found LFはイディオムを含むことが多く、LFテンプレート間で頻繁に共有されるサブツリーである。
0.80
normalization template procedure apply Thus we a al., et similar manner (Iyer as a in pre2019) to collapses all LF templates.
このように正規化テンプレートプロシージャは、すべてのlfテンプレートを崩壊させるために、al., et 類似の方法(pre2019 の iyer as a)を適用する。
0.65
process It idioms into such that all LF templates are converted single units form.
プロセス すべてのLFテンプレートが単一のユニット形式に変換されるようなイディオムです。
0.81
compact into a transition The of neural enan system consists coder and decoder for a probabilestimating action |a|(cid:89) ities.
P (at|a<t, x) P (a|x) = t=1 Encoder We Shortbidirectional Long apply a 1999) term Memory (Gers (LSTM) network et al., a to map of sequence into of n words sequence contextual word representations {e}n i=1.
P (at|a<t, x) P (a|x) = t=1 Encoder We Shortbidirectional Long apply a 1999) term Memory (Gers (LSTM) network et al., a to map of sequence into of n words sequence contextual word representations {e}n i=1。
0.96
Template Decoder The decoder a applies stackseto action LSTM (Dyer generate 2015) quences.
A stack-LSTM is an unidirectional LSTM The pointer points augmented with a pointer.
stack-LSTM は一方向 LSTM ポインタで拡張されたポインタポイントです。
0.81
to a the LSTM, which repreparticular hidden state of a sents to stack.
LSTMは、スタックへの送信の、特に隠れた状態を再現する。
0.67
of a It moves the state a different different hidden state state of to indicate stack.
a の場合、スタックを示すために異なる異なる隠れ状態の状態が状態を動かします。
0.81
the At stack-LSTM produces a hidden time t, the t−1), where µt = LSTM(µt, hd by hd state hd a is t t action cat−1 the of concatenation of the embedding t− 1 and the estimated at time representation hyt−1 history by partial the of at actions generated tree t − 1. time (cid:21) (cid:20)hd common practice, hd As a is concatenated with t an attended representation ha over encoder hidden t to yield ht, with ht = W , where W is states t ha t n(cid:88) is a weight matrix created by soft attention, and ha t P (ei|hd ha (2) t )ei = t i=1 product to compute the normalized We apply dot P (ei|hd attention 2015).
the At stack-LSTM produces a hidden time t, the t−1), where µt = LSTM(µt, hd by hd state hd a is t t action cat−1 the of concatenation of the embedding t− 1 and the estimated at time representation hyt−1 history by partial the of at actions generated tree t − 1. time (cid:21) (cid:20)hd common practice, hd As a is concatenated with t an attended representation ha over encoder hidden t to yield ht, with ht = W , where W is states t ha t n(cid:88) is a weight matrix created by soft attention, and ha t P (ei|hd ha (2) t )ei = t i=1 product to compute the normalized We apply dot P (ei|hd attention 2015).
0.98
al., et (Luong scores t ) supervised attention 2017; al., (Rabinovich et The Yin and Neubig, 2018) is also applied to facilitate the learning of attention weights.
al., et (Luong scores t )supervised attention 2017; al., (Rabinovich et The Yin and Neubig, 2018)も注目重量の学習を促進するために適用されます。
0.85
Given ht, the an action is estimated by: of probability (cid:80) (cid:124) P (at|ht) = atht) exp(c (cid:124) a(cid:48)∈At exp(c a(cid:48)ht) action a, of embedding the denotes ca denotes of applicable the set actions at
確率 (cid:80) (cid:124) P (at|ht) = atht) exp(c (cid:124) a(cid:48)∈At exp(c a(cid:48)ht) action a, of the denotes ca denotes of applicable the set action at
0.82
(3) and time where At
(3)と時間 どこだ?
0.70
be initialization t. The those of embeddings will explained in the section.
な 初期化 t. 埋め込みのものは、この節で説明します。
0.59
following Slot Filling tree may semantic in node A tree a contain more due variables than one slot to template normalization.
ノード A ツリーには、1 つのスロットからテンプレートの正規化よりも多くのデュー変数が含まれています。
0.63
Since there are two types of a variables, variables, slot slot node with tree given we employ a LSTM-based decoder with the same Template architecture as the decoder to fill each output respectively.
slot type The variables, of of a a such value is decoder sequence of the same length as the number of slot variables of that type in the given tree node.
slot type このような値の変数は、指定されたツリーノード内のその型のスロット変数の数と同じ長さのデコーダシーケンスである。
0.75
Few-Shot Model Training 4 supervised from the setting The few-shot differs addiin set in testing a setting by having support support The sets.
設定から監督されるFew-Shot Model Training 4 少数ショットは、そのセットをサポートすることで設定をテストする際の追加セットが異なる。
0.70
train/test contains tion set to k while the predicate, new pairs utterance-LF per predicates.
Its predictive performance is measured approach two-steps the set.
その予測性能はセットの2ステップのアプローチ測定されます。
0.68
test on take We the because i) our experiments show that this approach performs better than training on the union the of any new support train set for and the ii) set; support is efficient sets, it computationally more time than training from scratch on the union of train set the and the support set.
a distribution discrepancy between the There is due support and the train set to new predicates, set al., et the meta-learning 2017; (Snell algorithms the simulate Finn et al., 2017) suggest to testing batch each by pre-training scenario into splitting in set a meta-support and a meta-test set.
A distribution discrepancy between the A There is due support and the train set to new predicates, set al., et the meta-learning 2017; (Snell algorithm thesimulate Finn et al., 2017) では、事前トレーニングシナリオによって各バッチをメタサポートとメタテストセットで分割することを推奨している。
0.87
The models utilize the information (e g prototype vectors) to minimize set from the meta-support acquired the meta-test this way, set.
traincannot directly apply such a However, we two following the procedure reasons.
traincanは直接適用しないが、手順上の理由に従っている。
0.61
to ing due and support the First, LF in sets is a each test predicates known mixture both and new prediof sets, and support the simulate cates.
To ing due and support the First, LF in set is a each test predicates known mixed both and new prediof sets, and support the simulate cates.
0.83
the test To meta-support and meta-test sets should include
メタサポートとメタテストセットへのテストは含めるべき
0.75
英語（論文から抽出）
日本語訳
スコア
set. and train
セット そして 列車
0.62
predicates are only training In pre-training procedure.
述語は訓練のみです 事前訓練の手順。
0.72
types of ascannot We both as well.
タイプのascannot 私達は両方同様に。
0.68
that there predicates.
そこに述語があります
0.47
of sume one type show that Second, our preliminary experiments if off it data, sufficient is there is better training action embeddings of known predicates c (Eq.
embeddings (3)) supervised way, while action a in a metric-based meta-learning by initialized algoal., 2017) perform better et rithm (Snell for rarely occurred new predicates.
埋め込み (3)) 監督された方法一方、初期化された algoal., 2017) によるメトリックベースのメタラーニングにおけるアクション a は、より良い et rithm (まれに発生する新しい述語に対するスネル) を実行する。 訳抜け防止モード: Embeddings (3 ) ) struct way, while action a in a metric - based meta - learning by initialized algoal ., 2017 ) performed better et rithm () 稀に新しい述語が発生した。
0.76
cope with Therefore, we new predicates the known between differences a initialization method customized by using in finetuning and a designated pre-training procedure to fine-tuning mimic on the the following, we it helps introduce fine-tuning first because understand our 4.1 Fine-tuning During fine-tuning, and the the model parameters action embeddings predicates in Eq known for (3) are obtained from the pre-trained model.
(Snell et al., protoThe type representations of regularization, act as a type the deep learning idea which shares as similar the techniques using pre-trained models.
(Snell et al., proto正規化の型表現は、事前訓練されたモデルを使用して同様のテクニックを共有するディープラーニングのアイデアのタイプとして機能する。 訳抜け防止モード: (Snell et al , protoThe type representations of regularization, 深層学習のアイデアとして機能し トレーニング済みのモデルを使って、同様のテクニックを共有できる。
0.79
A prototype vector of an action at constructed is template decoder by using the hidden states of the time of predicting at support on a the collected at al., et Following 2017), set.
構築されたアクションのプロトタイプベクトルはテンプレートデコーダであり、al., et following 2017)で収集されたアクションのサポートで予測する時の隠れた状態を使用して設定される。
0.79
prototype (Snell a is built by taking the mean of vector such a set of (cid:88) states ht.
プロトタイプ (snell a) は、ベクトルの平均を (cid:88) 状態 ht で表すことで構築される。
0.77
hidden 1 |M| ht∈M hidden states the at the where M denotes all set of initialization, of action at.
hidden 1 |m| hthtmlm hidden states at the at the where m signs all set of initialization, of action at。
0.75
After the time applying action emthe whole model the parameters and beddings by fine-tuning improved further the are support model on the set with a supervised training objective Lf .
Lf = Ls + λΩ (5) Ls is where is and an loss cross-entropy Ω term explained The below.
Lf = Ls + λΩ (5) Ls はそこで、損失-エントロピー Ω 項が説明される。
0.82
regularization attention adjusted by λ ∈ R+.
λ ∈ R+ で調整された正規化注意。
0.59
degree of regularization is Regularization Attention the address We P (ei|hd scores learned poorly attention of t ) actions infrequent introducing a novel attention by the that observe We regularization.
正規化の度合いは正規化である アドレス We P (ei|hd scores learned poor attention of t ) アクションは、頻繁に We 正規化を観察する人によって新しい注意を喚起される。 訳抜け防止モード: 正規化の度合いは正規化 アドレス We P ( ei|hd scores learned less attention of t ) action infrequent 規則化を観察する人たちによって、新たな注目がもたらされる。
0.74
probabilP (aj|xi) count(aj ,xi) character ity and the count(xi)
probabilP (aj|xi) count(aj ,xi) character ity and the count(xi)
0.94
cat = the (4)
猫= はあ? (4)
0.66
ht = (6) similar regularization during
ht = (6) 類似の規則化
0.80
between the predicates generated similarity and the token are often strong by action aj xi indicators The alignment.
述語の生成した類似性とトークンの間は、しばしばアクションaj xiインジケータによって強くなる。
0.66
their of indicators further strengthened by manually annocan be predicates with their corresponding the tating In tokens.
手動で注釈を付けることでさらに強化されたインジケーターは、対応するイントークンで述語化できます。
0.55
natural our work, we language adopt − similarity, as the character dist(aj, xi) 1 where is normalized Levenshtein dist(aj, xj) distance (Levenshtein, 1966).
Both measures apply [0, 1], the are thus we range in g(aj, xi) = σ(·)P (aj|xi) − σ(·)char to + (1 sim(aj, xi) compute alignment scores, where the sigmoid func(cid:124) constant tion combines two measures phd σ(w t ) corresponding normalized The score.
single into a attention scores is given by (cid:80)n P (cid:48)(xi|ak) = g(ak, xi) j=1 g(ak, xj) P (xi|ak) = (cid:80) should scores be attention The to P (cid:48)(xi|ak).
注意スコアは (cid:80)n P (cid:48)(xi|ak) = g(ak, xi) j=1 g(ak, xj) P (xi|ak) = (cid:80) should scores be attention The to P (cid:48)(xi|ak)。
0.94
Thus, we define the |P (xi|aj) − P (cid:48)(xi|aj)| term as Ω ij training.
したがって、|P (xi|aj) − P (cid:48)(xi|aj)| を Ω ij トレーニングと定義する。
0.76
Pre-training 4.2 learn two-folds: objective i) pre-training The are superin a for known predicates action embeddings can quickly adapt vised way, ii) ensure our model to embedwhose predicates, new of actions the vectors before initialized dings are prototype by fine-tuning.
事前学習 4.2 事前学習 事前学習 a 既知の述語のためのスーパーイン アクション埋め込みは、素早く視覚的に適応することができる; i) 述語を埋め込むためのモデル、初期化前のベクトルの新たなアクションは、微調整によってプロトタイプ化される。
0.67
Predicate-dropout iniStarting with alternately use one tialized model parameters, we the meta-loss Lm and optione batch for batch loss Ls.
predicate-dropout inistarting with another use one tialized model parameters, we the meta-loss lm and optione batch for batch loss ls。
0.78
mizing the supervised In a batch for Lm, we into a metathe data split set.
Mizing the supervised Lm のバッチでは、メタセデータ分割セットにします。
0.61
set support and a meta-test In order to simulate existence of new predicates, we randomly select a subset of predicates ”new”, thus their action emas conreplaced by prototype vectors beddings c are applying Eq over structed the meta-support (4) by remaining set.
set support and a meta-test 新しい述語の存在をシミュレートするために、我々は「new」という述語のサブセットをランダムに選択する。 訳抜け防止モード: 新しい述語の存在をシミュレートするために、サポートとメタ-テストを設定する。 のサブセットをランダムに選択するので、プロトタイプベクターのベディングcで置き換えられたそれらのアクションemasがeqを上書きしている。 meta (複数形 metas) - 残りセットによるサポート(4つ)。
0.60
actions of keep their The predicates embeddings learned from previous batches.
前のバッチから学んだ述語埋め込みを保持するアクション。
0.57
The resulted action embedding matrix C is combithe both.
得られた作用埋め込み行列cは両立する。
0.63
nation C = (1 − m )Cs + m embedding matrix the where Cs is and Cm is pervised way, constructed totype vectors on the meta-support vector m is generated by setting the tions ones
国 C = (1 − m ) Cs + m 埋め込み行列 Cs と Cm がパービジョニングされる方法、メタ支持ベクトル m 上に構築されたtotype ベクトルは、 tions 行列を設定することによって生成される。
0.79
(7) Cm a learned suin prousing by set.
(7) Cmはセットによる学習されたスーインプロージングです。
0.72
The mask acindices of and other the
マスクの接尾辞と他のもの
0.59
randomly for predicates ”new”
ランダムに 述語 新たなもの」
0.68
the of of
はあ? ですから ですから
0.47
to (cid:124)
へ (cid:124)
0.71
(cid:124)
(cid:124)
0.78
英語（論文から抽出）
日本語訳
スコア
training set D Algorithm 1: Predicate-Dropout set D, :Training Input trained action supervisely embedding Cs, number of meta-support of meta-test examples k, examples number example, n per support one predicate-dropout ratio r loss Lm.
訓練 セットD Algorithm 1: Predicate-Dropout set D, :Training Input training action supervisely embedded Cs, number of meta- supported of meta-test example k, example number example, n per support one predicate-dropout ratio r loss Lm。
0.79
Output :The set T Extract from the template a subset Ti from T of Sample size k a S := ∅ # meta-support set Q := ∅ # meta-test set in Ti for do t example a meta-support Sample from D without replacement t set Q(cid:48) Sample of a meta-test from D t S = S ∪ s(cid:48) Q = Q ∪ Q(cid:48) end prototype matrix Cm on S Build a set P from S Extract a predicate subset Ps r × |P| Sample a of size predicates a mask m using Ps Build With Cs, Cm and m, to compute C apply Eq Compute Lm, cross-entropy on Q with C the
Output :The set T Extract from the template a subset Ti from T of Sample size k a S := ∅ # meta-support set Q := ∅ # meta-test set in Ti for do t example a meta-support Sample from D without replacement t set Q(cid:48) Sample of a meta-test from D t S = S ∪ s(cid:48) Q = Q ∪ Q(cid:48) end prototype matrix Cm on S Build a set P from S Extract a predicate subset Ps r × |P| Sample a of size predicates a mask m using Ps Build With Cs, Cm and m, to compute C apply Eq Compute Lm, cross-entropy on Q with C the
0.97
s(cid:48) with template size n with template
s(cid:48) with template size n with template
1.00
from P as new
p から new へ
0.73
(7) to predicateas the meta-loss the model paramecross-entropy attention regularization.
(7) へ predicate as the meta-loss the model paramecross-entropy attention regularization.
0.75
zeros. We to refer operation this dropout.
ゼロだ 私たちはこのドロップアウトの操作を参照する。
0.63
The training algorithm for summarised in Algorithm 1. is In a batch for Ls, we update ters and all action embeddings with a loss Ls, together with the overall the Thus, training objective becomes Lp = Lm + Ls + λΩ (8) smoothing Pre-training new prediDue to the the cates, the during actions candidate of number larger than prediction of fine-tuning and testing is to distribuone the during pre-training.
アルゴリズム1で要約されたトレーニングアルゴリズムは、Lsのバッチで、我々は損失Lsでterとすべてのアクション埋め込みを更新します。したがって、トレーニング目的はLp = Lm + Ls + λ*(8)平滑化新しいプリディを訓練するネコのために、細かい調整とテストの予測よりも大きい数のアクション候補は、プリトレーニング中にトリビュートすることです。
0.64
That leads tion discrepancy between pre-training and testing.
これにより、事前トレーニングとテストの差が生じます。
0.56
assume differences, the To minimize a we prior actions new predon knowledge the number of for icates adding a constant to the denominator by k (3) when of Eq probability action the estimating P (at|ht) during pre-training.
違いを仮定すると、We の先行行動を最小化するために新しいプリドン知識は、前訓練中の推定 P (at|ht) を Eq 確率作用のとき k (3) で分母に定数を追加するフォーイケートの数を最小化する。
0.68
(cid:80) (cid:124) P (at|ht) = exp(c atht) (cid:124) a(cid:48)∈At a(cid:48)ht) + k exp(c consider this smoothing technique dursimplicity, significant perforshow a results experimental
(cid:80) (cid:124) p (at|ht) = exp(c atht) (cid:124) a(cid:48) gilbertat a(cid:48)ht) + k exp(c) この平滑化技法は単純性に優れており、実験的な結果を示している。
0.70
We do not ing fine-tuning and testing.
微調整やテストは行いません。
0.61
Despite the mance gain on benchmark datasets.
ベンチマークデータセットの上昇にもかかわらず。
0.52
Experiments 5 Datasets.
実験 5 データセット。
0.86
JOBS, We use three semantic parsing datasets: contains
JOBS 3つのセマンティック解析データセットを使用します。
0.47
GEOQUERY, ATIS.
GEOQUERY。 ATIS。
0.72
(9) its and
(9) その そして
0.75
JOBS about in listjob Prolog pairs 640 question-LF and and Mooney, 1996) ings.
JOBS について in listjob Prolog pairs 640 question-LF and Mooney, 1996) ings。
0.71
GEOQUERY (Zelle include 880 and 5,410 utterance(Price, 1990) ATIS in about US geography LF pairs calculas lambda flight and respectively.
geoquery (zelle には 880 と 5,410 utterance (price, 1990) atis in about us geography lf pairs calculas lambda flight がある。 訳抜け防止モード: GEOQUERY (Zelle include 880 and 5,410 utterance (Price, 1990 ) ATIS in about US geography LF pairs calculas lambda flight そしてそれぞれ。
0.82
of booking, number The JOBS, GEOQUERY, ATIS in predicates 24, 15, is atoms respectively.
予約では、数 the jobs, geoquery, atis in predicates 24, 15 はそれぞれ原子である。
0.71
All in the datasets and 88, are and Lapata, 2016).
データセットと88のすべてが、2016年のand lapataである。
0.69
anonymized as in (Dong each dataset, we For randomly selected m predJOBS, is for 3 new predicates, which the as icates Then we and GEOQUERY split for and 5 ATIS.
aonymized as in (Dong each dataset, we For randomly selected m predJOBS, is for 3 new predicates, which the as icates Then we and GEOQUERY split for and 5 ATIS. (英語)
0.83
evaluation each dataset into an set.
データセットをセットに評価します
0.73
train set and a of the removed And template the instances, we dataset.
train set and a of the removed And template the instance, we dataset。
0.80
which is unique in each The number of such instances is around 100, 150 and 600 in JOBS, GEOQUERY, and ATIS.
The ratios between the and 2:5, 1:4, set train and set evaluation are the JOBS, GEOQUERY, and ATIS, 1:7 in respectively.
2:5, 1:4, set train と set evaluation の比率は、それぞれ jobs, geoquery, atis, 1:7 である。
0.76
Each LF in an evaluation set contains at least a new predicate, while contains an LF in a train set To predicates.
評価セット内の各LFは、少なくとも新しい述語を含み、列車セット内のLFを述語とする。
0.62
only learnk-shot evaluate known set by randomly sampling k ing, we build a support pairs per new predicate without replacement from an evaluation set, remaining pairs and keep the as caused To evaluation test the bias by avoid set.
k ing をランダムにサンプリングして既知のセットを評価する場合のみ、評価セットから置き換えることなく、新しい述語ごとにサポートペアを構築し、残りのペアを作成し、引き起こされるように評価テストを回避してバイアスを保ちます。
0.69
randomness, we repeat above process six times the build to different splits support and test set of six for hyperparameter from each evaluation set.
One consider evaluation. We rest tuning and the for at to the most 2-shot limited number of learning due instances per new predicate each evaluation set.
{3, 6}. The smoothing term is number set to The of meta-support examples is 30 and the number of 15. example support per examples meta-test The is attention coefficient of regularization is set to 0.01 on JOBS and on the other datasets.
We employ 1 the 200-dimensional GLOVE embedding (Penningthe word embeddings al., 2014) ton et to initialize all LSTM size of for utterances.
我々は,200 次元グローブ埋め込み (penning the word embeddeds al., 2014) ton et を用いて,発話のlstm サイズを初期化する。
0.68
The hidden state models (Hochreiter and Schmidhuber, 1997) is 256. size During batch is 2, the learnfine-tuning, selected from {0.001, ing rates are and the 0.0005} and {20, 30, 40, 60, 120}, respectively.
隠れ状態モデル (Hochreiter と Schmidhuber, 1997) は 256 サイズで、バッチ中は {0.001, ing rate is and the 0.0005} と {20, 30, 40, 60, 120} から選択された学習ファインチューニングが 2 である。
0.80
the epochs
時代は
0.57
英語（論文から抽出）
日本語訳
スコア
SEQ2SEQ (pt) SEQ2SEQ (cb) SEQ2SEQ (os) COARSE2FINE COARSE2FINE COARSE2FINE (pt) IRNET (cb) IRNET (os) IRNET DA PT-MAML Ours Evaluation of
Baselines. with methods our We compared attenSEQ2SEQ with baselines, competitive five (Dong COARSE2FINE al., et 2015), (Luong tion (Guo et and Lapata, 2018), al., 2019), PTIRNET et (Li 2018) DA al., and (Huang MAML et al., the best performing super2020).
ベースライン。 私たちはattenSEQ2SEQをベースライン、競争力5(Dong COARSE2FINE al., et 2015), (Luong tion (Guo et and Lapata, 2018), al., 2019), PTIRNET et (Li 2018), DA al., (Huang MAML et al., the best performing super2020)と比較しました。
0.74
COARSE2FINE is standard vised model on the split of GEOQUERY learnfew-shot PT-MAML is datasets.
In list of predicates case, we consider a in support sets as the columns of a new database schema and incorinto porate IRNET encoding module of schema the our base parser.
述語のリストでは、a のサポートセットを新しいデータベーススキーマの列として、incorinto porate irnet encoding module of schema the our base parserとして捉えています。
0.86
We encoder of the IRNET choose al., 2019) because over RATSQL (Wang et IRNET achieves superior performance on our datasets.
IRNETのエンコーダは、RATSQL(Wang et IRNETはデータセット上で優れたパフォーマンスを実現するため、al., 2019)を選択します。
0.73
three different supervised learning consider We settings.
3つの異なる教師付き学習が設定を検討します。
0.49
First, we pre-train a model on a train set, followed by fine-tuning it on the corresponding coined pt.
まず、列車セットでモデルを事前トレーニングし、その後、それに対応するptで微調整します。
0.68
is a model Second, support trained set, support set the and train on combination of a a coined cb.
We include normality in data the corresponding p-values our in tables.
テーブル内の対応するp-値のデータに正規性を含める。
0.66
result Results and Discussion 5.1 signifand accuracies average shows 2 Table the on all compared parsers icance test results of all outperthree datasets.
resultResult and Discussion 5.1 signifand accuracies average show 2 Table 比較されたすべてのパーサーのニュアンステスト結果を表に示します。
0.82
Overall, ProtoParser average least at baselines with all forms 2% on and one-shot accuracy twoof both in terms in signifstatistically results shot settings.
corresponding p-values COARSE2FINE The 0.00276 Given respectively.
対応するp値 COARSE2FINE 0.00276 それぞれ与えられる。
0.57
are and 0.000148, one-shot example on JOBS, achieves our parser baseline, best the than accuracy 7% higher and is the gap 4% on GEOQUERY with two-shots examples.
In addition, none of the SOTA baseline parsers can consistently SOTA outperform other for data few parallel are there parsers when new setting, predicates.
In one-shot supervised the best (cb) baseline IRNET can achieve the best results on GEOQUERY and JOBS among all baselines, and it setting, two-shot on on GEOonly best performs to achieve good perforalso difficult It QUERY.
one-shot supervised the best (cb) base irnet can achieve the best results on geoquery and jobs between all baselines, and it setting, two-shot on on on geoonly best perform to achieve good perfor also difficult it query. (英語)
0.80
is existing meta-learning mance by adapting the or transfer learning algorithms to evias dent by the moderate performance of PT-MAML and DA on The problems of few-shot learning demonstrate imposed predicates.
PT-MAML と DA の適度な性能によってデントを回避し, or transfer learning アルゴリズムを適応させることにより,既存のメタラーニングのマンスである。
0.75
infrequent by challenges the There are significant proportions of infrequent predexisting icates on the datasets.
infrequent by challenge the データセットに不定期な前指数のイケートの割合が有意に多い。
0.62
For example, on there GEOQUERY, are 10 predicates contributing 24 the 4% of to prediall of frequency total only the cates, while top two frequent predicates amount
IRNET achieve SOTA parsers result, As the a to 42%.
IRNET SOTAパーサーの結果を達成し、aから42%になります。
0.76
accuracy with 44% of 25% and than less merely respectively.
精度は25%の44%で、単にそれ以下である。
0.76
In examples, and two-shots one-shot contrast, than 84% acachieve more those parsers standard splits of the curacy on the same datasets supervised setting.
the in Infrequent predicates in can semantic parsing imbalance problem, when class a also be viewed as combined sets train sets and support cerare a in tain manner.
infrequent predicates in in can semantic parsing unbalance problem クラスaをセットのトレーニングセットとサポートのセラーレaの組み合わせと見なす場合。
0.64
In this work, the the ratio between JOBS, GEOQUERY, set the support in train and set respectively.
best option for and sets The to pre-train on a and SEQ2SEQ is COARSE2FINE train set followed by fine-tuning on the correspondsupport ing favors oversampling set, while two-shot in setting.
a と SEQ2SEQ のプリトレインは COARSE2FINE の列車セットに続き、対応するing の微調整はオーバーサンプリングセット、設定は2ショットである。
0.74
Ablation Study We of examine differeffect the ent components our of each removing by parser of them individually and reporting the corresponding removTable accuracy.
The than 0.00327. corresponding p-values less all are eiexclude predicate-dropout, we To investigate during pre-training supervised-loss ther (-sup) or new predicate initialization of by proembeddings totype vectors before fine-tuning (-proto).
is clear It from Table 3 that ablating either supervisely trained action embeddings or prototype vectors hurts performance attention regularefficacy of study the further We (-reg), removing removing it completely by ization (-strsim), or constring similarity feature only the the (-cond).
Support Set Analysis We observe that all modcertain on high achieve els consistently accuracy sets dataset, while obtaining support the same of other low accuracies on the ones.
We applied T-SNE (Maaten and Hinton, 2008) for dimension used sets two reduction.
t-sne (maaten and hinton, 2008) を2次元の減算に応用した。
0.70
We in support GEOQUERY.
GEOQUERYをサポートいたします。
0.81
the examone-shot All
テストワンショットのすべて
0.57
highlight setting on
ハイライト設定。
0.52
英語（論文から抽出）
日本語訳
スコア
to tend set in the highest performing support ples in regions evenly and cover different dense scatter the examples the space, while the lowest in feature significant support set are from a performing far examples number of dense regions.
When we leave out each example in the highest and re-evaluate our performing support set each parser ones good that the observe time, we the green box in Figure 1) (e g locate either in or close to some of regions.
6 Conclusion and Future Work learning based semanfew-shot We propose a novel cope with ProtoParser, to tic parser, coined challenges the address To in LFs.
6 結論と今後のワークラーニングベース semanfew-shot では、lfs のアドレスに対する挑戦としてprotoparser を用いた tic パーサ を提案する。 訳抜け防止モード: 6 Conclusion and Future Work Learning based semanfew - shot 私たちはProtoParserで新しい対処法を提案します。 to tic parser, coined the address To in LFs.
0.80
new predicates to train the parser learning, we few-shot in propose with a pre-training procedure involving predicateattention regularization, dropout, and pre-training smoothing.
The resulted model achieves superior results over competitive baselines on three benchmark datasets.
得られたモデルは、3つのベンチマークデータセットの競合ベースラインよりも優れた結果が得られます。
0.59
the dense and Yoshua Benby translation jointly arXiv preprint
密に ヨシュア・ベンビー訳はarxivプレプリントで
0.51
References Dzmitry Bahdanau, Kyunghyun Cho, gio.
Dzmitry Bahdanau,kyunghyun Cho, gioを参照。
0.81
Neural 2014. machine translate.
ニューラル2014の機械翻訳。
0.71
and align learning to arXiv:1409.0473.
arXiv:1409.0473に学習を合わせます
0.66
Cai, Shu Madalina Claire Bonial, Laura Banarescu, Griffitt, Kevin Hermjakob, Kira Georgescu, Ulf and Nathan Palmer, Philipp Koehn, Martha Knight, representation Abstract meaning 2013.
Cai, Shu Madalina Claire Bonial, Laura Banarescu, Griffitt, Kevin Hermjakob, Kira Georgescu, Ulf and Nathan Palmer, Philipp Koehn, Martha Knight, representation Abstract, 2013
0.81
Schneider. of the 7th LinguisIn Proceedings sembanking.
シュナイダー。 第7回LinguisIn Proceedings sembankingの。
0.52
for Annotation Workshop tic with Interoperability and Discourse, pages 178–186.
For Annotation Workshop tic with Interoperability and Discourse, page 178–186.
0.96
Yujia Bao, Shiyu Chang, Mo Yu, 2018. attention Deriving machine In Proceedings of the tionales.
Vijay Siva Jianpeng Cheng, and Reddy, Mirella Lapata.
Vijay Siva Jianpeng Cheng, and Reddy, Mirella Lapata
0.70
2019. Learning an neuLinguistics, Computational ral parser.
2019. NeuLinguistics、Computational ral parserを学ぶ。
0.79
semantic 45(1):59–94.
意味45(1):59-94。
0.67
Dadashkarimi, Alexander Javid and Dragomir R Radev.
Dadashkarimi、Alexander Javid、Dragomir R Radev。
0.71
Tatikonda, learning semantic for transfer arXiv:1808.09889. preprint Demˇsar.
Tatikonda, learning semantic for transfer arXiv:1808.09889。
0.87
2006. Statistical Janez over multiple sets.
2006. 複数の集合上の統計ジャネス。
0.79
data sifiers research, 7(Jan):1–30.
data sifiers research, 7(Jan):1–30。
0.85
learning 2016. and Mirella Li Dong form ical with attention.
2016年、ミレラ・リ・ドン(Mirrella Li Dong)がICを結成。
0.51
arXiv:1601.01280.
arXiv:1601.01280。
0.49
Li Dong coding neural arXiv:1805.04793.
Li Dong code neural arXiv:1805.04793
0.76
Afshar, Hadi Duong, Long Philip Pink, Cohen, and Mark learning tive for deep semantic 56th ings Annual Meeting the of for Computational Linguistics pers), pages 43–48.
Afshar, Hadi Duong, Long Philip Pink, Cohen, Mark Learning tive for Deep semantic 56th ings Annual Meetings of for Computational Linguistics pers), page 43–48。
0.79
Wang Ballesteros, Dyer, Chris Miguel and Matthews, A Noah Smith.
Wang Ballesteros、Dyer、Chris Miguel、Matthews、A Noah Smith。
0.78
2015. with parsing stack dependency based arXiv preprint arXiv:1505.08075. term memory.
2015年、 arXiv preprint arXiv:1505.08075. term memoryを解析した。
0.66
Catherine Zhang, Li Rui Zhang, text-to-sql arXiv:1806.09029. and Sergey Levine.
Catherine Zhang, Li Rui Zhang, text-to-sql arXiv:1806.09029. and Sergey Levine
0.72
2017. Chelsea Finn, for adaptation of Model-agnostic meta-learning 34th InternaIn Proceedings of deep networks.
2017. Chelsea Finn, for adapt of Model-Agnostic meta-learning 34th InternaIn Proceedings of Deep Network
0.83
on Machine Learning-Volume tional Conference 70, pages org.
on Machine Learning-Volume tional Conference 70, page org。
0.92
JMLR. 1126–1135.
JMLR所属。 1126–1135.
0.71
J¨urgen Schmidhuber, Felix A Gers, 1999. lstm.
J surgen Schmidhuber, Felix A Gers, 1999. lstm.
0.97
Xiao, Yan Gao, Yan Guo, Jiaqi Zhan, Zecheng Zhang.
Xiao, Yan Gao, Yan Guo, Jiaqi Zhan, Zecheng Zhang
0.75
Dongmei and Jian-Guang Ting Liu, Lou, Towards 2019. cross-domain complex in text-to-sql arXiv representation.
DongmeiとJian-Guang Ting Liu, Lou, Towards 2019. text-to-sql arXiv表現におけるクロスドメインコンプレックス。
0.69
intermediate database with preprint arXiv:1905.08205.
プレプリント arXiv:1905.08205 の中間データベース。
0.63
Herzig Jonathan Jonathan 2018. lexicon zero-shot pling structure arXiv preprint arXiv:1804.07918. parsing.
Po-Sen Huang, Chenglong Wang, Rishabh Singh, Wenand Xiaodong He
0.84
language Natural 2018. tau Yih, In generation to structured via meta-learning.
言語 自然言語 2018. tau Yih, In generation to structured via meta-learning
0.87
of Conference 2018 Proceedings the North of the American ComputaAssociation for Human tional Language Technologies, Volume 2 (Short Papers), pages 732–738.
会議2018 Proceedings the North of the American ComputaAssociation for Human tional Language Technologies, Volume 2 (Short Papers), Page 732–738。
0.79
Luke Srinivasan Zettlemoyer.
ルークSrinivasan Zettlemoyer。
0.65
and Iyer, Alvin Cheung, programmatic for 2019. idioms Learning scalable arXiv:1904.09086. arXiv preprint semantic parsing.
2017. parsing with Neural straints semi-structured tables.
2017年、ニューラルひずみ半構造テーブルで解析。
0.58
for the 2017 Conference on Empirical Methods ral Language Processing, pages 1516–1526.
2017 Conference on Empirical Methods ral Language Processing, page 1516–1526。
0.79
decodrecursive and Dongjun Lee.
decodrecursiveとDongjun Lee。
0.72
Clause-wise 2019. genercross-domain and complex for ing text-to-sql on Proceedings the 2019 Conference of In ation.
Clause-wise 2019. genercross-domain and complex for ing text-to-sql on Proceedings 2019 Conference of In ation
0.81
in Natural Language Processing Empirical Methods and on NatuJoint Conference International 9th the Language Processing ral (EMNLP-IJCNLP), pages 6047–6053.
In Natural Language Processing Empirical Methods and on NatuJoint Conference International 9th the Language Processing ral (EMNLP-IJCNLP), Page 6047–6053。
0.95
Jaesik Yoon, Dongjun Lee, and Sungroh Yoon.
Jaesik Yoon、Dongjun Lee、Sungroh Yoon。
0.66
2019. arXiv to-sql
2019. arXiv to-sql
0.67
Grefenstette, Edward and Blunsom, Phil parsing with Semantic arXiv autoencoders.
Grefenstette, Edward and Blunsom, Phil parsing with Semantic arXiv autoencoders。
for semantic parsarXiv preprint arXiv:2006.13071. ing.
for semantic parsarXiv preprint arXiv:2006.13071.
0.85
Neumann, Mark Ben Lin, Kevin Bogin, Berant, Matt and 2019.
Neumann, Mark Ben Lin, Kevin Bogin, Berant, Matt, 2019。
0.80
Gardner. arXiv text-to-sql neural based generation.
Gardner arXivテキストからsqlのニューラルベース生成。
0.61
arXiv:1905.13326.
arXiv:1905.13326。
0.49
Liu, Masao Lemao Utiyama, 2016.
Liu, Masao Lemao Utiyama, 2016年。
0.90
Sumita. ichiro with supervised attention.
墨田。 一郎は気を付けて見張った。
0.39
ING 2016, the 26th Computational Linguistics: 3093–3102.
2016年、26th Computational Linguistics: 3093–3102。
0.85
Shulin Liu, Yubo Chen, Kang Liu, argument Exploiting information via detection supervised 55th Proceedings of the sociation for Computational Long Papers), pages 1789–1798.
Shulin Liu, Yubo Chen, Kang Liu, argument Exploiting information via detection supervised supervised 55th Proceedings of the sociation for Computational Long Papers), page 1789–1798。
0.90
Pham, Luong, Minh-Thang Hieu and approaches Manning.
Pham, Luong, Minh-Thang Hieu は Manning に近づいた。
0.72
2015. Effective based neural machine translation.
2015. 効果的なニューラルマシン翻訳。
0.80
arXiv:1508.04025. van Laurens Geoffrey der Maaten Journal data using Visualizing research, 9(Nov):2579–2605.
arXiv:1508.04025. van Laurens Geoffrey der Maaten Journal data using Visualizing research, 9(Nov):2579–2605
0.84
learning Christopher D Manning, Prabhakar Raghavan, and Hinrerich trieval.
Christopher D Manning, Prabhakar Raghavan, Hinrerich trieval を学ぶ。
0.79
Cambridge university The Montague.
ケンブリッジ大学モンタギュー校教授。
0.65
1973. Richard ordinary in english.
1973. リチャードは通常英語です。
0.82
quantification language, pages natural 221–242.
Quantification Language, page natural 221–242。
0.80
Springer. and Graham Neubig.
Springer そしてGraham Neubig。
0.61
2019. Ansong Ni, Pengcheng Yin, for Merging weak active supervision semantic arXiv preprint arXiv:1911.12986. parsing.
2019. Ansong Ni, Pengcheng Yin, for Merging weak active supervision semantic arXiv preprint arXiv:1911.12986. parsing.
0.89
Socher, Jeffrey Richard Pennington, and Manning.
ソッチャー、ジェフリー・リチャード・ペニントン、マニング
0.60
2014. Glove: Global vectors of the In Proceedings 2014 resentation.
2014. glove: global vectors of the in proceedings 2014 resentation (英語)
0.83
on in natural language empirical methods ing (EMNLP), pages 1532–1543.
in natural language empirical method ing (EMNLP), page 1532–1543.
0.83
spoken Patti of Evaluation 1990.
評価のパティ1990を話しました。
0.51
J Price. Speech In tems: domain.
Jプライス。 Speech In tems: ドメイン。
0.63
The atis Proceedings guage: of a Workshop Held Valley, Pennsylvania, June 24-27, 1990. and Rabinovich, Maxim Mitchell 2017.
The atis Proceedings guage: of a Workshop Held Valley, Pennsylvania, June 24-27, and Rabinovich, Maxim Mitchell 2017
0.75
Abstract code Introduction to press.
抽象コード プレスの紹介。
0.69
proper treatment In Approaches
適切な治療 : アプローチ
0.77
language sysLanand Natural at Hidden
英語 sysLanand Natural at Hidden
0.76
Christopher for word repconference process-
Christopher for word reconference process-
0.91
Christopher D attentionto arXiv preprint
Christopher D attentionto arXiv preprint
0.85
Stern, networks for
Stern、ネットワーク。
0.68
Dan Klein. generation
ダン・クライン。 世代
0.59
2008. machine Hinton.
2008年 マシン Hinton
0.67
of information and t-sne.
ですから 情報 t-sne。
0.60
Sch¨utze. syntax 2008.
という。 構文 2008.
0.61
of to and
of ～ そして
0.71
英語（論文から抽出）
日本語訳
スコア
2018. models The 56th
2018年モデル56位。
0.66
and GraJunxian He, Pengcheng Yin, Chunting Zhou, laTree-structured ham Neubig.
そしてGraJunxian He, Pengcheng Yin, Chunting Zhou, laTree-structured ham Neubig。
0.88
StructVAE: semi-supervised variable tent semantic for the AssoAnnual Meeting of parsing.
StructVAE:AssoAnnual Meeting of parsingのための半教師付き可変テントセマンティック。
0.70
In ciation for Computational Linguistics (ACL).
計算言語学 (Computational Linguistics, ACL) の略。
0.71
Yasunaga, Tao Yu, Rui Zhang, Kai Yang, Michihiro Irene Li, QingnJames Ma, Dongxu Wang, Zifan Li, A 2018.
Building a semantic Association Annual Meeting of the ings of the 53rd 7th Internaand the Linguistics for Computational tional ProLanguage on Conference Joint Natural 1: Long cessing (Volume Papers), volume 1, pages 1332–1342.
第53回第7回internaand the languages for computational tional prolanguage on conference joint natural 1: long cessing ( volume papers), volume 1, pages 1332–1342 (英語)
We a frequent if are complete collapsed into the abstract subtrees corresponding tree nodes.
頻繁なifは、対応するツリーノードの抽象サブツリーに完全に崩壊します。
0.61
the The pseudocode of tree in Algoprovided algorithm is normalization rithm 2.
Algoprovidedアルゴリズムの木の擬符号は正規化rithm 2である。
0.87
One Example Transition Sequence B transition example an provide 4, we Table As in and states the corstack the sequence to display the utsequence when responding action parsing is ”how much Introduction terance ground the in transportation between and downtown?”.
1つの例 遷移シーケンスb遷移例 a provide 4, we table as in and states corstack the sequence to display the utsequence when response action parsing is "どの程度のテラスの導入が、市内とダウンタウンの間の輸送を基礎にしているか?
0.74
atlanta Algorithm 2: Template Normalization trees T , Input :A set a minimal abstract of τ Output :A set of normalized trees their subtrees of O := mapping to in T for do tree t update occurrence of leaf all end while O updated with new trees do in O do tree occur list t, l l(cid:48) for list occurrence build size(l(cid:48)) ≥ size(l) then if O[t(cid:48)] = l(cid:48) end
アトランタ アルゴリズム 2: テンプレート正規化木 T , Input :A set a minimal abstract of ^ Output :A set a set a normalized trees their subtrees of O := mapping to to in T for do tree t update occurrence of leaf all end while O do new tree do in O do tree occur list t, l l(cid:48) for list occurrence build size(l(cid:48)) ^ size(l) then if O[t(cid:48)] = l(cid:48) end
0.75
occurrences nodes v of t
t のノード v の発生
0.88
supertree for
スーパーツリー ですから
0.54
support in T . to O[v]
tでのサポート。 to O[v]
0.78
t(cid:48) of
t(cid:48) ですから
0.69
t end in O do t, occur tree list l size(l) ≥ τ if then collapse into a node t end
t end in o do t, occur tree list l size(l) ≥ τ if then collapse into a node t end
0.85
end for end for
終止符 終わり ですから
0.64
all t(cid:48)
全部 t(cid:48)
0.77
in l. Template Normalization A the in templates LF Many existing have corpora shared corresponding the subtrees in abstract setree mantic trees.
で l. テンプレートの正規化 LF 既存の多くのコーパスは抽象的なセトリーマンティックツリーで対応するサブツリーを共有している。
0.77
The normalization algorithm aims to treat those subtrees single as units.
正規化アルゴリズムは、これらのサブツリーを単位として扱うことを目指している。
0.43
The structured shared such of identification conis subtrees.
構造は、識別コニサブツリーのような共有である。
0.59
Given an LF ducted by finding frequent support dataset, the of a tree is the number of LFs t frequent that it occurs as a subtree.
t6 t7 t8 Stack [] va)] transport [(ground city va va), (to ve)] transport [(ground (from airport city va (to transport ve), va), ve), city va (to va), transport [(ground ve), (from airport (= (ground fare va)] va) va (ground [(and transport va) (to city va ve) (from airport ve) (= (ground fare va) va))] va ve) city va va) transport (ground [(exists (to (and va va)))] va) fare (= (ground ve) (from airport va e (exists (and (ground transport va) (to city va va va (from airport ve) (= (ground fare va) va))))] va
t6 t7 t8 Stack [] va)] transport [(ground city va va), (to ve)] transport [(ground (from airport city va (to transport ve), va), ve), city va (to va), transport [(ground ve), (from airport (= (ground fare va)] va) va (ground [(and transport va) (to city va ve) (from airport ve) (= (ground fare va) va))] va ve) city va va) transport (ground [(exists (to (and va va)))] va) fare (= (ground ve) (from airport va e (exists (and (ground transport va) (to city va va va (from airport ve) (= (ground fare va) va))))] va
0.77
va ve)] ve)
ヴァ ve)] ve)
0.68
[(ground [(lambda
[(グラウンド) [(ラムダ)]
0.64
Action GEN [(ground transport va)] GEN [(to ve)] city va GEN [(from airport ve)] va GEN [(= (ground fare va)] va) :- NT NT NT NT] REDUCE [and REDUCE [exists :- va NT] :REDUCE [lambda va
Action GEN [(ground transport va)] GEN [(to ve)] city va GEN [(from airport ve)] va GEN [(= (ground fare va)] va) :- NT NT NT NT] REDUCE [ and REDUCE [exists :- va NT] :REDUCE [lambda va] 訳抜け防止モード: アクション GEN [ ( Ground Transport va ) gen [ (to ve ) ] City va GEN [ (from Airport ve ) ] va GEN [ (= ( Ground fare va ) ) va ) : - NT NT NT NT NT ] REDUCE [in Japanese] REDUCE [ exists : - va NT ] : REDUCE [ lambda va]