Erik Velldal

David Samuel, Andrey Kutuzov, Lilja Øvrelid et al.

While modern masked language models (LMs) are trained on ever larger corpora, we here explore the effects of down-scaling training to a modestly-sized but representative, well-balanced, and publicly available English text source -- the British National Corpus. We show that pre-training on this carefully curated corpus can reach better performance than the original BERT model. We argue that this type of corpora has great potential as a language modeling benchmark. To showcase this potential, we present fair, reproducible and data-efficient comparative studies of LMs, in which we evaluate several training objectives and model architectures and replicate previous empirical results in a systematic way. We propose an optimized LM architecture called LTG-BERT.

Andrey Kutuzov, Erik Velldal, Lilja Øvrelid

We present a qualitative analysis of the (potentially erroneous) outputs of contextualized embedding-based methods for detecting diachronic semantic change. First, we introduce an ensemble method outperforming previously described contextualized approaches. This method is used as a basis for an in-depth analysis of the degrees of semantic change predicted for English words across 5 decades. Our findings show that contextualized methods can often predict high change scores for words which are not undergoing any real diachronic semantic shift in the lexicographic sense of the term (or at least the status of these shifts is questionable). Such challenging cases are discussed in detail with examples, and their linguistic categorization is proposed. Our conclusion is that pre-trained contextualized language models are prone to confound changes in lexicographic senses and changes in contextual variance, which naturally stem from their distributional nature, but is different from the types of issues observed in methods based on static embeddings. Additionally, they often merge together syntactic and semantic aspects of lexical entities. We propose a range of possible future solutions to these issues.

David Samuel, Jeremy Barnes, Robin Kurtz et al.

This paper demonstrates how a graph-based semantic parser can be applied to the task of structured sentiment analysis, directly predicting sentiment graphs from text. We advance the state of the art on 4 out of 5 standard benchmark sets. We release the source code, models and predictions.

Egil Rønningstad, Erik Velldal, Lilja Øvrelid

This paper explores the task of identifying the overall sentiment expressed towards volitional entities (persons and organizations) in a document -- what we refer to as Entity-Level Sentiment Analysis (ELSA). While identifying sentiment conveyed towards an entity is well researched for shorter texts like tweets, we find little to no research on this specific task for longer texts with multiple mentions and opinions towards the same entity. This lack of research would be understandable if ELSA can be derived from existing tasks and models. To assess this, we annotate a set of professional reviews for their overall sentiment towards each volitional entity in the text. We sample from data already annotated for document-level, sentence-level, and target-level sentiment in a multi-domain review corpus, and our results indicate that there is no single proxy task that provides this overall sentiment we seek for the entities at a satisfactory level of performance. We present a suite of experiments aiming to assess the contribution towards ELSA provided by document-, sentence-, and target-level sentiment analysis, and provide a discussion of their shortcomings. We show that sentiment in our dataset is expressed not only with an entity mention as target, but also towards targets with a sentiment-relevant relation to a volitional entity. In our data, these relations extend beyond anaphoric coreference resolution, and our findings call for further research of the topic. Finally, we also present a survey of previous relevant work.

Samia Touileb, Lilja Øvrelid, Erik Velldal

We investigate in this paper how distributions of occupations with respect to gender is reflected in pre-trained language models. Such distributions are not always aligned to normative ideals, nor do they necessarily reflect a descriptive assessment of reality. In this paper, we introduce an approach for measuring to what degree pre-trained language models are aligned to normative and descriptive occupational distributions. To this end, we use official demographic information about gender--occupation distributions provided by the national statistics agencies of France, Norway, United Kingdom, and the United States. We manually generate template-based sentences combining gendered pronouns and nouns with occupations, and subsequently probe a selection of ten language models covering the English, French, and Norwegian languages. The scoring system we introduce in this work is language independent, and can be used on any combination of template-based sentences, occupations, and languages. The approach could also be extended to other dimensions of national census data and other demographic variables.

Sondre Wold, Lilja Øvrelid, Erik Velldal

In contrast to large text corpora, knowledge graphs (KG) provide dense and structured representations of factual information. This makes them attractive for systems that supplement or ground the knowledge found in pre-trained language models with an external knowledge source. This has especially been the case for classification tasks, where recent work has focused on creating pipeline models that retrieve information from KGs like ConceptNet as additional context. Many of these models consist of multiple components, and although they differ in the number and nature of these parts, they all have in common that for some given text query, they attempt to identify and retrieve a relevant subgraph from the KG. Due to the noise and idiosyncrasies often found in KGs, it is not known how current methods compare to a scenario where the aligned subgraph is completely relevant to the query. In this work, we try to bridge this knowledge gap by reviewing current approaches to text-to-KG alignment and evaluating them on two datasets where manually created graphs are available, providing insights into the effectiveness of current methods.

Dilara Torunoğlu-Selamet, Dogukan Arslan, Rodrigo Wilkens et al.

Potentially idiomatic expressions (PIEs) construe meanings inherently tied to the everyday experience of a given language community. As such, they constitute an interesting challenge for assessing the linguistic (and to some extent cultural) capabilities of NLP systems. In this paper, we present XMPIE, a parallel multilingual and multimodal dataset of potentially idiomatic expressions. The dataset, containing 34 languages and over ten thousand items, allows comparative analyses of idiomatic patterns among language-specific realisations and preferences in order to gather insights about shared cultural aspects. This parallel dataset allows to evaluate model performance for a given PIE in different languages and whether idiomatic understanding in one language can be transferred to another. Moreover, the dataset supports the study of PIEs across textual and visual modalities, to measure to what extent PIE understanding in one modality transfers or implies in understanding in another modality (text vs. image). The data was created by language experts, with both textual and visual components crafted under multilingual guidelines, and each PIE is accompanied by five images representing a spectrum from idiomatic to literal meanings, including semantically related and random distractors. The result is a high-quality benchmark for evaluating multilingual and multimodal idiomatic language understanding.

David Samuel, Lilja Øvrelid, Erik Velldal et al.

We propose a post-training method for lower-resource languages that preserves fluency of language models even when aligned by disfluent reward models. Preference-optimization is now a well-researched topic, but previous work has mostly addressed models for English and Chinese. Lower-resource languages lack both datasets written by native speakers and language models capable of generating fluent synthetic data. Thus, in this work, we focus on developing a fluent preference-aligned language model without any instruction-tuning data in the target language. Our approach uses an on-policy training method, which we compare with two common approaches: supervised finetuning on machine-translated data and multilingual finetuning. We conduct a case study on Norwegian Bokmål and evaluate fluency through native-speaker assessments. The results show that the on-policy aspect is crucial and outperforms the alternatives without relying on any hard-to-obtain data.

Egil Rønningstad, Roman Klinger, Lilja Øvrelid et al.

In sentiment analysis of longer texts, there may be a variety of topics discussed, of entities mentioned, and of sentiments expressed regarding each entity. We find a lack of studies exploring how such texts express their sentiment towards each entity of interest, and how these sentiments can be modelled. In order to better understand how sentiment regarding persons and organizations (each entity in our scope) is expressed in longer texts, we have collected a dataset of expert annotations where the overall sentiment regarding each entity is identified, together with the sentence-level sentiment for these entities separately. We show that the reader's perceived sentiment regarding an entity often differs from an arithmetic aggregation of sentiments at the sentence level. Only 70\% of the positive and 55\% of the negative entities receive a correct overall sentiment label when we aggregate the (human-annotated) sentiment labels for the sentences where the entity is mentioned. Our dataset reveals the complexity of entity-specific sentiment in longer texts, and allows for more precise modelling and evaluation of such sentiment expressions.

Vladislav Mikhailov, Tita Enstad, David Samuel et al.

This paper introduces NorEval, a new and comprehensive evaluation suite for large-scale standardized benchmarking of Norwegian generative language models (LMs). NorEval consists of 24 high-quality human-created datasets -- of which five are created from scratch. In contrast to existing benchmarks for Norwegian, NorEval covers a broad spectrum of task categories targeting Norwegian language understanding and generation, establishes human baselines, and focuses on both of the official written standards of the Norwegian language: Bokmål and Nynorsk. All our datasets and a collection of over 100 human-written prompts are integrated into LM Evaluation Harness, ensuring flexible and reproducible evaluation. We describe the NorEval design and present the results of benchmarking 19 open-source pre-trained and instruction-tuned LMs for Norwegian in various scenarios. Our benchmark, evaluation framework, and annotation materials are publicly available.

Jeremy Barnes, Lilja Øvrelid, Erik Velldal

Neural methods for SA have led to quantitative improvements over previous approaches, but these advances are not always accompanied with a thorough analysis of the qualitative differences. Therefore, it is not clear what outstanding conceptual challenges for sentiment analysis remain. In this work, we attempt to discover what challenges still prove a problem for sentiment classifiers for English and to provide a challenging dataset. We collect the subset of sentences that an (oracle) ensemble of state-of-the-art sentiment classifiers misclassify and then annotate them for 18 linguistic and paralinguistic phenomena, such as negation, sarcasm, modality, etc. The dataset is available at https://github.com/ltgoslo/assessing_and_probing_sentiment. Finally, we provide a case study that demonstrates the usefulness of the dataset to probe the performance of a given sentiment classifier with respect to linguistic phenomena.

Petter Mæhlum, David Samuel, Rebecka Maria Norman et al.

Sentiment analysis is an important tool for aggregating patient voices, in order to provide targeted improvements in healthcare services. A prerequisite for this is the availability of in-domain data annotated for sentiment. This article documents an effort to add sentiment annotations to free-text comments in patient surveys collected by the Norwegian Institute of Public Health (NIPH). However, annotation can be a time-consuming and resource-intensive process, particularly when it requires domain expertise. We therefore also evaluate a possible alternative to human annotation, using large language models (LLMs) as annotators. We perform an extensive evaluation of the approach for two openly available pretrained LLMs for Norwegian, experimenting with different configurations of prompts and in-context learning, comparing their performance to human annotators. We find that even for zero-shot runs, models perform well above the baseline for binary sentiment, but still cannot compete with human annotators on the full dataset.

David Samuel, Vladislav Mikhailov, Erik Velldal et al.

Training large language models requires vast amounts of data, posing a challenge for less widely spoken languages like Norwegian and even more so for truly low-resource languages like Northern Sámi. To address this issue, we present a novel three-stage continual training approach that substantially improves the downstream performance together with the inference efficiency for the target languages. Based on our findings, we train, evaluate, and openly release a new generative language model for Norwegian Bokmål, Nynorsk, and Northern Sámi with 11.4 billion parameters: NorMistral-11B.

Vladislav Mikhailov, Petter Mæhlum, Victoria Ovedie Chruickshank Langø et al.

This paper introduces a new suite of question answering datasets for Norwegian; NorOpenBookQA, NorCommonSenseQA, NorTruthfulQA, and NRK-Quiz-QA. The data covers a wide range of skills and knowledge domains, including world knowledge, commonsense reasoning, truthfulness, and knowledge about Norway. Covering both of the written standards of Norwegian - Bokmål and Nynorsk - our datasets comprise over 10k question-answer pairs, created by native speakers. We detail our dataset creation approach and present the results of evaluating 11 language models (LMs) in zero- and few-shot regimes. Most LMs perform better in Bokmål than Nynorsk, struggle most with commonsense reasoning, and are often untruthful in generating answers to questions. All our datasets and annotation materials are publicly available.

Samia Touileb, Vladislav Mikhailov, Marie Kroka et al.

We introduce a dataset of high-quality human-authored summaries of news articles in Norwegian. The dataset is intended for benchmarking the abstractive summarisation capabilities of generative language models. Each document in the dataset is provided with three different candidate gold-standard summaries written by native Norwegian speakers, and all summaries are provided in both of the written variants of Norwegian -- Bokmål and Nynorsk. The paper describes details on the data creation effort as well as an evaluation of existing open LLMs for Norwegian on the dataset. We also provide insights from a manual human evaluation, comparing human-authored to model-generated summaries. Our results indicate that the dataset provides a challenging LLM benchmark for Norwegian summarisation capabilities

Javier de la Rosa, Vladislav Mikhailov, Lemei Zhang et al.

The use of copyrighted materials in training language models raises critical legal and ethical questions. This paper presents a framework for and the results of empirically assessing the impact of publisher-controlled copyrighted corpora on the performance of generative large language models (LLMs) for Norwegian. When evaluated on a diverse set of tasks, we found that adding both books and newspapers to the data mixture of LLMs tend to improve their performance, while the addition of fiction works seems to be detrimental. Our experiments could inform the creation of a compensation scheme for authors whose works contribute to AI development.

Egil Rønningstad, Lilja Charlotte Storset, Petter Mæhlum et al.

Sentiment analysis of patient feedback from the public health domain can aid decision makers in evaluating the provided services. The current paper focuses on free-text comments in patient surveys about general practitioners and psychiatric healthcare, annotated with four sentence-level polarity classes -- positive, negative, mixed and neutral -- while also attempting to alleviate data scarcity by leveraging general-domain sources in the form of reviews. For several different architectures, we compare in-domain and out-of-domain effects, as well as the effects of training joint multi-domain models.

Huiling You, Samia Touileb, Erik Velldal et al.

An abstractive summary of a news article contains its most important information in a condensed version. The evaluation of automatically generated summaries by generative language models relies heavily on human-authored summaries as gold references, by calculating overlapping units or similarity scores. News articles report events, and ideally so should the summaries. In this work, we propose to evaluate the quality of abstractive summaries by calculating overlapping events between generated summaries, reference summaries, and the original news articles. We experiment on a richly annotated Norwegian dataset comprising both events annotations and summaries authored by expert human annotators. Our approach provides more insight into the event information contained in the summaries.

Sondre Wold, Étienne Simon, Lucas Georges Gabriel Charpentier et al.

Grounded language models use external sources of information, such as knowledge graphs, to meet some of the general challenges associated with pre-training. By extending previous work on compositional generalization in semantic parsing, we allow for a controlled evaluation of the degree to which these models learn and generalize from patterns in knowledge graphs. We develop a procedure for generating natural language questions paired with knowledge graphs that targets different aspects of compositionality and further avoids grounding the language models in information already encoded implicitly in their weights. We evaluate existing methods for combining language models with knowledge graphs and find them to struggle with generalization to sequences of unseen lengths and to novel combinations of seen base components. While our experimental results provide some insight into the expressive power of these models, we hope our work and released datasets motivate future research on how to better combine language models with structured knowledge representations.

David Samuel, Andrey Kutuzov, Samia Touileb et al.

We present NorBench: a streamlined suite of NLP tasks and probes for evaluating Norwegian language models (LMs) on standardized data splits and evaluation metrics. We also introduce a range of new Norwegian language models (both encoder and encoder-decoder based). Finally, we compare and analyze their performance, along with other existing LMs, across the different benchmark tests of NorBench.

Jeremy Barnes, Robin Kurtz, Stephan Oepen et al.

Structured sentiment analysis attempts to extract full opinion tuples from a text, but over time this task has been subdivided into smaller and smaller sub-tasks, e,g,, target extraction or targeted polarity classification. We argue that this division has become counterproductive and propose a new unified framework to remedy the situation. We cast the structured sentiment problem as dependency graph parsing, where the nodes are spans of sentiment holders, targets and expressions, and the arcs are the relations between them. We perform experiments on five datasets in four languages (English, Norwegian, Basque, and Catalan) and show that this approach leads to strong improvements over state-of-the-art baselines. Our analysis shows that refining the sentiment graphs with syntactic dependency information further improves results.

Andrey Kutuzov, Jeremy Barnes, Erik Velldal et al.

We present the ongoing NorLM initiative to support the creation and use of very large contextualised language models for Norwegian (and in principle other Nordic languages), including a ready-to-use software environment, as well as an experience report for data preparation and training. This paper introduces the first large-scale monolingual language models for Norwegian, based on both the ELMo and BERT frameworks. In addition to detailing the training process, we present contrastive benchmark results on a suite of NLP tasks for Norwegian. For additional background and access to the data, models, and software, please see http://norlm.nlpl.eu

Jeremy Barnes, Lilja Øvrelid, Erik Velldal

Fine-grained sentiment analysis attempts to extract sentiment holders, targets and polar expressions and resolve the relationship between them, but progress has been hampered by the difficulty of annotation. Targeted sentiment analysis, on the other hand, is a more narrow task, focusing on extracting sentiment targets and classifying their polarity.In this paper, we explore whether incorporating holder and expression information can improve target extraction and classification and perform experiments on eight English datasets. We conclude that jointly predicting target and polarity BIO labels improves target extraction, and that augmenting the input text with gold expressions generally improves targeted polarity classification. This highlights the potential importance of annotating expressions for fine-grained sentiment datasets. At the same time, our results show that performance of current models for predicting polar expressions is poor, hampering the benefit of this information in practice.

Jeremy Barnes, Vinit Ravishankar, Lilja Øvrelid et al.

Documents are composed of smaller pieces - paragraphs, sentences, and tokens - that have complex relationships between one another. Sentiment classification models that take into account the structure inherent in these documents have a theoretical advantage over those that do not. At the same time, transfer learning models based on language model pretraining have shown promise for document classification. However, these two paradigms have not been systematically compared and it is not clear under which circumstances one approach is better than the other. In this work we empirically compare hierarchical models and transfer learning for document-level sentiment classification. We show that non-trivial hierarchical models outperform previous baselines and transfer learning on document-level sentiment classification in five languages.

Lilja Øvrelid, Petter Mæhlum, Jeremy Barnes et al.

We introduce NoReC_fine, a dataset for fine-grained sentiment analysis in Norwegian, annotated with respect to polar expressions, targets and holders of opinion. The underlying texts are taken from a corpus of professionally authored reviews from multiple news-sources and across a wide variety of domains, including literature, games, music, products, movies and more. We here present a detailed description of this annotation effort. We provide an overview of the developed annotation guidelines, illustrated with examples, and present an analysis of inter-annotator agreement. We also report the first experimental results on the dataset, intended as a preliminary benchmark for further experiments.

Fredrik Jørgensen, Tobias Aasmoe, Anne-Stine Ruud Husevåg et al.

This paper presents NorNE, a manually annotated corpus of named entities which extends the annotation of the existing Norwegian Dependency Treebank. Comprising both of the official standards of written Norwegian (Bokmål and Nynorsk), the corpus contains around 600,000 tokens and annotates a rich set of entity types including persons, organizations, locations, geo-political entities, products, and events, in addition to a class corresponding to nominals derived from names. We here present details on the annotation effort, guidelines, inter-annotator agreement and an experimental analysis of the corpus using a neural sequence labeling architecture.

Andrey Kutuzov, Erik Velldal, Lilja Øvrelid

We extend the well-known word analogy task to a one-to-X formulation, including one-to-none cases, when no correct answer exists. The task is cast as a relation discovery problem and applied to historical armed conflicts datasets, attempting to predict new relations of type `location:armed-group' based on data about past events. As the source of semantic information, we use diachronic word embedding models trained on English news texts. A simple technique to improve diachronic performance in such task is demonstrated, using a threshold based on a function of cosine distance to decrease the number of false positives; this approach is shown to be beneficial on two different corpora. Finally, we publish a ready-to-use test set for one-to-X analogy evaluation on historical armed conflicts data.

Jeremy Barnes, Erik Velldal, Lilja Øvrelid

Sentiment analysis is directly affected by compositional phenomena in language that act on the prior polarity of the words and phrases found in the text. Negation is the most prevalent of these phenomena and in order to correctly predict sentiment, a classifier must be able to identify negation and disentangle the effect that its scope has on the final polarity of a text. This paper proposes a multi-task approach to explicitly incorporate information about negation in sentiment analysis, which we show outperforms learning negation implicitly in a data-driven manner. We describe our approach, a cascading neural architecture with selective sharing of LSTM layers, and show that explicitly training the model with negation as an auxiliary task helps improve the main task of sentiment analysis. The effect is demonstrated across several different standard English-language data sets for both tasks and we analyze several aspects of our system related to its performance, varying types and amounts of input data and different multi-task setups.

Vinit Ravishankar, Lilja Øvrelid, Erik Velldal

This paper extends the task of probing sentence representations for linguistic insight in a multilingual domain. In doing so, we make two contributions: first, we provide datasets for multilingual probing, derived from Wikipedia, in five languages, viz. English, French, German, Spanish and Russian. Second, we evaluate six sentence encoders for each language, each trained by mapping sentence representations to English sentence representations, using sentences in a parallel corpus. We discover that cross-lingually mapped representations are often better at retaining certain linguistic information than representations derived from English encoders trained on natural language inference (NLI) as a downstream task.

Murhaf Fares, Stephan Oepen, Erik Velldal

In this paper, we empirically evaluate the utility of transfer and multi-task learning on a challenging semantic classification task: semantic interpretation of noun--noun compounds. Through a comprehensive series of experiments and in-depth error analysis, we show that transfer learning via parameter initialization and multi-task learning via parameter sharing can help a neural classification model generalize over a highly skewed distribution of relations. Further, we demonstrate how dual annotation with two distinct sets of relations over the same set of compounds can be exploited to improve the overall accuracy of a neural classifier and its F1 scores on the less frequent, but more difficult relations.

Andrey Kutuzov, Lilja Øvrelid, Terrence Szymanski et al.

Recent years have witnessed a surge of publications aimed at tracing temporal changes in lexical semantics using distributional methods, particularly prediction-based word embedding models. However, this vein of research lacks the cohesion, common terminology and shared practices of more established areas of natural language processing. In this paper, we survey the current state of academic research related to diachronic word embeddings and semantic shifts detection. We start with discussing the notion of semantic shifts, and then continue with an overview of the existing methods for tracing such time-related shifts with word embedding models. We propose several axes along which these methods can be compared, and outline the main challenges before this emerging subfield of NLP, as well as prospects and possible applications.

Erik Velldal, Lilja Øvrelid, Eivind Alexander Bergem et al.

This paper presents the Norwegian Review Corpus (NoReC), created for training and evaluating models for document-level sentiment analysis. The full-text reviews have been collected from major Norwegian news sources and cover a range of different domains, including literature, movies, video games, restaurants, music and theater, in addition to product reviews across a range of categories. Each review is labeled with a manually assigned score of 1-6, as provided by the rating of the original author. This first release of the corpus comprises more than 35,000 reviews. It is distributed using the CoNLL-U format, pre-processed using UDPipe, along with a rich set of metadata. The work reported in this paper forms part of the SANT initiative (Sentiment Analysis for Norwegian Text), a project seeking to provide resources and tools for sentiment analysis and opinion mining for Norwegian. As resources for sentiment analysis have so far been unavailable for Norwegian, NoReC represents a highly valuable and sought-after addition to Norwegian language technology.

Andrey Kutuzov, Erik Velldal, Lilja Øvrelid

This paper deals with using word embedding models to trace the temporal dynamics of semantic relations between pairs of words. The set-up is similar to the well-known analogies task, but expanded with a time dimension. To this end, we apply incremental updating of the models with new training texts, including incremental vocabulary expansion, coupled with learned transformation matrices that let us map between members of the relation. The proposed approach is evaluated on the task of predicting insurgent armed groups based on geographical locations. The gold standard data for the time span 1994--2010 is extracted from the UCDP Armed Conflicts dataset. The results show that the method is feasible and outperforms the baselines, but also that important work still remains to be done.

Andrey Kutuzov, Erik Velldal, Lilja Øvrelid

This paper studies how word embeddings trained on the British National Corpus interact with part of speech boundaries. Our work targets the Universal PoS tag set, which is currently actively being used for annotation of a range of languages. We experiment with training classifiers for predicting PoS tags for words based on their embeddings. The results show that the information about PoS affiliation contained in the distributional vectors allows us to discover groups of words with distributional patterns that differ from other words of the same part of speech. This data often reveals hidden inconsistencies of the annotation process or guidelines. At the same time, it supports the notion of `soft' or `graded' part of speech affiliations. Finally, we show that information about PoS is distributed among dozens of vector components, not limited to only one or two features.