Can language models read biomedical texts and explain the biomedical
mechanisms discussed? In this work we introduce a biomedical mechanism
summarization task. Biomedical studies often investigate the mechanisms behind
how one entity (e.g., a protein or a chemical) affects another in a biological
context. The abstracts of these publications often include a focused set of
sentences that present relevant supporting statements regarding such
relationships, associated experimental evidence, and a concluding sentence that
summarizes the mechanism underlying the relationship. We leverage this
structure and create a summarization task, where the input is a collection of
sentences and the main entities in an abstract, and the output includes the
relationship and a sentence that summarizes the mechanism. Using a small amount
of manually labeled mechanism sentences, we train a mechanism sentence
classifier to filter a large biomedical abstract collection and create a
summarization dataset with 22k instances. We also introduce conclusion sentence
generation as a pretraining task with 611k instances. We benchmark the
performance of large bio-domain language models. We find that while the
pretraining task help improves performance, the best model produces acceptable
mechanism outputs in only 32% of the instances, which shows the task presents
significant challenges in biomedical language understanding and summarization.
The abstracts of these publications often include a focused set of sentences that present relevant supporting statements regarding such relationships, associated experimental evidence, and a concluding sentence that summarizes the mechanism underlying the relationship.
We leverage this structure and create a summarization task, where the input is a collection of sentences and the main entities in an abstract, and the output includes the relationship and a sentence that summarizes the mechanism.
Using a small amount of manually labeled mechanism sentences, we train a mechanism sentence classifier to filter a large biomedical abstract collection and create a summarization dataset with 22k instances.
We also introduce conclusion sentence generation as a pretraining task with 611k instances.
また,611kインスタンスの事前学習タスクとして結論文生成を導入する。
0.58
We benchmark the performance of large bio-domain language models.
大規模生物ドメイン言語モデルの性能をベンチマークする。
0.65
We find that while the pretraining task help improves performance, the best model produces acceptable mechanism outputs in only 32% of the instances, which shows the task presents significant challenges in biomedical language understanding and summarization.
Keywords: Explanation Generation, Text Generation, Summarization, Biomedical NLP, Relation Extraction
キーワード:説明生成、テキスト生成、要約、生物医学的nlp、関係抽出
0.74
1. Introduction Understanding biochemical mechanisms such as protein signaling pathways is one of the central pursuits of biomedical research (Str¨otgen and Gertz, 2012) (Arighi et al , 2011; Krallinger et al , 2017; Demner-Fushman et al , 2020).
1. はじめに タンパク質シグナリング経路などの生化学的メカニズムを理解することは、生物医学研究の中心的研究の1つである(Arighi et al , 2011; Krallinger et al , 2017; Demner-Fushman et al , 2020)。
0.59
Biomedical research has advanced tremendously in the past few decades, to the point where we now suffer from “an embarrassment of riches”.
Publications are generated at such a rapid pace (PubMed 1 has indexed more than 1 million publications per year in the past 8 years!) that we need information access applications which can help extract and organize biomedical relations and summarize the biomedical mechanisms underlying them.
The task requires models to read text that presents information about the connection between two target entities and generate a summary sentence that explains the underlying mechanism and the relation between the entities.
First, summarizing biomedical mechanisms can be seen as part of the broader efforts in extracting (Czarnecki et al , 2012), organizing (Kemper et al , 2010; Kemper et al , 2010; Miwa et al , 2013; Subramani et al , 2015; Poon et al , 2014), and summarizing (Azadani et al , 2018) biomedical literature that are aimed at providing information access tools for domain experts.
第一に、バイオメディカルメカニズムの要約は、(Czarnecki et al , 2012)、組織化(Kemper et al , 2010; Kemper et al , 2010; Miwa et al , 2013; Subramani et al , 2015; Poon et al , 2014)、および(Azadani et al , 2018)、ドメインエキスパートに情報アクセスツールを提供することを目的としたバイオメディカル文献の要約(Azadani et al , 2018)の幅広い取り組みの一部として見ることができる。
0.80
1https://pubmed.ncbi .nlm.nih.gov
1https://pubmed.ncbi .nlm.nih.gov
0.16
Figure 1: Biomedical Mechanism Summarization Task: Example of an entry from the SuMe dataset.
図1: 生体医学的メカニズム 要約タスク:sumeデータセットからのエントリの例。
0.77
Some supporting text was removed to save space.
スペースを節約するためにいくつかのサポートテキストが削除された。
0.52
The input is the supporting sentences with the main two entities.
入力はメイン2つのエンティティを持つサポート文である。
0.70
The output is the relation type and a sentence concluding the mechanism underlying the relationship.
出力は関係型と関係の基盤となるメカニズムを構成する文である。
0.78
Second, from an NLP perspective this task can be seen as an explainable relation extraction in a biomedical context, where the explanation is the mechanism that provides information about why the relation holds or how it comes about.
To address this, we turn to the structure that exists
これに対処するために 存在する構造に目を向けます
0.69
Biomedical AbstractThis study re-examined the hyperactivity and disruption of prepulse inhibitioninduced by Nmethyl-D-aspartate stimulation ... of the rat ventral hippocampusand compared how both effects were affected by pretreatment with eitherhaloperidol or clozapine.
While the hyperactivity is thought to depend ondopamine receptor activation in the nucleus accumbens, the dopamine D2-classreceptor blocker haloperidol failed to antagonize the disruption of prepulseinhibition in previous studies.
However, an ameliorative effect of the atypicalneuroleptic clozapine on disruption of prepulse inhibition was suggested by ...In the present study, bilateral infusion of Nmethyl-D-aspartate ... into theventral hippocampus of Wistar rats increased ... disrupted prepulse inhibition.Both effects were observed immediately after infusion but disappeared 24h later.Injection of .
We conclude that dopaminergic mechanisms are differentially involved in thehyperactivity and disruption of prepulse inhibition induced by Nmethyl-D-aspartate stimulation of the ventral hippocampus.
in biomedical abstracts, make use of related datasets, and devise a semi-automatic bootstrapping process that builds on a relatively small amount of labeling effort from domain experts.
For a given abstract, we create a task instance that consists of a pair of biochemical entities (regulated and regulator), the relationship between them (positive/negative activation), and supporting sentences that provide information about this relationship, and a sentence that summarizes the mechanism underlying the relation (see Figure 1).
Creating such an instance would require a domain expert to read through an abstract and assess if it contains a biomedical mechanism and locate it if so.
To address this issue, we introduce a semi-automated annotation process to create a large-scale set for development and automatic evaluation purposes and a clean small-scale manually curated subset of instances for manual evaluation.
In particular, the necessary entities and relations are extracted using an existing biomedical information extraction system (Valenzuela-Esc´arcega et al , 2018).
特に、既存の生物医学情報抽出システム(valenzuela-esc ́arcega et al, 2018)を用いて必要な実体と関係を抽出する。 訳抜け防止モード: 特に 必要な実体と関係が抽出され 既存の生物医学情報抽出システム(valenzuela - esc ́arcega et al, 2018)の使用。
0.79
To extract mechanism summaries we first collected a small set of mechanism sentences with the help of domain experts.
We use this to bootstrap a larger sample by training a mechanism sentence classifier with a biomedical language model (LM) (Kanakarajan et al , 2021) and apply it to a large collection of about 611K abstracts that contained a conclusion sentence about the relationship between a pair of entities.
これを用いて,生物医学的言語モデル(LM)を用いたメカニズム文分類器(Kanakarajan et al , 2021)を訓練し,一対の実体の関係に関する結論文を含む約611Kの抽象文の大規模なコレクションに適用することにより,より大きなサンプルをブートストラップする。
0.79
The subset that the classifier identifies as containing mechanism sentences is used to create 22K mechanism summarization instances.
• We benchmark several state-of-the-art language models for the task of generating the underlying biochemical relations and the corresponding mechanism sentences.
We train general domain LMs (GPT2 (Radford et al , 2019), T5 (Raffel et al , 2020a), BART (Lewis et al , 2019)), as well as science domain adapted versions(scientific GPT2 (Papanikolaou and Pierleoni, 2020), and SciFive (Phan et al , 2021)) and benchmark their performance through both automatic evaluation and manual evaluation on curated evaluation samples.
GPT2(Radford et al , 2019)、T5(Raffel et al , 2020a)、BART(Lewis et al , 2019)、科学分野に適応したバージョン(Scientific GPT2(Papanikolaou and Pierleoni, 2020)、SciFive(Phan et al , 2021)を訓練し、キュレートされた評価サンプルの自動評価と手作業によるパフォーマンスのベンチマークを行う。
0.74
• The evaluation by domain experts suggests that this is a high quality dataset coupled with a challenging task, which deserves further investigation.
2. Related Work Deep learning models have been widely used in different NLP applications (Gaonkar et al , 2020; Bastan et al., 2020; Keymanesh et al , 2021; Heidari et al , 2021).
2.関連業務 ディープラーニングモデルはさまざまなNLPアプリケーションで広く使われている(Gaonkar et al , 2020; Bastan et al., 2020; Keymanesh et al , 2021; Heidari et al , 2021)。
0.78
Amongst these applications, biomedical NLP is using these models that looks at extracting (Alam et al , 2018; Mulyar et al , 2021; Giorgi and Bader, 2020), organizing (Yuan et al , 2020; Zhao et al , 2020; Lauriola et al., 2021), and summarizing information (Cohan et al , 2018) from scientific literature.
これらの応用のうち、生物医学的nlpは、これらのモデルを用いて、(alam et al , 2018; mulyar et al , 2021; giorgi and bader, 2020)、組織化(yuan et al , 2020; zhao et al , 2020; lauriola et al., 2021)、科学文献からの要約情報(cohan et al , 2018)を抽出している。
0.86
Within this broad context, the mechanism summarization task we introduce broadly relates to previous work in reading and generating information from scientific texts.
Most work in this area focus on generating summaries using scientific publication and some times in combination with external information (Yasunaga et al , 2019; DeYoung et al , 2020; Collins et al , 2017) Some works even seek to generate part of the scientific papers.
この分野のほとんどの研究は、科学出版物を使った要約や、外部情報(Yasunaga et al , 2019; DeYoung et al , 2020; Collins et al , 2017)と組み合わせて作成することに焦点を当てている。 訳抜け防止モード: この分野のほとんどの仕事は 科学論文を用いた要約の作成 また、外部情報(yasunaga et al, 2019 ; deyoung et al, 2020 ; collins et al, 2017)と組み合わせて、いくつかの研究は科学論文の一部を生成しようとしている。
0.82
For example, TLDR (Cachola et al , 2020) introduces a task and a dataset to generate TLDRs (Too Long; Didn’t Read) for papers.
例えば、TLDR(Cachola et al , 2020)では、論文のためにTLDRを生成するタスクとデータセット(Too Long; Didn’t Read)が導入されている。
0.70
They exploit titles and an auxiliary training signal in their model.
彼らは自身のモデルでタイトルと補助訓練信号を利用する。
0.71
ScisummNet (Yasunaga et al , 2019) introduces a large manually
annotated dataset for generating paper summaries by utilizing their abstracts and citations.
抄録と引用を利用して用紙要約を生成するための注釈付きデータセット。
0.56
TalkSumm (Lev et al , 2019) generates summaries for scientific papers by utilizing videos of talks at scientific conferences.
TalkSumm (Lev et al , 2019)は、科学会議での講演のビデオを利用して、科学論文の要約を生成する。
0.72
PaperRobot (Wang et al , 2019) generates a paper’s abstract, title, and conclusion using a knowledge graph.
PaperRobot(Wang et al , 2019)は、知識グラフを使用して論文の要約、タイトル、結論を生成する。
0.77
FacetSum (Cohan et al , 2018) used Emerald journal articles to generate 4 different abstractive summaries, each targeted at specific sections of scientific documents.
FacetSum (Cohan et al , 2018) はエメラルド誌の論文を用いて4つの異なる抽象的な要約を作成し、それぞれが特定の科学文書のセクションをターゲットにしている。 訳抜け防止モード: facetsum (cohan et al, 2018) emerald journal の記事 科学文書の特定のセクションを対象とする4つの異なる抽象要約を生成する。
0.77
In addition to the specifics of the output that we target, our work is different from all these other works because our proposed summarization task is grounded with the underlying biomedical event discussed, rather than focusing on generic summarization, which may lose the connection to the underlying biology that is the core material discussed in these papers.
There is a huge body of work that addresses explainable methods (e g , relation extraction (Shahbazi et al , 2020) or explainable QA (Thayaparan et al , 2020)).
説明可能な方法(例えば、関係抽出(Shahbazi et al , 2020)、説明可能なQA(Thayaparan et al , 2020))に対処する膨大な作業体が存在する。
0.77
Many prior works in relation and event extraction treat explanations as the task of selecting or ranking sentences that support a relation (e g , (Shahbazi et al , 2020; C¸ ano and Bojar, 2020; Yasunaga et al , 2019)).
関連性やイベント抽出に関する多くの先行研究は、関係を支持する文の選択やランク付けのタスクとして説明を扱っている(例えば、Shahbazi et al , 2020; C > ano and Bojar, 2020; Yasunaga et al , 2019)。
0.80
Our work differs from these in that it focuses on generating mechanisms underlying a relation from supporting sentences, rather than identifying existing sentences.
Our goal is to develop a task and a dataset that pushes models towards distilling the mechanisms that underlie the relationships between entities from biomedical literature.
From a biomedical science perspective, a mechanism provides two types of explanatory information, which we use to characterize mechanism sentences: Why is the relation true?
tions, the findings, which can be used to substantiate the conclusions that establish the relation of interest, and the mechanism underlying the relation.
This suggests a language processing task that tests for ability to understand biomedical mechanisms: given the preceding sentences in the abstract can a model accurately generate the underlying mechanism?
3.1. Task Definition Given a set of sentences from a scientific abstract (referred to as supporting sentences) and a pair of entities (ei, ej) that are the focus of the abstract (referred to as focus entities), generate the conclusion sentence that explains the mechanism behind the pair entities and output a relation that connects these entities (e g , positive activation(ei, ej)).
Figure 1 shows an example of such a tuple of supporting sentences, focus entities, relation, and mechanism sentence.
図1は、このような支持文、焦点エンティティ、関係、機構文のタプルの例を示しています。
0.68
As the example illustrates, mechanism sentences describe some pathway often involving another entity or a process (e g , dopaminergic mechanism), require identifying and combining information from multiple relevant sentences, and nontrivial inferences regarding the relationship between the entities (e g , recognizing that the different effects on prepulse inhibition imply differential involvement).
The task definition suggests what we need to build a dataset.
タスク定義は、データセットを構築するために必要なものを示しています。
0.57
Given an abstract of a scientific literature we need four pieces of information: (1) the two focus entities of the abstract; (2) the relation between entities; (3) sentences from the abstract in support of this relation; and (4) the conclusion sentence where the mechanism underlying the relation is summarized.
Instead, here we employ a bootstrapping process, where we first annotate a small amount of data to build a mechanism sentence classifier that can then helps us collect a large scale dataset for mechanism summarization.
The key observation here is that identifying sentences that express a mechanism is a simpler task than the targeted mechanism summarization task, and, thus, should be learnable from smaller amounts of data.
Figure 3: The bootstrapping pipeline for SuMe collection and human evaluation.
図3: SuMeコレクションと人的評価のためのブートストラップパイプライン。
0.78
The main idea behind the pipeline is to collect relatively easy to acquire judgments from domain experts to then bootstrap and generate a weakly-labeled large training corpus.
We further assess the quality of the resulting dataset through another round of human evaluation, which also yields a smaller curated evaluation dataset.
We use this matching process to also split the abstracts into the set of supporting sentences (the ones that lead up to the conclusion) and one conclusion sentence (the one that includes the conclude word).
2. Extracting Main Entities & Relation Starting with the abstracts which are now in the form of (supporting sentences, conclusion sentence), we then run a biomedical relation extractor, REACH (Valenzuela-Esc´arcega et al , 2018), which can identify protein-protein and chemical-protein relations between entities.
2) 現在(支持文,結論文)の抽象的な部分から主エンティティと関係を抽出し,生物医学的関係抽出装置REACH(Valenzuela-Esc ́arcega et al, 2018)を運用し,エンティティ間のタンパク質タンパク質および化学タンパク質の関係を同定する。
0.86
In this work, we focus on the relations where one entity is the controller and another entity is the controlled entity and the relation between them is either positive/negative activation or positive/negative regulation.
If an abstract does not contain any such relation, we keep it for the pretraining step (as described in Section 5.3); otherwise we use it for the main task.
To this end, we devised a bootstrapping process where we first collect supervised data to train a classifier.
そこで我々は,まず教師付きデータを収集し,分類器を訓練するブートストラッピングプロセスを開発した。
0.65
To collect likely mechanism sentences we made use of the ChemProt (Peng et al , 2019) relation extraction dataset which contains sentences annotated with positive and negative regulation relations between entities.
そこで我々は,ChemProt(Peng et al , 2019)関係抽出データセットを用いて,エンティティ間の正および負の規則関係を付加した文を抽出した。
0.71
However, not all of these sentences necessarily explain the mechanism behind these relations.
しかし、これらすべての文が必ずしもこれらの関係のメカニズムを説明するわけではない。
0.55
We asked 21 experts (grad students in a biomedical department) to inspect each sentence and rate whether it explains the mechanism behind the ChemProt annotated relation on
Each sentence is annotated by three experts and we find the inter-annotator agreement between users to be κ = 73% (Fleiss Kappa (Landis and Koch, 1977)).
各文は3人の専門家によって注釈付けされており、ユーザ間のアノテーション間の合意はκ = 73%である(Fleiss Kappa (Landis and Koch, 1977))。
0.75
The final label for a sentence is selected based on the majority voting after combining Clearly a Mechanism and Plausible a Mechanism labels.
Using this small scale mechanism sentence dataset, we train binary classifiers to identify mechanism sentences, where the positive label indicates that the underlying sentence is a mechanism sentence.
We fine-tuned multiple transformer-based models: BioBERT (Lee et al , 2020), SciBERT (Beltagy et al , 2019), BiomedNLP (Gu et al , 2020), and BioELECTRA (Kanakarajan et al , 2021) models.
BioBERT (Lee et al , 2020), SciBERT (Beltagy et al , 2019), BiomedNLP (Gu et al , 2020), BioELECTRA (Kanakarajan et al , 2021) モデルを微調整した。
0.65
Each model is fitted with a non-linear classification layer that takes the output representation for the [CLS] token.
各モデルは、[CLS]トークンの出力表現を取る非線形の分類層を備える。
0.69
The classification layer and top three layers of the transformer are finetuned using the annotated data4.
アノテーション付きデータ4を用いて、トランスの分類層と上位3層を微調整する。
0.76
We used 80%-20% split for train-test.
電車の試験に80%～20%の分割を使用した。
0.58
BioELECTRA performed the best with 74% macro F1 for mechanism sentence classification.
BioELECTRA は 74% のマクロ F1 で最善を尽くした。
0.66
We use this trained mechanism sentence classifier to label all conclusion sentences from the previous step and instances which are predicted to be mechanism sentences are used to create the mechanism generation of the SuMe dataset.
Table 1: Dataset Statistics: Each dataset contains a number of unique abstracts, a supporting set (supp.), a mechanism sentence (conc.) a pair of entities.
can define a broader conclusion generation task, which can be be used as a pre-training task for the generative models that eventually use for the mechanism summarization task (as we describe in Section 5.3).
The above procedure results in a dataset that allows us to define the following mechanism summarization task: Given a set of supporting sentences from an abstract and a pair of entities (ei, ej), generate a relation that connects these entities and a sentence that explains the mechanism that was the focus of the study.
抽象と一対のエンティティ(ei, ej)からのサポート文の集合が与えられたとき、これらのエンティティと研究の焦点となったメカニズムを説明する文とを関連付ける関係を生成します。 訳抜け防止モード: 上述の手順は、以下のメカニズムの要約タスクを定義できるデータセットとなる: 抽象と一対のエンティティ(ei, ei)からのサポート文の集合が与えられる。 ej ) これらの実体と文をつなぐ関係を生成する 研究の焦点となったメカニズムを説明します。
0.82
The statistics of the dataset are shown in Table 1.
データセットの統計はテーブル1に示されています。
0.83
The dataset consists of three subsets, the training set with about 20k instances which the parameters of the model are trained with, the validation set (Dev) for tuning hyper parameters and choosing the best model, and the test set which is not used until the final evaluation.
The experts were given the set of input supporting sentences, the potential mechanism sentence, and the relation between main entities.
専門家は、入力支援文のセット、潜在的メカニズム文、主エンティティ間の関係を与えられた。
0.55
Our aim is two fold, first to evaluate the quality of the data collection process, second to collect a clean human evaluated dataset which can be used as an extra test set.
This evaluation shows that the generated dataset is of reasonable quality, and can serve as a meaningful resource for training models for mechanism summarization.
2. Effect of pretraining: What is the impact of using
2.プレトレーニングの効果:使用の影響について
0.81
the additional data via pretraining?
事前トレーニングによる追加データ?
0.82
3. Effect of modeling supporting sentences: What is the impact of selecting a subset of supporting sentences?
3.支援文のモデル化の効果:支援文のサブセットの選択が与える影響について
0.73
4. Error analysis: What are the main failure modes of
4. エラー分析: 主な障害モードは何か。
0.74
language generation models? 5.1.
言語生成モデル? 5.1.
0.55
Experimental Setup We use SuMe to benchmark language generation models and measure their ability to correctly identify the relation between the focus entities and to summarize the mechanism behind the relation based on the input sentences from the abstract.
Models: We compare pretrained GPT-2 (Radford et al , 2019), T5 (Raffel et al , 2020b), BART (Lewis et al , 2019) models and two domain-adapted models, GPT2Pubmed (Papanikolaou and Pierleoni, 2020), and SciFive (Phan et al , 2021), which were trained on scientific literature.
モデル: GPT-2 (Radford et al , 2019), T5 (Raffel et al , 2020b), BART (Lewis et al , 2019) モデルと2つのドメイン適応モデル, GPT2Pubmed (Papanikolaou and Pierleoni, 2020), SciFive (Phan et al , 2021) を比較した。 訳抜け防止モード: モデル : 事前訓練GPT-2(Radford et al, 2019)の比較 T5 (Raffel et al, 2020b ), BART (Lewis et al, 2019 ) モデル そして2つのドメイン対応モデル、GPT2Pubmed(Ppaanikol aouとPierleoni、2020年)。 そしてSciFive(Phan et al, 2021)は科学文献の教育を受けた。
0.85
Evaluation Metrics: We conduct both automatic and manual evaluation of the model outputs.
評価指標: モデル出力の自動評価と手動評価の両方を行う。
0.77
Relation Generation (RG): The models are supposed to first generate the relation type (positive or negative regulation) and then generate the mechanism that underlies this relation.
We evaluate the model’s output as we would for a corresponding classification task, i.e., the generated relation is deemed correct if it exactly matches the correct relation name.
We report F1 numbers for this binary classification task.
このバイナリ分類タスクのF1番号を報告する。
0.75
Mechanism Generation: We evaluate the quality of the generated explanations using two language generation
メカニズム生成:2つの言語生成を用いて生成した説明の質を評価する
0.81
英語（論文から抽出）
日本語訳
スコア
Model BART GPT2
Model BART GPT2
0.47
T5 GPT2-Pubmed
T5 GPT2-Pubmed
0.31
SciFive RG (F1) BLEURT Rouge-1 Rouge-2 Rouge-L
はさみ RG (F1) BLEURT Rouge-1 Rouge-2 Rouge-L
0.32
76 74 72 78 79
76 74 72 78 79
0.43
42.49 44.19 44.41 46.33 47.81
42.49 44.19 44.41 46.33 47.81
0.22
46.54 46.54 48.26 48.37 52.10
46.54 46.54 48.26 48.37 52.10
0.22
25.92 28.32 27.63 29.55 32.62
25.92 28.32 27.63 29.55 32.62
0.22
35.34 38.78 38.77 40.19 43.31
35.34 38.78 38.77 40.19 43.31
0.22
Table 3: Benchmarking performance of strong language generation models and some domain-adapted models.
表3: 強い言語生成モデルといくつかのドメイン適応モデルのベンチマークパフォーマンス。
0.78
We present standard automatic evaluations measures for the mechanism sentence generation task along with F1 for the generated relations.
本稿では,生成関係のf1とともに,機構文生成タスクのための標準自動評価手法を提案する。
0.64
The science domain versions of both GPT2 and T5 work better than the original versions.
GPT2とT5の科学領域バージョンは、オリジナルのバージョンよりもうまく動作する。
0.81
metrics: the widely-used ROUGE (Lin, 2004) scores that rely on lexical overlap, and BLEURT scores (Sellam et al , 2020) which aim to capture semantic similarity between the generated and the gold reference.
メトリクス:広く使用されているROUGE(Lin, 2004)とBLEURTスコア(Sellam et al , 2020)は、生成されたものと金の参照とのセマンティックな類似性を捉えることを目的としている。
0.72
We use a recent version, the BLEURT-20 model that has been shown to be more effective (Pu et al , 2021) .
我々は、より効果的であることが示されているBLEURT-20モデル(Pu et al , 2021)を用いている。
0.81
We compare the generated text as the hypothesis against the actual text as the reference.
生成されたテキストを、実際のテキストを参照とする仮説と比較する。
0.72
Fine-tuning and Training Details: All models are original base models published by HuggingFace that were fine-tuned on the training portion of SuMe for 20 epochs.
For each model, we evaluate the average of BLEURT and Rouge-L score on the validation (Dev) set and the one with the highest average is chosen for prediction.
The learning rate is set to 6e-5, we use AdamW (Loshchilov and Hutter, 2017) optimizer with = 1e − 8.
学習速度は 6e-5 に設定され、adamw (loshchilov and hutter, 2017)オプティマイザは 1e − 8 である。
0.74
The input token is limited to 512 tokens, and the generated token is maxed out at 128.
入力トークンは512トークンに制限され、生成されたトークンは128で最大化される。
0.77
We select batch size of 8 with gradient accumulation steps of two.
8のバッチサイズを2の勾配累積ステップで選択する。
0.61
5.2. Automatic Evaluation Results Table 3 compares the performance of the five language generation models on both the relation generation (RG) and mechanism generation tasks.
The domain-adapted models, GPT2-Pubmed and SciFive, fare better than fine-tuning the standard pre-trained models for both relation and mechanism generation tasks.
SciFive achieves the best performance with more than a 7.5% increase in BLEURT score and more than 9.7% increase in RG F1 over the standard T5 model, highlighting the importance of domain adaptation for the SuMe tasks defined over scientific literature.
The models achieve better performance on the relation generation task but there is still a substantial room for improvement here with the best model achieving an F1 of 79.
If the model is unable to generate the relation correctly, then the mechanism it generates is not useful.
モデルが関係を正しく生成できない場合、それが生成するメカニズムは役に立たない。
0.72
Ideally we want models to correctly generate both the relation and the mechanism that underlies it.
理想的には、モデルが関係とそれを支えるメカニズムの両方を正しく生成することを望んでいます。
0.65
We also evaluated the correlation between BLEURT score and relation generation classification score.
また,BLEURTスコアと関係生成分類スコアの相関について検討した。
0.71
Our analysis shows that when the model generates an accurate relation, it gets higher BLEURT score while when it generates an incorrect relation, its gets a 10% lower
Figure 4: Comparison of relation generation (RG) F1 (left y-axis/blue bars) and the mechanism generation measures (right y-axis/teal+blue curves) against the amount of pretraining.
As we increase the size of the pretraining data, the model performance improves.
事前トレーニングデータのサイズが大きくなると、モデルのパフォーマンスが向上します。
0.73
BLEURT score (50.02 vs 45.08)
BLEURTスコア(50.02対45.08)
0.66
5.3. Pretraining with Conclusion Generation Next we analyze the impact of pre-training the models on the related task of generating conclusion (instead of mechanism) sentences, for which we can obtain data at scale without any manual labeling effort.
SuMe includes 611K instances of this kind which is an order of magnitude larger than the mechanism summarization instances and can be seen as a form of data augmentation.
Pretraining Data Size: We pretrain the SciFive model on the conclusion generation task with increasing amount of data (100K increments), and measure the performance of finetuning the pretrained models on the mechanism summarization task.
Figure 4 shows that performance increases with more data available for pre-
図4は、事前のデータによってパフォーマンスが向上していることを示している。
0.51
Model Performance Vs Size of Pretraining DataBLEURT/ROUGE Score4043464952F1 Score7879808182Size of Pretraining Data0100k200k300k400 k500k600kRG (F1)BLEURT Rouge-L
プレトレーニングDataBLEURT/ROUGE Score4043464952F1 Score7879808182Size of PretrainingData0100k 200k200k300k500k500k 600kRG (F1)BLEURT Rouge-L
0.49
英語（論文から抽出）
日本語訳
スコア
Supporting Set BLEURT Rouge-L
サポートセット ブルールトルージュl
0.62
SciFive +Oracle
SciFive + Oracle
0.42
+Pretraining +Pretraining+Oracle
+予習 +プリトレーニング+oracle
0.33
47.81 49 49.05 49.64
47.81 49 49.05 49.64
0.32
43.31 43.07 43.72 43.81
43.31 43.07 43.72 43.81
0.23
Figure 5: Number of pretraining epochs vs. fine-tuning epochs for each pretrained model until convergence.
図5: 事前トレーニングエポック数 収束するまで各事前トレーニングモデル毎の微調整エポック数。
0.71
training, suggesting that pretraining is beneficial for learning to generate mechanisms.
事前訓練は、学習がメカニズムを生成するのに有益であることを示す訓練。
0.60
Number of Epochs: We also compare the impact of the amount of pretraining on the number of epochs needed for convergence in fine-tuning.
Epochs の数: 微調整における収束に必要なエポックの数に対する事前学習の量の影響も比較する。
0.73
Figure 5 compares pretrained models with different number of pretraining epochs (x-axis) in terms of their overall effectiveness (BLEURT score bars) and the number of epochs to convergence (Finetuning epochs curve).
The figure shows that when we continue pretraining, not only does the resulting model performs better, but it also converges sooner taking fewer number of epochs to reach higher effectiveness.
This kind of an extractive step has been used previously in summarization tasks to reduce the amount of irrelevant information in the input (Narayan et al , 2018; Liu and Lapata, 2019).
この種の抽出ステップは、これまで、入力における無関係な情報の量を減らすために、要約タスクで用いられてきた(Narayan et al , 2018; Liu and Lapata, 2019)。
0.77
To understand the utility of this, we built a pseudo-oracle that finds the sentences that have the best overlap (measured via BLEURT score (Sellam et al , 2020)) with the output mechanism sentence.
本手法の有用性を理解するために, BLEURTスコア(Sellam et al , 2020)と出力機構文とを最も重なり合う文(BLEURTスコア(Sellam et al , 2020)を抽出する擬似オークルを構築した。
0.63
Then, we trained the SciFive model and pretrained version to only use the top few sentences according to BLEURT score such that input size is now half of the original input size.
Using this subset instead of the entire subset provides BLEURT score improvements only for the basic SciFive model and the gains reduce when we use the pretrained model.
Unlike standard summarization tasks there are fewer completely unrelated sentences in the abstracts and generating the mechanism sentences remains challenging even when we are able to identify the most relevant sentences within this set.
Table 4: The effect of selecting supporting sentences with highest BLEURT score.
表4: BLEURTスコアの高いサポート文を選択する効果。
0.65
5.5. Manual Evaluation We also conduct a manual evaluation of the outputs from the best model — the SciFive model that was pretrained with the conclusion generation task.
We asked three biomedical experts to evaluate output sentences for 100 instances and answer three questions (It took ∼ 5 minutes per expert per instance): 1.
This again highlights the significant challenge posed by this task.
これは、このタスクによってもたらされる重要な課題を再び強調する。
0.49
5.6. Error Analysis To understand the frequent failure modes of the model, we manually categorized the errors in 100 outputs that had the worst BLEURT scores with the reference mechanism sentences.
Despite this being a necessary feature in all of the mechanism sentences in the training data, the prevalence of this error shows that models find it difficult to track the main entities during generation.
2. Incorrect Mechanism (24%) – The model is unable to generate the correct mechanism even though it is able to identify the correct relation and fills in some information that is either unrelated to or unsupported by the input sentences.
コンバージェンス vs # プリトレーニングepochsbleurt score4747.648.248.84 9.450# 微調整epochs02468101214161 820# プリトレーニングepochs0123456789 微調整epochsbleurtスコア
0.45
英語（論文から抽出）
日本語訳
スコア
Gold Mechanism On the basis of these observations, we conclude that IL18 induces MCP-1 production through the PI3K/Akt and MEK/ERK1/2 pathways in macrophages.
Taken together, we conclude that DeltaNp73 negatively regulates NGF-mediated neuronal differentiation by transrepressing TrkA.
DeltaNp73はTrkAを転写することでNGFを介する神経分化を負に制御する。
0.59
In conclusion, the increase in SNGFR caused by ANF is associated with an increase in glomerular capillary hydraulic pressure and with a blunted maximal tubuloglomerular feedback response.
We conclude that, without modulatory factors which play a role in vivo, NGF can enhance the synthesis of tyrosine hydroxylase n sympathetic ganglia in vitro, provided organ culture conditions which permit optimal survival of adrenergic neurons are selected.
In vivoでの役割を担う調節因子がなければ、NGFはチロシン水酸化酵素n交感神経節の合成を促進することができ、アドレナリンニューロンの最適な生存を可能にする臓器培養条件が選択される。
0.76
We conclude that 20 mm alcohol/submaximal CCK as well supramaximal CCK stimulation can trigger pathologic basolateral exocytosis in pancreatic acinar cells via PKC alphamediated activation of Munc18c , which enables Syntaxin-4 to become receptive in forming a SNARE complex in the BPM.
We conclude that in the presence of high doses of insulin, FSH decreases aromatase activity, and an uncoupling of P450 aromatase mRNA and aromatase activity occurs.
In conclusion, our results indicate that DeltaNp73 negatively regulates NGF-mediated neuronal differentiation by transcriptionally repressing the expression of TrkA.
This conclusion was further supported by pulselabeling of tyrosine hydroxylase with [3H]leucine, which showed that NGF increased synthesis of tyrosine in sympathetic ganglia by approximately 50%.
We conclude that alcohol can induce a clinically relevant form of pancreatitis by blocking apical exocytosis and redirecting exocytosis to less efficient BPM, mimicking supramaximal CCK stimulation.
In conclusion, insulin stimulates aromatase activity in bovine granulosa cells at low doses but fails to stimulate activity at higher doses of insulin.
Table 5: Examples of the generated outputs by the model.
表5: モデルによって生成された出力の例。
0.87
The first three are good outputs where the mechanism is a simple paraphrase of the expected gold mechanism, while the next three illustrate the types of semantic errors we observe.
The model generates only a part of such mechanisms.
モデルはそのようなメカニズムの一部だけを生成する。
0.81
5.7. Word Analysis We further analyzed the unigrams of the supporting sentences corresponding to the instances where the model was most confident in its generated mechanism and where it was least confident.
The analysis shows that when the words ’binding’, ’caused’, ’demonstrated’, ’dose dependent’, ’investigated’, ’result’, and ’performed’ are available in the supporting sentences the model can generate explanation sentences with higher quality.
PRESSING. In the last three examples with incorrect information, the first shows a bad output which contains a mechanism but not of the relation connecting the main entities.
The last one is an example one of the entities are missing (FSH) and the generated text is about another relation.
最後のものはエンティティの欠落(FSH)の例であり、生成されたテキストは別の関係に関するものです。
0.76
6. Conclusions We introduced SuMe, a dataset for biomedical mechanism summarization.
6.結論 バイオメディカルメカニズム要約のためのデータセットSuMeを導入した。
0.69
This dataset is coupled with a challenging summarization task, which requires the generation of the relation between main entities as well as a textual summary of the mechanism which explains the reason behind the underlying relation.
The expert analysis also suggests the difficulty and importance of the task.
専門家分析は、タスクの難しさと重要性も示唆している。
0.65
All in all, we believe that SuMe dataset and associated task are a useful step towards building true informationaccess applications for the biomedical literature.
Acknowledgments This work was supported in part by the National Science Foundation under grants IIS-1815358 and IIS-1815948.
この研究は国立科学財団によってIIS-1815358とIIS-1815948の助成を受けた。
0.75
7. Bibliographical References Alam, F., Joty, S., and Imran, M. (2018).
書誌的参考文献 Alam, F., Joty, S. and Imran, M. (2018)。
0.45
Domain adaptation with adversarial training and graph embeddings.
敵のトレーニングとグラフ埋め込みによるドメイン適応。
0.74
arXiv preprint arXiv:1805.05151.
arXiv preprint arXiv:1805.05151
0.36
Arighi, C. N., Lu, Z., Krallinger, M., Cohen, K. B., Wilbur, W. J., Valencia, A., Hirschman, L., and Wu, C. H. (2011).
Arighi, C. N., Lu, Z., Krallinger, M., Cohen, K. B., Wilbur, W. J., Valencia, A., Hirschman, L., Wu, C. H. (2011)。
0.91
Overview of the biocreative iii workshop.
バイオクリエーションiiiワークショップの概要
0.53
BMC bioinformatics, 12(8):1–9.
BMCバイオインフォマティクス 12(8):1-9。
0.71
Azadani, M. N., Ghadiri, N., and Davoodijam, E. (2018).
Azadani, M. N., Ghadiri, N., Davoodijam, E. (2018)。
0.41
Graph-based biomedical text summarization: An itemset mining and sentence clustering approach.
グラフに基づく生物医学的テキスト要約:アイテムセットマイニングと文クラスタリングアプローチ
0.80
Journal of biomedical informatics, 84:42–58.
journal of biomedical informatics, 84:42-58 を参照。
0.58
Bastan, M., Koupaee, M., Son, Y., Sicoli, R., and Balasubramanian, N. (2020).
Bastan, M., Koupaee, M., Son, Y., Sicoli, R., Balasubramanian, N. (2020)。
0.79
Author’s sentiment prediction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 604–615, Barcelona, Spain (Online), December.
International Committee on Computational Linguistics.
計算言語学国際委員会委員。
0.69
Beltagy, I., Lo, K., and Cohan, A. (2019).
Beltagy, I., Lo, K. and Cohan, A. (2019)。
0.82
Scibert: Pretrained language model for scientific text.
Scibert: 科学テキストのための事前訓練された言語モデル。
0.67
In EMNLP. Cachola, I., Lo, K., Cohan, A., and Weld, D. S. (2020).
略称はemnlp。 Cachola, I., Lo, K., Cohan, A., and Weld, D. S. (2020)。
0.64
Tldr: Extreme summarization of scientific documents.
Tldr: 科学的文書の極端な要約。
0.78
arXiv preprint arXiv:2004.15011.
arXiv preprint arXiv:2004.15011
0.36
C¸ ano, E. and Bojar, O. (2020).
ano, E. and Bojar, O. (2020)。
0.78
Two huge title and keyword generation corpora of research articles.
2つの巨大なタイトルとキーワード生成コーパスの研究記事。
0.72
arXiv preprint arXiv:2002.04689.
arXiv preprint arXiv:2002.04689
0.36
Cohan, A., Dernoncourt, F., Kim, D. S., Bui, T., Kim, S., Chang, W., and Goharian, N. (2018).
Cohan, A., Dernoncourt, F., Kim, D. S., Bui, T., Kim, S., Chang, W., and Goharian, N. (2018)。 訳抜け防止モード: Cohan, A., Dernoncourt, F., Kim, D. S. Bui, T., Kim, S., Chang, W. とGoharian氏(2018年)。
0.80
A discourseaware attention model for abstractive summarization of long documents.
長文の抽象的要約のための難聴注意モデル
0.52
CoRR, abs/1804.05685.
CoRR, abs/1804.05685。
0.30
Collins, E., Augenstein, I., and Riedel, S. (2017).
Collins, E., Augenstein, I. and Riedel, S. (2017)。
0.44
A supervised approach to extractive summarisation of scientific papers.
科学的論文の抽出要約のための教師付きアプローチ
0.66
arXiv preprint arXiv:1706.03946.
arXiv preprint arXiv:1706.03946
0.36
Czarnecki, J., Nobeli, I., Smith, A. M., and Shepherd, A. J. (2012).
Czarnecki, J., Nobeli, I., Smith, A. M., and Shepherd, A. J. (2012)。
0.91
A text-mining system for extracting metabolic reactions from full-text articles.
全文記事から代謝反応を抽出するテキストマイニングシステム
0.65
BMC bioinformatics, 13(1):1–14.
BMCバイオインフォマティクス 13(1):1-14。
0.71
Dina Demner-Fushman, et al , editors.
Dina Demner-Fushman, et al , editors.
0.48
(2020). Proceedings of the 19th SIGBioMed Workshop on Biomedical Language Processing, Online, July.
(2020). 第19回SIGBioMed Workshop on Biomedical Language Processing, July に参加して
0.59
Association for Computational Linguistics. DeYoung, J., Lehman, E., Nye, B., Marshall, I. J., (2020).
計算言語学会会員。 DeYoung, J., Lehman, E., Nye, B., Marshall, I. J., (2020)。
0.70
Evidence inference and Wallace, B. C. 2.0: More data, better models.
Evidence Inference and Wallace, B. C. 2.0: より多くのデータ、より良いモデル。
0.78
arXiv preprint arXiv:2005.04177.
arXiv preprint arXiv:2005.04177。
0.63
Gaonkar, R., Kwon, H., Bastan, M., Balasubramanian, N., and Chambers, N. (2020).
Gaonkar, R., Kwon, H., Bastan, M., Balasubramanian, N., and Chambers, N. (2020)。
0.85
Modeling label semantics for predicting emotional reactions.
感情反応予測のためのラベルセマンティクスのモデル化
0.71
In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4687–4692, Online, July.
第58回計算言語学会年次大会紀要4687-4692頁、オンライン, 7月。
0.50
Association for Computational Linguistics. Giorgi, J. M. and Bader, G. D. (2020).
計算言語学会会員。 Giorgi, J. M. and Bader, G. D. (2020)。
0.48
Towards reliable named entity recognition in the biomedical domain.
生物医学領域における信頼性の高い実体認識を目指して
0.58
Bioinformatics, 36(1):280–286.
バイオインフォマティクス 36(1):280-286。
0.66
Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., and Poon, H. (2020).
Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H. (2020)。
0.42
Domain-specific language model pretraining for biomedical natural language processing.
バイオメディカル自然言語処理のためのドメイン固有言語モデル
0.83
Heidari, M., Zad, S., Hajibabaee, P., Malekzadeh, M., HekmatiAthar, S., Uzuner, O., and Jones, J. H. (2021).
Heidari, M., Zad, S., Hajibabaee, P., Malekzadeh, M., HekmatiAthar, S., Uzuner, O., Jones, J. H. (2021)。
0.84
Bert model for fake news detection based on social bot activities in the covid-19 pandemic.
In 2021 IEEE 12th Annual Ubiquitous Computing, Electronics Mobile Communication Conference (UEMCON), pages 0103–0109.
2021年、IEEE 12th Annual Ubiquitous Computing, Electronics Mobile Communication Conference (UEMCON)、ページ0103-0109。
0.86
Kanakarajan, K. r.
カナカラジャン k. r.
0.58
, Kundumani, B., and Sankarasubbu, M. (2021).
出典: kundumani, b., sankarasubbu, m. (2021)。
0.74
BioELECTRA:pretraine d biomedical text encoder using discriminators.
BioELECTRA:識別器を用いたバイオメディカルテキストエンコーダ
0.59
In Proceedings of the 20th Workshop on Biomedical Language Processing, pages 143–154, Online, June.
第20回バイオメディカル言語処理ワークショップ"Proceedings of the 20th Workshop on Biomedical Language Processing, page 143-154, June
0.66
Association for Computational Linguistics. Kemper, B., Matsuzaki, T., Matsuoka, Y., Tsuruoka, Y., Kitano, H., Ananiadou, S., and Tsujii, J. (2010).
計算言語学会会員。 Kemper, B., Matsuzaki, T., Matsuoka, Y., Tsuruoka, Y., Kitano, H., Ananiadou, S., and Tsujii, J. (2010)。 訳抜け防止モード: 計算言語学会会員。 Kemper, B., Matsuzaki, T., Matsuoka, Y. 鶴岡, Y., 北野, H., Ananiadou, S. そして、辻井(2010年)。
0.62
Pathtext: a text mining integrator for biological pathway visualizations.
Pathtext: 生物学的経路可視化のためのテキストマイニングインテグレータ。
0.71
Bioinformatics, 26(12):i374– i381.
バイオインフォマティクス26(12):i374-i381。
0.68
Keymanesh, M., Elsner, M., and Parthasarathy, S. (2021).
Keymanesh, M., Elsner, M., Parthasarathy, S. (2021)。
0.80
Privacy policy question answering assistant: A query-guided extractive summarization approach.
プライバシポリシ質問応答アシスタント:クエリ誘導抽出要約アプローチ
0.55
CoRR, abs/2109.14638.
corr, abs/2109.14638。
0.49
Krallinger, M., P´erez-P´erez, M., P´erez-Rodr´ıguez, G., Blanco-M´ıguez, A., Fdez-Riverola, F., CapellaGutierrez, S., Lourenc¸o, A., and Valencia, A. (2017).
Krallinger, M., P ́erez-P ́erez, M., P ́erez-Rodr ́ıguez, G., Blanco-M ́ıguez, A., Fdez-Riverola, F., CapellaGutierrez, S., Lourenc 'o, A., Valencia, A. (2017)。
0.76
The biocreative v. 5 evaluation workshop: tasks, organization, sessions and topics.
バイオ創造性v.5評価ワークショップ:タスク、組織、セッション、トピック。
0.67
Landis, J. R. and Koch, G. G. (1977).
Landis, J. R. and Koch, G. G. (1977)。
0.89
The measurement of observer agreement for categorical data.
分類データのオブザーバ合意度の測定
0.60
biometrics, pages 159–174.
バイオメトリックス、159-174頁。
0.53
Lauriola, I., Aiolli, F., Lavelli, A., and Rinaldi, F. (2021).
Lauriola, I., Aiolli, F., Lavelli, A., Rinaldi, F. (2021)。
0.80
Learning adaptive representations for entity recognition in the biomedical domain.
生物医学領域におけるエンティティ認識のための適応表現の学習
0.72
Journal of biomedical semantics, 12(1):1–13.
journal of biomedical semantics, 12(1):1–13を参照。
0.76
Lee, J., Yoon, W., Kim, S., Kim, D., Kim, S., So, C. H., and Kang, J. (2020).
Lee, J., Yoon, W., Kim, S., Kim, D., Kim, S., So, C. H., Kang, J. (2020)。
0.83
Biobert: a pre-trained biomedical language representation model for biomedical text mining.
Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., and Zettlemoyer, L. (2019).
Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L. (2019)。
0.41
Bart: Denoising sequence-tosequence pre-training 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列の事前学習。
0.70
arXiv preprint arXiv:1910.13461.
arXiv preprint arXiv:1910.13461
0.36
Lin, C. -Y. (2004).
リン、C。 -y。 (2004).
0.48
Rouge: A package for automatic In Text summarization
Rouge: 自動インテキスト要約のためのパッケージ
0.86
evaluation of summaries. branches out, pages 74–81.
要約の評価。 74-81頁。
0.49
Liu, Y. and Lapata, M.
Liu, Y. and Lapata, M。
0.46
(2019). Text summa-
(2019). テキスト要約
0.54
英語（論文から抽出）
日本語訳
スコア
standards. In Nicoletta Calzolari (Conference Chair), et al , editors, Proceedings of the Eight International Conference on Language Resources and Evaluation (LREC’12), pages 3746–3753, Istanbul, Turkey, may.
基準だ Nicoletta Calzolari (Conference Chair), et al , editors, Proceedings of the Eight International Conference on Language Resources and Evaluation (LREC’12), page 3746–3753, Istanbul, Turkey, may。 訳抜け防止モード: 基準だ nicoletta calzolari (カンファレンスチェア) など、編集者。 第8回言語資源評価国際会議(lrec'12)参加報告 3746-3753頁、イスタンブール、トルコ、5月。
0.59
European Language Resource Association (ELRA).
欧州言語資源協会 (ELRA) の略。
0.79
Subramani, S., Kalpana, R., Monickaraj, P. M., and Natarajan, J. (2015).
Subramani, S., Kalpana, R., Monickaraj, P. M., Natarajan, J. (2015)。
0.80
Hpiminer: A text mining system for building and visualizing human protein interaction networks and pathways.
Thayaparan, M., Valentino, M., and Freitas, A. (2020).
Tayaparan, M., Valentino, M. and Freitas, A. (2020)。
0.76
A survey on explainability in machine reading comprehension.
機械読解における説明可能性に関する調査
0.72
arXiv preprint arXiv:2010.00389.
arXiv preprint arXiv:2010.00389
0.36
Valenzuela-Esc´arcega, M. A., Babur, ¨O.
Valenzuela-Esc ́arcega, M. A。
0.78
, Hahn-Powell, G., Bell, D., Hicks, T., Noriega-Atala, E., Wang, X., Surdeanu, M., Demir, E., and Morrison, C. T. (2018).
Hahn-Powell, G., Bell, D., Hicks, T., Noriega-Atala, E., Wang, X., Surdeanu, M., Demir, E., Morrison, C. T. (2018)。 訳抜け防止モード: 、Hahn - Powell, G., Bell, D., Hicks, T,ノリエガ-アタラ、E、王、X Surdeanu, M., Demir, E. and Morrison, C. T. (2018)。
0.84
Large-scale automated machine reading discovers new cancer driving mechanisms.
大規模自動機械読み取りは、新しいがん駆動機構を発見する。
0.63
Database: The Journal of Biological Databases and Curation.
データベース:journal of biological database and curation。
0.68
Wang, Q., Huang, L., Jiang, Z., Knight, K., Ji, H., Bansal, M., and Luan, Y. (2019).
Wang, Q., Huang, L., Jiang, Z., Knight, K., Ji, H., Bansal, M., and Luan, Y. (2019) 訳抜け防止モード: Wang, Q., Huang, L., Jiang, Z. Knight, K., Ji, H., Bansal, M. とLuan氏(2019年)。
0.76
Paperrobot: Incremental draft generation of scientific ideas.
Paperrobot: 科学的アイデアのインクリメンタルドラフト生成。
0.71
arXiv preprint arXiv:1905.07870.
arXiv preprint arXiv:1905.07870。
0.32
Yao, L., Riedel, S., and McCallum, A. (2010).
Yao, L., Riedel, S. and McCallum, A. (2010)。
0.41
Collective cross-document relation extraction without labelled data.
ラベル付きデータのない集合文書間関係抽出
0.74
In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1013–1023.
2010年自然言語処理における経験的手法に関する会議の議題1013-1023頁。
0.74
Yasunaga, M., Kasai, J., Zhang, R., Fabbri, A. R., Li, I., Friedman, D., and Radev, D. R. (2019).
Yasunaga, M., Kasai, J., Zhang, R., Fabbri, A. R., Li, I., Friedman, D., Radev, D. R. (2019)。
0.88
Scisummnet: A large annotated corpus and content-impact models for scientific paper summarization with citation networks.
Scisummnet: 科学論文要約のための大規模な注釈付きコーパスとコンテンツインパクトモデル。
0.75
In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 7386–7393.
The Proceedings of the AAAI Conference on Artificial Intelligence, Volume 33, page 7386–7393。
0.43
Yuan, J., Jin, Z., Guo, H., Jin, H., Zhang, X., Smith, T., and Luo, J. (2020).
Yuan, J., Jin, Z., Guo, H., Jin, H., Zhang, X., Smith, T., and Luo, J. (2020)。
0.83
Constructing biomedical domainspecific knowledge graph with minimum supervision.
最小限の監督によるバイオメディカルドメイン固有知識グラフの構築
0.60
Knowledge and Information Systems, 62(1):317–336.
知識・情報システム 62(1):317-336。
0.76
Zhao, L., Wang, J., Cheng, L., and Wang, C. (2020).
Zhao, L., Wang, J., Cheng, L., Wang, C. (2020)。
0.81
Ontosem: an ontology semantic representation methodology for biomedical domain.
ontosem: 生物医学領域のためのオントロジー意味表現方法論。
0.68
In 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 523–527.
2020年、IEEE International Conference on Bioinformatics and Biomedicine (BIBM) 523-527頁。
0.81
IEEE. rization with pretrained encoders.
IEEE。 プリトレーニングエンコーダによるrization。
0.52
arXiv preprint arXiv:1908.08345.
arXiv preprint arXiv:1908.08345
0.36
Loshchilov, I. and Hutter, F. (2017).
loshchilov, i. and hutter, f. (2017)。
0.89
Fixing weight decay regularization in adam.
adam における重み減衰正規化の固定
0.65
CoRR, abs/1711.05101.
corr、abs/1711.05101。
0.38
Miwa, M., Ohta, T., Rak, R., Rowley, A., Kell, D. B., Pyysalo, S., and Ananiadou, S. (2013).
Miwa, M., Ohta, T., Rak, R., Rowley, A., Kell, D. B., Pysalo, S., and Ananiadou, S. (2013)。
0.85
A method for integrating and ranking the evidence for biochemical pathways by mining reactions from text.
テキストからの鉱毒反応による生化学的経路の証拠の統合とランキングの方法。
0.69
Bioinformatics, 29(13):i44–i52.
バイオインフォマティクス29(13):i44-i52。
0.64
Mulyar, A., Uzuner, O., and McInnes, B. (2021).
Mulyar, A., Uzuner, O. and McInnes, B. (2021)。
0.82
Mtclinical bert: scaling clinical information extraction with multitask learning.
Mtclinical bert: マルチタスク学習による臨床情報抽出のスケーリング。
0.88
Journal of the American Medical Informatics Association, 28(10):2108–2115.
Journal of the American Medical Informatics Association, 28(10):2108–2115
0.45
Narayan, S., Cohen, S. B., and Lapata, M. (2018).
Narayan, S., Cohen, S. B. and Lapata, M. (2018)。
0.47
Don’t give me the details, just the summary! topicaware convolutional neural networks for extreme summarization.
極端に要約するためのトピックを意識した畳み込みニューラルネットワーク。
0.39
arXiv preprint arXiv:1808.08745.
arXiv preprint arXiv:1808.08745
0.36
Papanikolaou, Y. and Pierleoni, A. (2020).
Papanikolaou, Y. and Pierleoni, A. (2020)。
0.43
Dare: Data augmented relation extraction with gpt-2.
Dare: gpt-2によるデータ拡張関係抽出。
0.73
arXiv preprint arXiv:2004.13845.
arXiv preprint arXiv:2004.13845
0.36
Peng, Y., Yan, S., and Lu, Z. (2019).
Peng, Y., Yan, S. and Lu, Z. (2019)。
0.82
Transfer learning in biomedical natural language processing: an evaluation of bert and elmo on ten benchmarking datasets.