Annemarie Friedrich

Sophie Henning, William Beluch, Alexander Fraser et al.

Many natural language processing (NLP) tasks are naturally imbalanced, as some target categories occur much more frequently than others in the real world. In such scenarios, current NLP models still tend to perform poorly on less frequent classes. Addressing class imbalance in NLP is an active research topic, yet, finding a good approach for a particular task and imbalance scenario is difficult. With this survey, the first overview on class imbalance in deep-learning based NLP, we provide guidance for NLP researchers and practitioners dealing with imbalanced data. We first discuss various types of controlled and real-world class imbalance. Our survey then covers approaches that have been explicitly proposed for class-imbalanced NLP tasks or, originating in the computer vision community, have been evaluated on them. We organize the methods by whether they are based on sampling, data augmentation, choice of loss function, staged learning, or model design. Finally, we discuss open problems such as dealing with multi-label scenarios, and propose systematic benchmarking and reporting in order to move forward on this problem as a community.

Qingyu Chen, Alexis Allot, Robert Leaman et al.

The COVID-19 pandemic has been severely impacting global society since December 2019. Massive research has been undertaken to understand the characteristics of the virus and design vaccines and drugs. The related findings have been reported in biomedical literature at a rate of about 10,000 articles on COVID-19 per month. Such rapid growth significantly challenges manual curation and interpretation. For instance, LitCovid is a literature database of COVID-19-related articles in PubMed, which has accumulated more than 200,000 articles with millions of accesses each month by users worldwide. One primary curation task is to assign up to eight topics (e.g., Diagnosis and Treatment) to the articles in LitCovid. Despite the continuing advances in biomedical text mining methods, few have been dedicated to topic annotations in COVID-19 literature. To close the gap, we organized the BioCreative LitCovid track to call for a community effort to tackle automated topic annotation for COVID-19 literature. The BioCreative LitCovid dataset, consisting of over 30,000 articles with manually reviewed topics, was created for training and testing. It is one of the largest multilabel classification datasets in biomedical scientific literature. 19 teams worldwide participated and made 80 submissions in total. Most teams used hybrid systems based on transformers. The highest performing submissions achieved 0.8875, 0.9181, and 0.9394 for macro F1-score, micro F1-score, and instance-based F1-score, respectively. The level of participation and results demonstrate a successful track and help close the gap between dataset curation and method development. The dataset is publicly available via https://ftp.ncbi.nlm.nih.gov/pub/lu/LitCovid/biocreative/ for benchmarking and further development.

Annemarie Friedrich, Nianwen Xue, Alexis Palmer

Aspectual meaning refers to how the internal temporal structure of situations is presented. This includes whether a situation is described as a state or as an event, whether the situation is finished or ongoing, and whether it is viewed as a whole or with a focus on a particular phase. This survey gives an overview of computational approaches to modeling lexical and grammatical aspect along with intuitive explanations of the necessary linguistic concepts and terminology. In particular, we describe the concepts of stativity, telicity, habituality, perfective and imperfective, as well as influential inventories of eventuality and situation types. We argue that because aspect is a crucial component of semantics, especially when it comes to reporting the temporal structure of situations in a precise way, future NLP approaches need to be able to handle and evaluate it systematically in order to achieve human-level language understanding.

Sophie Henning, Nicole Macher, Stefan Grünewald et al.

Modal verbs (e.g., "can", "should", or "must") occur highly frequently in scientific articles. Decoding their function is not straightforward: they are often used for hedging, but they may also denote abilities and restrictions. Understanding their meaning is important for various NLP tasks such as writing assistance or accurate information extraction from scientific text. To foster research on the usage of modals in this genre, we introduce the MIST (Modals In Scientific Text) dataset, which contains 3737 modal instances in five scientific domains annotated for their semantic, pragmatic, or rhetorical function. We systematically evaluate a set of competitive neural architectures on MIST. Transfer experiments reveal that leveraging non-scientific data is of limited benefit for modeling the distinctions in MIST. Our corpus analysis provides evidence that scientific communities differ in their usage of modal verbs, yet, classifiers trained on scientific data generalize to some extent to unseen scientific domains.

Timo Pierre Schrader, Teresa Bürkle, Sophie Henning et al.

Scientific publications follow conventionalized rhetorical structures. Classifying the Argumentative Zone (AZ), e.g., identifying whether a sentence states a Motivation, a Result or Background information, has been proposed to improve processing of scholarly documents. In this work, we adapt and extend this idea to the domain of materials science research. We present and release a new dataset of 50 manually annotated research articles. The dataset spans seven sub-topics and is annotated with a materials-science focused multi-label annotation scheme for AZ. We detail corpus statistics and demonstrate high inter-annotator agreement. Our computational experiments show that using domain-specific pre-trained transformer-based text encoders is key to high classification performance. We also find that AZ categories from existing datasets in other domains are transferable to varying degrees.

Valentin Knappich, Simon Razniewski, Annemarie Friedrich

Automatic simplification can help laypeople to comprehend complex scientific text. Language models are frequently applied to this task by translating from complex to simple language. In this paper, we describe our system based on Llama 2, which ranked first in the PLABA shared task addressing the simplification of biomedical text. We find that the large portion of shared tokens between input and output leads to weak training signals and conservatively editing models. To mitigate these issues, we propose sentence-level and token-level loss weights. They give higher weight to modified tokens, indicated by edit distance and edit operations, respectively. We conduct an empirical evaluation on the PLABA dataset and find that both approaches lead to simplifications closer to those created by human annotators (+1.8% / +3.5% SARI), simpler language (-1 / -1.1 FKGL) and more edits (1.6x / 1.8x edit distance) compared to the same model fine-tuned with standard cross entropy. We furthermore show that the hyperparameter $λ$ in token-level loss weights can be used to control the edit distance and the simplicity level (FKGL).

Timo Pierre Schrader, Lukas Lange, Tobias Kaminski et al.

The rise of large language models (LLMs) has sparked interest in coding assistants. While general-purpose programming languages are well supported, generating code for domain-specific languages remains a challenging problem for LLMs. In this paper, we focus on the LLM-based generation of code for Answer Set Programming (ASP), a particularly effective approach for finding solutions to combinatorial search problems. The effectiveness of LLMs in ASP code generation is currently hindered by the limited number of examples seen during their initial pre-training phase. In this paper, we introduce a novel ASP-solver-in-the-loop approach for solver-guided instruction-tuning of LLMs to addressing the highly complex semantic parsing task inherent in ASP code generation. Our method only requires problem specifications in natural language and their solutions. Specifically, we sample ASP statements for program continuations from LLMs for unriddling logic puzzles. Leveraging the special property of declarative ASP programming that partial encodings increasingly narrow down the solution space, we categorize them into chosen and rejected instances based on solver feedback. We then apply supervised fine-tuning to train LLMs on the curated data and further improve robustness using a solver-guided search that includes best-of-N sampling. Our experiments demonstrate consistent improvements in two distinct prompting settings on two datasets.

Steffen Kleinle, Jakob Prange, Annemarie Friedrich

When immigrating to a new country, it is easy to feel overwhelmed by the need to obtain information on financial support, housing, schooling, language courses, and other issues. If relocation is rushed or even forced, the necessity for high-quality answers to such questions is all the more urgent. Official immigration counselors are usually overbooked, and online systems could guide newcomers to the requested information or a suitable counseling service. To this end, we present OMoS-QA, a dataset of German and English questions paired with relevant trustworthy documents and manually annotated answers, specifically tailored to this scenario. Questions are automatically generated with an open-source large language model (LLM) and answer sentences are selected by crowd workers with high agreement. With our data, we conduct a comparison of 5 pretrained LLMs on the task of extractive question answering (QA) in German and English. Across all models and both languages, we find high precision and low-to-mid recall in selecting answer sentences, which is a favorable trade-off to avoid misleading users. This performance even holds up when the question language does not match the document language. When it comes to identifying unanswerable questions given a context, there are larger differences between the two languages.

Christian Jaumann, Andreas Wiedholz, Annemarie Friedrich

The scientific literature is growing rapidly, making it hard to keep track of the state-of-the-art. Systematic literature reviews (SLRs) aim to identify and evaluate all relevant papers on a topic. After retrieving a set of candidate papers, the abstract screening phase determines initial relevance. To date, abstract screening methods using large language models (LLMs) focus on binary classification settings; existing question answering (QA) based ranking approaches suffer from error propagation. LLMs offer a unique opportunity to evaluate the SLR's inclusion and exclusion criteria, yet, existing benchmarks do not provide them exhaustively. We manually extract these criteria as well as research questions for 57 SLRs, mostly in the medical domain, enabling principled comparisons between approaches. Moreover, we propose LGAR, a zero-shot LLM Guided Abstract Ranker composed of an LLM based graded relevance scorer and a dense re-ranker. Our extensive experiments show that LGAR outperforms existing QA-based methods by 5-10 pp. in mean average precision. Our code and data is publicly available.

Timo Pierre Schrader, Matteo Finco, Stefan Grünewald et al.

Keeping track of all relevant recent publications and experimental results for a research area is a challenging task. Prior work has demonstrated the efficacy of information extraction models in various scientific areas. Recently, several datasets have been released for the yet understudied materials science domain. However, these datasets focus on sub-problems such as parsing synthesis procedures or on sub-domains, e.g., solid oxide fuel cells. In this resource paper, we present MuLMS, a new dataset of 50 open-access articles, spanning seven sub-domains of materials science. The corpus has been annotated by domain experts with several layers ranging from named entities over relations to frame structures. We present competitive neural models for all tasks and demonstrate that multi-task training with existing related resources leads to benefits.

Christian Jaumann, Annemarie Friedrich, Rainer Lienhart

This paper describes our system for the SciVQA 2025 Shared Task on Scientific Visual Question Answering. Our system employs an ensemble of two Multimodal Large Language Models and various few-shot example retrieval strategies. The model and few-shot setting are selected based on the figure and question type. We also select answers based on the models' confidence levels. On the blind test data, our system ranks third out of seven with an average F1 score of 85.12 across ROUGE-1, ROUGE-L, and BERTS. Our code is publicly available.

Valentin Knappich, Anna Hätty, Simon Razniewski et al.

Novelty assessment is a critical yet complex task in the examination process for patent acceptance, requiring examiners to determine whether an invention is disclosed in a prior art document. The process involves intricate matching between specific features of a patent claim and passages in the prior art. While prior work has approached novelty prediction primarily as a binary classification task at the claim level, we argue that this formulation is susceptible to spurious correlations and lacks the granularity required for practical application. In this work, we introduce FiNE-Patents (Fine-grained Novelty Examination of Patents), a novel dataset comprising 3,658 first patent claims annotated with fine-grained, feature-level prior art references extracted from European Search Opinion (ESOP) documents. We propose shifting the evaluation paradigm from simple binary classification to a joint retrieval and abstract reasoning task at the feature level, requiring models to identify specific passages from a prior art document that disclose individual claim features, and to identify which features of a claim make it novel. We implement and evaluate LLM-based workflows that decompose claims into features, analyze each feature against prior art, and finally derive a claim-level novelty prediction. Our experiments demonstrate that these workflows outperform embedding-based baselines on passage retrieval and novel feature identification. Furthermore, we show that unlike trained classifiers, LLMs are robust against spurious correlations present in the claim-level novelty classification task. We release the dataset and code to foster further research into transparent and granular patent analysis.

Wei Zhou, Mohsen Mesgar, Heike Adel et al.

Table Question Answering (TQA) aims at composing an answer to a question based on tabular data. While prior research has shown that TQA models lack robustness, understanding the underlying cause and nature of this issue remains predominantly unclear, posing a significant obstacle to the development of robust TQA systems. In this paper, we formalize three major desiderata for a fine-grained evaluation of robustness of TQA systems. They should (i) answer questions regardless of alterations in table structure, (ii) base their responses on the content of relevant cells rather than on biases, and (iii) demonstrate robust numerical reasoning capabilities. To investigate these aspects, we create and publish a novel TQA evaluation benchmark in English. Our extensive experimental analysis reveals that none of the examined state-of-the-art TQA systems consistently excels in these three aspects. Our benchmark is a crucial instrument for monitoring the behavior of TQA systems and paves the way for the development of robust TQA systems. We release our benchmark publicly.

Bolei Ma, Yuting Li, Wei Zhou et al.

Understanding pragmatics-the use of language in context-is crucial for developing NLP systems capable of interpreting nuanced language use. Despite recent advances in language technologies, including large language models, evaluating their ability to handle pragmatic phenomena such as implicatures and references remains challenging. To advance pragmatic abilities in models, it is essential to understand current evaluation trends and identify existing limitations. In this survey, we provide a comprehensive review of resources designed for evaluating pragmatic capabilities in NLP, categorizing datasets by the pragmatic phenomena they address. We analyze task designs, data collection methods, evaluation approaches, and their relevance to real-world applications. By examining these resources in the context of modern language models, we highlight emerging trends, challenges, and gaps in existing benchmarks. Our survey aims to clarify the landscape of pragmatic evaluation and guide the development of more comprehensive and targeted benchmarks, ultimately contributing to more nuanced and context-aware NLP models.

Wei Zhou, Mohsen Mesgar, Annemarie Friedrich et al.

Complex table question answering (TQA) aims to answer questions that require complex reasoning, such as multi-step or multi-category reasoning, over data represented in tabular form. Previous approaches demonstrated notable performance by leveraging either closed-source large language models (LLMs) or fine-tuned open-weight LLMs. However, fine-tuning LLMs requires high-quality training data, which is costly to obtain, and utilizing closed-source LLMs poses accessibility challenges and leads to reproducibility issues. In this paper, we propose Multi-Agent Collaboration with Tool use (MACT), a framework that requires neither closed-source models nor fine-tuning. In MACT, a planning agent and a coding agent that also make use of tools collaborate to answer questions. Our experiments on four TQA benchmarks show that MACT outperforms previous SoTA systems on three out of four benchmarks and that it performs comparably to the larger and more expensive closed-source model GPT-4 on two benchmarks, even when using only open-weight models without any fine-tuning. We conduct extensive analyses to prove the effectiveness of MACT's multi-agent collaboration in TQA.

Lukas Lange, Marc Müller, Ghazaleh Haratinezhad Torbati et al.

Monitoring the threat landscape to be aware of actual or potential attacks is of utmost importance to cybersecurity professionals. Information about cyber threats is typically distributed using natural language reports. Natural language processing can help with managing this large amount of unstructured information, yet to date, the topic has received little attention. With this paper, we present AnnoCTR, a new CC-BY-SA-licensed dataset of cyber threat reports. The reports have been annotated by a domain expert with named entities, temporal expressions, and cybersecurity-specific concepts including implicitly mentioned techniques and tactics. Entities and concepts are linked to Wikipedia and the MITRE ATT&CK knowledge base, the most widely-used taxonomy for classifying types of attacks. Prior datasets linking to MITRE ATT&CK either provide a single label per document or annotate sentences out-of-context; our dataset annotates entire documents in a much finer-grained way. In an experimental study, we model the annotations of our dataset using state-of-the-art neural models. In our few-shot scenario, we find that for identifying the MITRE ATT&CK concepts that are mentioned explicitly or implicitly in a text, concept descriptions from MITRE ATT&CK are an effective source for training data augmentation.

Wei Zhou, Mohsen Mesgar, Heike Adel et al.

In table question answering (TQA), tables are encoded as either texts or images. Prior work suggests that passing images of tables to multi-modal large language models (MLLMs) performs comparably to or even better than using textual input with large language models (LLMs). However, the lack of controlled setups limits fine-grained distinctions between these approaches. In this paper, we conduct the first controlled study on the effectiveness of several combinations of table representations and models from two perspectives: question complexity and table size. We build a new benchmark based on existing TQA datasets. In a systematic analysis of seven pairs of MLLMs and LLMs, we find that the best combination of table representation and model varies across setups. We propose FRES, a method selecting table representations dynamically, and observe a 10% average performance improvement compared to using both representations indiscriminately.

Timo Pierre Schrader, Lukas Lange, Simon Razniewski et al.

Reasoning is key to many decision making processes. It requires consolidating a set of rule-like premises that are often associated with degrees of uncertainty and observations to draw conclusions. In this work, we address both the case where premises are specified as numeric probabilistic rules and situations in which humans state their estimates using words expressing degrees of certainty. Existing probabilistic reasoning datasets simplify the task, e.g., by requiring the model to only rank textual alternatives, by including only binary random variables, or by making use of a limited set of templates that result in less varied text. In this work, we present QUITE, a question answering dataset of real-world Bayesian reasoning scenarios with categorical random variables and complex relationships. QUITE provides high-quality natural language verbalizations of premises together with evidence statements and expects the answer to a question in the form of an estimated probability. We conduct an extensive set of experiments, finding that logic-based models outperform out-of-the-box large language models on all reasoning types (causal, evidential, and explaining-away). Our results provide evidence that neuro-symbolic models are a promising direction for improving complex reasoning. We release QUITE and code for training and experiments on Github.

Julian Lorenz, Mrunmai Phatak, Robin Schön et al.

2D scene graphs provide a structural and explainable framework for scene understanding. However, current work still struggles with the lack of accurate scene graph data. To overcome this data bottleneck, we present CoPa-SG, a synthetic scene graph dataset with highly precise ground truth and exhaustive relation annotations between all objects. Moreover, we introduce parametric and proto-relations, two new fundamental concepts for scene graphs. The former provides a much more fine-grained representation than its traditional counterpart by enriching relations with additional parameters such as angles or distances. The latter encodes hypothetical relations in a scene graph and describes how relations would form if new objects are placed in the scene. Using CoPa-SG, we compare the performance of various scene graph generation models. We demonstrate how our new relation types can be integrated in downstream applications to enhance planning and reasoning capabilities.

Timo Pierre Schrader, Simon Razniewski, Lukas Lange et al.

Understanding causality is a core aspect of intelligence. The Event Causality Identification with Causal News Corpus Shared Task addresses two aspects of this challenge: Subtask 1 aims at detecting causal relationships in texts, and Subtask 2 requires identifying signal words and the spans that refer to the cause or effect, respectively. Our system, which is based on pre-trained transformers, stacked sequence tagging, and synthetic data augmentation, ranks third in Subtask 1 and wins Subtask 2 with an F1 score of 72.8, corresponding to a margin of 13 pp. to the second-best system.

Wei Zhou, Bolei Ma, Annemarie Friedrich et al.

Table Question Answering (TQA) aims to answer natural language questions about tabular data, often accompanied by additional contexts such as text passages. The task spans diverse settings, varying in table representation, question/answer complexity, modality involved, and domain. While recent advances in large language models (LLMs) have led to substantial progress in TQA, the field still lacks a systematic organization and understanding of task formulations, core challenges, and methodological trends, particularly in light of emerging research directions such as reinforcement learning. This survey addresses this gap by providing a comprehensive and structured overview of TQA research with a focus on LLM-based methods. We provide a comprehensive categorization of existing benchmarks and task setups. We group current modeling strategies according to the challenges they target, and analyze their strengths and limitations. Furthermore, we highlight underexplored but timely topics that have not been systematically covered in prior research. By unifying disparate research threads and identifying open problems, our survey offers a consolidated foundation for the TQA community, enabling a deeper understanding of the state of the art and guiding future developments in this rapidly evolving area.

Valentin Knappich, Annemarie Friedrich, Anna Hätty et al.

Patent claims define the scope of protection for an invention. If there are ambiguities in a claim, it is rejected by the patent office. In the US, this is referred to as indefiniteness (35 U.S.C § 112(b)) and is among the most frequent reasons for patent application rejection. The development of automatic methods for patent definiteness examination has the potential to make patent drafting and examination more efficient, but no annotated dataset has been published to date. We introduce PEDANTIC (Patent Definiteness Examination Corpus), a novel dataset of 14k US patent claims from patent applications relating to Natural Language Processing (NLP), annotated with reasons for indefiniteness. We construct PEDANTIC using a fully automatic pipeline that retrieves office action documents from the USPTO and uses Large Language Models (LLMs) to extract the reasons for indefiniteness. A human validation study confirms the pipeline's accuracy in generating high-quality annotations. To gain insight beyond binary classification metrics, we implement an LLM-as-Judge evaluation that compares the free-form reasoning of every model-cited reason with every examiner-cited reason. We show that LLM agents based on Qwen 2.5 32B and 72B struggle to outperform logistic regression baselines on definiteness prediction, even though they often correctly identify the underlying reasons. PEDANTIC provides a valuable resource for patent AI researchers, enabling the development of advanced examination models. We will publicly release the dataset and code.

Wei Zhou, Mohsen Mesgar, Heike Adel et al.

Table question answering (TQA) focuses on answering questions based on tabular data. Developing TQA systems targets effective interaction with tabular data for tasks such as cell retrieval and data analysis. While recent work has leveraged fine-tuning to improve TQA systems, existing approaches often under-utilize available data and neglect the potential of post-training for further gains. In this work, we introduce p2-TQA, a process-based preference learning framework for TQA post-training. p2-TQA automatically constructs process-based preference data via a table-specific pipeline, eliminating the need for manual or costly data collection. It then optimizes models through contrastive learning on the collected data. Experiments show that p2-TQA effectively improves TQA models by up to 5% on in-domain datasets and 2.4% on out-of-domain datasets with only 8,000 training instances. Furthermore, models enhanced with p2-TQA achieve competitive results against larger, more complex state-of-the-art TQA systems, while maintaining up to five times higher efficiency.

Elizaveta Sineva, Stefan Grünewald, Annemarie Friedrich et al.

In this paper, we revisit the task of negation resolution, which includes the subtasks of cue detection (e.g. "not", "never") and scope resolution. In the context of previous shared tasks, a variety of evaluation metrics have been proposed. Subsequent works usually use different subsets of these, including variations and custom implementations, rendering meaningful comparisons between systems difficult. Examining the problem both from a linguistic perspective and from a downstream viewpoint, we here argue for a negation-instance based approach to evaluating negation resolution. Our proposed metrics correspond to expectations over per-instance scores and hence are intuitively interpretable. To render research comparable and to foster future work, we provide results for a set of current state-of-the-art systems for negation resolution on three English corpora, and make our implementation of the evaluation scripts publicly available.

Stefan Grünewald, Prisca Piccirilli, Annemarie Friedrich

In this paper, we address the representation of coordinate constructions in Enhanced Universal Dependencies (UD), where relevant dependency links are propagated from conjunction heads to other conjuncts. English treebanks for enhanced UD have been created from gold basic dependencies using a heuristic rule-based converter, which propagates only core arguments. With the aim of determining which set of links should be propagated from a semantic perspective, we create a large-scale dataset of manually edited syntax graphs. We identify several systematic errors in the original data, and propose to also propagate adjuncts. We observe high inter-annotator agreement for this semantic annotation task. Using our new manually verified dataset, we perform the first principled comparison of rule-based and (partially novel) machine-learning based methods for conjunction propagation for English. We show that learning propagation rules is more effective than hand-designing heuristic rules. When using automatic parses, our neural graph-parser based edge predictor outperforms the currently predominant pipelinesusing a basic-layer tree parser plus converters.

Stefan Grünewald, Annemarie Friedrich, Jonas Kuhn

The introduction of pre-trained transformer-based contextualized word embeddings has led to considerable improvements in the accuracy of graph-based parsers for frameworks such as Universal Dependencies (UD). However, previous works differ in various dimensions, including their choice of pre-trained language models and whether they use LSTM layers. With the aims of disentangling the effects of these choices and identifying a simple yet widely applicable architecture, we introduce STEPS, a new modular graph-based dependency parser. Using STEPS, we perform a series of analyses on the UD corpora of a diverse set of languages. We find that the choice of pre-trained embeddings has by far the greatest impact on parser performance and identify XLM-R as a robust choice across the languages in our study. Adding LSTM layers provides no benefits when using transformer-based embeddings. A multi-task training setup outputting additional UD features may contort results. Taking these insights together, we propose a simple but widely applicable parser architecture and configuration, achieving new state-of-the-art results (in terms of LAS) for 10 out of 12 diverse languages.

Annemarie Friedrich, Heike Adel, Federico Tomazic et al.

This paper presents a new challenging information extraction task in the domain of materials science. We develop an annotation scheme for marking information on experiments related to solid oxide fuel cells in scientific publications, such as involved materials and measurement conditions. With this paper, we publish our annotation guidelines, as well as our SOFC-Exp corpus consisting of 45 open-access scholarly articles annotated by domain experts. A corpus and an inter-annotator agreement study demonstrate the complexity of the suggested named entity recognition and slot filling tasks as well as high annotation quality. We also present strong neural-network based models for a variety of tasks that can be addressed on the basis of our new data set. On all tasks, using BERT embeddings leads to large performance gains, but with increasing task complexity, adding a recurrent neural network on top seems beneficial. Our models will serve as competitive baselines in future work, and analysis of their performance highlights difficult cases when modeling the data and suggests promising research directions.

Anna Senina, Marcus Rohrbach, Wei Qiu et al.

Humans can easily describe what they see in a coherent way and at varying level of detail. However, existing approaches for automatic video description are mainly focused on single sentence generation and produce descriptions at a fixed level of detail. In this paper, we address both of these limitations: for a variable level of detail we produce coherent multi-sentence descriptions of complex videos. We follow a two-step approach where we first learn to predict a semantic representation (SR) from video and then generate natural language descriptions from the SR. To produce consistent multi-sentence descriptions, we model across-sentence consistency at the level of the SR by enforcing a consistent topic. We also contribute both to the visual recognition of objects proposing a hand-centric approach as well as to the robust generation of sentences using a word lattice. Human judges rate our multi-sentence descriptions as more readable, correct, and relevant than related work. To understand the difference between more detailed and shorter descriptions, we collect and analyze a video description corpus of three levels of detail.