Det svenska språket ska finnas tillgängligt för alla som bor och verkar i Sverige. Därförär det viktigt att det finns lättlästa alternativ för dem som har svårighet att läsa svensktext. Detta arbete bygger vidare på att visa att det är möjligt att skapa ett automatisktomskrivningsprogram som gör texter mer lättlästa. Till grund för arbetet liggerCogFLUX som är ett verktyg för automatisk omskrivning till lätt svenska. CogFLUXinnehåller funktioner för att syntaktiskt skriva om texter till mer lättläst svenska.Omskrivningarna görs med hjälp av omskrivningsregler framtagna i ett tidigare projekt.I detta arbete implementeras ytterligare omskrivningsregler och även en ny modul förhantering av synonymer. Med dessa nya regler och modulen ska arbetet undersöka omdet är det är möjligt att skapa system som ger en mer lättläst text enligt etableradeläsbarhetsmått som LIX, OVIX och Nominalkvot. Omskrivningsreglerna ochsynonymhanteraren testas på tre olika texter med en total lägnd på ungefär hundra tusenord. Arbetet visar att det går att sänka både LIX-värdet och Nominalkvoten signifikantmed hjälp av omskrivningsregler och synonymhanterare. Arbetet visar även att det finnsfler saker kvar att göra för att framställa ett riktigt bra program för automatiskomskrivning till lätt svenska.

This paper profiles the Europarl part of an English-Swedish parallel corpus and compares it with three other subcorpora of the sameparallel corpus. We first describe our method for comparison which is based on alignments, both at the token level and the structurallevel. Although two of the other subcorpora contains fiction, it is found that the Europarl part is the one having the highest proportion ofmany types of restructurings, including additions, deletions and long distance reorderings. We explain this by the fact that the majorityof Europarl segments are parallel translations.

As machine translation technology improves comparisons to human performance are often made in quite general and exaggerated terms. Thus, it is important to be able to account for differences accurately. This paper reports a simple, descriptive scheme for comparing translations and applies it to two translations of a British opinion article published in March, 2017. One is a human translation (HT) into Swedish, and the other a machine translation (MT). While the comparison is limited to one text, the results are indicative of current limitations in MT.

The paper reports experiences of automatically converting the dependency analysis of the LinES English-Swedish parallel treebank to universal dependencies (UD). The most tangible result is a version of the treebank that actually employs the relations and parts-of-speech categories required by UD, and no other. It is also more complete in that punctuation marks have received dependencies, which is not the case in the original version. We discuss our method in the light of problems that arise from the desire to keep the syntactic analyses of a parallel treebank internally consistent, while available monolingual UD treebanks for English and Swedish diverge somewhat in their use of UD annotations. Finally, we compare the output from the conversion program with the existing UD treebanks.

In principle the CLARIN research infrastructure provides a good environment to support research on translation. In reality, the progress within CLARIN in this area seems to be fairly slow. In this paper I will give examples of the resources currently available, and suggest what is needed to achieve a relevant research infrastructure for translation studies. Also, I argue that translation studies has more to gain from language technology, and statistical machine translation in particular, than what is generally assumed, and give some examples.

The parataxis relation as defined for Universal Dependencies 2.0 is general and, for this reason,sometimes hard to distinguish from competing analyses, such as coordination, conj, or apposi-tion, appos. The specific subtypes that are listed for parataxis are also quite different in character.In this study we first show that the actual practice by UD-annotators is varied, using the parallelUD (PUD-) treebanks as data. We then review the current definitions and guidelines and suggestimprovements.

In Swedish, as in other languages, the words used to refer to disabilities and people with disabilities are manifold. Recommendations as to which terms to use have been changed several times over the last hundred years. In this exploratory paper we have used textual resources provided by Swe-Clarin to study such changes quantitatively. We demonstrate that old and new recommendations co-exist for long periods of time, and that usage sometimes converges.

We show how the tools and language resources developed within the SweClarin infrastructure can be used to investigate changes in the use and understanding of the Swedish related words arkitektur, design, form, and formgivning. Specifically, we compare their use in two governmental public reports on design, one from 1999 and the other from 2015. We test the hypothesis that their meaning has developed in a way that blurs distinctions that may be important to stakeholders in the respective fields.

We investigate changes in the use of four Swedish words from the fields of design and archi- tecture. It has been suggested that their meanings have been blurred, especially in governmental reports and policy documents, so that distinctions between them that are important to stakeholders in the respective fields are lost. Specifically, we compare usage in two governmental public reports on design, one from 1999 and the other from 2015, and additionally in opinion responses to the 2015 report. Our approach is to contextualise occurrences of the words in different representations of the texts using word embeddings, topic modelling and sentiment analysis. Tools and language resources developed within the SweClarin infrastructure have been crucial for the implementation of the study.

The most promising approach to word alignment is to combine statistical methods with non-statistical information sources. Some of the proposed non-statistical sources, including bilingual dictionaries, POS-taggers and lemmatizers, rely on considerable linguistic knowledge, while other knowledge-lite sources such as cognate heuristics and word order heuristics can be implemented relatively easy. While knowledge-heavy sources might be expected to give better performance, knowledge-lite systems are easier to port to new language pairs and text types, and they can give sufficiently good results for many purposes, e.g. if the output is to be used by a human user for the creation of a complete word-aligned bitext. In this paper we describe the current status of the Linköping Word Aligner (LWA), which combines the use of statistical measures of co-occurrence with four knowledge-lite modules for (i)) word categorization, (ii) morphological variation, (iii) word order, and (iv) phrase recognition. We demonstrate the portability of the system (from English-Swedish texts to French-English texts) and present results for these two language-pairs. Finally, we will report observations from an error analysis of system output, and identify the major strengths and weaknesses of the system.

As empirical methods have come to the fore in multilingual language technology and translation studies, the processing of parallel texts and parallel corpora have become a major research area in computational linguistics. In this article we review the state of the art in alignment and data extraction techniques for parallel texts, and give an overview of current work in Sweden in this area. In a final section, we summarize the results achieved so far and make some proposals for future research.

Comparable or parallel corpora are beneficial for many NLP tasks.  The automatic collection of corpora enables large-scale resources, even for less-resourced languages, which in turn can be useful for deducing rules and patterns for text rewriting algorithms, a subtask of automatic text simplification. We present two methods for the alignment of Swedish easy-to-read text segments to text segments from a reference corpus.  The first method (M1) was originally developed for the task of text reuse detection, measuring sentence similarity by a modified version of a TF-IDF vector space model. A second method (M2), also accounting for part-of-speech tags, was devel- oped, and the methods were compared.  For evaluation, a crowdsourcing platform was built for human judgement data collection, and preliminary results showed that cosine similarity relates better to human ranks than the Dice coefficient. We also saw a tendency that including syntactic context to the TF-IDF vector space model is beneficial for this kind of paraphrase alignment task.

Finding the appropriate level of text for students with varying reading abilities is an important and demanding task for teachers. Radar chart visualisations of text complexity could potentially be an aid in that process, but they need to be evaluated to see if they are intuitive and if they say something about the complexity of a text. This study explores how visualisations of text complexity, in the format of radar charts, are interpreted, what measures they should include and what information they should contain in order to be intelligible for teachers who work with people who have language and/or reading diffi- culties. A preliminary study and three focus group sessions were conducted with teachers from special education schools for adults and gymnasium level. Through thematic analysis of the sessions, five themes were generated and it was found that the visualisations are intelligible to some extent, but they need to be adapted to the target group by making sure the measures are relevant, and that the scale, colours, categories and measures are clearly explained.

Developments in deep learning and machine learning overall has created a plethora of opportunities for easier training of automatic text summarization (ATS) models for producing summaries with higher quality. ATS can be split into extractive and abstractive tasks; extractive models extract sentences from the original text to create summaries. On the contrary, abstractive models generate novel sentences to create summaries. While extractive summaries are often preferred over abstractive ones, summaries created by extractive models trained on Swedish texts often lack cohesion, which affects the readability and overall quality of the summary. Therefore, there is a need to improve the process of training ATS models in terms of cohesion, while maintaining other text qualities such as content coverage. This thesis explores and implements methods at the data-processing stage aimed at improving cohesion of generated summaries. The methods are based around Sentence-BERT for creating advanced sentence embeddings that can be used to rank sentences in a text in terms of if it should be included in the extractive summary or not. Three models are trained using different methods and evaluated using ROUGE, BERTScore for measuring content coverage and Coh-Metrix for measuring cohesion. The results of the evaluation suggest that the methods can indeed be used to create more cohesive summaries, although content coverage was reduced, which gives rise to the potential for extensive future exploration of further implementation. 

Topic models are useful tools for exploring large data sets of textual content by exposing a generative process from which the text was produced. Anchor-based topic models utilize the anchor word assumption to define a set of algorithms with provable guarantees which recover the underlying topics with a run time practically independent of corpus size. A number of extensions to the initial anchor word-based algorithms, and enhancements made to tangential models, have been proposed which improve the intrinsic characteristics of the model making them more interpretable by humans. This thesis evaluates improvements to human interpretability due to: low-dimensional word embeddings in combination with a regularized objective function, automatic topic merging using tandem anchors, and utilizing word embeddings to synthetically increase corpus density. Results show that tandem anchors are viable vehicles for automatic topic merging, and that using word embeddings significantly improves the original anchor method across all measured metrics. Combining low-dimensional embeddings and a regularized objective results in computational downsides with small or no improvements to the metrics measured.

In this paper, Finite-State-Automata (FSA) and theorem-proving approaches to spoken dialogue systems (SLDS) are contrasted to each other. FSA are too rigid to deal with unpredictable user reactions, such as corrections or counter-questions, whereas plan-based approaches are usually too complex to be effectively used, given the unreliability of word recognition and the elliptical and unconventional nature of spontaneous speech. As an alternative, a Dialogue Manager architecture is proposed which uses knowledge on both the possible sequences of dialogue acts and the dynamic representation of the task and requirements for its fulfillment. The behaviour of the specific user is taken into consideration, including their expectations about the system and the service offered, as are instances of miscommunication and disagreement in the course of the dialogue, and the successful completion of sub-plans relevant to the task and the dialogue flow.

Access to various texts in different languages causes an increasing demand for fast, multi-purpose, and cheap translators. Pervasive internet use intensifies the necessity for intelligent and cheap translators, since traditional translation methods are excessively slow to translate different texts. During the past years, scientists carried out much research in order to add human and artificial intelligence into the old machine translation systems and the idea of developing a machine translation system came into existence during the days of World War (Kohenn, 2010). The new invention was useful in order to help the human translators and many other people who need to translate different types of texts according to their needs. The new translation systems are useful in meeting people’s needs. Since the machine translation systems vary according to the quality of the systems outputs, their performance should be evaluated from the linguistic point of view in order to reach a fair judgment about the quality of the systems outputs. To achieve this goal, two various Swedish texts were translated by two different machine translation systems in the thesis. The translated texts were evaluated to examine the extent to which errors affect the comprehensibility of the translations. The performances of the systems were evaluated using three approaches. Firstly, most common linguistically errors, which appear in the machine translation systems outputs, were analyzed (e.g. word alignment of the translated texts). Secondly, the influence of different types of errors on the cohesion chains were evaluated. Finally, the effect of the errors on the comprehensibility of the translations were investigated.

Numerical results showed that some types of errors have more effects on the comprehensibility of the systems’ outputs. The obtained data illustrated that the subjects’ comprehension of the translated texts depend on the type of error, but not frequency. The analyzing depicted which translation system had best performance.

Humans receive a constant stream of input that potentially influence their affective experience. Despite intensive research on affect, it is still largely unknown how various sources of information are integrated into the single, unified affective features that accompany consciousness. Here, we aimed to investigate how a stream of evocative input we receive is dynamically represented in self-reported affect. In 4 experiments, participants viewed a number of sequentially presented images and reported their momentary affective experience on valence and arousal scales. The number and duration of images in a trial varied across studies. In Study 4, we also measured participants physiological responses while they viewed images. We formulated and compared several models with respect to their capacity to predict self-reported affect based on normative image ratings, physiological measurements, and prior affective experience (measured in the previous trial). Our data best supported a model incorporating a temporally sensitive averaging mechanism for affective integration that assigns higher weights to effectively more potent and recently represented stimuli. Crucially, affective averaging of sensory information and prior affect accounted for distinct contributions to currently experienced affect. Taken together, the current study provides evidence that prior affect and integrated affective impact of stimuli partly shape currently experienced affect.

Affect fluctuates in a moment-to-moment fashion, reflecting the continuous relationship between the individual and the environment. Despite substantial research, there remain important open questions regarding how a stream of sensory input is dynamically represented in experienced affect. Here, approaching affect as a temporally dependent process, we show that momentary affect is shaped by a combination of the affective impact of stimuli (i.e., visual images for the current studies) and previously experienced affect. We also found that this temporal dependency is influenced by uncertainty of the affective context. Participants in each trial viewed sequentially presented images and subsequently reported their affective experience, which was modeled based on images normative affect ratings and participants previously reported affect. Study 1 showed that self-reported valence and arousal in a given trial is partly shaped by the affective impact of the given images and previously experienced affect. In Study 2, we manipulated context uncertainty by controlling occurrence probabilities for normatively pleasant and unpleasant images in separate blocks. Increasing context uncertainty (i.e., random occurrence of pleasant and unpleasant images) was associated with increased negative affect. In addition, the relative contribution of the most recent image to experienced pleasantness increased with increasing context uncertainty. Taken together, these findings provide clear behavioral evidence that momentary affect is a temporally dependent and continuous process, which reflects the affective impact of recent input variables and the previous internal state, and that this process is sensitive to the affective context and its uncertainty.

This paper is concerned with the extraction of the surface topology of tensor fields on 2D triangulated manifoldsembedded in 3D. In scientific visualization topology is a meaningful instrument to get a hold on the structure of agiven dataset. Due to the discontinuity of tensor fields on a piecewise planar domain, standard topology extractionmethods result in an incomplete topological skeleton. In particular with regard to the high computational costs ofthe extraction this is not satisfactory. This paper provides a method for topology extraction of tensor fields thatleads to complete results. The core idea is to include the locations of discontinuity into the topological analysis.For this purpose the model of continuous transition bridges is introduced, which allows to capture the entiretopology on the discontinuous field. The proposed method is applied to piecewise linear three-dimensional tensorfields defined on the vertices of the triangulation and for piecewise constant two or three-dimensional tensor fieldsgiven per triangle, e.g. rate of strain tensors of piecewise linear flow fields.

Dependency parsing can be a useful tool to allow computers to parse text. In 2015, Kuhlmann and Jonsson proposed a logical deduction system that parsed to non-crossing dependency graphs with an asymptotic time complexity of O(n³), where “n” is the length of the sentence to parse. This thesis extends the deduction system by Kuhlmann and Jonsson; the extended deduction system introduces certain crossing edges, while maintaining an asymptotic time complexity of O(n⁴).

In order to extend the deduction system by Kuhlmann and Jonsson, fifteen logical item types are added to the five proposed by Kuhlmann and Jonsson. These item types allow the deduction system to intro-duce crossing edges while acyclicity can be guaranteed. The number of inference rules in the deduction system is increased from the 19 proposed by Kuhlmann and Jonsson to 172, mainly because of the larger number of combinations of the 20 item types.

The results are a modest increase in coverage on test data (by roughly 10% absolutely, i.e. approx. from 70% to 80%), and a comparable placement to that of Kuhlmann and Jonsson by the SemEval 2015 task 18 metrics. By the method employed to introduce crossing edges, derivational uniqueness is impossible to maintain. It is hard to defien the graph class to which the extended algorithm, QAC, parses, and it is therefore empirically compared to 1-endpoint crossing and graphs with a page number of two or less, compared to which it achieves lower coverage on test data. The QAC graph class is not limited by page number or crossings.

The takeaway of the thesis is that extending a very minimal deduction system is not necessarily the best approach, and that it may be better to start off with a strong idea of to which graph class the extended algorithm should parse. Additionally, several alternative ways of extending Kuhlmann and Jonsson are proposed.

Aligning parallell terms in a parallell corpus can be done by aligning all words and phrases in the corpus and then performing term extraction on the aligned set of word pairs. Alternatively, term extraction in the source and target text can be made separately and then the resulting term candidates can be aligned, forming aligned parallell terms. This thesis describes an implementation of a word aligner that is applied on extracted term candidates in both the source and the target texts. The term aligner uses statistical measures, the tool Giza++ and heuristics in the search for alignments. The evaluation reveals that the best results are obtained when the term alignment relies heavily on the Giza++ tool and Levenshtein heuristic.

This study conducts an experimental evaluation of two hypotheses about the contributions of formal and semantic features to the grammatical gender assignment of nouns. One of the hypotheses (Corbett and Fraser 2000) claims that semantic features dominate formal ones. The other hypothesis, formulated within the optimal gender assignment theory (Rice 2006), states that form and semantics contribute equally. Both hypotheses claim that the combination of formal and semantic features yields the most accurate gender identification. In this paper, we operationalize and test these hypotheses by trying to predict grammatical gender using only character-based embeddings (that capture only formal features), only context-based embeddings (that capture only semantic features) and the combination of both. We performed the experiment using data from three languages with different gender systems (French, German and Russian). Formal features are a significantly better predictor of gender than semantic ones, and the difference in prediction accuracy is very large. Overall, formal features are also significantly better than the combination of form and semantics, but the difference is very small and the results for this comparison are not entirely consistent across languages.

A primary goal with this thesis work has been to investigate the consequences of a token-based program representation. Among the results which are presented here are an incremental scanning algorithm together with a token-based syntax sensitive editing approach for program editing.The design and implementation of an incremental scanner and a practically useful syntax-sensitive editor is described in some detail. The language independent incremental scanner converts textual edit operations to corresponding operations on the token sequence. For example, user input is converted to tokens as it is typed in. This editor design makes it possible to edit programs with almost the same flexibility as with a conventional text editor and also provides some features offered by a syntax-directed editor, such as template instantiation, automatic indentation and prettyprinting, lexical and syntactic error handling.We have found that a program represented as a token sequence can on the average be represented in less than half the storage space required for a program in text form. Also, interactive syntax checking is speeded up since rescanning is not needed.The current implementation, called TOSSED - Token-based Syntax Sensitive Editor, supports editing and development of programs written in Pascal. The user is guaranteed a lexically and syntactically correct program on exit from the editor, which avoids many unnecessary compilations. The scanner, parser, prettyprinter, and syntactic error recovery are table-driven and language independent template specification is supported. Thus, editors supporting other languages can be generated.

One of the largest challenges in the field of artificial intelligence and machine learning isthe acquisition of a large quantity of quality data to train models on.This thesis investigates and evaluates approaches to data generation in a telecom domain for the task of extractive QA. To do this a pipeline was built using a combination ofBERT-like models and T5 models for data generation. We then evaluated our generateddata using the downstream task of QA on a telecom domain data set. We measured theperformance using EM and F1-scores. We achieved results that are state of the art on thetelecom domain data set.We found that synthetic data generation is a viable approach to obtaining synthetictelecom QA data with the potential of improving model performance when used in addition to human-annotated data. We also found that using models from the general domainprovided results that are on par or better than domain-specific models for the generation, which provides possibilities to use a single generation pipeline for many differentdomains. Furthermore, we found that increasing the amount of synthetic data providedlittle benefit for our models on the downstream task, with diminishing returns setting inquickly. We were unable to pinpoint the reason for this. In short, our approach works butmuch more work remains to understand and optimize it for greater results

This thesis investigates how pre-trained contextualized language models can be adapted for multi-label text classification of Swedish news articles. Various classifiers are built on pre-trained BERT and ELECTRA models, exploring global and local classifier approaches. Furthermore, the effects of domain specialization, using additional metadata features and model compression are investigated. Several hundred thousand news articles are gathered to create unlabeled and labeled datasets for pre-training and fine-tuning, respectively. The findings show that a local classifier approach is superior to a global classifier approach and that BERT outperforms ELECTRA significantly. Notably, a baseline classifier built on SVMs yields competitive performance. The effect of further in-domain pre-training varies; ELECTRA’s performance improves while BERT’s is largely unaffected. It is found that utilizing metadata features in combination with text representations improves performance. Both BERT and ELECTRA exhibit robustness to quantization and pruning, allowing model sizes to be cut in half without any performance loss.

We describe an implemented spoken-language dialogue system for a travel-planning domain, which accesses a commercially available travel-information web-server and supports a flexible mixed-initiative dialogue strategy. We argue, based on data from initial Wizard-of-Oz experiments, that mixed-initiative strategies are appropriate for many types of user, but require more sophisticated architectures for processing of language and dialogue; we then use these observations to motivate an architecture which combines parallel deep and shallow natural language analysis engines and an agenda-driven dialogue manager. We outline the top-level processing strategy used by the dialogue manager, and also a novel formalism, which we call Flat Utterance Description, that allows us to reduce the output of the deep and shallow language-processing engines to a common representation.

Evaluation of machine translation (MT) is a difficult task, both for humans, and using automatic metrics. The main difficulty lies in the fact that there is not one single correct translation, but many alternative good translation options.MT systems are often evaluated using automatic metrics, which commonly rely on comparing a translation to only a single human reference translation. An alternative is different types of human evaluations, commonly ranking be-tween systems or estimations of adequacy and fluency on some scale, or error analyses.

We have explored four different evaluation methods on output from three different statistical MT systems. The main focus is on different types of human evaluation. We compare two conventional evaluation methods, human error analysis and automatic metrics, to two lesser used evaluation methods based on reading comprehension and eye-tracking. These two methods of evaluations are performed without the subjects seeing the source sentence. There have been few previous attempts of using reading comprehension and eye-tracking for MT evaluation.

One example of a reading comprehension study is Fuji (1999) who conducted an experiment to compare English-to-Japanese MT to several versions of manual corrections of the system output. He found significant differences be-tween texts with large differences on reading comprehension questions. Doherty and O’Brien (2009) is the only study we are aware of using eye-tracking for MT evaluation. They found that the average gaze time and fixation counts were significantly lower for sentences judged as excellent in an earlier evaluation, than for bad sentences.

Like previous research we find that both reading comprehension and eye-tracking can be useful for MT evaluation.

The results of these methods are consistent with the other methods for comparison between systems with a big quality difference. For systems with similar quality, however, the different evaluation methods often does not show any significant differences.

The Spoken Language Translator (SLT) is a multi-lingual speech-to-speech translation prototype supporting English, Swedish and French within the air traffic information system (ATIS) domain. The design of SLT is characterized by a strongly corpus-driven approach, which accentuates the need for cost-efficient collection procedures to obtain training data. This paper discusses various approaches to the data collection issue pursued within a speech translation framework. Original American English speech and language data have been collected using traditional Wizard-of-Oz (WOZ) techniques, a relatively costly procedure yielding high-quality results. The resulting corpus has been translated textually into Swedish by a large number of native speakers (427) and used as prompts for training the target language speech model. This ￜbudgetￜ collection method is compared to the accepted method, i.e., gathering data by means of a full-blown WOZ simulation. The results indicate that although translation in this case proved economical and produced considerable data, the method is not sensitive to certain features typical of spoken language, for which WOZ is superior

Sharing data is an important part of the progress of science in many fields. In the largely deep learning dominated field of natural language processing, textual resources are in high demand. In certain domains, such as that of medical records, the sharing of data is limited by ethical and legal restrictions and therefore requires anonymization. The process of manual anonymization is tedious and expensive, thus automated anonymization is of great value. Since medical records consist of unstructured text, pieces of sensitive information have to be identified in order to be masked for anonymization. Named-entity recognition (NER) is the subtask of information extraction named entities, such as person names or locations, are identified and categorized. Recently, models that leverage unsupervised training on large quantities of unlabeled training data have performed impressively on the NER task, which shows promise in their usage for the problem of anonymization. In this study, a small set of medical records was annotated with named-entity tags. Because of the lack of any training data, a BERT model already fine-tuned for NER was then evaluated on the evaluation set. The aim was to find out how well the model would perform on NER on medical records, and to explore the possibility of using the model to anonymize medical records. The most positive result was that the model was able to identify all person names in the dataset. The average accuracy for identifying all entity types was however relatively low. It is discussed that the success of identifying person names shows promise in the model’s application for anonymization. However, because the overall accuracy is significantly worse than that of models fine-tuned on domain-specific data, it is suggested that there might be better methods for anonymization in the absence of relevant training data.

Many approaches to solving tasks in the field of Natural Language Processing (NLP) use syntactic dependency trees (SDTs) as a feature to represent the latent nonlinear structure within sentences. Recently, work in parsing sentences to graph-based structures which encode semantic relationships between words—called semantic dependency graphs (SDGs)—has gained interest. This thesis seeks to explore the use of SDGs in place of and alongside SDTs within a relation classification system based on long short-term memory (LSTM) neural networks. Two methods for handling the information in these graphs are presented and compared between two SDG formalisms. Three new relation extraction system architectures have been created based on these methods and are compared to a recent state-of-the-art LSTM-based system, showing comparable results when semantic dependencies are used to enhance syntactic dependencies, but with significantly fewer training parameters.

This thesis describes the work of creating an automatic system for identifying synonyms and semantically related words in medical texts. Before this work, as a part of the project E-care@home, medical texts have been classified as either lay or specialized by both a lay annotator and an expert annotator. The lay annotator, in this case, is a person without any medical knowledge, whereas the expert annotator has professional knowledge in medicine. Using these texts made it possible to create co-occurrences matrices from which the related words could be identified. Fifteen medical terms were chosen as system input. The Dice similarity of these words in a context window of ten words around them was calculated. As output, five candidate related terms for each medical term was returned. Only unigrams were considered. The candidate related terms were evaluated using a questionnaire, where 223 healthcare professionals rated the similarity using a scale from one to five. A Fleiss kappa test showed that the agreement among these raters was 0.28, which is a fair agreement. The evaluation further showed that there was a significant correlation between the human ratings and the relatedness score (Dice similarity). That is, words with higher Dice similarity tended to get a higher human rating. However, the Dice similarity interval in which the words got the highest average human rating was 0.35-0.39. This result means that there is much room for improving the system. Further developments of the system should remove the unigram limitation and expand the corpus the provide a more accurate and reliable result.

Text data is one of the most abundant types of data available, produced every day across all domains of society. Understanding the contents of this data can support important policy decisions, help us understand society and culture, and improve business processes. While machine learning techniques are growing in their power for analyzing text data, there is still a clear role for human analysis and decision-making. This seminar explored the use of visual analytics applied to text data as a means to bridge the complementary strengths of people and computers. The field of visual text analytics applies visualization and interaction approaches which are tightly coupled to natural language processing systems to create analysis processes and systems for examining text and multimedia data. During the seminar, interdisciplinary working groups of experts from visualization, natural language processing, and machine learning examined seven topic areas to reflect on the state of the field, identify gaps in knowledge, and create an agenda for future cross-disciplinary research. This report documents the program and the outcomes of Dagstuhl Seminar 22191 "Visual Text Analytics".

Reflecting on three papers included in this issue, we suggest that research on memory and conversation could benefit by making more use of analyzing real-life situations or close to real-life scenarios, full speech and body interactions, and the interaction with the physical environment. We also suggest that the process of remembering during conversation is investigated on a level of detail and sequence that allow for locating actual functions of different actions. Finally, we suggest that a life-span perspective on transactive memory systems must also model the development, maintenance, breakdown, and reestablishment of such systems.

When it comes to the task of classification the data used for training is the most crucial part. It follows that how this data is processed and presented for the classifier plays an equally important role. This thesis attempts to investigate the performance of multiple classifiers depending on the features that are used, the type of classes to classify and the optimization of said classifiers. The classifiers of interest are support-vector machines (SMO) and multilayer perceptron (MLP), the features tested are word vector spaces and text complexity measures, along with principal component analysis on the complexity measures. The features are created based on the Stockholm-Umeå-Corpus (SUC) and DigInclude, a dataset containing standard and easy-to-read sentences. For the SUC dataset the classifiers attempted to classify texts into nine different text categories, while for the DigInclude dataset the sentences were classified into either standard or simplified classes. The classification tasks on the DigInclude dataset showed poor performance in all trials. The SUC dataset showed best performance when using SMO in combination with word vector spaces. Comparing the SMO classifier on the text complexity measures when using or not using PCA showed that the performance was largely unchanged between the two, although not using PCA had slightly better performance

In this thesis, the application of transformer-based models on the real-world task of identifying patients as implant bearing is investigated. The task is approached as a classification task and five transformer-based models relying on the BERT architecture are implemented, along with a Support Vector Machine (SVM) as a baseline for comparison. The models are fine-tuned with Swedish medical texts, i.e. patients’ medical histories.

The five transformer-based models in question makes use of two pre-trained BERT models, one released by the National Library of Sweden and a second one using the same pre-trained model but which has also been further pre-trained on domain specific language. These are in turn fine-tuned using five different types of architectures. These are: (1) a typical BERT model, (2) GAN-BERT, (3) RoBERT, (4) chunkBERT, (5) a frequency based optimized BERT. The final classifier, an SVM baseline, is trained using TF-IDF as the feature space.

The data used in the thesis comes from a subset of an unreleased corpus from four Swedish clinics that cover a span of five years. The subset contains electronic medical records of patients belonging to the radiology, and cardiology clinics. Four training sets were created, respectively containing 100, 200, 300, and 903 labelled records. The test set, containing 300 labelled samples, was also created from said subset. The labels upon which the models are trained are created by labelling the patients as implant bearing based on the amount of implant terms each patient history contain.

The results are promising, and show favourable performance when classifying the patient histories. Models trained on 903 and 300 samples are able to outperform the baseline, and at their peak, BERT, chunkBERT and the frequency based optimized BERT achieves an F1-measure of 0.97. When trained using 100 and 200 labelled records all of the transformerbased models are outperformed by the baseline, except for the semi-supervised GAN-BERT which is able to achieve competitive scores with 200 records.

There is not a clear delineation between using the pre-trained BERT or the BERT model that has additional pre-training on domain specific language. However, it is believed that further research could shed additional light on the subject since the results are inconclusive.

Syntactic representations based on word-to-word dependencies have a long tradition in descriptive linguistics [29]. In recent years, they have also become increasingly used in computational tasks, such as information extraction [5], machine translation [43], and parsing [42]. Among the purported advantages of dependency over phrase structure representations are conciseness, intuitive appeal, and closeness to semantic representations such as predicate-argument structures. On the more practical side, dependency representations are attractive due to the increasing availability of large corpora of dependency analyses, such as the Prague Dependency Treebank [19].

Extensible Dependency Grammar (XDG; Duchier and Debusmann (2001)) is a recently developed dependency grammar formalism that allows the characterization of linguistic structures along multiple dimensions of description. It can be implemented efficiently using constraint programming (CP; Koller and Niehren 2002). In the CP context, parsing is cast as a search problem: The states of the search are partial parse trees, successful end states are complete and valid parses. In this paper, we propose a probability model for XDG dependency trees and an A-Star search control regime for the XDG parsing algorithm that guarantees the best parse to be found first. Extending XDG with a statistical component has the benefit of bringing the formalism further into the grammatical mainstream; it also enables XDG to efficiently deal with large, corpus-induced grammars that come with a high degree of ambiguity.

Transformer-based models trained on text and vision modalities try to improve the performance on multimodal downstream tasks or tackle the problem Transformer-based models trained on text and vision modalities try to improve the performance on multimodal downstream tasks or tackle the problem of lack of grounding, e.g., addressing issues like models’ insufficient commonsense knowledge. While it is more straightforward to evaluate the effects of such models on multimodal tasks, such as visual question answering or image captioning, it is not as well-understood how these tasks affect the model itself, and its internal linguistic representations. In this work, we experiment with language models grounded in videos and measure the models’ performance on predicting masked words chosen based on their imageability. The results show that the smaller model benefits from video grounding in predicting highly imageable words, while the results for the larger model seem harder to interpret.of lack of grounding, e.g., addressing issues like models’ insufficient commonsense knowledge. While it is more straightforward to evaluate the effects of such models on multimodal tasks, such as visual question answering or image captioning, it is not as well-understood how these tasks affect the model itself, and its internal linguistic representations. In this work, we experiment with language models grounded in videos and measure the models’ performance on predicting masked words chosen based on their imageability. The results show that the smaller model benefits from video grounding in predicting highly imageable words, while the results for the larger model seem harder to interpret.

In natural language processing (NLP) there is an increasing interest in formal models for processing graphs rather than more restricted structures such as strings or trees. Such models of graph transformation have previously been studied and applied in various other areas of computer science, including formal language theory, term rewriting, theory and implementation of programming languages, concurrent processes, and software engineering. However, few researchers from NLP are familiar with this work, and at the same time, few researchers from the theory of graph transformation are aware of the specific desiderata, possibilities and challenges that one faces when applying the theory of graph transformation to NLP problems. The Dagstuhl Seminar 15122 “Formal Models of Graph Transformation in Natural Language Processing” brought researchers from the two areas together. It initiated an interdisciplinary exchange about existing work, open problems, and interesting applications.

Quality in translations depends on the correct use of specialized terms, which can make the translation easier to understand as well as reduce the required time and costs for the translation (Lommel, 2007). Consistent use of terminology is important, and should be taken into account during quality checks of for example translated documentation (Esselink, 2000). Today, several commercial programs have functions for automatic quality checking of terminology. The aim of this study is to evaluate such functions since no earlier major study of this has been found.

To get some insight into quality checking in practice, two qualitative interviews were initially carried out with individuals involved in this at a translation agency. The results were compared to current theories in the subject field and revealed a general agreement with for example the recommendations of Bass (2006).

The evaluations started with an examination of the recall for a genuine terminology database compared to subjectively marked terms in a test corpus based on an authentic translation memory. The examination however revealed a relatively low recall. To increase the recall the terminology database was modified, it was for example extended with longer terms from the test corpus.

After that, the function for checking terminology in four different commercial programs was run on the test corpus using the modified terminology database. Finally, the test corpus was also modified, by planting out a number of errors to produce a more idealized evaluation. The results from the programs, in the form of alarms for potential errors, were categorized and judged as true or false alarms. This constitutes a base for measures of precision of the checks, and in the last evaluation also of their recall.

The evaluations showed that for terminology in translations of English to Swedish, it was advantageous to match terms from the terminology database using partial matching of words in the source and target segments of the translation. In that way, terms with different inflected forms could be matched without support for language﷓specific morphology. A cause of many problems in the matching process was the form of the entries in the terminology database, which were more suited for being read by human translators than by a machine.

Recommendations regarding the introduction of tools for automatic checking of terminology were formulated, based on the results from the interviews and evaluations. Due to factors of uncertainty in the automatic checking, a manual review of its results is motivated. By running the check on a sample that has already been manually checked in other aspects, a reasonable number of results to manually review can be obtained. The quality of the terminology database is crucial for its recall on translations, and in the long run also for the value of using it for automatic checking.