Most NLP tasks are modeled as supervised learning and thus require labeled training data to train effective models. However, manually producing such data at sufficient quality and quantity is known to be costly and time-intensive. Current research addresses this bottleneck by exploring a novel paradigm called zero-shot learning via dataset generation. Here, a powerful LLM is prompted with a task description to generate labeled data that can be used to train a downstream NLP model. For instance, an LLM might be prompted to "generate 500 movie reviews with positive overall sentiment, and another 500 with negative sentiment." The generated data could then be used to train a binary sentiment classifier, effectively leveraging an LLM as a teacher to a smaller student model. With this demo, we introduce Fabricator, an open-source Python toolkit for dataset generation. Fabricator implements common dataset generation workflows, supports a wide range of downstream NLP tasks (such as text classification, question answering, and entity recognition), and is integrated with well-known libraries to facilitate quick experimentation. With Fabricator, we aim to support researchers in conducting reproducible dataset generation experiments using LLMs and help practitioners apply this approach to train models for downstream tasks.

Automatically estimating the complexity of texts for readers has a variety of applications, such as recommending texts with an appropriate complexity level to language learners or supporting the evaluation of text simplification approaches. In this paper, we present our submission to the Text Complexity DE Challenge 2022, a regression task where the goal is to predict the complexity of a German sentence for German learners at level B. Our approach relies on more than 220,000 pseudo-labels created from the German Wikipedia and other corpora to train Transformer-based models, and refrains from any feature engineering or any additional, labeled data. We find that the pseudo-label-based approach gives impressive results yet requires little to no adjustment to the specific task and therefore could be easily adapted to other domains and tasks.

The evaluation of question answering models compares ground-truth annotations with model predictions. However, as of today, this comparison is mostly lexical-based and therefore misses out on answers that have no lexical overlap but are still semantically similar, thus treating correct answers as false. This underestimation of the true performance of models hinders user acceptance in applications and complicates a fair comparison of different models. Therefore, there is a need for an evaluation metric that is based on semantics instead of pure string similarity. In this short paper, we present SAS, a cross-encoder-based metric for the estimation of semantic answer similarity, and compare it to seven existing metrics. To this end, we create an English and a German three-way annotated evaluation dataset containing pairs of answers along with human judgment of their semantic similarity, which we release along with an implementation of the SAS metric and the experiments. We find that semantic similarity metrics based on recent transformer models correlate much better with human judgment than traditional lexical similarity metrics on our two newly created datasets and one dataset from related work.

Open-domain extractive question answering works well on textual data by first retrieving candidate texts and then extracting the answer from those candidates. However, some questions cannot be answered by text alone but require information stored in tables. In this paper, we present an approach for retrieving both texts and tables relevant to a question by jointly encoding texts, tables and questions into a single vector space. To this end, we create a new multi-modal dataset based on text and table datasets from related work and compare the retrieval performance of different encoding schemata. We find that dense vector embeddings of transformer models outperform sparse embeddings on four out of six evaluation datasets. Comparing different dense embedding models, tri-encoders with one encoder for each question, text and table increase retrieval performance compared to bi-encoders with one encoder for the question and one for both text and tables. We release the newly created multi-modal dataset to the community so that it can be used for training and evaluation.

Patent document collections are an immense source of knowledge for research and innovation communities worldwide. The rapid growth of the number of patent documents poses an enormous challenge for retrieving and analyzing information from this source in an effective manner. Based on deep learning methods for natural language processing, novel approaches have been developed in the field of patent analysis. The goal of these approaches is to reduce costs by automating tasks that previously only domain experts could solve. In this article, we provide a comprehensive survey of the application of deep learning for patent analysis. We summarize the state-of-the-art techniques and describe how they are applied to various tasks in the patent domain. In a detailed discussion, we categorize 40 papers based on the dataset, the representation, and the deep learning architecture that were used, as well as the patent analysis task that was targeted. With our survey, we aim to foster future research at the intersection of patent analysis and deep learning and we conclude by listing promising paths for future work.

With the rise of research on toxic comment classification, more and more annotated datasets have been released. The wide variety of the task (different languages, different labeling processes and schemes) has led to a large amount of heterogeneous datasets that can be used for training and testing very specific settings. Despite recent efforts to create web pages that provide an overview, most publications still use only a single dataset. They are not stored in one central database, they come in many different data formats and it is difficult to interpret their class labels and how to reuse these labels in other projects. To overcome these issues, we present a collection of more than thirty datasets in the form of a software tool that automatizes downloading and processing of the data and presents them in a unified data format that also offers a mapping of compatible class labels. Another advantage of that tool is that it gives an overview of properties of available datasets, such as different languages, platforms, and class labels to make it easier to select suitable training and test data.

A major challenge of research on non-English machine reading for question answering (QA) is the lack of annotated datasets. In this paper, we present GermanQuAD, a dataset of 13,722 extractive question/answer pairs. To improve the reproducibility of the dataset creation approach and foster QA research on other languages, we summarize lessons learned and evaluate reformulation of question/answer pairs as a way to speed up the annotation process. An extractive QA model trained on GermanQuAD significantly outperforms multilingual models and also shows that machine-translated training data cannot fully substitute hand-annotated training data in the target language. Finally, we demonstrate the wide range of applications of GermanQuAD by adapting it to GermanDPR, a training dataset for dense passage retrieval (DPR), and train and evaluate one of the first non-English DPR models.

Current document embeddings require large training corpora but fail to learn high-quality representations when confronted with a small number of domain-specific documents and rare terms. Further, they transform each document into a single embedding vector, making it hard to capture different notions of document similarity or explain why two documents are considered similar. In this work, we propose our Faceted Domain Encoder, a novel approach to learn multifaceted embeddings for domain-specific documents. It is based on a Siamese neural network architecture and leverages knowledge graphs to further enhance the embeddings even if only a few training samples are available. The model identifies different types of domain knowledge and encodes them into separate dimensions of the embedding, thereby enabling multiple ways of finding and comparing related documents in the vector space. We evaluate our approach on two benchmark datasets and find that it achieves the same embedding quality as state-of-the-art models while requiring only a tiny fraction of their training data. An interactive demo, our source code, and the evaluation datasets are available online: https://hpi.de/naumann/s/multifaceted-embeddings and a screencast is available on YouTube: https://youtu.be/HHcsX2clEwg

With the rise of research on toxic comment classification, more and more annotated datasets have been released. The wide variety of the task (different languages, different labeling processes and schemes) has led to a large amount of heterogeneous datasets that can be used for training and testing very specific settings. Despite recent efforts to create web pages that provide an overview, most publications still use only a single dataset. They are not stored in one central database, they come in many different data formats and it is difficult to interpret their class labels and how to reuse these labels in other projects. To overcome these issues, we present a collection of more than forty datasets in the form of a software tool that automatizes downloading and processing of the data and presents them in a unified data format that also offers a mapping of compatible class labels. Another advantage of that tool is that it gives an overview of properties of available datasets, such as different languages, platforms, and class labels to make it easier to select suitable training and test data.

The comment sections of online news platforms have shaped the way in which people express their opinion online. However, due to the overwhelming number of comments, no in-depth discussions emerge. To foster more interactive and engaging discussions, we propose our ComEx interface for the exploration of reader comments on online news platforms. Potential discussion participants can get a quick overview and are not discouraged by an abundance of comments. It is our goal to represent the discussion in a graph of comments that can be used in an interactive user interface for exploration. To this end, a processing pipeline fetches comments from several different platforms and adds edges in the graph based on topical similarity or meta-data and ranks nodes on metrics such as controversy or toxicity. By interacting with the graph, users can explore and react to single comments or entire threads they are interested in.

We present the GermEval 2021 shared task on the identification of toxic, engaging, and fact-claiming comments. This shared task comprises three binary classification subtasks with the goal to identify: toxic comments, engaging comments, and comments that include indications of a need for fact-checking, here referred to as fact-claiming comments. Building on the two previous GermEval shared tasks on the identification of offensive language in 2018 and 2019, we extend this year’s task definition to meet the demand of moderators and community managers to also highlight comments that foster respectful communication, encourage in-depth discussions, and check facts that lines of arguments rely on. The dataset comprises 4,188 posts extracted from the Facebook page of a German political talk show of a national public television broadcaster. A theoretical framework and additional reliability tests during the data annotation process ensure particularly high data quality. The shared task had 15 participating teams submitting 31 runs for the subtask on toxic comments, 25 runs for the subtask on engaging comments, and 31 for the subtask on fact-claiming comments. The shared task website can be found at https://germeval2021toxic.github.io/SharedTask/.

Many online discussion platforms use a content moderation process, where human moderators check user comments for offensive language and other rule violations. It is the moderator’s decision which comments to remove from the platform because of violations and which ones to keep. Research so far focused on automating this decision process in the form of supervised machine learning for a classification task. However, even with machine-learned models achieving better classification accuracy than human experts, there is still a reason why human moderators are preferred. In contrast to black-box models, such as neural networks, humans can give explanations for their decision to remove a comment. For example, they can point out which phrase in the comment is offensive or what subtype of offensiveness applies. In this paper, we analyze and compare four explanation methods for different offensive language classifiers: an interpretable machine learning model (naive Bayes), a model-agnostic explanation method (LIME), a model-based explanation method (LRP), and a self-explanatory model (LSTM with an attention mechanism). We evaluate these approaches with regard to their explanatory power and their ability to point out which words are most relevant for a classifier’s decision. We find that the more complex models achieve better classification accuracy while also providing better explanations than the simpler models.

Hierarchical classification schemes are an effective and natural way to organize large document collections. However, complex schemes make the manual classification time-consuming and require domain experts. Current machine learning approaches for hierarchical classification do not exploit all the information contained in the hierarchical schemes. During training, they do not make full use of the inherent parent-child relation of classes. For example, they neglect to tailor document representations, such as embeddings, to each individual hierarchy level. Our model overcomes these problems by addressing hierarchical classification as a sequence generation task. To this end, our neural network transforms a sequence of input words into a sequence of labels, which represents a path through a tree-structured hierarchy scheme. The evaluation uses a patent corpus, which exhibits a complex class hierarchy scheme and high-quality annotations from domain experts and comprises millions of documents. We re-implemented five models from related work and show that our basic model achieves competitive results in comparison with the best approach. A variation of our model that uses the recent Transformer architecture outperforms the other approaches. The error analysis reveals that the encoder of our model has the strongest influence on its classification performance.

Comment sections below online news articles enjoy growing popularity among readers. However, the overwhelming number of comments makes it infeasible for the average news consumer to read all of them and hinders engaging discussions. Most platforms display comments in chronological order, which neglects that some of them are more relevant to users and are better conversation starters.In this paper, we systematically analyze user engagement in the form of the upvotes and replies that a comment receives. Based on comment texts, we train a model to distinguish comments that have either a high or low chance of receiving many upvotes and replies. Our evaluation on user comments from TheGuardian.com compares recurrent and convolutional neural network models, and a traditional feature-based classifier. Further, we investigate what makes some comments more engaging than others. To this end, we identify engagement triggers and arrange them in a taxonomy. Explanation methods for neural networks reveal which input words have the strongest influence on our model's predictions. In addition, we evaluate on a dataset of product reviews, which exhibit similar properties as user comments, such as featuring upvotes for helpfulness.

The comment sections of online news platforms are an important space to indulge in political conversations and to discuss opinions. Although primarily meant as forums where readers discuss amongst each other, they can also spark a dialog with the journalists who authored the article. A small but important fraction of comments address the journalists directly, e.g., with questions, recommendations for future topics, thanks and appreciation, or article corrections. However, the sheer number of comments makes it infeasible for journalists to follow discussions around their articles in extenso. A better understanding of this data could support journalists in gaining insights into their audience and fostering engaging and respectful discussions. To this end, we present a dataset of dialogs in which journalists of The Guardian replied to reader comments and identify the reasons why. Based on this data, we formulate the novel task of recommending reader comments to journalists that are worth reading or replying to, i.e., ranking comments in such a way that the top comments are most likely to require the journalists' reaction. As a baseline, we trained a neural network model with the help of a pair-wise comment ranking task. Our experiment reveals the challenges of this task and we outline promising paths for future work. The data and our code are available for research purposes from: https://hpi.de/naumann/projects/repeatability/text-mining.html.

Many online news platforms provide comment sections for reader discussions below articles. While users of these platforms often read comments, only a minority of them regularly write comments. To encourage and foster more frequent engagement, we study the task of personalized recommendation of reader discussions to users. We present a neural network model that jointly learns embeddings for users and comments encoding general properties. Based on explicit and implicit user feedback, we sample relevant and irrelevant reader discussions to build a representative training dataset. We compare to several baselines and state-of-the-art approaches in an evaluation on two datasets from The Guardian and Daily Mail. Experimental results show that the recommendations of our approach are superior in terms of precision and recall. Further, the learned user embeddings are of general applicability because they preserve the similarity of users who share interests in similar topics.

Comment sections of online news platforms are an essential space to express opinions and discuss political topics. In contrast to other online posts, news discussions are related to particular news articles, comments refer to each other, and individual conversations emerge. However, the misuse by spammers, haters, and trolls makes costly content moderation necessary. Sentiment analysis can not only support moderation but also help to understand the dynamics of online discussions. A subtask of content moderation is the identification of toxic comments. To this end, we describe the concept of toxicity and characterize its subclasses. Further, we present various deep learning approaches, including datasets and architectures, tailored to sentiment analysis in online discussions. One way to make these approaches more comprehensible and trustworthy is fine-grained instead of binary comment classification. On the downside, more classes require more training data. Therefore, we propose to augment training data by using transfer learning. We discuss real-world applications, such as semi-automated comment moderation and troll detection. Finally, we outline future challenges and current limitations in light of most recent research publications.

Patent examiners need to solve a complex information retrieval task when they assess the novelty and inventive step of claims made in a patent application. Given a claim, they search for prior art, which comprises all relevant publicly available information. This time-consuming task requires a deep understanding of the respective technical domain and the patent-domain-specific language. For these reasons, we address the computer-assisted search for prior art by creating a training dataset for supervised machine learning called PatentMatch. It contains pairs of claims from patent applications and semantically corresponding text passages of different degrees from cited patent documents. Each pair has been labeled by technically-skilled patent examiners from the European Patent Office. Accordingly, the label indicates the degree of semantic correspondence (matching), i.e., whether the text passage is prejudicial to the novelty of the claimed invention or not. Preliminary experiments using a baseline system show that PatentMatch can indeed be used for training a binary text pair classifier on this challenging information retrieval task. The dataset is available online: https://hpi.de/naumann/s/patentmatch.

Accessible and reusable datasets are a necessity to accomplish repeatable research. This requirement poses a problem particularly for web science, since scraped data comes in various formats and can change due to the dynamic character of the web. Further, usage of web data is typically restricted by copyright-protection or privacy regulations, which hinder publication of datasets. To alleviate these problems and reach what we define as “partial data repeatability”, we present a process that consists of multiple components. Researchers need to distribute only a scraper and not the data itself to comply with legal limitations. If a dataset is re-scraped for repeatability after some time, the integrity of different versions can be checked based on fingerprints. Moreover, fingerprints are sufficient to identify what parts of the data have changed and how much. We evaluate an implementation of this process with a dataset of 250 million online comments collected from five different news discussion platforms. We re-scraped the dataset after pausing for one year and show that less than ten percent of the data has actually changed. These experiments demonstrate that providing a scraper and fingerprints enables recreating a dataset and supports the repeatability of web science experiments.