In this paper, we present AfricaPOS, the largest part-of-speech (POS) dataset for 20 typologically diverse African languages. We discuss the challenges in annotating POS for these languages using the universal dependencies (UD) guidelines. We conducted extensive POS baseline experiments using both conditional random field and several multilingual pre-trained language models. We applied various cross-lingual transfer models trained with data available in the UD. Evaluating on the AfricaPOS dataset, we show that choosing the best transfer language(s) in both single-source and multi-source setups greatly improves the POS tagging performance of the target languages, in particular when combined with parameter-fine-tuning methods. Crucially, transferring knowledge from a language that matches the language family and morphosyntactic properties seems to be more effective for POS tagging in unseen languages.

Weakly supervised learning is a popular approach for training machine learning models in low-resource settings. Instead of requesting high-quality yet costly human annotations, it allows training models with noisy annotations obtained from various weak sources. Recently, many sophisticated approaches have been proposed for robust training under label noise, reporting impressive results. In this paper, we revisit the setup of these approaches and find that the benefits brought by these approaches are significantly overestimated. Specifically, we find that the success of existing weakly supervised learning approaches heavily relies on the availability of clean validation samples which, as we show, can be leveraged much more efficiently by simply training on them. After using these clean labels in training, the advantages of using these sophisticated approaches are mostly wiped out. This remains true even when reducing the size of the available clean data to just five samples per class, making these approaches impractical. To understand the true value of weakly supervised learning, we thoroughly analyze diverse NLP datasets and tasks to ascertain when and why weakly supervised approaches work. Based on our findings, we provide recommendations for future research.

This paper investigates the performance of massively multilingual neural machine translation (NMT) systems in translating Yorùbá greetings (kú mask), which are a big part of Yorùbá language and culture, into English. To evaluate these models, we present IkiniYorùbá, a Yorùbá-English translation dataset containing some Yorùbá greetings, and sample use cases. We analysed the performance of different multilingual NMT systems including Google and NLLB and show that these models struggle to accurately translate Yorùbá greetings into English. In addition, we trained a Yorùbá-English model by fine-tuning an existing NMT model on the training split of IkiniYorùbá and this achieved better performance when compared to the pre-trained multilingual NMT models, although they were trained on a large volume of data.

Training deep neural networks (DNNs) under weak supervision has attracted increasing research attention as it can significantly reduce the annotation cost. However, labels from weak supervision can be noisy, and the high capacity of DNNs enables them to easily overfit the label noise, resulting in poor generalization. Recent methods leverage self-training to build noise-resistant models, in which a teacher trained under weak supervision is used to provide highly confident labels for teaching the students. Nevertheless, the teacher derived from such frameworks may have fitted a substantial amount of noise and therefore produce incorrect pseudo-labels with high confidence, leading to severe error propagation. In this work, we propose Meta Self-Refinement (MSR), a noise-resistant learning framework, to effectively combat label noise from weak supervision. Instead of relying on a fixed teacher trained with noisy labels, we encourage the teacher to refine its pseudo-labels. At each training step, MSR performs a meta gradient descent on the current mini-batch to maximize the student performance on a clean validation set. Extensive experimentation on eight NLP benchmarks demonstrates that MSR is robust against label noise in all settings and outperforms state-of-the-art methods by up to 11.4% in accuracy and 9.26% in F1 score.

The detection and normalization of temporal expressions is an important task and preprocessing step for many applications. However, prior work on normalization is rule-based, which severely limits the applicability in real-world multilingual settings, due to the costly creation of new rules. We propose a novel neural method for normalizing temporal expressions based on masked language modeling. Our multilingual method outperforms prior rule-based systems in many languages, and in particular, for low-resource languages with performance improvements of up to 33 F1 on average compared to the state of the art.

We present a cross-linguistic study of vowel harmony that aims to quantifies this phenomenon using data-driven computational modeling. Concretely, we define an information-theoretic measure of harmonicity based on the predictability of vowels in a natural language lexicon, which we estimate using phoneme-level language models (PLMs). Prior quantitative studies have heavily relied on inflected word-forms in the analysis on vowel harmony. On the contrary, we train our models using cross-linguistically comparable lemma forms with little or no inflection, which enables us to cover more under-studied languages. Training data for our PLMs consists of word lists offering a maximum of 1000 entries per language. Despite the fact that the data we employ are substantially smaller than previously used corpora, our experiments demonstrate the neural PLMs capture vowel harmony patterns in a set of languages that exhibit this phenomenon. Our work also demonstrates that word lists are a valuable resource for typological research, and offers new possibilities for future studies on low-resource, under-studied languages.

Self-supervision has emerged as an effective paradigm for learning representations of spoken language from raw audio without explicit labels or transcriptions. Self-supervised speech models, such as wav2vec 2.0 (Baevski et al., 2020) and HuBERT (Hsu et al., 2021), have shown significant promise in improving the performance across different speech processing tasks. One of the main advantages of self-supervised speech models is that they can be pre-trained on a large sample of languages (Conneau et al., 2020; Babu et al.,2022), which facilitates cross-lingual transfer for low-resource languages (San et al., 2021). State-of-the-art self-supervised speech models include a quantization module that transforms the continuous acoustic input into a sequence of discrete units. One of the key questions in this area is whether the discrete representations learned via self-supervision are language-specific or language-universal. In other words, we ask: do the discrete units learned by a multilingual speech model represent the same speech sounds across languages or do they differ based on the specific language being spoken? From the practical perspective, this question has important implications for the development of speech models that can generalize across languages, particularly for low-resource languages. Furthermore, examining the level of linguistic abstraction in speech models that lack symbolic supervision is also relevant to the field of human language acquisition (Dupoux, 2018).

Few-shot fine-tuning and in-context learning are two alternative strategies for task adaptation of pre-trained language models. Recently, in-context learning has gained popularity over fine-tuning due to its simplicity and improved out-of-domain generalization, and because extensive evidence shows that fine-tuned models pick up on spurious correlations.Unfortunately, previous comparisons of the two approaches were done using models of different sizes. This raises the question of whether the observed weaker out-of-domain generalization of fine-tuned models is an inherent property of fine-tuning or a limitation of the experimental setup. In this paper, we compare the generalization of few-shot fine-tuning and in-context learning to challenge datasets, while controlling for the models used, the number of examples, and the number of parameters, ranging from 125M to 30B. Our results show that fine-tuned language models can in fact generalize well out-of-domain. We find that both approaches generalize similarly; they exhibit large variation and depend on properties such as model size and the number of examples, highlighting that robust task adaptation remains a challenge.

Analyzing ethnic or religious bias is important for improving fairness, accountability, and transparency of natural language processing models. However, many techniques rely on human-compiled lists of bias terms, which are expensive to create and are limited in coverage. In this study, we present a fully data-driven pipeline for generating a knowledge graph (KG) of cultural knowledge and stereotypes. Our resulting KG covers 5 religious groups and 5 nationalities and can easily be extended to more entities. Our human evaluation shows that the majority (59.2%) of non-singleton entries are coherent and complete stereotypes. We further show that performing intermediate masked language model training on the verbalized KG leads to a higher level of cultural awareness in the model and has the potential to increase classification performance on knowledge-crucial samples on a related task, i.e., hate speech detection.

Multilingual pre-trained language models (PLMs) have demonstrated impressive performance on several downstream tasks for both high-resourced and low-resourced languages. However, there is still a large performance drop for languages unseen during pre-training, especially African languages. One of the most effective approaches to adapt to a new language is language adaptive fine-tuning (LAFT) — fine-tuning a multilingual PLM on monolingual texts of a language using the pre-training objective. However, adapting to target language individually takes large disk space and limits the cross-lingual transfer abilities of the resulting models because they have been specialized for a single language. In this paper, we perform multilingual adaptive fine-tuning on 17 most-resourced African languages and three other high-resource languages widely spoken on the African continent to encourage cross-lingual transfer learning. To further specialize the multilingual PLM, we removed vocabulary tokens from the embedding layer that corresponds to non-African writing scripts before MAFT, thus reducing the model size by around 50%. Our evaluation on two multilingual PLMs (AfriBERTa and XLM-R) and three NLP tasks (NER, news topic classification, and sentiment classification) shows that our approach is competitive to applying LAFT on individual languages while requiring significantly less disk space. Additionally, we show that our adapted PLM also improves the zero-shot cross-lingual transfer abilities of parameter efficient fine-tuning methods.

Recent advances in the pre-training for language models leverage large-scale datasets to create multilingual models. However, low-resource languages are mostly left out in these datasets. This is primarily because many widely spoken languages that are not well represented on the web and therefore excluded from the large-scale crawls for datasets. Furthermore, downstream users of these models are restricted to the selection of languages originally chosen for pre-training. This work investigates how to optimally leverage existing pre-trained models to create low-resource translation systems for 16 African languages. We focus on two questions: 1) How can pre-trained models be used for languages not included in the initial pretraining? and 2) How can the resulting translation models effectively transfer to new domains? To answer these questions, we create a novel African news corpus covering 16 languages, of which eight languages are not part of any existing evaluation dataset. We demonstrate that the most effective strategy for transferring both additional languages and additional domains is to leverage small quantities of high-quality translation data to fine-tune large pre-trained models.

Learning semantically meaningful sentence embeddings is an open problem in natural language processing. In this work, we propose a sentence embedding learning approach that exploits both visual and textual information via a multimodal contrastive objective. Through experiments on a variety of semantic textual similarity tasks, we demonstrate that our approach consistently improves the performance across various datasets and pre-trained encoders. In particular, combining a small amount of multimodal data with a large text-only corpus, we improve the state-of-the-art average Spearman’s correlation by 1.7%. By analyzing the properties of the textual embedding space, we show that our model excels in aligning semantically similar sentences, providing an explanation for its improved performance.

In this paper the authors detail the various legal and ethical issues faced during the ATCO2 project. This project is aimed at developing tools to automatically collect and transcribe air traffic conversations, especially conversations between pilots and air controls towers. In this paper the authors will develop issues related to intellectual property, public data, privacy, and general ethics issues related to the collection of air-traffic control speech.

Acoustic word embeddings (AWEs) are fixed-dimensionality vector representations in a vector space such that different acoustic exemplars of the same word are projected nearby in the embedding space. In addition to their use in speech technology applications such as spoken term discovery and keyword spotting, AWE models have been adopted as models of spoken-word processing in several cognitively motivated studies and they have shown to exhibit a human-like performance in some auditory processing tasks. Nevertheless, the representation geometry of AWEs remains an under-explored topic that has not been studied in the literature. In this paper, we take a closer analytical look at AWEs and study how the choice of the learning objective and the architecture shapes their representational profile. Our main findings highlight the prominent role of the learning objective on the representational geometry over the architecture.

Incorrect labels in training data occur when human annotators make mistakes or when the data is generated via weak or distant supervision. It has been shown that complex noise-handling techniques - by modeling, cleaning or filtering the noisy instances - are required to prevent models from fitting this label noise. However, we show in this work that, for text classification tasks with modern NLP models like BERT, over a variety of noise types, existing noise-handling methods do not always improve its performance, and may even deteriorate it, suggesting the need for further investigation. We also back our observations with a comprehensive analysis.

Recently neural network based approaches to knowledge-intensive NLP tasks, such as question answering, started to rely heavily on the combination of neural retrievers and readers. Retrieval is typically performed over a large textual knowledge base (KB) which requires significant memory and compute resources, especially when scaled up. On HotpotQA we systematically investigate reducing the size of the KB index by means of dimensionality (sparse random projections, PCA, autoencoders) and numerical precision reduction. Our results show that PCA is an easy solution that requires very little data and is only slightly worse than autoencoders, which are less stable. All methods are sensitive to pre- and post-processing and data should always be centered and normalized both before and after dimension reduction. Finally, we show that it is possible to combine PCA with using 1bit per dimension. Overall we achieve (1) 100× compression with 75%, and (2) 24× compression with 92% original retrieval performance.

Closely related languages are often mutually intelligible to various degrees. Therefore, speakers of closely related languages are usually capable of (partially) comprehending each other’s speech without explicitly learning the target, second language. The cross-linguistic intelligibility among closely related languages is mainly driven by linguistic factors such as lexical similarities. This paper presents a computational model of spoken-word recognition and investigates its ability to recognize word forms from different languages than its native, training language. Our model is based on a recurrent neural network that learns to map a word’s phonological sequence onto a semantic representation of the word. Furthermore, we present a case study on the related Slavic languages and demonstrate that the cross-lingual performance of our model not only predicts mutual intelligibility to a large extent but also reflects the genetic classification of the languages in our study.

Call centres endeavour to achieve the highest possible level of transparency with regard to the factors influencing sales success. Existing approaches to the quality assessment of customer-agent sales negotiations do not enable in-depths analysis of sales behaviour. This study addresses this gap and presents a conceptual and operational framework applying the ISO 24617-2 dialogue act annotation scheme, a multidimensional taxonomy of interoperable semantic concepts. We hypothesise that the ISO 24617-2 dialogue act annotation framework adequately supports sales negotiation assessment in the domain of call centre conversations. Authentic call centre conversations are annotated and a range of extensions/modifications are proposed making the annotation scheme better fit this new domain. We concluded that ISO 24617-2 serves as a powerful instrument for the analysis and assessment of sales negotiation and strategies applied by a call centre agent.

Even though hate speech (HS) online has been an important object of research in the last decade, most HS-related corpora over-simplify the phenomenon of hate by attempting to label user comments as “hate” or “neutral”. This ignores the complex and subjective nature of HS, which limits the real-life applicability of classifiers trained on these corpora. In this study, we present the M-Phasis corpus, a corpus of ~9k German and French user comments collected from migration-related news articles. It goes beyond the “hate”-“neutral” dichotomy and is instead annotated with 23 features, which in combination become descriptors of various types of speech, ranging from critical comments to implicit and explicit expressions of hate. The annotations are performed by 4 native speakers per language and achieve high (0.77 <= k <= 1) inter-annotator agreements. Besides describing the corpus creation and presenting insights from a content, error and domain analysis, we explore its data characteristics by training several classification baselines.

In recent years, voice-controlled personal assistants have revolutionized the interaction with smart devices and mobile applications. The collected data are then used by system providers to train language models (LMs). Each spoken message reveals personal information, hence removing private information from the input sentences is necessary. Our data sanitization process relies on recognizing and replacing named entities by other words from the same class. However, this may harm LM training because privacy-transformed data is unlikely to match the test distribution. This paper aims to fill the gap by focusing on the adaptation of LMs initially trained on privacy-transformed sentences using a small amount of original untransformed data. To do so, we combine class-based LMs, which provide an effective approach to overcome data sparsity in the context of n-gram LMs, and neural LMs, which handle longer contexts and can yield better predictions. Our experiments show that training an LM on privacy-transformed data result in a relative 11% word error rate (WER) increase compared to training on the original untransformed data, and adapting that model on a limited amount of original untransformed data leads to a relative 8% WER improvement over the model trained solely on privacy-transformed data.

We focus on the syntactic variation and measure syntactic distances between nine Slavic languages (Belarusian, Bulgarian, Croatian, Czech, Polish, Slovak, Slovene, Russian, and Ukrainian) using symmetric measures of insertion, deletion and movement of syntactic units in the parallel sentences of the fable “The North Wind and the Sun”. Additionally, we investigate phonetic and orthographic asymmetries between selected languages by means of the information theoretical notion of surprisal. Syntactic distance and surprisal are, thus, considered as potential predictors of mutual intelligibility between related languages. In spoken and written cloze test experiments for Slavic native speakers, the presented predictors will be validated as to whether variations in syntax lead to a slower or impeded intercomprehension of Slavic texts.

African languages are spoken by over a billion people, but they are under-represented in NLP research and development. Multiple challenges exist, including the limited availability of annotated training and evaluation datasets as well as the lack of understanding of which settings, languages, and recently proposed methods like cross-lingual transfer will be effective. In this paper, we aim to move towards solutions for these challenges, focusing on the task of named entity recognition (NER). We present the creation of the largest to-date human-annotated NER dataset for 20 African languages. We study the behaviour of state-of-the-art cross-lingual transfer methods in an Africa-centric setting, empirically demonstrating that the choice of source transfer language significantly affects performance. While much previous work defaults to using English as the source language, our results show that choosing the best transfer language improves zero-shot F1 scores by an average of 14% over 20 languages as compared to using English.

Recent research on style transfer takes inspiration from unsupervised neural machine translation (UNMT), learning from large amounts of non-parallel data by exploiting cycle consistency loss, back-translation, and denoising autoencoders. By contrast, the use of selfsupervised NMT (SSNMT), which leverages (near) parallel instances hidden in non-parallel data more efficiently than UNMT, has not yet been explored for style transfer. In this paper we present a novel Self-Supervised Style Transfer (3ST) model, which augments SSNMT with UNMT methods in order to identify and efficiently exploit supervisory signals in non-parallel social media posts. We compare 3ST with state-of-the-art (SOTA) style transfer models across civil rephrasing, formality and polarity tasks. We show that 3ST is able to balance the three major objectives (fluency, content preservation, attribute transfer accuracy) the best, outperforming SOTA models on averaged performance across their tested tasks in automatic and human evaluation.

Theories and models of spoken word recognition aim to explain the process of accessing lexical knowledge given an acoustic realization of a word form. There is consensus that phonological and semantic information is crucial for this process. However, there is accumulating evidence that orthographic information could also have an impact on auditory word recognition. This paper presents two models of spoken word recognition that instantiate different hypotheses regarding the influence of orthography on this process. We show that these models reproduce human-like behavior in different ways and provide testable hypotheses for future research on the source of orthographic effects in spoken word recognition.

We present a deep neural model of spoken word recognition which is trained to retrieve the meaning of a word (in the form of a word embedding) given its spoken form, a task which resembles that faced by a human listener. Furthermore, we investigate the influence of variability in speech signals on the model’s performance. To this end, we conduct of set of controlled experiments using word-aligned read speech data in German. Our experiments show that (1) the model is more sensitive to dialectical variation than gender variation, and (2) recognition performance of word cognates from related languages reflect the degree of relatedness between languages in our study. Our work highlights the feasibility of modeling human speech perception using deep neural networks.

Sentiment tasks such as hate speech detection and sentiment analysis, especially when performed on languages other than English, are often low-resource. In this study, we exploit the emotional information encoded in emojis to enhance the performance on a variety of sentiment tasks. This is done using a transfer learning approach, where the parameters learned by an emoji-based source task are transferred to a sentiment target task. We analyse the efficacy of the transfer under three conditions, i.e. i) the emoji content and ii) label distribution of the target task as well as iii) the difference between monolingually and multilingually learned source tasks. We find i.a. that the transfer is most beneficial if the target task is balanced with high emoji content. Monolingually learned source tasks have the benefit of taking into account the culturally specific use of emojis and gain up to F1 +0.280 over the baseline.

For most language combinations and parallel data is either scarce or simply unavailable. To address this and unsupervised machine translation (UMT) exploits large amounts of monolingual data by using synthetic data generation techniques such as back-translation and noising and while self-supervised NMT (SSNMT) identifies parallel sentences in smaller comparable data and trains on them. To this date and the inclusion of UMT data generation techniques in SSNMT has not been investigated. We show that including UMT techniques into SSNMT significantly outperforms SSNMT (up to +4.3 BLEU and af2en) as well as statistical (+50.8 BLEU) and hybrid UMT (+51.5 BLEU) baselines on related and distantly-related and unrelated language pairs.

How do neural networks “perceive” speech sounds from unknown languages? Does the typological similarity between the model’s training language (L1) and an unknown language (L2) have an impact on the model representations of L2 speech signals? To answer these questions, we present a novel experimental design based on representational similarity analysis (RSA) to analyze acoustic word embeddings (AWEs)—vector representations of variable-duration spoken-word segments. First, we train monolingual AWE models on seven Indo-European languages with various degrees of typological similarity. We then employ RSA to quantify the cross-lingual similarity by simulating native and non-native spoken-word processing using AWEs. Our experiments show that typological similarity indeed affects the representational similarity of the models in our study. We further discuss the implications of our work on modeling speech processing and language similarity with neural networks.

Hate speech and profanity detection suffer from data sparsity, especially for languages other than English, due to the subjective nature of the tasks and the resulting annotation incompatibility of existing corpora. In this study, we identify profane subspaces in word and sentence representations and explore their generalization capability on a variety of similar and distant target tasks in a zero-shot setting. This is done monolingually (German) and cross-lingually to closely-related (English), distantly-related (French) and non-related (Arabic) tasks. We observe that, on both similar and distant target tasks and across all languages, the subspace-based representations transfer more effectively than standard BERT representations in the zero-shot setting, with improvements between F1 +10.9 and F1 +42.9 over the baselines across all tested monolingual and cross-lingual scenarios.

Combining several embeddings typically improves performance in downstream tasks as different embeddings encode different information. It has been shown that even models using embeddings from transformers still benefit from the inclusion of standard word embeddings. However, the combination of embeddings of different types and dimensions is challenging. As an alternative to attention-based meta-embeddings, we propose feature-based adversarial meta-embeddings (FAME) with an attention function that is guided by features reflecting word-specific properties, such as shape and frequency, and show that this is beneficial to handle subword-based embeddings. In addition, FAME uses adversarial training to optimize the mappings of differently-sized embeddings to the same space. We demonstrate that FAME works effectively across languages and domains for sequence labeling and sentence classification, in particular in low-resource settings. FAME sets the new state of the art for POS tagging in 27 languages, various NER settings and question classification in different domains.

Documents as short as a single sentence may inadvertently reveal sensitive information about their authors, including e.g. their gender or ethnicity. Style transfer is an effective way of transforming texts in order to remove any information that enables author profiling. However, for a number of current state-of-the-art approaches the improved privacy is accompanied by an undesirable drop in the down-stream utility of the transformed data. In this paper, we propose a simple, zero-shot way to effectively lower the risk of author profiling through multilingual back-translation using off-the-shelf translation models. We compare our models with five representative text style transfer models on three datasets across different domains. Results from both an automatic and a human evaluation show that our approach achieves the best overall performance while requiring no training data. We are able to lower the adversarial prediction of gender and race by up to 22% while retaining 95% of the original utility on downstream tasks.

In low-resource settings, model transfer can help to overcome a lack of labeled data for many tasks and domains. However, predicting useful transfer sources is a challenging problem, as even the most similar sources might lead to unexpected negative transfer results. Thus, ranking methods based on task and text similarity — as suggested in prior work — may not be sufficient to identify promising sources. To tackle this problem, we propose a new approach to automatically determine which and how many sources should be exploited. For this, we study the effects of model transfer on sequence labeling across various domains and tasks and show that our methods based on model similarity and support vector machines are able to predict promising sources, resulting in performance increases of up to 24 F1 points.

We present an extended version of a tool developed for calculating linguistic distances and asymmetries in auditory perception of closely related languages. Along with evaluating the metrics available in the initial version of the tool, we introduce word adaptation entropy as an additional metric of linguistic asymmetry. Potential predictors of speech intelligibility are validated with human performance in spoken cognate recognition experiments for Bulgarian and Russian. Special attention is paid to the possibly different contributions of vowels and consonants in oral intercomprehension. Using incom.py 2.0 it is possible to calculate, visualize, and validate three measurement methods of linguistic distances and asymmetries as well as carrying out regression analyses in speech intelligibility between related languages.

The paper presents a novel discourse-based approach to argument quality assessment defined as a graph classification task, where the depth of reasoning (argumentation) is evident from the number and type of detected discourse units and relations between them. We successfully applied state-of-the-art discourse parsers and machine learning models to reconstruct argument graphs with the identified and classified discourse units as nodes and relations between them as edges. Then Graph Neural Networks were trained to predict the argument quality assessing its acceptability, relevance, sufficiency and overall cogency. The obtained accuracy ranges from 74.5% to 85.0% and indicates that discourse-based argument structures reflect qualitative properties of natural language arguments. The results open many interesting prospects for future research in the field of argumentation mining.

Deep neural networks and huge language models are becoming omnipresent in natural language applications. As they are known for requiring large amounts of training data, there is a growing body of work to improve the performance in low-resource settings. Motivated by the recent fundamental changes towards neural models and the popular pre-train and fine-tune paradigm, we survey promising approaches for low-resource natural language processing. After a discussion about the different dimensions of data availability, we give a structured overview of methods that enable learning when training data is sparse. This includes mechanisms to create additional labeled data like data augmentation and distant supervision as well as transfer learning settings that reduce the need for target supervision. A goal of our survey is to explain how these methods differ in their requirements as understanding them is essential for choosing a technique suited for a specific low-resource setting. Further key aspects of this work are to highlight open issues and to outline promising directions for future research.

The paper presents a discourse-based approach to the analysis of argumentative texts departing from the assumption that the coherence of a text should capture argumentation structure as well and, therefore, existing discourse analysis tools can be successfully applied for argument segmentation and annotation tasks. We tested the widely used Penn Discourse Tree Bank full parser (Lin et al., 2010) and the state-of-the-art neural network NeuralEDUSeg (Wang et al., 2018) and XLNet (Yang et al., 2019) models on the two-stage discourse segmentation and discourse relation recognition. The two-stage approach outperformed the PDTB parser by broad margin, i.e. the best achieved F1 scores of 21.2 % for PDTB parser vs 66.37% for NeuralEDUSeg and XLNet models. Neural network models were fine-tuned and evaluated on the argumentative corpus showing a promising accuracy of 60.22%. The complete argument structures were reconstructed for further argumentation mining tasks. The reference Dagstuhl argumentative corpus containing 2,222 elementary discourse unit pairs annotated with the top-level and fine-grained PDTB relations will be released to the research community.

We present our submission and results for SemEval-2020 Task 12: Multilingual Offensive Language Identification in Social Media (OffensEval 2020) where we participated in offensive tweet classification tasks in English, Arabic, Greek, Turkish and Danish. Our approach included classical machine learning architectures such as support vector machines and logistic regression combined in an ensemble with a multilingual transformer-based model (XLM-R). The transformer model is trained on all languages combined in order to create a fully multilingual model which can leverage knowledge between languages. The machine learning model hyperparameters are fine-tuned and the statistically best performing ones included in the final ensemble.

Deep neural networks have been employed for various spoken language recognition tasks, including tasks that are multilingual by definition such as spoken language identification (LID). In this paper, we present a neural model for Slavic language identification in speech signals and analyze its emergent representations to investigate whether they reflect objective measures of language relatedness or non-linguists’ perception of language similarity. While our analysis shows that the language representation space indeed captures language relatedness to a great extent, we find perceptual confusability to be the best predictor of the language representation similarity.

Fine-tuning pre-trained contextualized embedding models has become an integral part of the NLP pipeline. At the same time, probing has emerged as a way to investigate the linguistic knowledge captured by pre-trained models. Very little is, however, understood about how fine-tuning affects the representations of pre-trained models and thereby the linguistic knowledge they encode. This paper contributes towards closing this gap. We study three different pre-trained models: BERT, RoBERTa, and ALBERT, and investigate through sentence-level probing how fine-tuning affects their representations. We find that for some probing tasks fine-tuning leads to substantial changes in accuracy, possibly suggesting that fine-tuning introduces or even removes linguistic knowledge from a pre-trained model. These changes, however, vary greatly across different models, fine-tuning and probing tasks. Our analysis reveals that while fine-tuning indeed changes the representations of a pre-trained model and these changes are typically larger for higher layers, only in very few cases, fine-tuning has a positive effect on probing accuracy that is larger than just using the pre-trained model with a strong pooling method. Based on our findings, we argue that both positive and negative effects of fine-tuning on probing require a careful interpretation.

With the increase in the use of AI systems, a need for explanation systems arises. Building an explanation system requires a definition of explanation. However, the natural language term explanation is difficult to define formally as it includes multiple perspectives from different domains such as psychology, philosophy, and cognitive sciences. We study multiple perspectives and aspects of explainability of recommendations or predictions made by AI systems, and provide a generic definition of explanation. The proposed definition is ambitious and challenging to apply. With the intention to bridge the gap between theory and application, we also propose a possible architecture of an automated explanation system based on our definition of explanation.

Transformer-based language models achieve high performance on various tasks, but we still lack understanding of the kind of linguistic knowledge they learn and rely on. We evaluate three models (BERT, RoBERTa, and ALBERT), testing their grammatical and semantic knowledge by sentence-level probing, diagnostic cases, and masked prediction tasks. We focus on relative clauses (in American English) as a complex phenomenon needing contextual information and antecedent identification to be resolved. Based on a naturalistic dataset, probing shows that all three models indeed capture linguistic knowledge about grammaticality, achieving high performance.Evaluation on diagnostic cases and masked prediction tasks considering fine-grained linguistic knowledge, however, shows pronounced model-specific weaknesses especially on semantic knowledge, strongly impacting models’ performance. Our results highlight the importance of (a)model comparison in evaluation task and (b) building up claims of model performance and the linguistic knowledge they capture beyond purely probing-based evaluations.

The neural attention model has achieved great success in data-to-text generation tasks. Though usually excelling at producing fluent text, it suffers from the problem of information missing, repetition and “hallucination”. Due to the black-box nature of the neural attention architecture, avoiding these problems in a systematic way is non-trivial. To address this concern, we propose to explicitly segment target text into fragment units and align them with their data correspondences. The segmentation and correspondence are jointly learned as latent variables without any human annotations. We further impose a soft statistical constraint to regularize the segmental granularity. The resulting architecture maintains the same expressive power as neural attention models, while being able to generate fully interpretable outputs with several times less computational cost. On both E2E and WebNLG benchmarks, we show the proposed model consistently outperforms its neural attention counterparts.

Multilingual transformer models like mBERT and XLM-RoBERTa have obtained great improvements for many NLP tasks on a variety of languages. However, recent works also showed that results from high-resource languages could not be easily transferred to realistic, low-resource scenarios. In this work, we study trends in performance for different amounts of available resources for the three African languages Hausa, isiXhosa and on both NER and topic classification. We show that in combination with transfer learning or distant supervision, these models can achieve with as little as 10 or 100 labeled sentences the same performance as baselines with much more supervised training data. However, we also find settings where this does not hold. Our discussions and additional experiments on assumptions such as time and hardware restrictions highlight challenges and opportunities in low-resource learning.

A lot of real-world phenomena are complex and cannot be captured by single task annotations. This causes a need for subsequent annotations, with interdependent questions and answers describing the nature of the subject at hand. Even in the case a phenomenon is easily captured by a single task, the high specialisation of most annotation tools can result in having to switch to another tool if the task only slightly changes. We introduce HUMAN, a novel web-based annotation tool that addresses the above problems by a) covering a variety of annotation tasks on both textual and image data, and b) the usage of an internal deterministic state machine, allowing the researcher to chain different annotation tasks in an interdependent manner. Further, the modular nature of the tool makes it easy to define new annotation tasks and integrate machine learning algorithms e.g., for active learning. HUMAN comes with an easy-to-use graphical user interface that simplifies the annotation task and management.

In air traffic control, assistant systems support air traffic controllers in their work. To improve the reactivity and accuracy of the assistant, automatic speech recognition can monitor the commands uttered by the controller. However, to provide sufficient training data for the speech recognition system, many hours of air traffic communications have to be transcribed and semantically annotated. For this purpose we developed the annotation tool ATC-ANNO. It provides a number of features to support the annotator in their task, such as auto-complete suggestions for semantic tags, access to preliminary speech recognition predictions, syntax highlighting and consistency indicators. Its core assistive feature, however, is its ability to automatically generate semantic annotations. Although it is based on a simple hand-written finite state grammar, it is also able to annotate sentences that deviate from this grammar. We evaluate the impact of different features on annotator efficiency and find that automatic annotation allows annotators to cover four times as many utterances in the same time.

Fine-tuning pre-trained contextualized embedding models has become an integral part of the NLP pipeline. At the same time, probing has emerged as a way to investigate the linguistic knowledge captured by pre-trained models. Very little is, however, understood about how fine-tuning affects the representations of pre-trained models and thereby the linguistic knowledge they encode. This paper contributes towards closing this gap. We study three different pre-trained models: BERT, RoBERTa, and ALBERT, and investigate through sentence-level probing how fine-tuning affects their representations. We find that for some probing tasks fine-tuning leads to substantial changes in accuracy, possibly suggesting that fine-tuning introduces or even removes linguistic knowledge from a pre-trained model. These changes, however, vary greatly across different models, fine-tuning and probing tasks. Our analysis reveals that while fine-tuning indeed changes the representations of a pre-trained model and these changes are typically larger for higher layers, only in very few cases, fine-tuning has a positive effect on probing accuracy that is larger than just using the pre-trained model with a strong pooling method. Based on our findings, we argue that both positive and negative effects of fine-tuning on probing require a careful interpretation.

Research on hate speech classification has received increased attention. In real-life scenarios, a small amount of labeled hate speech data is available to train a reliable classifier. Semi-supervised learning takes advantage of a small amount of labeled data and a large amount of unlabeled data. In this paper, label propagation-based semi-supervised learning is explored for the task of hate speech classification. The quality of labeling the unlabeled set depends on the input representations. In this work, we show that pre-trained representations are label agnostic, and when used with label propagation yield poor results. Neural network-based fine-tuning can be adopted to learn task-specific representations using a small amount of labeled data. We show that fully fine-tuned representations may not always be the best representations for the label propagation and intermediate representations may perform better in a semi-supervised setup.

Many text generation tasks naturally contain two steps: content selection and surface realization. Current neural encoder-decoder models conflate both steps into a black-box architecture. As a result, the content to be described in the text cannot be explicitly controlled. This paper tackles this problem by decoupling content selection from the decoder. The decoupled content selection is human interpretable, whose value can be manually manipulated to control the content of generated text. The model can be trained end-to-end without human annotations by maximizing a lower bound of the marginal likelihood. We further propose an effective way to trade-off between performance and controllability with a single adjustable hyperparameter. In both data-to-text and headline generation tasks, our model achieves promising results, paving the way for controllable content selection in text generation.

In low-resource settings, the performance of supervised labeling models can be improved with automatically annotated or distantly supervised data, which is cheap to create but often noisy. Previous works have shown that significant improvements can be reached by injecting information about the confusion between clean and noisy labels in this additional training data into the classifier training. However, for noise estimation, these approaches either do not take the input features (in our case word embeddings) into account, or they need to learn the noise modeling from scratch which can be difficult in a low-resource setting. We propose to cluster the training data using the input features and then compute different confusion matrices for each cluster. To the best of our knowledge, our approach is the first to leverage feature-dependent noise modeling with pre-initialized confusion matrices. We evaluate on low-resource named entity recognition settings in several languages, showing that our methods improve upon other confusion-matrix based methods by up to 9%.

Pointer Generators have been the de facto standard for modern summarization systems. However, this architecture faces two major drawbacks: Firstly, the pointer is limited to copying the exact words while ignoring possible inflections or abstractions, which restricts its power of capturing richer latent alignment. Secondly, the copy mechanism results in a strong bias towards extractive generations, where most sentences are produced by simply copying from the source text. In this paper, we address these problems by allowing the model to “edit” pointed tokens instead of always hard copying them. The editing is performed by transforming the pointed word vector into a target space with a learned relation embedding. On three large-scale summarization dataset, we show the model is able to (1) capture more latent alignment relations than exact word matches, (2) improve word alignment accuracy, allowing for better model interpretation and controlling, (3) generate higher-quality summaries validated by both qualitative and quantitative evaluations and (4) bring more abstraction to the generated summaries.

This work explores cross-lingual transfer learning (TL) for named entity recognition, focusing on bootstrapping Japanese from English. A deep neural network model is adopted and the best combination of weights to transfer is extensively investigated. Moreover, a novel approach is presented that overcomes linguistic differences between this language pair by romanizing a portion of the Japanese input. Experiments are conducted on external datasets, as well as internal large-scale real-world ones. Gains with TL are achieved for all evaluated cases. Finally, the influence on TL of the target dataset size and of the target tagset distribution is further investigated.

In this paper, we address the problem of effectively self-training neural networks in a low-resource setting. Self-training is frequently used to automatically increase the amount of training data. However, in a low-resource scenario, it is less effective due to unreliable annotations created using self-labeling of unlabeled data. We propose to combine self-training with noise handling on the self-labeled data. Directly estimating noise on the combined clean training set and self-labeled data can lead to corruption of the clean data and hence, performs worse. Thus, we propose the Clean and Noisy Label Neural Network which trains on clean and noisy self-labeled data simultaneously by explicitly modelling clean and noisy labels separately. In our experiments on Chunking and NER, this approach performs more robustly than the baselines. Complementary to this explicit approach, noise can also be handled implicitly with the help of an auxiliary learning task. To such a complementary approach, our method is more beneficial than other baseline methods and together provides the best performance overall.

Languages may be differently distant from each other and their mutual intelligibility may be asymmetric. In this paper we introduce incom.py, a toolbox for calculating linguistic distances and asymmetries between related languages. incom.py allows linguist experts to quickly and easily perform statistical analyses and compare those with experimental results. We demonstrate the efficacy of incom.py in an incomprehension experiment on two Slavic languages: Bulgarian and Russian. Using incom.py we were able to validate three methods to measure linguistic distances and asymmetries: Levenshtein distance, word adaptation surprisal, and conditional entropy as predictors of success in a reading intercomprehension experiment.

The processing of medical information is not a trivial task for medical non-experts. The paper presents an artificial assistant designed to facilitate a reliable access to medical online contents. Interactions are modelled as doctor-patient Question Answering sessions within a pre-operative patient education scenario where the system addresses patient’s information needs explaining medical events and procedures. This implies an accurate medical information extraction from and reasoning with available medical knowledge and large amounts of unstructured multilingual online data. Bridging the gap between medical knowledge and data, we explore a language-agnostic approach to medical concepts mining from the standard terminologies, and the data-driven collection of the corresponding seed terms in a distant supervision setting for German. Experimenting with different terminologies, features and term matching strategies, we achieved a promising F-score of 0.91 on the medical term extraction task. The concepts and terms are used to search and retrieve definitions from the verified online free resources. The proof-of-concept definition retrieval system is designed and evaluated showing promising results, acceptable by humans in 92% of cases.

Popular word embedding methods such as word2vec and GloVe assign a single vector representation to each word, even if a word has multiple distinct meanings. Multi-sense embeddings instead provide different vectors for each sense of a word. However, they typically cannot serve as a drop-in replacement for conventional single-sense embeddings, because the correct sense vector needs to be selected for each word. In this work, we study the effect of multi-sense embeddings on the task of reverse dictionaries. We propose a technique to easily integrate them into an existing neural network architecture using an attention mechanism. Our experiments demonstrate that large improvements can be obtained when employing multi-sense embeddings both in the input sequence as well as for the target representation. An analysis of the sense distributions and of the learned attention is provided as well.

We apply hyperbolic embeddings to trace the dynamics of change of conceptual-semantic relationships in a large diachronic scientific corpus (200 years). Our focus is on emerging scientific fields and the increasingly specialized terminology establishing around them. Reproducing high-quality hierarchical structures such as WordNet on a diachronic scale is a very difficult task. Hyperbolic embeddings can map partial graphs into low dimensional, continuous hierarchical spaces, making more explicit the latent structure of the input. We show that starting from simple lists of word pairs (rather than a list of entities with directional links) it is possible to build diachronic hierarchical semantic spaces which allow us to model a process towards specialization for selected scientific fields.

Manually labeled corpora are expensive to create and often not available for low-resource languages or domains. Automatic labeling approaches are an alternative way to obtain labeled data in a quicker and cheaper way. However, these labels often contain more errors which can deteriorate a classifier’s performance when trained on this data. We propose a noise layer that is added to a neural network architecture. This allows modeling the noise and train on a combination of clean and noisy data. We show that in a low-resource NER task we can improve performance by up to 35% by using additional, noisy data and handling the noise.

Many natural and formal languages contain words or symbols that require a matching counterpart for making an expression well-formed. The combination of opening and closing brackets is a typical example of such a construction. Due to their commonness, the ability to follow such rules is important for language modeling. Currently, recurrent neural networks (RNNs) are extensively used for this task. We investigate whether they are capable of learning the rules of opening and closing brackets by applying them to synthetic Dyck languages that consist of different types of brackets. We provide an analysis of the statistical properties of these languages as a baseline and show strengths and limits of Elman-RNNs, GRUs and LSTMs in experiments on random samples of these languages. In terms of perplexity and prediction accuracy, the RNNs get close to the theoretical baseline in most cases.

Developing conventional natural language generation systems requires extensive attention from human experts in order to craft complex sets of sentence planning rules. We propose a Bayesian nonparametric approach to learn sentence planning rules by inducing synchronous tree substitution grammars for pairs of text plans and morphosyntactically-specified dependency trees. Our system is able to learn rules which can be used to generate novel texts after training on small datasets.

Sequence-to-Sequence (seq2seq) models have become overwhelmingly popular in building end-to-end trainable dialogue systems. Though highly efficient in learning the backbone of human-computer communications, they suffer from the problem of strongly favoring short generic responses. In this paper, we argue that a good response should smoothly connect both the preceding dialogue history and the following conversations. We strengthen this connection by mutual information maximization. To sidestep the non-differentiability of discrete natural language tokens, we introduce an auxiliary continuous code space and map such code space to a learnable prior distribution for generation purpose. Experiments on two dialogue datasets validate the effectiveness of our model, where the generated responses are closely related to the dialogue context and lead to more interactive conversations.

Word embeddings are high-dimensional vector representations of words and are thus difficult to interpret. In order to deal with this, we introduce an unsupervised parameter free method for creating a hierarchical graphical clustering of the full ensemble of word vectors and show that this structure is a geometrically meaningful representation of the original relations between the words. This newly obtained representation can be used for better understanding and thus improving the embedding algorithm and exhibits semantic meaning, so it can also be utilized in a variety of language processing tasks like categorization or measuring similarity.

This paper describes a method to automatically create dialogue resources annotated with dialogue act information by reusing existing dialogue corpora. Numerous dialogue corpora are available for research purposes and many of them are annotated with dialogue act information that captures the intentions encoded in user utterances. Annotated dialogue resources, however, differ in various respects: data collection settings and modalities used, dialogue task domains and scenarios (if any) underlying the collection, number and roles of dialogue participants involved and dialogue act annotation schemes applied. The presented study encompasses three phases of data-driven investigation. We, first, assess the importance of various types of features and their combinations for effective cross-domain dialogue act classification. Second, we establish the best predictive model comparing various cross-corpora training settings. Finally, we specify models adaptation procedures and explore late fusion approaches to optimize the overall classification decision taking process. The proposed methodology accounts for empirically motivated and technically sound classification procedures that may reduce annotation and training costs significantly.

In an intercomprehension scenario, typically a native speaker of language L1 is confronted with output from an unknown, but related language L2. In this setting, the degree to which the receiver recognizes the unfamiliar words greatly determines communicative success. Despite exhibiting great string-level differences, cognates may be recognized very successfully if the receiver is aware of regular correspondences which allow to transform the unknown word into its familiar form. Modeling L1-L2 intercomprehension then requires the identification of all the regular correspondences between languages L1 and L2. We here present a set of linguistic orthographic correspondences manually compiled from comparative linguistics literature along with a set of statistically-inferred suggestions for correspondence rules. In order to do statistical inference, we followed the Minimum Description Length principle, which proposes to choose those rules which are most effective at describing the data. Our statistical model was able to reproduce most of our linguistic correspondences (88.5% for Czech-Polish and 75.7% for Bulgarian-Russian) and furthermore allowed to easily identify many more non-trivial correspondences which also cover aspects of morphology.

Training good word embeddings requires large amounts of data. Out-of-vocabulary words will still be encountered at test-time, leaving these words without embeddings. To overcome this lack of embeddings for rare words, existing methods leverage morphological features to generate embeddings. While the existing methods use computationally-intensive rule-based (Soricut and Och, 2015) or tool-based (Botha and Blunsom, 2014) morphological analysis to generate embeddings, our system applies a computationally-simpler sub-word search on words that have existing embeddings. Embeddings of the sub-word search results are then combined using string similarity functions to generate rare word embeddings. We augmented pre-trained word embeddings with these novel embeddings and evaluated on a rare word similarity task, obtaining up to 3 times improvement in correlation over the original set of embeddings. Applying our technique to embeddings trained on larger datasets led to on-par performance with the existing state-of-the-art for this task. Additionally, while analysing augmented embeddings in a log-bilinear language model, we observed up to 50% reduction in rare word perplexity in comparison to other more complex language models.

This paper describes the data collection and annotation carried out within the DBOX project ( Eureka project, number E! 7152). This project aims to develop interactive games based on spoken natural language human-computer dialogues, in 3 European languages: English, German and French. We collect the DBOX data continuously. We first start with human-human Wizard of Oz experiments to collect human-human data in order to model natural human dialogue behaviour, for better understanding of phenomena of human interactions and predicting interlocutors actions, and then replace the human Wizard by an increasingly advanced dialogue system, using evaluation data for system improvement. The designed dialogue system relies on a Question-Answering (QA) approach, but showing truly interactive gaming behaviour, e.g., by providing feedback, managing turns and contact, producing social signals and acts, e.g., encouraging vs. downplaying, polite vs. rude, positive vs. negative attitude towards players or their actions, etc. The DBOX dialogue corpus has required substantial investment. We expect it to have a great impact on the rest of the project. The DBOX project consortium will continue to maintain the corpus and to take an interest in its growth, e.g., expand to other languages. The resulting corpus will be publicly released.

We present a gold standard for semantic relation extraction in the food domain for German. The relation types that we address are motivated by scenarios for which IT applications present a commercial potential, such as virtual customer advice in which a virtual agent assists a customer in a supermarket in finding those products that satisfy their needs best. Moreover, we focus on those relation types that can be extracted from natural language text corpora, ideally content from the internet, such as web forums, that are easy to retrieve. A typical relation type that meets these requirements are pairs of food items that are usually consumed together. Such a relation type could be used by a virtual agent to suggest additional products available in a shop that would potentially complement the items a customer has already in their shopping cart. Our gold standard comprises structural data, i.e. relation tables, which encode relation instances. These tables are vital in order to evaluate natural language processing systems that extract those relations.

In this paper we explore a task-driven approach to interfacing NLP components, where language processing is guided by the end-task that each application requires. The core idea is to generalize feature values into feature value distributions, representing under-specified feature values, and to fit linguistic pipelines with a back-channel of specification requests through which subsequent components can declare to preceding ones the importance of narrowing the value distribution of particular features that are critical for the current task.

Creating more fine-grained annotated data than previously relevent document sets is important for evaluating individual components in automatic question answering systems. In this paper, we describe using the Amazon's Mechanical Turk (AMT) to judge whether paragraphs in relevant documents answer corresponding list questions in TREC QA track 2004. Based on AMT results, we build a collection of 1300 gold-standard supporting paragraphs for list questions. Our online experiments suggested that recruiting more people per task assures better annotation quality. In order to learning true labels from AMT annotations, we investigated three approaches on two datasets with different levels of annotation errors. Experimental studies show that the Naive Bayesian model and EM-based GLAD model can generate results highly agreeing with gold-standard annotations, and dominate significantly over the majority voting method for true label learning. We also suggested setting higher HIT approval rate to assure better online annotation quality, which leads to better performance of learning methods.

In recent years, text classification in sentiment analysis has mostly focused on two types of classification, the distinction between objective and subjective text, i.e. subjectivity detection, and the distinction between positive and negative subjective text, i.e. polarity classification. So far, there has been little work examining the distinction between definite polar subjectivity and indefinite polar subjectivity. While the former are utterances which can be categorized as either positive or negative, the latter cannot be categorized as either of these two categories. This paper presents a small set of domain independent features to detect indefinite polar sentences. The features reflect the linguistic structure underlying these types of utterances. We give evidence for the effectiveness of these features by incorporating them into an unsupervised rule-based classifier for sentence-level analysis and compare its performance with supervised machine learning classifiers, i.e. Support Vector Machines (SVMs) and Nearest Neighbor Classifier (kNN). The data used for the experiments are web-reviews collected from three different domains.

Named Entity Recognition is a relatively well-understood NLP task, with many publicly available training resources and software for processing English data. Other languages tend to be underserved in this area. For German, CoNLL-2003 Shared Task provided training data, but there are no publicly available, ready-to-use tools. We fill this gap and develop a German NER system with state-of-the-art performance. In addition to CoNLL 2003 labeled training data, we use two additional resources: (i) 32 million words of unlabeled news article text and (ii) infobox labels from German Wikipedia articles. From the unlabeled text we derive distributional word clusters. Then we use cluster membership features and Wikipedia infobox label features to train a supervised model on the labeled training data. This approach allows us to deal better with word-types unseen in the training data and achieve good performance on German with little engineering effort.

One problem of data-driven answer extraction in open-domain factoid question answering is that the class distribution of labeled training data is fairly imbalanced. In an ordinary training set, there are far more incorrect answers than correct answers. The class-imbalance is, thus, inherent to the classification task. It has a deteriorating effect on the performance of classifiers trained by standard machine learning algorithms. They usually have a heavy bias towards the majority class, i.e. the class which occurs most often in the training set. In this paper, we propose a method to tackle class imbalance by applying some form of cost-sensitive learning which is preferable to sampling. We present a simple but effective way of estimating the misclassification costs on the basis of class distribution. This approach offers three benefits. Firstly, it maintains the distribution of the classes of the labeled training data. Secondly, this form of meta-learning can be applied to a wide range of common learning algorithms. Thirdly, this approach can be easily implemented with the help of state-of-the-art machine learning software.

The growing interest in open-domain question answering is limited by the lack of evaluation and training resources. To overcome this resource bottleneck for German, we propose a novel methodology to acquire new question-answer pairs for system evaluation that relies on volunteer collaboration over the Internet. Utilizing Wikipedia, a popular free online encyclopedia available in several languages, we show that the data acquisition problem can be cast as a Web experiment. We present a Web-based annotation tool and carry out a distributed data collection experiment. The data gathered from the mostly anonymous contributors is compared to a similar dataset produced in-house by domain experts on the one hand, and the German questions from the from the CLEF QA 2004 effort on the other hand. Our analysis of the datasets suggests that using our novel method a medium-scale evaluation resource can be built at very small cost in a short period of time. The technique and software developed here is readily applicable to other languages where free online encyclopedias are available, and our resulting corpus is likewise freely available.