Identifying direct speech in literary fiction is challenging for cases that do not mark speech segments with quotation marks. Such efforts have previously been based either on smaller manually annotated gold data or larger automatically annotated silver data, extracted from works with quotation marks. However, no direct comparison has so far been made between the performance of these two types of training data. In this work, we address this gap. We further explore the effect of different types of typographical speech marking and of using evaluation metrics of different granularity. We perform experiments on Swedish literary texts and find that using gold and silver data has different strengths, with gold data having stronger results on token-level metrics, whereas silver data overall has stronger results on span-level metrics. If the training data contains some data that matches the typographical speech marking of the target, that is generally sufficient for achieving good results, but it does not seem to hurt if the training data also contains other types of marking.

Matching genre in training and test data has been shown to improve dependency parsing. However, it is not clear whether the used methods capture only the genre feature. We hypothesize that successful transfer may also depend on topic similarity. Using topic modelling, we assess whether cross-genre transfer in dependency parsing is stable with respect to topic distribution. We show that LAS scores in cross-genre transfer within and across treebanks typically align with topic distances. This indicates that topic is an important explanatory factor for genre transfer.

We present version 1.3 of the PARSEME multilingual corpus annotated with verbal multiword expressions. Since the previous version, new languages have joined the undertaking of creating such a resource, some of the already existing corpora have been enriched with new annotated texts, while others have been enhanced in various ways. The PARSEME multilingual corpus represents 26 languages now. All monolingual corpora therein use Universal Dependencies v.2 tagset. They are (re-)split observing the PARSEME v.1.2 standard, which puts impact on unseen VMWEs. With the current iteration, the corpus release process has been detached from shared tasks; instead, a process for continuous improvement and systematic releases has been introduced.

Extracting statements about causality from text documents is a challenging task in the absence of annotated training data. We create a search system for causal statements about user-specified concepts by combining pattern matching of causal connectives with semantic similarity ranking, using a language model fine-tuned for semantic textual similarity. Preliminary experiments on a small test set from Swedish governmental reports show promising results in comparison to two simple baselines.

Multiword expressions (MWEs) are challenging and pervasive phenomena whose idiosyncratic properties show notably at the levels of lexicon, morphology, and syntax. Thus, they should best be annotated jointly with morphosyntax. We discuss two multilingual initiatives, Universal Dependencies and PARSEME, addressing these annotation layers in cross-lingually unified ways. We compare the annotation principles of these initiatives with respect to MWEs, and we put forward a roadmap towards their gradual unification. The expected outcomes are more consistent treebanking and higher universality in modeling idiosyncrasy.

In this paper, we present principles of constructing and resolving ambiguity in implicit discourse relations. Following these principles, we created a dataset in both English and Egyptian Arabic that controls for semantic disambiguation, enabling the investigation of prosodic features in future work. In these datasets, examples are two-part sentences with an implicit discourse relation that can be ambiguously read as either causal or concessive, paired with two different preceding context sentences forcing either the causal or the concessive reading. We also validated both datasets by humans and language models (LMs) to study whether context can help humans or LMs resolve ambiguities of implicit relations and identify the intended relation. As a result, this task posed no difficulty for humans, but proved challenging for BERT/CamelBERT and ELECTRA/AraELECTRA models.

In this study, we aim to find a parser for accurately identifying different types of subordinate clauses, and related phenomena, in 19th–20th-century Swedish literature. Since no test set is available for parsing from this time period, we propose a lightweight annotation scheme for annotating a single relation of interest per sentence. We train a variety of parsers for Swedish and compare evaluations on standard modern test sets and our targeted test set. We find clear trends in which parser types perform best on the standard test sets, but that performance is considerably more varied on the targeted test set. We believe that our proposed annotation scheme can be useful for complementing standard evaluations, with a low annotation effort.

We investigate the effectiveness of multilingual automatic speech recognition models for Scandinavian languages by further fine-tuning a Swedish model on Swedish, Danish, and Norwegian. We first explore zero-shot models, which perform poorly across the three languages. However, we show that a multilingual model based on a strong Swedish model, further fine-tuned on all three languages, performs well for Norwegian and Danish, with a relatively low decrease in the performance for Swedish. With a language classification module, we improve the performance of the multilingual model even further.

In this paper, we describe version 2.0 of the SLäNDa corpus. SLäNDa, the Swedish Literary corpus of Narrative and Dialogue, now contains excerpts from 19 novels, written between 1809–1940. The main focus of the SLäNDa corpus is to distinguish between direct speech and the main narrative. In order to isolate the narrative, we also annotate everything else which does not belong to the narrative, such as thoughts, quotations, and letters. SLäNDa version 2.0 has a slightly updated annotation scheme from version 1.0. In addition, we added new texts from eleven authors and performed quality control on the previous version. We are specifically interested in different ways of marking speech segments, such as quotation marks, dashes, or no marking at all. To allow a detailed evaluation of this aspect, we added dedicated test sets to SLäNDa for these different types of speech marking. In a pilot experiment, we explore the impact of typographic speech marking by using these test sets, as well as artificially stripping the training data of speech markers.

We present the Uppsala University system for SemEval-2022 Task 1: Comparing Dictionaries and Word Embeddings (CODWOE). We explore the performance of multilingual reverse dictionaries as well as the possibility of utilizing annotated data in other languages to improve the quality of a reverse dictionary in the target language. We mainly focus on character-based embeddings.In our main experiment, we train multilingual models by combining the training data from multiple languages. In an additional experiment, using resources beyond the shared task, we use the training data in Russian and French to improve the English reverse dictionary using unsupervised embeddings alignment and machine translation. The results show that multilingual models occasionally but not consistently can outperform the monolingual baselines. In addition, we demonstrate an improvement of an English reverse dictionary using translated entries from the Russian training data set.

Causality detection is the task of extracting information about causal relations from text. It is an important task for different types of document analysis, including political impact assessment. We present two new data sets for causality detection in Swedish. The first data set is annotated with binary relevance judgments, indicating whether a sentence contains causality information or not. In the second data set, sentence pairs are ranked for relevance with respect to a causality query, containing a specific hypothesized cause and/or effect. Both data sets are carefully curated and mainly intended for use as test data. We describe the data sets and their annotation, including detailed annotation guidelines. In addition, we present pilot experiments on cross-lingual zero-shot and few-shot causality detection, using training data from English and German.

We describe the Uppsala NLP submission to SemEval-2021 Task 2 on multilingual and cross-lingual word-in-context disambiguation. We explore the usefulness of three pre-trained multilingual language models, XLM-RoBERTa (XLMR), Multilingual BERT (mBERT) and multilingual distilled BERT (mDistilBERT). We compare these three models in two setups, fine-tuning and as feature extractors. In the second case we also experiment with using dependency-based information. We find that fine-tuning is better than feature extraction. XLMR performs better than mBERT in the cross-lingual setting both with fine-tuning and feature extraction, whereas these two models give a similar performance in the multilingual setting. mDistilBERT performs poorly with fine-tuning but gives similar results to the other models when used as a feature extractor. We submitted our two best systems, fine-tuned with XLMR and mBERT.

Choosing a transfer language is a crucial step in transfer learning. In much previous research on dependency parsing, related languages have successfully been used. However, when parsing Latin, it has been suggested that languages such as ancient Greek could be helpful. In this work we parse Latin in a low-resource scenario, with the main goal to investigate if Greek languages are more helpful for parsing Latin than related Italic languages, and show that this is indeed the case. We further investigate the influence of other factors including training set size and content as well as linguistic distances. We find that one explanatory factor seems to be the syntactic similarity between Latin and Ancient Greek. The influence of genres or shared annotation projects seems to have a smaller impact.

In this paper we explore PoS tagging for the Scots language. Scots is spoken in Scotland and Northern Ireland, and is closely related to English. As no linguistically annotated Scots data were available, we manually PoS tagged a small set that is used for evaluation and training. We use English as a transfer language to examine zero-shot transfer and transfer learning methods. We find that training on a very small amount of Scots data was superior to zero-shot transfer from English. Combining the Scots and English data led to further improvements, with a concatenation method giving the best results. We also compared the use of two different English treebanks and found that a treebank containing web data was superior in the zero-shot setting, while it was outperformed by a treebank containing a mix of genres when combined with Scots data.

Exploring aspects of sentential meaning that are implicit or underspecified in context is important for sentence understanding. In this paper, we propose a novel architecture based on mentions for revision requirements detection. The goal is to improve understandability, addressing some types of revisions, especially for the Replaced Pronoun type. We show that our mention-based system can predict replaced pronouns well on the mention-level. However, our combined sentence-level system does not improve on the sentence-level BERT baseline. We also present additional contrastive systems, and show results for each type of edit.

We describe a new corpus, SLäNDa, the Swedish Literary corpus of Narrative and Dialogue. It contains Swedish literary fiction, which has been manually annotated for cited materials, with a focus on dialogue. The annotation covers excerpts from eight Swedish novels written between 1879–1940, a period of modernization of the Swedish language. SLäNDa contains annotations for all cited materials that are separate from the main narrative, like quotations and signs. The main focus is on dialogue, for which we annotate speech segments, speech tags, and speakers. In this paper we describe the annotation protocol and procedure and show that we can reach a high inter-annotator agreement. In total, SLäNDa contains annotations of 44 chapters with over 220K tokens. The annotation identified 4,733 instances of cited material and 1,143 named speaker–speech mappings. The corpus is useful for developing computational tools for different types of analysis of literary narrative and speech. We perform a small pilot study where we show how our annotation can help in analyzing language change in Swedish. We find that a number of common function words have their modern version appear earlier in speech than in narrative.

There is a growing interest in investigating what neural NLP models learn about language. A prominent open question is the question of whether or not it is necessary to model hierarchical structure. We present a linguistic investigation of a neural parser adding insights to this question. We look at transitivity and agreement information of auxiliary verb constructions (AVCs) in comparison to finite main verbs (FMVs). This comparison is motivated by theoretical work in dependency grammar and in particular the work of Tesnière (1959), where AVCs and FMVs are both instances of a nucleus, the basic unit of syntax. An AVC is a dissociated nucleus; it consists of at least two words, and an FMV is its non-dissociated counterpart, consisting of exactly one word. We suggest that the representation of AVCs and FMVs should capture similar information. We use diagnostic classifiers to probe agreement and transitivity information in vectors learned by a transition-based neural parser in four typologically different languages. We find that the parser learns different information about AVCs and FMVs if only sequential models (BiLSTMs) are used in the architecture but similar information when a recursive layer is used. We find explanations for why this is the case by looking closely at how information is learned in the network and looking at what happens with different dependency representations of AVCs. We conclude that there may be benefits to using a recursive layer in dependency parsing and that we have not yet found the best way to integrate it in our parsers.

We discuss a set of methods for the creation of IESTAC: a English-Italian speech and text parallel corpus designed for the training of end-to-end speech-to-text machine translation models and publicly released as part of this work. We first mapped English LibriVox audiobooks and their corresponding English Gutenberg Project e-books to Italian e-books with a set of three complementary methods. Then we aligned the English and the Italian texts using both traditional Gale-Church based alignment methods and a recently proposed tool to perform bilingual sentences alignment computing the cosine similarity of multilingual sentence embeddings. Finally, we forced the alignment between the English audiobooks and the English side of our textual parallel corpus with a text-to-speech and dynamic time warping based forced alignment tool. For each step, we provide the reader with a critical discussion based on detailed evaluation and comparison of the results of the different methods.

Code-mixed texts are abundant, especially in social media, and poses a problem for NLP tools, which are typically trained on monolingual corpora. In this paper, we explore and evaluate different types of word embeddings for Indonesian–English code-mixed text. We propose the use of code-mixed embeddings, i.e. embeddings trained on code-mixed text. Because large corpora of code-mixed text are required to train embeddings, we describe a method for synthesizing a code-mixed corpus, grounded in literature and a survey. Using sentiment analysis as a case study, we show that code-mixed embeddings trained on synthesized data are at least as good as cross-lingual embeddings and better than monolingual embeddings.

We present edition 1.2 of the PARSEME shared task on identification of verbal multiword expressions (VMWEs). Lessons learned from previous editions indicate that VMWEs have low ambiguity, and that the major challenge lies in identifying test instances never seen in the training data. Therefore, this edition focuses on unseen VMWEs. We have split annotated corpora so that the test corpora contain around 300 unseen VMWEs, and we provide non-annotated raw corpora to be used by complementary discovery methods. We released annotated and raw corpora in 14 languages, and this semi-supervised challenge attracted 7 teams who submitted 9 system results. This paper describes the effort of corpus creation, the task design, and the results obtained by the participating systems, especially their performance on unseen expressions.

We present the Uppsala system for the CoNLL 2018 Shared Task on universal dependency parsing. Our system is a pipeline consisting of three components: the first performs joint word and sentence segmentation; the second predicts part-of-speech tags and morphological features; the third predicts dependency trees from words and tags. Instead of training a single parsing model for each treebank, we trained models with multiple treebanks for one language or closely related languages, greatly reducing the number of models. On the official test run, we ranked 7th of 27 teams for the LAS and MLAS metrics. Our system obtained the best scores overall for word segmentation, universal POS tagging, and morphological features.

How to make the most of multiple heterogeneous treebanks when training a monolingual dependency parser is an open question. We start by investigating previously suggested, but little evaluated, strategies for exploiting multiple treebanks based on concatenating training sets, with or without fine-tuning. We go on to propose a new method based on treebank embeddings. We perform experiments for several languages and show that in many cases fine-tuning and treebank embeddings lead to substantial improvements over single treebanks or concatenation, with average gains of 2.0–3.5 LAS points. We argue that treebank embeddings should be preferred due to their conceptual simplicity, flexibility and extensibility.

We present an analysis of a number of coreference phenomena in English-Croatian human and machine translations. The aim is to shed light on the differences in the way these structurally different languages make use of discourse information and provide insights for discourse-aware machine translation system development. The phenomena are automatically identified in parallel data using annotation produced by parsers and word alignment tools, enabling us to pinpoint patterns of interest in both languages. We make the analysis more fine-grained by including three corpora pertaining to three different registers. In a second step, we create a test set with the challenging linguistic constructions and use it to evaluate the performance of three MT systems. We show that both SMT and NMT systems struggle with handling these discourse phenomena, even though NMT tends to perform somewhat better than SMT. By providing an overview of patterns frequently occurring in actual language use, as well as by pointing out the weaknesses of current MT systems that commonly mistranslate them, we hope to contribute to the effort of resolving the issue of discourse phenomena in MT applications.

We provide a comprehensive analysis of the interactions between pre-trained word embeddings, character models and POS tags in a transition-based dependency parser. While previous studies have shown POS information to be less important in the presence of character models, we show that in fact there are complex interactions between all three techniques. In isolation each produces large improvements over a baseline system using randomly initialised word embeddings only, but combining them quickly leads to diminishing returns. We categorise words by frequency, POS tag and language in order to systematically investigate how each of the techniques affects parsing quality. For many word categories, applying any two of the three techniques is almost as good as the full combined system. Character models tend to be more important for low-frequency open-class words, especially in morphologically rich languages, while POS tags can help disambiguate high-frequency function words. We also show that large character embedding sizes help even for languages with small character sets, especially in morphologically rich languages.

We present the Uppsala submission to the CoNLL 2017 shared task on parsing from raw text to universal dependencies. Our system is a simple pipeline consisting of two components. The first performs joint word and sentence segmentation on raw text; the second predicts dependency trees from raw words. The parser bypasses the need for part-of-speech tagging, but uses word embeddings based on universal tag distributions. We achieved a macro-averaged LAS F1 of 65.11 in the official test run, which improved to 70.49 after bug fixes. We obtained the 2nd best result for sentence segmentation with a score of 89.03.

We describe the design, the setup, and the evaluation results of the DiscoMT 2017 shared task on cross-lingual pronoun prediction. The task asked participants to predict a target-language pronoun given a source-language pronoun in the context of a sentence. We further provided a lemmatized target-language human-authored translation of the source sentence, and automatic word alignments between the source sentence words and the target-language lemmata. The aim of the task was to predict, for each target-language pronoun placeholder, the word that should replace it from a small, closed set of classes, using any type of information that can be extracted from the entire document. We offered four subtasks, each for a different language pair and translation direction: English-to-French, English-to-German, German-to-English, and Spanish-to-English. Five teams participated in the shared task, making submissions for all language pairs. The evaluation results show that most participating teams outperformed two strong n-gram-based language model-based baseline systems by a sizable margin.

We describe the Uppsala system for the 2017 DiscoMT shared task on cross-lingual pronoun prediction. The system is based on a lower layer of BiLSTMs reading the source and target sentences respectively. Classification is based on the BiLSTM representation of the source and target positions for the pronouns. In addition we enrich our system with dependency representations from an external parser and character representations of the source sentence. We show that these additions perform well for German and Spanish as source languages. Our system is competitive and is in first or second place for all language pairs.

In this paper, we extend the arc-hybrid system for transition-based parsing with a swap transition that enables reordering of the words and construction of non-projective trees. Although this extension breaks the arc-decomposability of the transition system, we show how the existing dynamic oracle for this system can be modified and combined with a static oracle only for the swap transition. Experiments on 5 languages show that the new system gives competitive accuracy and is significantly better than a system trained with a purely static oracle.

We present a preliminary study where we use eye tracking as a complement to machine translation (MT) error analysis, the task of identifying and classifying MT errors. We performed a user study where subjects read short texts translated by three MT systems and one human translation, while we gathered eye tracking data. The subjects were also asked comprehension questions about the text, and were asked to estimate the text quality. We found that there are a longer gaze time and a higher number of fixations on MT errors, than on correct parts. There are also differences in the gaze time of different error types, with word order errors having the longest gaze time. We also found correlations between eye tracking data and human estimates of text quality. Overall our study shows that eye tracking can give complementary information to error analysis, such as aiding in ranking error types for seriousness.

Error analysis is a means to assess machine translation output in qualitative terms, which can be used as a basis for the generation of error profiles for different systems. As for other subjective approaches to evaluation it runs the risk of low inter-annotator agreement, but very often in papers applying error analysis to MT, this aspect is not even discussed. In this paper, we report results from a comparative evaluation of two systems where agreement initially was low, and discuss the different ways we used to improve it. We compared the effects of using more or less fine-grained taxonomies, and the possibility to restrict analysis to short sentences only. We report results on inter-annotator agreement before and after measures were taken, on error categories that are most likely to be confused, and on the possibility to establish error profiles also in the absence of a high inter-annotator agreement.

We present a method for improving word alignment quality for phrase-based statistical machine translation by reordering the source text according to the target word order suggested by an initial word alignment. The reordered text is used to create a second word alignment which can be an improvement of the first alignment, since the word order is more similar. The method requires no other pre-processing such as part-of-speech tagging or parsing. We report improved Bleu scores for English-to-German and English-to-Swedish translation. We also examined the effect on word alignment quality and found that the reordering method increased recall while lowering precision, which partly can explain the improved Bleu scores. A manual evaluation of the translation output was also performed to understand what effect our reordering method has on the translation system. We found that where the system employing reordering differed from the baseline in terms of having more words, or a different word order, this generally led to an improvement in translation quality.

One problem in statistical machine translation (SMT) is that the output often is ungrammatical. To address this issue, we have investigated the use of a grammar checker for two purposes in connection with SMT: as an evaluation tool and as a postprocessing tool. To assess the feasibility of the grammar checker on SMT output, we performed an error analysis, which showed that the precision of error identification in general was higher on SMT output than in previous studies on human texts. Using the grammar checker as an evaluation tool gives a complementary picture to standard metrics such as Bleu, which do not account well for grammaticality. We use the grammar checker as a postprocessing tool by automatically applying the error correction suggestions it gives. There are only small overall improvements of the postprocessing on automatic metrics, but the sentences that are affected by the changes are improved, as shown both by automatic metrics and by a human error analysis. These results indicate that grammar checker techniques are a useful complement to SMT.