The goal of a press release is to have the information spread as widely as possible. A suitable approach to distribute the information is to target journalists who are likely to distribute the information further. Deciding which journalists to target has traditionally been performed manually without intelligent digital assistance and therefore has been a time consuming task. Machine learning can be used to assist the user by predicting a ranking of journalists based on their most semantically similar written article to the press release. The purpose of this thesis was to compare different methods in their ability to compare semantic similarity between articles and press releases when used for the task of ranking journalists. Three methods were chosen for comparison: (1.) TF-IDF together with cosine similarity, (2.) TF-IDF together with soft-cosine similarity and (3.) sentence mover’s distance (SMD) together with SBERT. Based on the proposed heuristic success metric, both TF-IDF methods outperformed the SMD method. The best performing method was TF-IDF with soft-cosine similarity. / Målet med ett pressmeddelande är att få informationen att spriddas till så många som möjligt. Ett lämpligt tillvägagångssätt för att sprida informationen är att rikta in sig på journalister som sannolikt kommer att sprida informationen vidare. Beslutet om vilka journalister man ska rikta sig till har traditionellt utförts manuellt utan intelligent digital assistans och har därför varit en tidskrävande uppgift. Maskininlärning kan användas för att hjälpa användaren genom att förutsäga en rankning av journalister baserat på deras mest semantiskt liknande skrivna artikel till pressmeddelandet. Syftet med denna uppsats var att jämföra olika metoder i deras förmåga att jämföra semantisk likhet mellan artiklar och pressmeddelanden när de används för att rangordna journalister. Tre metoder valdes för jämförelse: (1.) TF-IDF tillsammans med cosinus likhet, (2.) TF-IDF tillsammans med mjuk-cosinus likhet och (3.) sentence mover’s distance (SMD) tillsammans med SBERT. Baserat på det föreslagna heuristiska framgångsmåttet överträffade båda TF-IDF-metoderna SMD-metoden. Den bäst presterande metoden var TF-IDF med mjuk-cosinus likhet.

The ability to automate the process of comparing and matching resumes with job adverts is a growing research field. This can be done through the use of the machine learning area Natural Language Processing (NLP), which enables a model to learn human language. This thesis explores and evaluates the application of the state-of-the-art NLP model, SBERT, on the task of comparing and calculating a measure of similarity between extracted text from resumes and adverts. This thesis also investigates what type of data that generates the best performing model on said task. The results show that SBERT quickly can be trained on unlabeled data from the HR domain with the usage of a Triplet network, and achieves high performance and good results when tested on various tasks. The models are shown to be bilingual, can tackle unseen vocabulary and understand the concept and descriptive context of entire sentences instead of solely single words. Thus, the conclusion is that the models have a neat understanding of semantic similarity and relatedness. However, in some cases the models are also shown to become binary in their calculations of similarity between inputs. Moreover, it is hard to tune a model that is exhaustively comprehensive of such diverse domain such as HR. A model fine-tuned on clean and generic data extracted from adverts shows the overall best performance in terms of loss and consistency.

This thesis investigates Natural Language Processing (NLP) techniques to extract relevant information from long and unstructured insurance policy documents. The goal is to reduce the amount of time required by readers to understand the coverage within the documents. The study uses predefined insurance policy coverage parameters, created by industry experts to represent what is covered in the policy documents. Three NLP approaches are used to classify the text sequences as insurance parameter classes. The thesis shows that using SBERT to create vector representations of text to allow cosine similarity calculations is an effective approach. The top scoring sequences for each parameter are assigned that parameter class. This approach shows a significant reduction in the number of sequences required to read by a user but misclassifies some positive examples. To improve the model, the parameter definitions and training data were combined into a support set. Similarity scores were calculated between all sequences and the support sets for each parameter using different pooling strategies. This few-shot classification approach performed well for the use case, improving the model’s performance significantly. In conclusion, this thesis demonstrates that NLP techniques can be applied to help understand unstructured insurance policy documents. The model developed in this study can be used to extract important information and reduce the time needed to understand the contents of aninsurance policy document. A human expert would however still be required to interpret the extracted text. The balance between the amount of relevant information and the amount of text shown would depend on how many of the top-scoring sequences are classified for each parameter. This study also identifies some limitations of the approach depending on available data. Overall, this research provides insight into the potential implications of NLP techniques for information extraction and the insurance industry.

The domain of rhetoric has changed dramatically since its inception as the art of persuasion. It has adapted to encompass many forms of digital media, including, for example, data visualization and coding as a form of literature, but the approach has frequently been that of an outsider looking in. The use of comprehensive computational tools as a part of rhetorical analysis has largely been lacking. In this report, we attempt to address this lack by means of three case studies in natural language processing tasks, all of which can be used as part of a computational approach to rhetoric. At this same moment in time, it is becoming all the more important to transition to renewable energy in order to keep global warming under 1.5 degrees Celsius and ensure that countries meet the conditions of the Paris Agreement. Thus, we make use of speech data on climate policy from the Swedish parliament to ground these three analyses in semantic textual similarity, topic modeling, and political party attribution. We find that speeches are, to a certain extent, consistent within parties, given that a slight majority of most semantically similar speeches come from the same party. We also find that some of the most common topics discussed in these speeches are nuclear energy and the Swedish Green party, purported environmental risks due to renewable energy sources, and the job market. Finally, we find that though pairs of speeches are semantically similar, party rhetoric on the whole is generally not unique enough for speeches to be distinguishable by party. These results then open the door for a broader exploration of computational rhetoric for Swedish political science in the future.

Le regroupement est une tâche non supervisée consistant à rassembler les éléments semblables sous un même groupe et les éléments différents dans des groupes distincts. Le regroupement de textes est effectué en représentant les textes dans un espace vectoriel et en étudiant leur similarité dans cet espace. Les meilleurs résultats sont obtenus à l’aide de modèles neuronaux qui affinent une représentation vectorielle contextuelle de manière non supervisée. Or, cette technique peuvent nécessiter un temps d’entraînement important et sa performance n’est pas comparée à des techniques plus simples ne nécessitant pas l’entraînement de modèles neuronaux. Nous proposons, dans ce mémoire, une étude de l’état actuel du domaine. Tout d’abord, nous étudions les meilleures métriques d’évaluation pour le regroupement de textes. Puis, nous évaluons l’état de l’art et portons un regard critique sur leur protocole d’entraînement. Nous proposons également une analyse de certains choix d’implémentation en regroupement de textes, tels que le choix de l’algorithme de regroupement, de la mesure de similarité, de la représentation vectorielle ou de l’affinage non supervisé de la représentation vectorielle. Finalement, nous testons la combinaison de certaines techniques ne nécessitant pas d’entraînement avec la représentation vectorielle contextuelle telles que le prétraitement des données, la réduction de dimensionnalité ou l’inclusion de Tf-idf. Nos expériences démontrent certaines lacunes dans l’état de l’art quant aux choix des métriques d’évaluation et au protocole d’entraînement. De plus, nous démontrons que l’utilisation de techniques simples permet d’obtenir des résultats meilleurs ou semblables à des méthodes sophistiquées nécessitant l’entraînement de modèles neuronaux. Nos expériences sont évaluées sur huit corpus issus de différents domaines. / Clustering is an unsupervised task of bringing similar elements in the same cluster and different elements in distinct groups. Text clustering is performed by representing texts in a vector space and studying their similarity in this space. The best results are obtained using neural models that fine-tune contextual embeddings in an unsupervised manner. However, these techniques require a significant amount of training time and their performance is not compared to simpler techniques that do not require training of neural models. In this master’s thesis, we propose a study of the current state of the art. First, we study the best evaluation metrics for text clustering. Then, we evaluate the state of the art and take a critical look at their training protocol. We also propose an analysis of some implementation choices in text clustering, such as the choice of clustering algorithm, similarity measure, contextual embeddings or unsupervised fine-tuning of the contextual embeddings. Finally, we test the combination of contextual embeddings with some techniques that don’t require training such as data preprocessing, dimensionality reduction or Tf-idf inclusion. Our experiments demonstrate some shortcomings in the state of the art regarding the choice of evaluation metrics and the training protocol. Furthermore, we demonstrate that the use of simple techniques yields better or similar results to sophisticated methods requiring the training of neural models. Our experiments are evaluated on eight benchmark datasets from different domains.