F1 Points Research Articles

Boosting multi-document summarization with hierarchical graph convolutional networks

The input of the multi-document summarization task is usually long and has high redundancy. Encoding multiple documents is a challenge for the Seq2Seq architecture. The way of concatenating multiple documents into a sequence ignores the relation between documents. Attention-based Seq2Seq architectures have slightly improved the cross-document relation modeling for multi-document summarization. However, these methods ignore the relation between sentences, and there is little improvement that can be achieved through the attention mechanism alone. This paper proposes a hierarchical approach to leveraging the relation between words, sentences, and documents for abstractive multi-document summarization. Our model employs the Graph Convolutional Networks (GCN) for capturing the cross-document and cross-sentence relations. The GCN module can enrich semantic representations by generating high-level hidden features. Our model achieves significant improvement over the attention-based baseline, beating the Hierarchical Transformer by 3.4/1.64, 1.92/1.44 ROUGE-1/2 F1 points on the Multi-News and WikiSum datasets, respectively. Experimental results demonstrate that our delivered method brings substantial improvements over several strong baselines.

Propagating variational model uncertainty for bioacoustic call label smoothing

Along with propagating the input toward making a prediction, Bayesian neural networks also propagate uncertainty. This has the potential to guide the training process by rejecting predictions of low confidence, and recent variational Bayesian methods can do so without Monte Carlo sampling of weights. Here, we apply sample-free methods for wildlife call detection on recordings made via passive acoustic monitoring equipment in the animals' natural habitats. We further propose uncertainty-aware label smoothing, where the smoothing probability is dependent on sample-free predictive uncertainty, in order to downweigh data samples that should contribute less to the loss value. We introduce a bioacoustic dataset recorded in Malaysian Borneo, containing overlapping calls from 30 species. On that dataset, our proposed method achieves an absolute percentage improvement of around 1.5 points on area under the receiver operating characteristic (AU-ROC), 13 points in F1, and 19.5 points in expected calibration error (ECE) compared to the point-estimate network baseline averaged across all target classes.

KC-GEE: knowledge-based conditioning for generative event extraction

AbstractEvent extraction is an important, but challenging task. Many existing techniques decompose it into event and argument detection/classification subtasks, which are complex structured prediction problems. Generation-based extraction techniques lessen the complexity of the problem formulation and are able to leverage the reasoning capabilities of large pretrained language models. However, they still suffer from poor zero-shot generalizability and are ineffective in handling long contexts such as documents. We propose a generative event extraction model, KC-GEE, that addresses these limitations. A key contribution of KC-GEE is a novel knowledge-based conditioning technique that injects the schema of candidate event types as the prefix into each layer of an encoder-decoder language model. This enables effective zero-shot learning and improves supervised learning. Our experiments on two benchmark datasets demonstrate the strong performance of our KC-GEE model. It achieves particularly strong results in the challenging document-level extraction task and in the zero-shot learning setting, outperforming state-of-the-art models by up to 5.4 absolute F1 points.

Tumor detection under cystoscopy with transformer-augmented deep learning algorithm

Objective. Accurate tumor detection is critical in cystoscopy to improve bladder cancer resection and decrease recurrence. Advanced deep learning algorithms hold the potential to improve the performance of standard white-light cystoscopy (WLC) in a noninvasive and cost-effective fashion. The purpose of this work is to develop a cost-effective, transformer-augmented deep learning algorithm for accurate detection of bladder tumors in WLC and to assess its performance on archived patient data. Approach. ‘CystoNet-T’, a deep learning-based bladder tumor detector, was developed with a transformer-augmented pyramidal CNN architecture to improve automated tumor detection of WLC. CystoNet-T incorporated the self-attention mechanism by attaching transformer encoder modules to the pyramidal layers of the feature pyramid network (FPN), and obtained multi-scale activation maps with global features aggregation. Features resulting from context augmentation served as the input to a region-based detector to produce tumor detection predictions. The training set was constructed by 510 WLC frames that were obtained from cystoscopy video sequences acquired from 54 patients. The test set was constructed based on 101 images obtained from WLC sequences of 13 patients. Main results. CystoNet-T was evaluated on the test set with 96.4 F1 and 91.4 AP (Average Precision). This result improved the benchmark of Faster R-CNN and YOLO by 7.3 points in F1 and 3.8 points in AP. The improvement is attributed to the strong ability of global attention of CystoNet-T and better feature learning of the pyramids architecture throughout the training. The model was found to be particularly effective in highlighting the foreground information for precise localization of the true positives while favorably avoiding false alarms Significance. We have developed a deep learning algorithm that accurately detects bladder tumors in WLC. Transformer-augmented AI framework promises to aid in clinical decision-making for improved bladder cancer diagnosis and therapeutic guidance.

VersaMatch: Ontology Matching with Weak Supervision

Ontology matching is crucial to data integration for across-silo data sharing and has been mainly addressed with heuristic and machine learning (ML) methods. While heuristic methods are often inflexible and hard to extend to new domains, ML methods rely on substantial and hard to obtain amounts of labeled training data. To overcome these limitations, we propose VersaMatch , a flexible, weakly-supervised ontology matching system. VersaMatch employs various weak supervision sources, such as heuristic rules, pattern matching, and external knowledge bases, to produce labels from a large amount of unlabeled data for training a discriminative ML model. For prediction, VersaMatch develops a novel ensemble model combining the weak supervision sources with the discriminative model to support generalization while retaining a high precision. Our ensemble method boosts end model performance by 4 points compared to a traditional weak-supervision baseline. In addition, compared to state-of-the-art ontology matchers, VersaMatch achieves an overall 4-point performance improvement in F1 score across 26 ontology combinations from different domains. For recently released, in-the-wild datasets, VersaMatch beats the next best matchers by 9 points in F1. Furthermore, its core weak-supervision logic can easily be improved by adding more knowledge sources and collecting more unlabeled data for training.

Generating EDU Extracts for Plan-Guided Summary Re-Ranking.

Two-step approaches, in which summary candidates are generated-then-reranked to return a single summary, can improve ROUGE scores over the standard single-step approach. Yet, standard decoding methods (i.e., beam search, nucleus sampling, and diverse beam search) produce candidates with redundant, and often low quality, content. In this paper, we design a novel method to generate candidates for re-ranking that addresses these issues. We ground each candidate abstract on its own unique content plan and generate distinct plan-guided abstracts using a model's top beam. More concretely, a standard language model (a BART LM) auto-regressively generates elemental discourse unit (EDU) content plans with an extractive copy mechanism. The top beams from the content plan generator are then used to guide a separate LM, which produces a single abstractive candidate for each distinct plan. We apply an existing re-ranker (BRIO) to abstractive candidates generated from our method, as well as baseline decoding methods. We show large relevance improvements over previously published methods on widely used single document news article corpora, with ROUGE-2 F1 gains of 0.88, 2.01, and 0.38 on CNN / Dailymail, NYT, and Xsum, respectively. A human evaluation on CNN / DM validates these results. Similarly, on 1k samples from CNN / DM, we show that prompting GPT-3 to follow EDU plans outperforms sampling-based methods by 1.05 ROUGE-2 F1 points. Code to generate and realize plans is available at https://github.com/griff4692/edu-sum.

An Exploratory Study on Utilising the Web of Linked Data for Product Data Mining

The Linked Open Data practice has led to a significant growth of structured data on the Web. While this has created an unprecedented opportunity for research in the field of Natural Language Processing, there is a lack of systematic studies on how such data can be used to support downstream NLP tasks. This work focuses on the e-commerce domain and explores how we can use such structured data to create language resources for product data mining tasks. To do so, we process billions of structured data points in the form of RDF n-quads, to create multi-million words of product-related corpora that are later used in three different ways for creating language resources: training word-embedding models, continued pre-training of BERT-like language models, and training machine translation models that are used as a proxy to generate product-related keywords. These language resources are then evaluated in three downstream tasks, product classification, linking, and fake review detection using an extensive set of benchmarks. Our results show word embeddings to be the most reliable and consistent method to improve the accuracy on all tasks (with up to 6.9% points in macro-average F1 on some datasets). Contrary to some earlier studies that suggest a rather simple but effective approach such as building domain-specific language models by pre-training using in-domain corpora, our work serves a lesson that adapting these methods to new domains may not be as easy as it seems. We further analyse our datasets and reflect on how our findings can inform future research and practice.

A Neighborhood Framework for Resource-Lean Content Flagging

Abstract We propose a novel framework for cross- lingual content flagging with limited target- language data, which significantly outperforms prior work in terms of predictive performance. The framework is based on a nearest-neighbor architecture. It is a modern instantiation of the vanilla k-nearest neighbor model, as we use Transformer representations in all its components. Our framework can adapt to new source- language instances, without the need to be retrained from scratch. Unlike prior work on neighborhood-based approaches, we encode the neighborhood information based on query– neighbor interactions. We propose two encoding schemes and we show their effectiveness using both qualitative and quantitative analysis. Our evaluation results on eight languages from two different datasets for abusive language detection show sizable improvements of up to 9.5 F1 points absolute (for Italian) over strong baselines. On average, we achieve 3.6 absolute F1 points of improvement for the three languages in the Jigsaw Multilingual dataset and 2.14 points for the WUL dataset.

CLIN-X: pre-trained language models and a study on cross-task transfer for concept extraction in the clinical domain.

The field of natural language processing (NLP) has recently seen a large change toward using pre-trained language models for solving almost any task. Despite showing great improvements in benchmark datasets for various tasks, these models often perform sub-optimal in non-standard domains like the clinical domain where a large gap between pre-training documents and target documents is observed. In this article, we aim at closing this gap with domain-specific training of the language model and we investigate its effect on a diverse set of downstream tasks and settings. We introduce the pre-trained CLIN-X (Clinical XLM-R) language models and show how CLIN-X outperforms other pre-trained transformer models by a large margin for 10 clinical concept extraction tasks from two languages. In addition, we demonstrate how the transformer model can be further improved with our proposed task- and language-agnostic model architecture based on ensembles over random splits and cross-sentence context. Our studies in low-resource and transfer settings reveal stable model performance despite a lack of annotated data with improvements of up to 47 F1 points when only 250 labeled sentences are available. Our results highlight the importance of specialized language models, such as CLIN-X, for concept extraction in non-standard domains, but also show that our task-agnostic model architecture is robust across the tested tasks and languages so that domain- or task-specific adaptations are not required. The CLIN-X language models and source code for fine-tuning and transferring the model are publicly available at https://github.com/boschresearch/clin_x/ and the huggingface model hub.

Glean

Extracting structured information from templatic documents is an important problem with the potential to automate many real-world business workflows such as payment, procurement, and payroll. The core challenge is that such documents can be laid out in virtually infinitely different ways. A good solution to this problem is one that generalizes well not only to known templates such as invoices from a known vendor, but also to unseen ones. We developed a system called Glean to tackle this problem. Given a target schema for a document type and some labeled documents of that type, Glean uses machine learning to automatically extract structured information from other documents of that type. In this paper, we describe the overall architecture of Glean, and discuss three key data management challenges : 1) managing the quality of ground truth data, 2) generating training data for the machine learning model using labeled documents, and 3) building tools that help a developer rapidly build and improve a model for a given document type. Through empirical studies on a real-world dataset, we show that these data management techniques allow us to train a model that is over 5 F1 points better than the exact same model architecture without the techniques we describe. We argue that for such information-extraction problems, designing abstractions that carefully manage the training data is at least as important as choosing a good model architecture.

Attention-Gated Graph Convolutions for Extracting Drug Interaction Information from Drug Labels

Preventable adverse events as a result of medical errors present a growing concern in the healthcare system. As drug-drug interactions (DDIs) may lead to preventable adverse events, being able to extract DDIs from drug labels into a machine-processable form is an important step toward effective dissemination of drug safety information. Herein, we tackle the problem of jointly extracting mentions of drugs and their interactions, including interaction outcome, from drug labels. Our deep learning approach entails composing various intermediate representations, including graph-based context derived using graph convolutions (GCs) with a novel attention-based gating mechanism (holistically called GCA), which are combined in meaningful ways to predict on all subtasks jointly. Our model is trained and evaluated on the 2018 TAC DDI corpus. Our GCA model in conjunction with transfer learning performs at 39.20% F1 and 26.09% F1 on entity recognition (ER) and relation extraction (RE), respectively, on the first official test set and at 45.30% F1 and 27.87% F1 on ER and RE, respectively, on the second official test set. These updated results lead to improvements over our prior best by up to 6 absolute F1 points. After controlling for available training data, the proposed model exhibits state-of-the-art performance for this task.

Creating Hardware Component Knowledge Bases with Training Data Generation and Multi-task Learning

Hardware component databases are vital resources in designing embedded systems. Since creating these databases requires hundreds of thousands of hours of manual data entry, they are proprietary, limited in the data they provide, and have random data entry errors. We present a machine learning based approach for creating hardware component databases directly from datasheets. Extracting data directly from datasheets is challenging because: (1) the data is relational in nature and relies on non-local context, (2) the documents are filled with technical jargon, and (3) the datasheets are PDFs, a format that decouples visual locality from locality in the document. Addressing this complexity has traditionally relied on human input, making it costly to scale. Our approach uses a rich data model, weak supervision, data augmentation, and multi-task learning to create these knowledge bases in a matter of days. We evaluate the approach on datasheets of three types of components and achieve an average quality of 77 F1 points—quality comparable to existing human-curated knowledge bases. We perform application studies that demonstrate the extraction of multiple data modalities including numerical properties and images. We show how different sources of supervision such as heuristics and human labels have distinct advantages that can be utilized together to improve knowledge base quality. Finally, we present a case study to show how this approach changes the way practitioners create hardware component knowledge bases.

Weakly Supervised Sequence Tagging from Noisy Rules

We propose a framework for training sequence tagging models with weak supervision consisting of multiple heuristic rules of unknown accuracy. In addition to supporting rules that vote on tags in the output sequence, we introduce a new type of weak supervision, called linking rules, that vote on how sequence elements should be grouped into spans with the same tag. These rules are an alternative to candidate span generators that require significantly more human effort. To estimate the accuracies of the rules and combine their conflicting outputs into training data, we introduce a new type of generative model, linked hidden Markov models (linked HMMs), and prove they are generically identifiable (up to a tag permutation) without any observed training labels. We find that linked HMMs provide an average 7 F1 point boost on benchmark named entity recognition tasks versus generative models that assume the tags are i.i.d. Further, neural sequence taggers trained with these structure-aware generative models outperform comparable state-of-the-art approaches to weak supervision by an average of 2.6 F1 points.

“Less is more”

Purpose This work studies automated user classification on Twitter in the public health domain, a task that is essential to many public health-related research works on social media but has not been addressed. The purpose of this paper is to obtain empirical knowledge on how to optimise the classifier performance on this task. Design/methodology/approach A sample of 3,100 Twitter users who tweeted about different health conditions were manually coded into six most common stakeholders. The authors propose new, simple features extracted from the short Twitter profiles of these users, and compare a large set of classification models (including state-of-the-art) that use more complex features and with different algorithms on this data set. Findings The authors show that user classification in the public health domain is a very challenging task, as the best result the authors can obtain on this data set is only 59 per cent in terms of F1 score. Compared to state-of-the-art, the methods can obtain significantly better (10 percentage points in F1 on a “best-against-best” basis) results when using only a small set of 40 features extracted from the short Twitter user profile texts. Originality/value The work is the first to study the different types of users that engage in health-related communication on social media, applicable to a broad range of health conditions rather than specific ones studied in the previous work. The methods are implemented as open source tools, and together with data, are the first of this kind. The authors believe these will encourage future research to further improve this important task.

Learning soft domain constraints in a factor graph model for template-based information extraction

The ability to accurately extract key information from textual documents is necessary in several downstream applications e.g., automatic knowledge base population from text, semantic information retrieval, question answering, or text summarization. However, information extraction (IE) systems are far from being errorless and in some cases commit errors that seem obvious to a human expert as they violate common sense or domain knowledge.Towards improving the performance of IE systems, we focus on the question of how domain knowledge can be incorporated into IE models to reduce the number of spurious extractions. Starting from the assumption that such domain knowledge cannot be incorporated explicitly and manually by domain experts due to the amount of effort and technical complexities involved, we propose a machine learning approach in which domain constraints are acquired as a byproduct of learning a model that learns to extract key information in a supervised setting. We frame the task as a template-based information extraction problem in which several dependent slots need to be automatically filled and propose a factor graph based approach to model the joint distribution of slot assignments given a text. Beyond using standard textual features in factors that score the compatibility of slot fillers in relation to the text, we use additional features that are text-independent and capture soft domain constraints. During the training process, these constraints receive a weight as part of the parameter learning process indicating how strongly a constraint should be enforced. These domain constraints are thus ‘soft’ in the sense that they can be violated, but the system learns to penalize solutions that violate them. The soft constraints we introduce come in two flavors: on the one hand we incorporate information about the mean of numerical attributes and use features that indicate how far a certain value is from the mean. We call these features single slot soft constraints. On the other hand, we model the pairwise compatibility between slot filler assignments independent of the textual context, thus modeling the (domain) compatibility of the slot assignments. We call the latter ones pairwise slot soft constraints. As main result of our work, we show that learning pairwise slot soft constraints improves the performance of our extraction model compared to single slot soft constraints by up to 6 points in F1, leading to an F1 score of 0.91 for individual template types. Further, the human readable output format of our model enables the extraction and interpretation of the learned soft constraints. Based on this, we show in an evaluation by domain experts that more than 68% of the learned soft constraints are regarded as plausible.

Hate speech detection: A solved problem? The challenging case of long tail on Twitter

In recent years, the increasing propagation of hate speech on social media and the urgent need for effective counter-measures have drawn significant investment from governments, companies, and researchers. A large number of methods have been developed for automated hate speech detection online. This aims to classify textual content into non-hate or hate speech, in which case the method may also identify the targeting characteristics (i.e., types of hate, such as race, and religion) in the hate speech. However, we notice significant difference between the performance of the two (i.e., non-hate v.s. hate). In this work, we argue for a focus on the latter problem for practical reasons. We show that it is a much more challenging task, as our analysis of the language in the typical datasets shows that hate speech lacks unique, discriminative features and therefore is found in the 'long tail' in a dataset that is difficult to discover. We then propose Deep Neural Network structures serving as feature extractors that are particularly effective for capturing the semantics of hate speech. Our methods are evaluated on the largest collection of hate speech datasets based on Twitter, and are shown to be able to outperform the best performing method by up to 5 percentage points in macro-average F1, or 8 percentage points in the more challenging case of identifying hateful content.

EUSKOR: End-to-end coreference resolution system for Basque.

This paper describes the process of adapting the Stanford Coreference resolution module to the Basque language, taking into account the characteristics of the language. The module has been integrated in a linguistic analysis pipeline obtaining an end-to-end coreference resolution system for the Basque language. The adaptation process explained can benefit and facilitate other languages with similar characteristics in the implementation of their coreference resolution systems. During the experimentation phase, we have demonstrated that language-specific features have a noteworthy effect on coreference resolution, obtaining a gain in CoNLL score of 7.07 with respect to the baseline system. We have also analysed the effect that preprocessing has in coreference resolution, comparing the results obtained with automatic mentions versus gold mentions. When gold mentions are provided, the results increase 11.5 points in CoNLL score in comparison with results obtained when automatic mentions are used. The contribution of each sieve is analysed concluding that morphology is essential for agglutinative languages to obtain good performance in coreference resolution. Finally, an error analysis of the coreference resolution system is presented which have revealed our system’s weak points and help to determine the improvements of the system. As a result of the error analysis, we have enriched the Basque coreference resolution adding new two sieves, obtaining an improvement of 0.24 points in CoNLL F1 when automatic mentions are used and of 0.39 points when the gold mentions are provided.

Snuba: Automating Weak Supervision to Label Training Data.

As deep learning models are applied to increasingly diverse problems, a key bottleneck is gathering enough high-quality training labels tailored to each task. Users therefore turn to weak supervision, relying on imperfect sources of labels like pattern matching and user-defined heuristics. Unfortunately, users have to design these sources for each task. This process can be time consuming and expensive: domain experts often perform repetitive steps like guessing optimal numerical thresholds and developing informative text patterns. To address these challenges, we present Snuba, a system to automatically generate heuristics using a small labeled dataset to assign training labels to a large, unlabeled dataset in the weak supervision setting. Snuba generates heuristics that each labels the subset of the data it is accurate for, and iteratively repeats this process until the heuristics together label a large portion of the unlabeled data. We develop a statistical measure that guarantees the iterative process will automatically terminate before it degrades training label quality. Snuba automatically generates heuristics in under five minutes and performs up to 9.74 F1 points better than the best known user-defined heuristics developed over many days. In collaborations with users at research labs, Stanford Hospital, and on open source datasets, Snuba outperforms other automated approaches like semi-supervised learning by up to 14.35 F1 points.

SemRe-Rank

Automatic Term Extraction (ATE) deals with the extraction of terminology from a domain specific corpus, and has long been an established research area in data and knowledge acquisition. ATE remains a challenging task as it is known that there is no existing ATE methods that can consistently outperform others in any domain. This work adopts a refreshed perspective to this problem: instead of searching for such a ‘one-size-fit-all’ solution that may never exist, we propose to develop generic methods to ‘enhance’ existing ATE methods. We introduce SemRe-Rank, the first method based on this principle, to incorporate semantic relatedness—an often overlooked venue—into an existing ATE method to further improve its performance. SemRe-Rank incorporates word embeddings into a personalised PageRank process to compute ‘semantic importance’ scores for candidate terms from a graph of semantically related words (nodes), which are then used to revise the scores of candidate terms computed by a base ATE algorithm. Extensively evaluated with 13 state-of-the-art base ATE methods on four datasets of diverse nature, it is shown to have achieved widespread improvement over all base methods and across all datasets, with up to 15 percentage points when measured by the Precision in the top ranked K candidate terms (the average for a set of K ’s), or up to 28 percentage points in F1 measured at a K that equals to the expected real terms in the candidates (F1 in short). Compared to an alternative approach built on the well-known TextRank algorithm, SemRe-Rank can potentially outperform by up to 8 points in Precision at top K , or up to 17 points in F1.

A scaffolding approach to coreference resolution integrating statistical and rule-based models

AbstractWe describe a scaffolding approach to the task of coreference resolution that incrementally combines statistical classifiers, each designed for a particular mention type, with rule-based models (for sub-tasks well-matched to determinism). We motivate our design by an oracle-based analysis of errors in a rule-based coreference resolution system, showing that rule-based approaches are poorly suited to tasks that require a large lexical feature space, such as resolving pronominal and common-noun mentions. Our approach combines many advantages: it incrementally builds clusters integrating joint information about entities, uses rules for deterministic phenomena, and integrates rich lexical, syntactic, and semantic features with random forest classifiers well-suited to modeling the complex feature interactions that are known to characterize the coreference task. We demonstrate that all these decisions are important. The resulting system achieves 63.2 F1 on the CoNLL-2012 shared task dataset, outperforming the rule-based starting point by over seven F1 points. Similarly, our system outperforms an equivalent sieve-based approach that relies on logistic regression classifiers instead of random forests by over four F1 points. Lastly, we show that by changing the coreference resolution system from relying on constituent-based syntax to using dependency syntax, which can be generated in linear time, we achieve a runtime speedup of 550 per cent without considerable loss of accuracy.