Computation and Language (cs.CL)

Abstract: Entity alignment is the task of identifying corresponding entities across different knowledge graphs (KGs). Although recent embedding-based entity alignment methods have shown significant advancements, they still struggle to fully utilize KG structural information. In this paper, we introduce FGWEA, an unsupervised entity alignment framework that leverages the Fused Gromov-Wasserstein (FGW) distance, allowing for a comprehensive comparison of entity semantics and KG structures within a joint optimization framework. To address the computational challenges associated with optimizing FGW, we devise a three-stage progressive optimization algorithm. It starts with a basic semantic embedding matching, proceeds to approximate cross-KG structural and relational similarity matching based on iterative updates of high-confidence entity links, and ultimately culminates in a global structural comparison between KGs. We perform extensive experiments on four entity alignment datasets covering 14 distinct KGs across five languages. Without any supervision or hyper-parameter tuning, FGWEA surpasses 21 competitive baselines, including cutting-edge supervised entity alignment methods. Our code is available at https://github.com/squareRoot3/FusedGW-Entity-Alignment.

Abstract: Large language models (LLMs) have shown surprisingly good performance in multilingual neural machine translation (MNMT) even when trained without parallel data. Yet, despite the fact that the amount of training data is gigantic, they still struggle with translating rare words, particularly for low-resource languages. Even worse, it is usually unrealistic to retrieve relevant demonstrations for in-context learning with low-resource languages on LLMs, which restricts the practical use of LLMs for translation -- how should we mitigate this problem? To this end, we present a novel method, CoD, which augments LLMs with prior knowledge with the chains of multilingual dictionaries for a subset of input words to elicit translation abilities for LLMs. Extensive experiments indicate that augmenting ChatGPT with CoD elicits large gains by up to 13x ChrF++ points for MNMT (3.08 to 42.63 for English to Serbian written in Cyrillic script) on FLORES-200 full devtest set. We further demonstrate the importance of chaining the multilingual dictionaries, as well as the superiority of CoD to few-shot demonstration for low-resource languages.

Abstract: We present the findings of SemEval-2023 Task 2 on Fine-grained Multilingual Named Entity Recognition (MultiCoNER 2). Divided into 13 tracks, the task focused on methods to identify complex fine-grained named entities (like WRITTENWORK, VEHICLE, MUSICALGRP) across 12 languages, in both monolingual and multilingual scenarios, as well as noisy settings. The task used the MultiCoNER V2 dataset, composed of 2.2 million instances in Bangla, Chinese, English, Farsi, French, German, Hindi, Italian., Portuguese, Spanish, Swedish, and Ukrainian. MultiCoNER 2 was one of the most popular tasks of SemEval-2023. It attracted 842 submissions from 47 teams, and 34 teams submitted system papers. Results showed that complex entity types such as media titles and product names were the most challenging. Methods fusing external knowledge into transformer models achieved the best performance, and the largest gains were on the Creative Work and Group classes, which are still challenging even with external knowledge. Some fine-grained classes proved to be more challenging than others, such as SCIENTIST, ARTWORK, and PRIVATECORP. We also observed that noisy data has a significant impact on model performance, with an average drop of 10% on the noisy subset. The task highlights the need for future research on improving NER robustness on noisy data containing complex entities.

Abstract: In real world applications, knowledge graphs (KG) are widely used in various domains (e.g. medical applications and dialogue agents). However, for fact verification, KGs have not been adequately utilized as a knowledge source. KGs can be a valuable knowledge source in fact verification due to their reliability and broad applicability. A KG consists of nodes and edges which makes it clear how concepts are linked together, allowing machines to reason over chains of topics. However, there are many challenges in understanding how these machine-readable concepts map to information in text. To enable the community to better use KGs, we introduce a new dataset, FactKG: Fact Verification via Reasoning on Knowledge Graphs. It consists of 108k natural language claims with five types of reasoning: One-hop, Conjunction, Existence, Multi-hop, and Negation. Furthermore, FactKG contains various linguistic patterns, including colloquial style claims as well as written style claims to increase practicality. Lastly, we develop a baseline approach and analyze FactKG over these reasoning types. We believe FactKG can advance both reliability and practicality in KG-based fact verification.

Abstract: The analysis of consumer sentiment, as expressed through reviews, can provide a wealth of insight regarding the quality of a product. While the study of sentiment analysis has been widely explored in many popular languages, relatively less attention has been given to the Bangla language, mostly due to a lack of relevant data and cross-domain adaptability. To address this limitation, we present BanglaBook, a large-scale dataset of Bangla book reviews consisting of 158,065 samples classified into three broad categories: positive, negative, and neutral. We provide a detailed statistical analysis of the dataset and employ a range of machine learning models to establish baselines including SVM, LSTM, and Bangla-BERT. Our findings demonstrate a substantial performance advantage of pre-trained models over models that rely on manually crafted features, emphasizing the necessity for additional training resources in this domain. Additionally, we conduct an in-depth error analysis by examining sentiment unigrams, which may provide insight into common classification errors in under-resourced languages like Bangla. Our codes and data are publicly available at https://github.com/mohsinulkabir14/BanglaBook.

Abstract: We show that the laws of autocorrelations decay in texts are closely related to applicability limits of language models. Using distributional semantics we empirically demonstrate that autocorrelations of words in texts decay according to a power law. We show that distributional semantics provides coherent autocorrelations decay exponents for texts translated to multiple languages. The autocorrelations decay in generated texts is quantitatively and often qualitatively different from the literary texts. We conclude that language models exhibiting Markov behavior, including large autoregressive language models, may have limitations when applied to long texts, whether analysis or generation.

Abstract: Continual few-shot relation extraction (RE) aims to continuously train a model for new relations with few labeled training data, of which the major challenges are the catastrophic forgetting of old relations and the overfitting caused by data sparsity. In this paper, we propose a new model, namely SCKD, to accomplish the continual few-shot RE task. Specifically, we design serial knowledge distillation to preserve the prior knowledge from previous models and conduct contrastive learning with pseudo samples to keep the representations of samples in different relations sufficiently distinguishable. Our experiments on two benchmark datasets validate the effectiveness of SCKD for continual few-shot RE and its superiority in knowledge transfer and memory utilization over state-of-the-art models.

Abstract: Continual relation extraction (RE) aims to learn constantly emerging relations while avoiding forgetting the learned relations. Existing works store a small number of typical samples to re-train the model for alleviating forgetting. However, repeatedly replaying these samples may cause the overfitting problem. We conduct an empirical study on existing works and observe that their performance is severely affected by analogous relations. To address this issue, we propose a novel continual extraction model for analogous relations. Specifically, we design memory-insensitive relation prototypes and memory augmentation to overcome the overfitting problem. We also introduce integrated training and focal knowledge distillation to enhance the performance on analogous relations. Experimental results show the superiority of our model and demonstrate its effectiveness in distinguishing analogous relations and overcoming overfitting.

Abstract: Though majority vote among annotators is typically used for ground truth labels in natural language processing, annotator disagreement in tasks such as hate speech detection may reflect differences among group opinions, not noise. Thus, a crucial problem in hate speech detection is whether a statement is offensive to the demographic group that it targets, which may constitute a small fraction of the annotator pool. We construct a model that predicts individual annotator ratings on potentially offensive text and combines this information with the predicted target group of the text to model the opinions of target group members. We show gains across a range of metrics, including raising performance over the baseline by 22% at predicting individual annotators' ratings and 33% at predicting variance among annotators, which provides a method of measuring model uncertainty downstream. We find that annotators' ratings can be predicted using their demographic information and opinions on online content, without the need to track identifying annotator IDs that link each annotator to their ratings. We also find that use of non-invasive survey questions on annotators' online experiences helps to maximize privacy and minimize unnecessary collection of demographic information when predicting annotators' opinions.

Abstract: Based on the remarkable achievements of pre-trained language models in abstractive summarization, the copying mechanism has proved helpful by improving the factuality, stability, and overall performance. This work proposes PROM, a new PhRase-level cOpying Mechanism that enhances attention on n-grams, which can be applied to zero-shot summarization with pre-training. PROM adds an indicator layer to explicitly pick up tokens in n-gram that can be copied from the source, and calculates an auxiliary loss for the copying prediction. Empirical studies show that PROM makes significant improvements in fine-tuning on benchmarks. In zero-shot setting, PROM is utilized in the self-supervised pre-training on raw corpora and provides new general baselines on a wide range of summarization datasets. Further analysis shows that PROM performs more reasonable copying and contributes to faithfulness.

Abstract: Context-dependent Text-to-SQL aims to translate multi-turn natural language questions into SQL queries. Despite various methods have exploited context-dependence information implicitly for contextual SQL parsing, there are few attempts to explicitly address the dependencies between current question and question context. This paper presents QURG, a novel Question Rewriting Guided approach to help the models achieve adequate contextual understanding. Specifically, we first train a question rewriting model to complete the current question based on question context, and convert them into a rewriting edit matrix. We further design a two-stream matrix encoder to jointly model the rewriting relations between question and context, and the schema linking relations between natural language and structured schema. Experimental results show that QURG significantly improves the performances on two large-scale context-dependent datasets SParC and CoSQL, especially for hard and long-turn questions.

Abstract: A salient characteristic of large pre-trained language models (PTLMs) is a remarkable improvement in their generalization capability and emergence of new capabilities with increasing model capacity and pre-training dataset size. Consequently, we are witnessing the development of enormous models pushing the state-of-the-art. It is, however, imperative to realize that this inevitably leads to prohibitively long training times, extortionate computing costs, and a detrimental environmental impact. Significant efforts are underway to make PTLM training more efficient through innovations in model architectures, training pipelines, and loss function design, with scant attention being paid to optimizing the utility of training data. The key question that we ask is whether it is possible to train PTLMs by employing only highly informative subsets of the training data while maintaining downstream performance? Building upon the recent progress in informative data subset selection, we show how we can employ submodular optimization to select highly representative subsets of the training corpora. Our results demonstrate that the proposed framework can be applied to efficiently train multiple PTLMs (BERT, BioBERT, GPT-2) using only a fraction of data while retaining up to $\sim99\%$ of the performance of the fully-trained models.

Abstract: This paper introduces a novel worker selection algorithm, enhancing annotation quality and reducing costs in challenging span-based sequence labeling tasks in Natural Language Processing (NLP). Unlike previous studies targeting simpler tasks, this study contends with the complexities of label interdependencies in sequence labeling tasks. The proposed algorithm utilizes a Combinatorial Multi-Armed Bandit (CMAB) approach for worker selection. The challenge of dealing with imbalanced and small-scale datasets, which hinders offline simulation of worker selection, is tackled using an innovative data augmentation method termed shifting, expanding, and shrinking (SES). The SES method is designed specifically for sequence labeling tasks. Rigorous testing on CoNLL 2003 NER and Chinese OEI datasets showcased the algorithm's efficiency, with an increase in F1 score up to 100.04% of the expert-only baseline, alongside cost savings up to 65.97%. The paper also encompasses a dataset-independent test emulating annotation evaluation through a Bernoulli distribution, which still led to an impressive 97.56% F1 score of the expert baseline and 59.88% cost savings. This research addresses and overcomes numerous obstacles in worker selection for complex NLP tasks.

Abstract: To advance the neural encoding of Portuguese (PT), and a fortiori the technological preparation of this language for the digital age, we developed a Transformer-based foundation model that sets a new state of the art in this respect for two of its variants, namely European Portuguese from Portugal (PT-PT) and American Portuguese from Brazil (PT-BR). To develop this encoder, which we named Albertina PT-*, a strong model was used as a starting point, DeBERTa, and its pre-training was done over data sets of Portuguese, namely over a data set we gathered for PT-PT and over the brWaC corpus for PT-BR. The performance of Albertina and competing models was assessed by evaluating them on prominent downstream language processing tasks adapted for Portuguese. Both Albertina PT-PT and PT-BR versions are distributed free of charge and under the most permissive license possible and can be run on consumer-grade hardware, thus seeking to contribute to the advancement of research and innovation in language technology for Portuguese.

Abstract: Kurdish Sign Language (KuSL) is the natural language of the Kurdish Deaf people. We work on automatic translation between spoken Kurdish and KuSL. Sign languages evolve rapidly and follow grammatical rules that differ from spoken languages. Consequently,those differences should be considered during any translation. We proposed an avatar-based automatic translation of Kurdish texts in the Sorani (Central Kurdish) dialect into the Kurdish Sign language. We developed the first parallel corpora for that pair that we use to train a Statistical Machine Translation (SMT) engine. We tested the outcome understandability and evaluated it using the Bilingual Evaluation Understudy (BLEU). Results showed 53.8% accuracy. Compared to the previous experiments in the field, the result is considerably high. We suspect the reason to be the similarity between the structure of the two pairs. We plan to make the resources publicly available under CC BY-NC-SA 4.0 license on the Kurdish-BLARK (https://kurdishblark.github.io/).

Abstract: Transformer architectures are complex and their use in NLP, while it has engendered many successes, makes their interpretability or explainability challenging. Recent debates have shown that attention maps and attribution methods are unreliable (Pruthi et al., 2019; Brunner et al., 2019). In this paper, we present some of their limitations and introduce COCKATIEL, which successfully addresses some of them. COCKATIEL is a novel, post-hoc, concept-based, model-agnostic XAI technique that generates meaningful explanations from the last layer of a neural net model trained on an NLP classification task by using Non-Negative Matrix Factorization (NMF) to discover the concepts the model leverages to make predictions and by exploiting a Sensitivity Analysis to estimate accurately the importance of each of these concepts for the model. It does so without compromising the accuracy of the underlying model or requiring a new one to be trained. We conduct experiments in single and multi-aspect sentiment analysis tasks and we show COCKATIEL's superior ability to discover concepts that align with humans' on Transformer models without any supervision, we objectively verify the faithfulness of its explanations through fidelity metrics, and we showcase its ability to provide meaningful explanations in two different datasets.

Abstract: This paper shines a light on the potential of definition-based semantic models for detecting idiomatic and semi-idiomatic multiword expressions (MWEs) in clinical terminology. Our study focuses on biomedical entities defined in the UMLS ontology and aims to help prioritize the translation efforts of these entities. In particular, we develop an effective tool for scoring the idiomaticity of biomedical MWEs based on the degree of similarity between the semantic representations of those MWEs and a weighted average of the representation of their constituents. We achieve this using a biomedical language model trained to produce similar representations for entity names and their definitions, called BioLORD. The importance of this definition-based approach is highlighted by comparing the BioLORD model to two other state-of-the-art biomedical language models based on Transformer: SapBERT and CODER. Our results show that the BioLORD model has a strong ability to identify idiomatic MWEs, not replicated in other models. Our corpus-free idiomaticity estimation helps ontology translators to focus on more challenging MWEs.

Abstract: This paper describes the approach of the THUIR team at the COLIEE 2023 Legal Case Entailment task. This task requires the participant to identify a specific paragraph from a given supporting case that entails the decision for the query case. We try traditional lexical matching methods and pre-trained language models with different sizes. Furthermore, learning-to-rank methods are employed to further improve performance. However, learning-to-rank is not very robust on this task. which suggests that answer passages cannot simply be determined with information retrieval techniques. Experimental results show that more parameters and legal knowledge contribute to the legal case entailment task. Finally, we get the third place in COLIEE 2023. The implementation of our method can be found at https://github.com/CSHaitao/THUIR-COLIEE2023.

Abstract: Suspense is an important tool in storytelling to keep readers engaged and wanting to read more. However, it has so far not been studied extensively in Computational Literary Studies. In this paper, we focus on one of the elements authors can use to build up suspense: dangerous situations. We introduce a corpus of texts annotated with dangerous situations, distinguishing between 7 types of danger. Additionally, we annotate parts of the text that describe fear experienced by a character, regardless of the actual presence of danger. We present experiments towards the automatic detection of these situations, finding that unsupervised baseline methods can provide valuable signals for the detection, but more complex methods are necessary for further analysis. Not unexpectedly, the description of danger and fear often relies heavily on the context, both local (e.g., situations where danger is only mentioned, but not actually present) and global (e.g., "storm" being used in a literal sense in an adventure novel, but metaphorically in a romance novel).

Abstract: While the Large Language Models (LLMs) dominate a majority of language understanding tasks, previous work shows that some of these results are supported by modelling spurious correlations of training datasets. Authors commonly assess model robustness by evaluating their models on out-of-distribution (OOD) datasets of the same task, but these datasets might share the bias of the training dataset. We propose a simple method for measuring a scale of models' reliance on any identified spurious feature and assess the robustness towards a large set of known and newly found prediction biases for various pre-trained models and debiasing methods in Question Answering (QA). We find that the reported OOD gains of debiasing methods can not be explained by mitigated reliance on biased features, suggesting that biases are shared among QA datasets. We further evidence this by measuring that performance of OOD models depends on bias features comparably to the ID model, motivating future work to refine the reports of LLMs' robustness to a level of known spurious features.

Abstract: Long-form question answering (LFQA) aims at answering complex, open-ended questions with detailed, paragraph-length responses. The de facto paradigm of LFQA necessitates two procedures: information retrieval, which searches for relevant supporting facts, and information synthesis, which integrates these facts into a coherent answer. In this paper, we introduce WebCPM, the first Chinese LFQA dataset. One unique feature of WebCPM is that its information retrieval is based on interactive web search, which engages with a search engine in real time. Following WebGPT, we develop a web search interface. We recruit annotators to search for relevant information using our interface and then answer questions. Meanwhile, the web search behaviors of our annotators would be recorded. In total, we collect 5,500 high-quality question-answer pairs, together with 14,315 supporting facts and 121,330 web search actions. We fine-tune pre-trained language models to imitate human behaviors for web search and to generate answers based on the collected facts. Our LFQA pipeline, built on these fine-tuned models, generates answers that are no worse than human-written ones in 32.5% and 47.5% of the cases on our dataset and DuReader, respectively.

Abstract: This paper describes our system on SemEval-2023 Task 10: Explainable Detection of Online Sexism (EDOS). This work aims to design an automatic system for detecting and classifying sexist content in online spaces. We propose a set of transformer-based pre-trained models with task-adaptive pretraining and ensemble learning. The main contributions of our system include analyzing the performance of different transformer-based pre-trained models and combining these models, as well as providing an efficient method using large amounts of unlabeled data for model adaptive pretraining. We have also explored several other strategies. On the test dataset, our system achieves F1-scores of 83%, 64%, and 47% on subtasks A, B, and C, respectively.

Abstract: African languages have far less in-language content available digitally, making it challenging for question answering systems to satisfy the information needs of users. Cross-lingual open-retrieval question answering (XOR QA) systems -- those that retrieve answer content from other languages while serving people in their native language -- offer a means of filling this gap. To this end, we create AfriQA, the first cross-lingual QA dataset with a focus on African languages. AfriQA includes 12,000+ XOR QA examples across 10 African languages. While previous datasets have focused primarily on languages where cross-lingual QA augments coverage from the target language, AfriQA focuses on languages where cross-lingual answer content is the only high-coverage source of answer content. Because of this, we argue that African languages are one of the most important and realistic use cases for XOR QA. Our experiments demonstrate the poor performance of automatic translation and multilingual retrieval methods. Overall, AfriQA proves challenging for state-of-the-art QA models. We hope that the dataset enables the development of more equitable QA technology.

Abstract: Despite the remarkable ability of large language models (LMs) to comprehend and generate language, they have a tendency to hallucinate and create factually inaccurate output. Augmenting LMs by retrieving information from external knowledge resources is one promising solution. Most existing retrieval-augmented LMs employ a retrieve-and-generate setup that only retrieves information once based on the input. This is limiting, however, in more general scenarios involving generation of long texts, where continually gathering information throughout the generation process is essential. There have been some past efforts to retrieve information multiple times while generating outputs, which mostly retrieve documents at fixed intervals using the previous context as queries. In this work, we provide a generalized view of active retrieval augmented generation, methods that actively decide when and what to retrieve across the course of the generation. We propose Forward-Looking Active REtrieval augmented generation (FLARE), a generic retrieval-augmented generation method which iteratively uses a prediction of the upcoming sentence to anticipate future content, which is then utilized as a query to retrieve relevant documents to regenerate the sentence if it contains low-confidence tokens. We test FLARE along with baselines comprehensively over 4 long-form knowledge-intensive generation tasks/datasets. FLARE achieves superior or competitive performance on all tasks, demonstrating the effectiveness of our method. Code and datasets are available at https://github.com/jzbjyb/FLARE.

Abstract: Lexical matching remains the de facto evaluation method for open-domain question answering (QA). Unfortunately, lexical matching fails completely when a plausible candidate answer does not appear in the list of gold answers, which is increasingly the case as we shift from extractive to generative models. The recent success of large language models (LLMs) for QA aggravates lexical matching failures since candidate answers become longer, thereby making matching with the gold answers even more challenging. Without accurate evaluation, the true progress in open-domain QA remains unknown. In this paper, we conduct a thorough analysis of various open-domain QA models, including LLMs, by manually evaluating their answers on a subset of NQ-open, a popular benchmark. Our assessments reveal that while the true performance of all models is significantly underestimated, the performance of the InstructGPT (zero-shot) LLM increases by nearly +60%, making it on par with existing top models, and the InstructGPT (few-shot) model actually achieves a new state-of-the-art on NQ-open. We also find that more than 50% of lexical matching failures are attributed to semantically equivalent answers. We further demonstrate that regex matching ranks QA models consistent with human judgments, although still suffering from unnecessary strictness. Finally, we demonstrate that automated evaluation models are a reasonable surrogate for lexical matching in some circumstances, but not for long-form answers generated by LLMs. The automated models struggle in detecting hallucinations in LLM answers and are thus unable to evaluate LLMs. At this time, there appears to be no substitute for human evaluation.

Abstract: The Smatch metric is a popular method for evaluating graph distances, as is necessary, for instance, to assess the performance of semantic graph parsing systems. However, we observe some issues in the metric that jeopardize meaningful evaluation. E.g., opaque pre-processing choices can affect results, and current graph-alignment solvers do not provide us with upper-bounds. Without upper-bounds, however, fair evaluation is not guaranteed. Furthermore, adaptions of Smatch for extended tasks (e.g., fine-grained semantic similarity) are spread out, and lack a unifying framework. For better inspection, we divide the metric into three modules: pre-processing, alignment, and scoring. Examining each module, we specify its goals and diagnose potential issues, for which we discuss and test mitigation strategies. For pre-processing, we show how to fully conform to annotation guidelines that allow structurally deviating but valid graphs. For safer and enhanced alignment, we show the feasibility of optimal alignment in a standard evaluation setup, and develop a lossless graph compression method that shrinks the search space and significantly increases efficiency. For improved scoring, we propose standardized and extended metric calculation of fine-grained sub-graph meaning aspects. Our code is available at https://github.com/flipz357/smatchpp

Abstract: Large language models (LLMs) demonstrate impressive multilingual capability, but their performance varies substantially across different languages. In this work, we introduce a simple yet effective method, called cross-lingual-thought prompting (XLT), to systematically improve the multilingual capability of LLMs. Specifically, XLT is a generic template prompt that stimulates cross-lingual and logical reasoning skills to enhance task performance across languages. We conduct comprehensive evaluations on 7 typical benchmarks related to reasoning, understanding, and generation tasks, covering both high-resource and low-resource languages. Experimental results show that XLT not only remarkably enhances the performance of various multilingual tasks but also significantly reduces the gap between the average performance and the best performance of each task in different languages. Notably, XLT brings over 10 points of average improvement in arithmetic reasoning and open-domain question-answering tasks.

Abstract: Subword segmenters like BPE operate as a preprocessing step in neural machine translation and other (conditional) language models. They are applied to datasets before training, so translation or text generation quality relies on the quality of segmentations. We propose a departure from this paradigm, called subword segmental machine translation (SSMT). SSMT unifies subword segmentation and MT in a single trainable model. It learns to segment target sentence words while jointly learning to generate target sentences. To use SSMT during inference we propose dynamic decoding, a text generation algorithm that adapts segmentations as it generates translations. Experiments across 6 translation directions show that SSMT improves chrF scores for morphologically rich agglutinative languages. Gains are strongest in the very low-resource scenario. SSMT also learns subwords that are closer to morphemes compared to baselines and proves more robust on a test set constructed for evaluating morphological compositional generalisation.

Abstract: Massively multilingual pretrained transformers (MMTs) have tremendously pushed the state of the art on multilingual NLP and cross-lingual transfer of NLP models in particular. While a large body of work leveraged MMTs to mine parallel data and induce bilingual document embeddings, much less effort has been devoted to training general-purpose (massively) multilingual document encoder that can be used for both supervised and unsupervised document-level tasks. In this work, we pretrain a massively multilingual document encoder as a hierarchical transformer model (HMDE) in which a shallow document transformer contextualizes sentence representations produced by a state-of-the-art pretrained multilingual sentence encoder. We leverage Wikipedia as a readily available source of comparable documents for creating training data, and train HMDE by means of a cross-lingual contrastive objective, further exploiting the category hierarchy of Wikipedia for creation of difficult negatives. We evaluate the effectiveness of HMDE in two arguably most common and prominent cross-lingual document-level tasks: (1) cross-lingual transfer for topical document classification and (2) cross-lingual document retrieval. HMDE is significantly more effective than (i) aggregations of segment-based representations and (ii) multilingual Longformer. Crucially, owing to its massively multilingual lower transformer, HMDE successfully generalizes to languages unseen in document-level pretraining. We publicly release our code and models at https://github.com/ogaloglu/pre-training-multilingual-document-encoders .