This thesis was done in collaboration with Ericsson AB with the goal of researching the possibility of creating a machine learning model that can transfer the style of a text into another arbitrary style depending on the data used. This had the purpose of making their technical documentation appear to have been written with one cohesive style for a better reading experience. Two approaches to solve this task were tested, the first one was to implement an encoder-decoder model from scratch, and the second was to use the pre-trained GPT-2 model created by a team from OpenAI and fine-tune the model on the specific task. Both of these models were trained on data provided by Ericsson, sentences were extracted from their documentation. To evaluate the models training loss, test sentences, and BLEU scores were used and these were compared to each other and with other state-of-the-art models. The models did not succeed in transforming text into a general technical documentation style but a good understanding of what would need to be improved and adjusted to improve the results were obtained. / <p>This thesis was presented on June 22, 2021, the presentation was done online on Microsoft teams. </p>

Backdoor (Trojan) attacks are a major threat to the security of deep neural network (DNN) models. They are created by an attacker who adds a certain pattern to a portion of given training dataset, causing the DNN model to misclassify any inputs that contain the pattern. These infected classifiers are called Trojan models and the added pattern is referred to as the trigger. In image domain, a trigger can be a patch of pixel values added to the images and in text domain, it can be a set of words. In this thesis, we propose Trojan-Miner (T-Miner), a defense scheme against such backdoor attacks on text classification deep learning models. The goal of T-Miner is to detect whether a given classifier is a Trojan model or not. To create T-Miner , our approach is based on a sequence-to-sequence text generation model. T-Miner uses feedback from the suspicious (test) classifier to perturb input sentences such that their resulting class label is changed. These perturbations can be different for each of the inputs. T-Miner thus extracts the perturbations to determine whether they include any backdoor trigger and correspondingly flag the suspicious classifier as a Trojan model. We evaluate T-Miner on three text classification datasets: Yelp Restaurant Reviews, Twitter Hate Speech, and Rotten Tomatoes Movie Reviews. To illustrate the effectiveness of T-Miner, we evaluate it on attack models over text classifiers. Hence, we build a set of clean classifiers with no trigger in their training datasets and also using several trigger phrases, we create a set of Trojan models. Then, we compute how many of these models are correctly marked by T-Miner. We show that our system is able to detect trojan and clean models with 97% overall accuracy over 400 classifiers. Finally, we discuss the robustness of T-Miner in the case that the attacker knows T-Miner framework and wants to use this knowledge to weaken T-Miner performance. To this end, we propose four different scenarios for the attacker and report the performance of T-Miner under these new attack methods. / M.S. / Backdoor (Trojan) attacks are a major threat to the security of predictive models that make use of deep neural networks. The idea behind these attacks is as follows: an attacker adds a certain pattern to a portion of given training dataset and in the next step, trains a predictive model over this dataset. As a result, the predictive model misclassifies any inputs that contain the pattern. In image domain this pattern that is called trigger, can be a patch of pixel values added to the images and in text domain, it can be a set of words. In this thesis, we propose Trojan-Miner (T-Miner), a defense scheme against such backdoor attacks on text classification deep learning models. The goal of T-Miner is to detect whether a given classifier is a Trojan model or not. T-Miner is based on a sequence-to-sequence text generation model that is connected to the given predictive model and determine if the predictive model is being backdoor attacked. When T-Miner is connected to the predictive model, it generates a set of words, called perturbations, and analyses these perturbations to determine whether they include any backdoor trigger. Hence if any part of the trigger is present in the perturbations, the predictive model is flagged as a Trojan model. We evaluate T-Miner on three text classification datasets: Yelp Restaurant Reviews, Twitter Hate Speech, and Rotten Tomatoes Movie Reviews. To illustrate the effectiveness of T-Miner, we evaluate it on attack models over text classifiers. Hence, we build a set of clean classifiers with no trigger in their training datasets and also using several trigger phrases, we create a set of Trojan models. Then, we compute how many of these models are correctly marked by T-Miner. We show that our system is able to detect Trojan models with 97% overall accuracy over 400 predictive models.

Previous works in Human-Robot Interaction have demonstrated the positive potential benefit of designing highly anthropomorphic robots. This includes physical appearance but also whether they can express emotions, behave in a congruent manner, etc. This work wants to explore the creation of a robot that is able to express a given personality consistently throughout a dialogue while also manifesting congruent emotional expressions. Personality defines many aspects of the character of a person and it can influence how one speaks, behaves, reacts to events, etc. Here, we only focus our attention on language and on how it changes depending on one particular personality trait, the extraversion. To this end, we tested different language models to automate the process of generating language according to a particular personality. We also compared large language models such as GPT-3 to smaller ones, to analyse how size can correlate to performance in this task. We initially evaluated these methods through a fairly small user study in order to confirm the correct manipulation of personality in a text-only context. Results suggest that personality manipulation and how well it is understood highly depend on the context of a dialogue, with a more ‘personal’ dialogue being more successful in manifesting personality. Also, the performance of GPT-3 is comparable to smaller models, specifically trained, with the main difference only given in the perceived fluency of the generations. We then conducted a follow-up study where we chose to use a robot that is capable of showing different facial expressions used to manifest different emotions, the Furhat robot. We integrated into the robot the generations from our language models together with an emotion classification method that is used to guide its facial expressions. Whilst the output of our models did trigger different emotional expressions, resulting in robots which differed both in their language and nonverbal behaviour, resultant perception of these robots’ personality only approached significance (p ∼ 0.08). In this study, GPT3 performed very similarly to much smaller models, with the difference in fluency also being much smaller than before. We did not see any particular change in the perception of the robots in terms of likeability nor uncanniness. / Tidigare arbeten inom Människa-robotinteraktion har visat den positiva potentiella fördelen med att designa mycket antropomorfa robotar. Detta inkluderar fysiskt utseende men också huruvida de kan uttrycka känslor, bete sig på ett kongruent sätt, etc. Detta arbete vill utforska skapandet av en robot som kan uttrycka en given personlighet konsekvent under en dialog samtidigt som den manifesterar kongruenta känslomässiga uttryck. Personlighet definierar många aspekter av en persons karaktär och den kan påverka hur man talar, beter sig, reagerar på händelser etc. Här fokuserar vi vår uppmärksamhet endast på språket och på hur det förändras beroende på ett särskilt personlighetsdrag, extraversion. För detta ändamål testade vi olika språkmodeller för att automatisera processen att skapa språk enligt en viss personlighet. Vi jämförde även stora språkmodeller som GPT-3 med mindre, för att analysera hur storlek kan relatera till prestanda i denna uppgift. Vi utvärderade inledningsvis dessa metoder genom en mindre användarstudie för att bekräfta att personligheten kan manipuleras på rätt sätt i en textbaserad kontext. Resultaten tyder på att personlighetsmanipulation och hur väl den förstås i hög grad beror på sammanhanget i en dialog, där en mer ‘personlig’ dialog är mer framgångsrik när det gäller att manifestera personlighet. Prestandan hos GPT-3 är också jämförbar med mindre modeller, specifikt tränade på en uppgift, där den största skillnaden var i den genererade textens upplevda flyt. Vi gjorde sedan en uppföljningsstudie där vi valde att använda en robot som är kapabel att visa olika ansiktsuttryck och därigenom kapabel att manifestera olika känslor, Furhat-roboten. Vi integrerade talet som genererades från våra språkmodeller i roboten tillsammans med en känsloklassificeringsmetod som används för att styra dess ansiktsuttryck. Medan resultatet av våra modeller framkallade olika känslomässiga uttryck, vilket resulterade i robotar som skilde sig åt både i språk och icke-verbal kommunikation, närmade sig endast den resulterande uppfattningen av dessa robotars personlighet signifikans (p ∼ 0.08). I denna studie presterade GPT-3 mycket likartat med mycket mindre modeller, med skillnaden i flyt också mycket mindre än tidigare. Vi såg ingen speciell förändring i uppfattningen av robotarna när det gäller sympati eller obehaglighet.

The topic of this thesis is the detoxification of language in social networks with a particular focus on style transfer techniques that combine deep learning and linguistic resources. In today’s digital landscape, social networks are rife with communication that can often be toxic, either intentionally or unintentionally. Given the pervasiveness of social media and the potential for toxic language to perpetuate negativity and polarization, this study addresses the problem of toxic language and its transformation into more neutral expressions. The importance of this issue is underscored by the need to promote non-toxic communication in the social networks that are an integral part of modern society. The complexity of natural language and the subtleties of what constitutes toxicity make this a challenging problem worthy of study. To address this problem, this research proposes two models, LexiconGST and MultiLexiconGST, developed based on the Delete&amp;Generate framework. These models integrate linguistic resources into the detoxification system to guide deep learning techniques. Experimental results show that the proposed models perform commendably in the detoxification task compared to stateof-the-art methods. The integration of linguistic resources with deep learning techniques is confirmed to improve the performance of detoxification systems. Finally, this research has implications for social media platforms and online communities, which can now implement more effective moderation tools to promote non-toxic communication. It also opens lines of further research to generalize our proposed method to other text styles. / Ämnet för denna avhandling är avgiftning av språk i sociala nätverk med särskilt fokus på stilöverföringstekniker som kombinerar djupinlärning och språkliga resurser. I dagens digitala landskap är sociala nätverk fulla av kommunikation som ofta kan vara giftig, antingen avsiktligt eller oavsiktligt. Med tanke på hur utbredda sociala medier är och hur giftigt språk kan bidra till negativitet och polarisering, tar den här studien upp problemet med giftigt språk och hur det kan omvandlas till mer neutrala uttryck. Vikten av denna fråga understryks av behovet av att främja giftfri kommunikation i de sociala nätverk som är en integrerad del av det moderna samhället. Komplexiteten i naturligt språk och de subtila aspekterna av vad som utgör toxicitet gör detta till ett utmanande problem som är värt att studera. För att ta itu med detta problem föreslår denna forskning två modeller, LexiconGST och MultiLexiconGST, som utvecklats baserat på ramverket Delete&amp;Generate. Dessa modeller integrerar språkliga resurser i avgiftningssystemet för att vägleda djupinlärningstekniker. Experimentella resultat visar att de föreslagna modellerna presterar lovvärt i avgiftningsuppgiften jämfört med toppmoderna metoder. Integrationen av språkliga resurser med djupinlärningstekniker bekräftas för att förbättra prestanda för avgiftningssystem. Slutligen har denna forskning konsekvenser för sociala medieplattformar och onlinegemenskaper, som nu kan implementera mer effektiva modereringsverktyg för att främja giftfri kommunikation. Det öppnar också för ytterligare forskning för att generalisera vår föreslagna metod till andra textstilar.