In this thesis, we investigate the problem of building a text classifier for transcribed voice recordings submitted by insurance clients. We compare different models in the context of two tasks. The first is a binary classification problem, where the models are tasked with determining if a transcript belongs to a particular type or not. The second is a multiclass problem, where the models have to choose between several types when labelling transcripts, resulting in a data set with a highly imbalanced class distribution. We evaluate four different models: pretrained BERT and three LSTMs with different word embeddings. The used word embeddings are ELMo, word2vec and a baseline model with randomly initialized embedding layer. In the binary task, we are more concerned with false positives than false negatives. Thus, we also use weighted cross entropy loss to achieve high precision for the positive class, while sacrificing recall. In the multiclass task, we use focal loss and weighted cross entropy loss to reduce bias toward majority classes. We find that BERT outperforms the other models and the baseline model is worst across both tasks. The difference in performance is greatest in the multiclass task on classes with fewer samples. This demonstrates the benefit of using large language models in data constrained scenarios. In the binary task, we find that weighted cross entropy loss provides a simple, yet effective, framework for conditioning the model to favor certain types of errors. In the multiclass task, both focal loss and weighted cross entropy loss are shown to reduce bias toward majority classes. However, we also find that BERT fine tuned with regular cross entropy loss does not show bias toward majority classes, having high recall across all classes. / I examensarbetet undersöks klassificering av transkriberade röstinspelningar från försäkringskunder. Flera modeller jämförs på två uppgifter. Den första är binär klassificering, där modellerna ska särskilja på inspelningar som tillhör en specifik klass av ärende från resterande inspelningar. I det andra inkluderas flera olika klasser som modellerna ska välja mellan när inspelningar klassificeras, vilket leder till en ojämn klassfördelning. Fyra modeller jämförs: förtränad BERT och tre LSTM-nätverk med olika varianter av förtränade inbäddningar. De inbäddningar som används är ELMo, word2vec och en basmodell som har inbäddningar som inte förtränats. I det binära klassificeringsproblemet ligger fokus på att minimera antalet falskt positiva klassificeringar, därför används viktad korsentropi. Utöver detta används även fokal förlustfunktion när flera klasser inkluderas, för att minska partiskhet mot majoritetsklasser. Resultaten indikerar att BERT är en starkare modell än de andra modellerna i båda uppgifterna. Skillnaden mellan modellerna är tydligast när flera klasser används, speciellt på de klasser som är underrepresenterade. Detta visar på fördelen av att använda stora, förtränade, modeller när mängden data är begränsad. I det binära klassificeringsproblemet ser vi även att en viktad förlustfunktion ger ett enkelt men effektivt sätt att reglera vilken typ av fel modellen ska vara partisk mot. När flera klasser inkluderas ser vi att viktad korsentropi, samt fokal förlustfunktion, kan bidra till att minska partiskhet mot överrepresenterade klasser. Detta var dock inte fallet för BERT, som visade bra resultat på minoritetsklasser även utan att modifiera förlustfunktionen.

The rise of social media and the use of mobile applications has led to increasing concerns regarding the content that is shared through these apps and whether they are being regulated or not. One of the problems that can arise due to a lack of regulation is that chat messages that are inappropriate or of profane nature can be allowed to be shared through these apps. Thus, it is vital to detect whenever these types of chat messages are shared through these mobile applications. In addition to that, there should also be detection of chat messages that can lead to the identity of the users being revealed as that is how the app in this thesis project was intended to be used. One of the most popular approaches to detect chat messages of this nature is to use machine learning techniques that can classify text. We were quick to discover that there were not many machine learning models that were built to classify short text messages in the Swedish language, thus the main problem of our thesis was the lack of evaluation and analysis of machine learning models for text classification in the context of the chat messages in Swedish. Thus, the purpose of our project was mainly to find the best performing models for text classification, implement these models and evaluate them to find the best among the ones we found. After the models were created, a hosting server, as well as an API, was required for the text classifying system to compute and communicate the prediction results to the mobile application in real-time. Therefore, the models were containerized and deployed as a REST API that serves requests upon arrival on a cloud server. The goal of this project was to help future work being done on text classification in the Swedish language by providing the results of this thesis to any parties that are interested in our line of work. From our own experience, we realized how challenging it can be to find and choose the best machine learning models when one has no previous data on which can be the best performing one. Thus, we believe that the results of this thesis project will greatly aid future projects in this area. The chosen research methodology was qualitative and dealt with quantitative data. The results we received showed that the BERT model was the best choice among the three models that we compared. With minute adjustments, this model should be more than capable of detecting the type of chat messages that it is required within the mobile application. / Uppkomsten av social media och användning av mobilapplikationer ledde till ökande oro om innehållet som är delad inom dessa appar och om dem är reglerad eller inte. Ett problem som uppstår på grund av bristande reglering kan vara att chatmeddelanden som är olämplig eller profan kan bli delad med dessa appar. Därför är det viktig att upptäcka när dessa typer av chatmeddelande är delad genom mobilapplikationer. Dessutom det måste finnas ett system som upptäcker chattmeddelanden som kan hjälpa att avslöja användarens identiteter, som den här appen i detta projekt avsedda att användas. En av mest populära sett att upptäcka den typen av chattmeddelanden är användning av mäskinlärning tekniker som kan klassificera text. Vi snart hittade att det finns inte så många mäskinlärning modeller som var byggt att klassificera texter på svenska, alltså huvudproblem med vår exam en var bistrande utvärdering och analys av mäskinlärning modeller för textklassificering i kontext av svenska språket. Så, syftet med vårt projekt var att hitta de bästa presenterande modeller för textklassifikation, genomföra dessa modeller själva och sedan utvärdera dem att hitta den bästa. Därtill, för att textklassificering ska beräkna och kommunicera den förutsägelseresultaten till mobila applikationer i realtid behövs en värdserver samt en API. Därför, modellerna containeriserades och distribuerad es som en REST API som betjänar begäran vid ankomst på en molnserver. Målet med det här projektet var att hjälpa framtidsarbete inom textklassifikation på svenska språket genom att tillhandahålla resultaten till partier som är intresserad i vår arbetslin je. Från vår egen erfarenhet, vi insåg att det var svårt att hitta och välja dem bästa mäskinlärning modeller, specifikt när man har inga data som tidigare visat den med bäst prestanda. Och därför vi anser att den resultaten av den har examen kommer att v ara stor hjälp till framtida projekt i det här området. Den valda forskningsmetodiken var kvalitativ och handlade om kvantitativ data. Resultaten visade att BERT modell var den bästa bland de tre modellerna som vi jämförde med. Med lite justeringen är mod ellen mer än kapable att detektera den typen av krävs inom mobilapplikationen.

The emergence of large language models, such as BERT and GPT-3, has revolutionized natural language processing tasks. However, the development and deployment of these models pose challenges, including concerns about computational resources and environmental impact. This study aims to compare discriminative language models for text classification based on their performance and usage cost. We evaluate the models using a hierarchical multi-label text classification task and assess their performance using primarly F1-score. Additionally, we analyze the usage cost by calculating the Floating Point Operations (FLOPs) required for inference. We compare a baseline model, which consists of a classifier chain with logistic regression models, with fine-tuned discriminative language models, including BERT with two different sequence lengths and DistilBERT, a distilled version of BERT. Results show that the DistilBERT model performs optimally in terms of performance, achieving an F1-score of 0.56 averaged on all classification layers. The baseline model and BERT with a maximal sequence length of 128 achieve F1-scores of 0.51. However, the baseline model outperforms the transformers at the most specific classification level with an F1-score of 0.33. Regarding usage cost, the baseline model significantly requires fewer FLOPs compared to the transformers. Furthermore, restricting BERT to a maximum sequence length of 128 tokens instead of 512 sacrifices some performance but offers substantial gains in usage cost. The code and dataset are available on GitHub. / Uppkomsten av stora språkmodeller, som BERT och GPT-3, har revolutionerat språkteknologi. Dock ger utvecklingen och implementeringen av dessa modeller upphov till utmaningar, bland annat gällande beräkningsresurser och miljöpåverkan. Denna studie syftar till att jämföra diskriminativa språkmodeller för textklassificering baserat på deras prestanda och användningskostnad. Vi utvärderar modellerna genom att använda en hierarkisk textklassificeringsuppgift och bedöma deras prestanda primärt genom F1-score. Dessutom analyserar vi användningskostnaden genom att beräkna antalet flyttalsoperationer (FLOPs) som krävs för inferens. Vi jämför en grundläggande modell, som består av en klassifikationskedja med logistisk regression, med finjusterande diskriminativa språkmodeller, inklusive BERT med två olika sekvenslängder och DistilBERT, en destillerad version av BERT. Resultaten visar att DistilBERT-modellen presterar optimalt i fråga om prestanda och uppnår en genomsnittlig F1-score på 0,56 för alla klassificeringsnivåer. Den grundläggande modellen och BERT med en maximal sekvenslängd på 128 uppnår ett F1-score på 0,51. Dock överträffar den grundläggande modellen transformermodellerna på den mest specifika klassificeringsnivån med en F1-score på 0,33. När det gäller användningskostnaden kräver den grundläggande modellen betydligt färre FLOPs jämfört med transformermodellerna. Att begränsa BERT till en maximal sekvenslängd av 128 tokens ger vissa prestandaförluster men erbjuder betydande besparingar i användningskostnaden. Koden och datamängden är tillgängliga på GitHub.

Le danmaku désigne les commentaires synchronisés qui s’affichent et défilent directement en surimpression sur des vidéos au fil du visionnement. Bien que les danmakus proposent à l’audience une manière originale de partager leur sentiments, connaissances, compréhensions et prédictions sur l’histoire d’une série, etc., et d’interagir entre eux, la façon dont les commentaires s’affichent peut nuire à l’expérience de visionnement, lorsqu’une densité excessive de commentaires dissimule complètement les images de la vidéo ou distrait l’audience. Actuellement, les sites de vidéo chinois emploient principalement des méthodes par mots-clés s’appuyant sur des expressions régulières pour éliminer les commentaires non désirés. Ces approches risquent fortement de surgénéraliser en supprimant involontairement des commentaires intéressants contenant certains mots-clés ou, au contraire, de sous-généraliser en étant incapables de détecter ces mots lorsqu’ils sont camouflés sous forme d’homophones. Par ailleurs, les recherches existantes sur la classification automatique du danmaku se consacrent principalement à la reconnaissance de la polarité des sentiments exprimés dans les commentaires. Ainsi, nous avons cherché à regrouper les commentaires par classes fonctionnelles, à évaluer la robustesse d’une telle classification et la possibilité de l’automatiser dans la perspective de développer de meilleurs systèmes de filtrage des commentaires. Nous avons proposé une nouvelle taxonomie pour catégoriser les commentaires en nous appuyant sur la théorie des actes de parole et la théorie des gratifications dans l’usage des médias, que nous avons utilisées pour produire un corpus annoté. Un fragment de ce corpus a été co-annoté pour estimer un accord inter-annotateur sur la classification manuelle. Enfin, nous avons réalisé plusieurs expériences de classification automatique. Celles-ci comportent trois étapes : 1) des expériences de classification binaire où l’on examine si la machine est capable de faire la distinction entre la classe majoritaire et les classes minoritaires, 2) des expériences de classification multiclasses à granularité grosse cherchant à classifier les commentaires selon les catégories principales de notre taxonomie, et 3) des expériences de classification à granularité fine sur certaines sous-catégories. Nous avons expérimenté avec des méthodes d’apprentissage automatique supervisé et semi-supervisé avec différents traits. / Danmaku denotes synchronized comments which are displayed and scroll directly on top of videos as they unfold. Although danmaku offers an innovative way to share their sentiments, knowledge, predictions on the plot of a series, etc., as well as to interact with each other, the way comments display can have a negative impact on the watching experience, when the number of comments displayed in a given timespan is so high that they completely hide the pictures, or distract audience. Currently, Chinese video websites mainly ressort to keyword approaches based on regular expressions to filter undesired comments. These approaches are at high risk to overgeneralize, thus deleting interesting comments coincidentally containing some keywords, or, to the contrary, undergeneralize due to their incapacity to detect occurrences of these keywords disguised as homophones. On another note, existing research focus essentially on recognizing the polarity of sentiments expressed within comments. Hence, we have sought to regroup comments into functional classes, evaluate the robustness of such a classification and the feasibility of its automation, under an objective of developping better comments filtering systems. Building on the theory of speech acts and the theory of gratification in media usage, we have proposed a new taxonomy of danmaku comments, and applied it to produce an annotated corpus. A fragment of the corpus has been co-annotated to estimate an interannotator agreement for human classification. Finally, we performed several automatic classification experiments. These involved three steps: 1) binary classification experiments evaluating whether the machine can distinguish the most frequent class from all others, 2) coarse-grained multi-class classification experiments aiming at classifying comments within the main categories of our taxonomy, and 3) fine-grained multi-class classification experiments on specific subcategories. We experimented both with supervised and semi-supervised learning algorithms with diffrent features.

This study aims to construct a state-of-the-art classifier model and compare it against a largelanguage model. A variation of SVM called LinearSVC was utilised and the BERT model usingbert-base-uncased was used. The data, provided by the Swedish Research Council, consisted ofresearch grant applications. The research grant applications were divided into two groups, whichwere further divided into several subgroups. The subgroups represented research fields such ascomputer science and applied physics. Significant class imbalances were present, with someclasses having only a tenth of the applications of the largest class. To address these imbalances,a new dataset was created using data that had been randomly oversampled. The models weretrained and tested on their ability to correctly assign a subgroup to a research grant application.Results indicate that the BERT model outperformed the SVM model on the original dataset,but not on the balanced dataset . Furthermore, the BERT model’s performance decreased whentransitioning from the original to the balanced dataset, due to overfitting or randomness. / Denna studie har som mål att bygga en state-of-the-art klassificerar model och sedan jämföraden mot en stor språkmodel. SVM modellen var en variation av SVM vid namn LinearSVC ochför BERT användes bert-base-uncased. Data erhölls från Vetenskapsrådet och bestod av forskn-ingsbidragsansökningar. Forskningsbidragsansökningarna var uppdelade i två grupper, som varytterligare uppdelade i ett flertal undergrupper. Dessa undergrupper representerar forsknings-fält såsom datavetenskap och tillämpad fysik. I den data som användes i studien fanns storaskillnader mellan klasserna, där somliga klasser hade en tiondel av ansökningarna som de storaklasserna hade. I syfte att lösa dessa klassbalanseringsproblem skapades en datamängd somundergått slumpmässig översampling. Modellerna tränades och testades på deras förmåga attkorrekt klassificera en forskningsbidragsansökan in i rätt undergrupp. Studiens fynd visade attBERT modellen presterade bättre än SVM modellen på både den ursprungliga datamängden,dock inte på den balanserade datamängden. Tilläggas kan, BERTs prestanda sjönk vid övergångfrån den ursprungliga datamängden till den balanserade datamängden, något som antingen berorpå överanpassning eller slump.

This thesis investigates the most effective way of performing classification of text labels and clustering of duplicate texts in technical documentation written in Simplified Technical English. Pre-trained language models from transformers (BERT) were tested against traditional methods such as tf-idf with cosine similarity (kNN) and SVMs on the classification task. For detecting duplicate texts, vector representations from pre-trained transformer and LSTM models were tested against tf-idf using the density-based clustering algorithms DBSCAN and HDBSCAN. The results show that traditional methods are comparable to pre-trained models for classification, and that using tf-idf vectors with a low distance threshold in DBSCAN is preferable for duplicate detection.

Algoritmos de aprendizado semissupervisionado aprendem a partir de uma combinação de dados rotulados e não rotulados. Assim, eles podem ser aplicados em domínios em que poucos exemplos rotulados e uma vasta quantidade de exemplos não rotulados estão disponíveis. Além disso, os algoritmos semissupervisionados podem atingir um desempenho superior aos algoritmos supervisionados treinados nos mesmos poucos exemplos rotulados. Uma poderosa abordagem ao aprendizado semissupervisionado, denominada aprendizado multidescrição, pode ser usada sempre que os exemplos de treinamento são descritos por dois ou mais conjuntos de atributos disjuntos. A classificação de textos é um domínio de aplicação no qual algoritmos semissupervisionados vêm obtendo sucesso. No entanto, o aprendizado semissupervisionado multidescrição ainda não foi bem explorado nesse domínio dadas as diversas maneiras possíveis de se descrever bases de textos. O objetivo neste trabalho é analisar o desempenho de algoritmos semissupervisionados multidescrição na classificação de textos, usando unigramas e bigramas para compor duas descrições distintas de documentos textuais. Assim, é considerado inicialmente o difundido algoritmo multidescrição CO-TRAINING, para o qual são propostas modificações a fim de se tratar o problema dos pontos de contenção. É também proposto o algoritmo COAL, o qual pode melhorar ainda mais o algoritmo CO-TRAINING pela incorporação de aprendizado ativo como uma maneira de tratar pontos de contenção. Uma ampla avaliação experimental desses algoritmos foi conduzida em bases de textos reais. Os resultados mostram que o algoritmo COAL, usando unigramas como uma descrição das bases textuais e bigramas como uma outra descrição, atinge um desempenho significativamente melhor que um algoritmo semissupervisionado monodescrição. Levando em consideração os bons resultados obtidos por COAL, conclui-se que o uso de unigramas e bigramas como duas descrições distintas de bases de textos pode ser bastante compensador / Semi-supervised learning algorithms learn from a combination of both labeled and unlabeled data. Thus, they can be applied in domains where few labeled examples and a vast amount of unlabeled examples are available. Furthermore, semi-supervised learning algorithms may achieve a better performance than supervised learning algorithms trained on the same few labeled examples. A powerful approach to semi-supervised learning, called multi-view learning, can be used whenever the training examples are described by two or more disjoint sets of attributes. Text classification is a domain in which semi-supervised learning algorithms have shown some success. However, multi-view semi-supervised learning has not yet been well explored in this domain despite the possibility of describing textual documents in a myriad of ways. The aim of this work is to analyze the effectiveness of multi-view semi-supervised learning in text classification using unigrams and bigrams as two distinct descriptions of text documents. To this end, we initially consider the widely adopted CO-TRAINING multi-view algorithm and propose some modifications to it in order to deal with the problem of contention points. We also propose the COAL algorithm, which further improves CO-TRAINING by incorporating active learning as a way of dealing with contention points. A thorough experimental evaluation of these algorithms was conducted on real text data sets. The results show that the COAL algorithm, using unigrams as one description of text documents and bigrams as another description, achieves significantly better performance than a single-view semi-supervised algorithm. Taking into account the good results obtained by COAL, we conclude that the use of unigrams and bigrams as two distinct descriptions of text documents can be very effective

全球智慧型手機的出貨量持續提升，且熱門市集的App下載次數紛紛突破500億次。而在iOS和Android手機App市集中，App的評價和評論對App在市集的排序有很大的影響；對於App開發者而言，透過評論確實可掌握使用者的需求，並在產生抱怨前能快速反應避免危機。然而，每日多達上百篇的評論，透過人力逐篇查看，不止耗費時間，更無法整合性的瞭解使用者的需求與問題。 文字情感分析通常會使用監督式或非監督式的方法分析文字評論，其中監督式方法被證實透過簡單的文件量化方法就可達到很高的正確率。但監督式方法有無法預期未知趨勢的限制，且需要進行耗費人力的文章類別標注工作。 本研究透過情感傾向和熱門關注議題兩個面向來分析App評論，提出一個混合非監督式與監督式的中文情感分析方法。我們先透過非監督式方法標注評論類別，並作視覺化整理呈現，最後再用監督式方法建立分類模型，並驗證其效果。 在實驗結果中，利用中文詞彙網路所建立的情感詞集，確實可用來判斷評論的正反情緒，唯判斷負面評論效果不佳需作改善。在議題擷取方面，嘗試使用兩種不同分群方法，其中使用NPMI衡量字詞間關係強度，再配合社群網路分析的Concor方法結果有不錯的成效。最後在使用監督式學習的分類結果中，情感傾向的分類正確率達到87%，關注議題的分類正確率達到96%，皆有不錯表現。 本研究利用中文詞彙網路與社會網路分析，來發展一個非監督式的中文類別判斷方法，並建立一個中文情感分析的範例。另外透過建立全面性的視覺化報告來瞭解使用者的正反回饋意見，並可透過分類模型來掌握新評論的內容，以提供App開發者在市場上之競爭智慧。 / While the number of smartphone shipment is continuesly growing, the number of App downloads from the popular app markets has been already over 50 billion. By Apple App Store and Google Play, ratings and reviews play a more important role in influencing app difusion. While app developers can realize users’ needs by app reviews, more than thousands of reviews produced by user everday become difficult to be read and collated. Sentiment Analysis researchs encompass supervised and unsupervised methods for analyzing review text. The supervised learning is proven as a useful method and can reach high accuracy, but there are limits where future trend can not be recognized and the labels of individual classes must be made manually. We concentrate on two issues, viz Sentiment Orientation and Popular Topic, to propose a Chinese Sentiment Analysis method which combines supervised and unsupervised learning. At First, we use unsupervised learning to label every review articles and produce visualized reports. Secondly, we employee supervised learning to build classification model and verify the result. In the experiment, the Chinese WordNet is used to build sentiment lexicon to determin review’s sentiment orientation, but the result shows it is weak to find out negative review opinions. In the Topic Extraction phase, we apply two clustering methods to extract Popular Topic classes and its result is excellent by using of NPMI Model with Social Network Analysis Method i.e. Concor. In the supervised learning phase, the accuracy of Sentiment Orientation class is 87% and the accuracy of Popular Topic class is 96%. In this research, we conduct an exemplification of the unsupervised method by means of Chinese WorkNet and Social Network Analysis to determin the review classes. Also, we build a comprehensive visualized report to realize users’ feedbacks and utilize classification to explore new comments. Last but not least, with Chinese Sentiment Analysis of this research, and the competitive intelligence in App market can be provided to the App develops.

過往對於情緒分析的研究顯少聚焦在讀者情緒，往往著眼於筆者情緒之研究。讀者情緒是指讀者閱讀文章後產生之情緒感受。然而相同一篇文章可能會引起讀者多種情緒反應，甚至產生與筆者迥異之情緒感受，也突顯其讀者情緒分析存在更複雜的問題。本研究之目的在於辨識讀者閱讀文章後之切確情緒，而文件分類的方法能有效地應用於讀者情緒偵測的研究，除了能辨識出正確的讀者情緒之外，並且能保留讀者情緒文件之相關內容。然而，目前的資訊檢索系統仍缺乏對隱含情緒之文件有效的辨識能力，特別是對於讀者情緒的辨識。除此之外，基於機器學習的方法難以讓人類理解，也很難查明辨識失敗的原因，進而無法了解何種文章引發讀者切確的情緒感受。有鑑於此，本研究提出一套基於語意框架(frame-based approach, FBA)之讀者情緒偵測研究的方法，FBA能模擬人類閱讀文章的方式外，並且可以有效地建構讀者情緒之基礎知識，以形成讀者情緒的知識庫。FBA具備高自動化抽取語意概念的基礎知識，除了利用語法結構的特徵，我們進一步考量周邊語境和語義關聯，將相似的知識整合成具有鑑別力之語意框架，並且透過序列比對(sequence alignment)的方式進行讀者情緒文件之匹配。經實驗結果顯示證明，本研究方法能有效地運用於讀者情緒偵測之相關研究。 / Previous studies on emotion classification mainly focus on the writer's emotional state. By contrast, this research emphasizes emotion detection from the readers' perspective. The classification of documents into reader-emotion categories can be applied in several ways, and one of the applications is to retain only the documents that cause desired emotions for enabling users to retrieve documents that contain relevant contents and at the same time instill proper emotions. However, current IR systems lack of ability to discern emotion within texts, reader-emotion has yet to achieve comparable performance. Moreover, the pervious machine learning-based approaches are generally not human understandable, thereby, it is difficult to pinpoint the reason for recognition failures and understand what emotions do articles trigger in their readers. We propose a flexible semantic frame-based approach (FBA) for reader's emotion detection that simulates such process in human perception. FBA is a highly automated process that incorporates various knowledge sources to learn semantic frames that characterize an emotion and is comprehensible for humans from raw text. Generated frames are adopted to predict readers' emotion through an alignment-based matching algorithm that allows a semantic frame to be partially matched through a statistical scoring scheme. Experiment results demonstrate that our approach can effectively detect readers' emotion by exploiting the syntactic structures and semantic associations in the context as well as outperforms currently well-known statistical text classification methods and the stat-of-the-art reader-emotion detection method.

Cette thèse porte sur l’impact de l’usage de la sémantique dans le processus de la classification supervisée de textes. Cet impact est évalué au travers d’une étude expérimentale sur des documents issus du domaine médical et en utilisant UMLS (Unified Medical Language System) en tant que ressource sémantique. Cette évaluation est faite selon quatre scénarii expérimentaux d’ajout de sémantique à plusieurs niveaux du processus de classification. Le premier scénario correspond à la conceptualisation où le texte est enrichi avant indexation par des concepts correspondant dans UMLS ; le deuxième et le troisième scénario concernent l’enrichissement des vecteurs représentant les textes après indexation dans un sac de concepts (BOC – bag of concepts) par des concepts similaires. Enfin le dernier scénario utilise la sémantique au niveau de la prédiction des classes, où les concepts ainsi que les relations entre eux, sont impliqués dans la prise de décision. Le premier scénario est testé en utilisant trois des méthodes de classification: Rocchio, NB et SVM. Les trois autres scénarii sont uniquement testés en utilisant Rocchio qui est le mieux à même d’accueillir les modifications nécessaires. Au travers de ces différentes expérimentations nous avons tout d’abord montré que des améliorations significatives pouvaient être obtenues avec la conceptualisation du texte avant l’indexation. Ensuite, à partir de représentations vectorielles conceptualisées, nous avons constaté des améliorations plus modérées avec d’une part l’enrichissement sémantique de cette représentation vectorielle après indexation, et d’autre part l’usage de mesures de similarité sémantique en prédiction. / The main interest of this research is the effect of using semantics in the process of supervised text classification. This effect is evaluated through an experimental study on documents related to the medical domain using the UMLS (Unified Medical Language System) as a semantic resource. This evaluation follows four scenarios involving semantics at different steps of the classification process: the first scenario incorporates the conceptualization step where text is enriched with corresponding concepts from UMLS; both the second and the third scenarios concern enriching vectors that represent text as Bag of Concepts (BOC) with similar concepts; the last scenario considers using semantics during class prediction, where concepts as well as the relations between them are involved in decision making. We test the first scenario using three popular classification techniques: Rocchio, NB and SVM. We choose Rocchio for the other scenarios for its extendibility with semantics. According to experiment, results demonstrated significant improvement in classification performance using conceptualization before indexing. Moderate improvements are reported using conceptualized text representation with semantic enrichment after indexing or with semantic text-to-text semantic similarity measures for prediction.