Question answering systems should help users to access knowledge on a broad range of topics and to answer a wide array of different questions. Most systems fall short of this expectation as they are only specialized in one particular setting, e.g., answering factual questions with Wikipedia data. To overcome this limitation, we propose composing multiple QA agents within a meta-QA system. We argue that there exist a wide range of specialized QA agents in literature. Thus, we address the central research question of how to effectively and efficiently identify suitable QA agents for any given question. We study both supervised and unsupervised approaches to address this challenge, showing that TWEAC - Transformer with Extendable Agent Classifiers - achieves the best performance overall with 94% accuracy. We provide extensive insights on the scalability of TWEAC, demonstrating that it scales robustly to over 100 QA agents with each providing just 1000 examples of questions they can answer.
Learning sentence embeddings often requires large amount of labeled data. However, for most tasks and domains, labeled data is seldom available and creating it is expensive. In this work, we present a new state-of-the-art unsupervised method based on pre-trained Transformers and Sequential Denoising Auto-Encoder (TSDAE) which outperforms previous approaches by up to 6.4 points. It can achieve up to 93.1% of the performance of in-domain supervised approaches. Further, we show that TSDAE is a strong pre-training method for learning sentence embeddings, significantly outperforming other approaches like Masked Language Model. A crucial shortcoming of previous studies is the narrow evaluation: Most work mainly evaluates on the single task of Semantic Textual Similarity (STS), which does not require any domain knowledge. It is unclear if these proposed methods generalize to other domains and tasks. We fill this gap and evaluate TSDAE and other recent approaches on four different datasets from heterogeneous domains.
Claim verification is the task of predicting the veracity of written statements against evidence. Previous large-scale datasets model the task as classification, ignoring the need to retrieve evidence, or are constructed for research purposes, and may not be representative of real-world needs. In this paper, we introduce a novel claim verification dataset with instances derived from search-engine queries, yielding 10,987 claims annotated with evidence that represent real-world information needs. For each claim, we annotate evidence from full Wikipedia articles with both section and sentence-level granularity. Our annotation allows comparison between two complementary approaches to verification: stance classification, and evidence extraction followed by entailment recognition. In our comprehensive evaluation, we find no significant difference in accuracy between these two approaches. This enables systems to use evidence extraction to summarize a rationale for an end-user while maintaining the accuracy when predicting a claim's veracity. With challenging claims and evidence documents containing hundreds of sentences, our dataset presents interesting challenges that are not captured in previous work -- evidenced through transfer learning experiments. We release code and data to support further research on this task.
Current state-of-the-art approaches to cross-modal retrieval process text and visual input jointly, relying on Transformer-based architectures with cross-attention mechanisms that attend over all words and objects in an image. While offering unmatched retrieval performance, such models: 1) are typically pretrained from scratch and thus less scalable, 2) suffer from huge retrieval latency and inefficiency issues, which makes them impractical in realistic applications. To address these crucial gaps towards both improved and efficient cross-modal retrieval, we propose a novel fine-tuning framework which turns any pretrained text-image multi-modal model into an efficient retrieval model. The framework is based on a cooperative retrieve-and-rerank approach which combines: 1) twin networks to separately encode all items of a corpus, enabling efficient initial retrieval, and 2) a cross-encoder component for a more nuanced (i.e., smarter) ranking of the retrieved small set of items. We also propose to jointly fine-tune the two components with shared weights, yielding a more parameter-efficient model. Our experiments on a series of standard cross-modal retrieval benchmarks in monolingual, multilingual, and zero-shot setups, demonstrate improved accuracy and huge efficiency benefits over the state-of-the-art cross-encoders.
Previous work on text generation from graph-structured data relies on pretrained language models (PLMs) and utilizes graph linearization heuristics rather than explicitly considering the graph structure. Efficiently encoding the graph structure in PLMs is challenging because they were pretrained on natural language, and modeling structured data may lead to catastrophic forgetting of distributional knowledge. In this paper, we propose StructAdapt, an adapter method to encode graph structure into PLMs. Contrary to prior work, StructAdapt effectively models interactions among the nodes based on the graph connectivity, only training graph structure-aware adapter parameters. In this way, we avoid catastrophic forgetting while maintaining the topological structure of the graph. We empirically show the benefits of explicitly encoding graph structure into PLMs using adapters and achieve state-of-the-art results on two AMR-to-text datasets, training only 5.1% of the PLM parameters.
Decision-making usually takes five steps: identifying the problem, collecting data, extracting evidence, identifying pro and con arguments, and making decisions. Focusing on extracting evidence, this paper presents a hybrid model that combines latent Dirichlet allocation and word embeddings to obtain external knowledge from structured and unstructured data. We study the task of sentence-level argument mining, as arguments mostly require some degree of world knowledge to be identified and understood. Given a topic and a sentence, the goal is to classify whether a sentence represents an argument in regard to the topic. We use a topic model to extract topic- and sentence-specific evidence from the structured knowledge base Wikidata, building a graph based on the cosine similarity between the entity word vectors of Wikidata and the vector of the given sentence. Also, we build a second graph based on topic-specific articles found via Google to tackle the general incompleteness of structured knowledge bases. Combining these graphs, we obtain a graph-based model which, as our evaluation shows, successfully capitalizes on both structured and unstructured data.
Text style transfer has gained increasing attention from the research community over the recent years. However, the proposed approaches vary in many ways, which makes it hard to assess the individual contribution of the model components. In style transfer, the most important component is the optimization technique used to guide the learning in the absence of parallel training data. In this work we empirically compare the dominant optimization paradigms which provide supervision signals during training: backtranslation, adversarial training and reinforcement learning. We find that backtranslation has model-specific limitations, which inhibits training style transfer models. Reinforcement learning shows the best performance gains, while adversarial training, despite its popularity, does not offer an advantage over the latter alternative. In this work we also experiment with Minimum Risk Training, a popular technique in the machine translation community, which, to our knowledge, has not been empirically evaluated in the task of style transfer. We fill this research gap and empirically show its efficacy.
In this work we provide a \textit{systematic empirical comparison} of pretrained multilingual language models versus their monolingual counterparts with regard to their monolingual task performance. We study a set of nine typologically diverse languages with readily available pretrained monolingual models on a set of five diverse monolingual downstream tasks. We first establish if a gap between the multilingual and the corresponding monolingual representation of that language exists, and subsequently investigate the reason for a performance difference. To disentangle the impacting variables, we train new monolingual models on the same data, but with different tokenizers, both the monolingual and the multilingual version. We find that while the pretraining data size is an important factor, the designated tokenizer of the monolingual model plays an equally important role in the downstream performance. Our results show that languages which are adequately represented in the multilingual model's vocabulary exhibit negligible performance decreases over their monolingual counterparts. We further find that replacing the original multilingual tokenizer with the specialized monolingual tokenizer improves the downstream performance of the multilingual model for almost every task and language.
The ability to reason about multiple references to a given entity is essential for natural language understanding and has been long studied in NLP. In recent years, as the format of Question Answering (QA) became a standard for machine reading comprehension (MRC), there have been data collection efforts, e.g., Dasigi et al. (2019), that attempt to evaluate the ability of MRC models to reason about coreference. However, as we show, coreference reasoning in MRC is a greater challenge than was earlier thought; MRC datasets do not reflect the natural distribution and, consequently, the challenges of coreference reasoning. Specifically, success on these datasets does not reflect a model's proficiency in coreference reasoning. We propose a methodology for creating reading comprehension datasets that better reflect the challenges of coreference reasoning and use it to show that state-of-the-art models still struggle with these phenomena. Furthermore, we develop an effective way to use naturally occurring coreference phenomena from annotated coreference resolution datasets when training MRC models. This allows us to show an improvement in the coreference reasoning abilities of state-of-the-art models across various MRC datasets. We will release all the code and the resulting dataset at https://github.com/UKPLab/coref-reasoning-in-qa.
Massively multilingual language models such as multilingual BERT (mBERT) and XLM-R offer state-of-the-art cross-lingual transfer performance on a range of NLP tasks. However, due to their limited capacity and large differences in pretraining data, there is a profound performance gap between resource-rich and resource-poor target languages. The ultimate challenge is dealing with under-resourced languages not covered at all by the models, which are also written in scripts \textit{unseen} during pretraining. In this work, we propose a series of novel data-efficient methods that enable quick and effective adaptation of pretrained multilingual models to such low-resource languages and unseen scripts. Relying on matrix factorization, our proposed methods capitalize on the existing latent knowledge about multiple languages already available in the pretrained model's embedding matrix. Furthermore, we show that learning of the new dedicated embedding matrix in the target language can be improved by leveraging a small number of vocabulary items (i.e., the so-called \textit{lexically overlapping} tokens) shared between mBERT's and target language vocabulary. Our adaptation techniques offer substantial performance gains for languages with unseen scripts. We also demonstrate that they can also yield improvements for low-resource languages written in scripts covered by the pretrained model.