Natural language processing has for a long time been a field of research and has been regarded as a thing of the future. Due to its complexity it stopped being featured in computer games in the early 2000s. It has however had a recent revival as a consequence of advancements made in speech recognition, making the possible applications of natural language processing much larger. One market that hasn’t seen much in the way of natural language interfaces recently is that of computer games. This report covers the basics of natural language processing needed to implement two versions of a simple text-based adventure game, one with a menu-based interface and one with a natural lan- guage interface. These were then played by a test group from which usability statistics were gathered to determine if it is likely that NLP will find its way back in to choice driven games in the future. The results showed that even though the menu-based interface has a faster rate of progression, the NLI version of the game was perceived as more enjoyable by users with experience in gaming. The reason being that the NLI al- lowed for more thinking on the user’s part and therefore the game presented a greater challenge, something that is perceived as attractive by users with experience in com- puter games. Also the measured usability was roughly the same for both interfaces while it was feared that it would be much lower for NLIs. Therefore, the conclusion was that it is highly plausible that NLI will find its way back into the gaming world, since it adds a new dimension to adventure games, which is something that attracts users. However, this is given that NLP development continues in the same fast pace as it is today, making it possible to implement a more accurate NLI.

This project investigates how to determine the credibility of a tweet without using human perception. Information about the user and the tweet is studied in search for correlations between their properties and the credibility of the tweet. An alternating decision tree is created to automatically determine the credibility of tweets. Some features are found to correlate to the credibility of the tweets, amongst which the number of previous tweets by a user and the use of uppercase characters are the most prominent.

Sentiment classification is the process of analyzing data and classifying it based on its sentiment conveying properties and the process has a multitude of applications in different industries. However, the different application areas also introduce diverse challenges in implementing the methods successfully. This report examines two of the main approaches commonly used for sentiment classification which entail the use of machine learning and a glossary of weighted words respectively. In addition, preprocessing is explored as an enhancement to the previously mentioned approaches. The approaches are tested on data collected from Twitter to examine their performance in social media. The results indicate that lexicon-based classifiers are the most performant, and that removal of emoticons increases the correctness of classification. / Att kategorisera text beroende på vilken känsla som uttrycks har fått många användningsområden i många industrier. De olika användningsområdena introducerar olika svårigheter att på ett korrekt och konsekvent sätt uppfylla de krav som ställs. Denna rapport avser utforska och bedöma två tillvägagångssätt, ett i form av maskininlärning samt en metod som jämför orden i en text med ordvikter från ett fördefinierat lexikon. Utöver detta analyseras emoji-borttagning som ett möjligt förbättringssätt till båda tillvägagångssätten. Metoderna är testade på data taget från Twitter i syfte att analysera prestandan när data från sociala medier används. Resultaten indikerar att den lexikon-baserade metoden presterar bättre, och att borttagning av emojis ökar korrektheten av klassificeringen.

This report aims to evaluate three different tools for web data extraction in Java for the company Top of Europe. The tools used in the evaluation was jArvest, Jaunt and Selenium WebDriver. Through a case implementation which wrapped parts of a specific web application, web document data was to be automatically identified, extracted and structured. By using the results of the case implementation, the tools was contrasted and evaluated. The results discovered jArvest as non-functioning while the other alternatives provided similar performance but also offering somewhat different strengths. Jaunt provides a good interface to the HTTP protocol and has bigger possibilities for wrapping DOM elements while Selenium Web-Driver supports JavaScript, AJAX and some graphical interface aspects.

Every semester universities around the world have to create new schedules. This task can be very complex considering that a number of constraints has to be taken into account, e.g. there should not exist any timetable clashes for students and a room cannot be double-booked. This can be very hard and time-consuming for a human to do by hand, which is why methods to automate this problem, the University Course Timetabling Problem, has been researched for many years. This report investigates the performance of a hybrid consisting of Genetic Algorithm and Simulated Annealing when solving the University Course Timetabling Problem. An implementation by Yamazaki & Pertoft (2014) was used for the Genetic Algorithm. Simulated Annealing used the Genetic Algorithm as base for its implementation. The hybrid runs the Genetic Algorithm until some breakpoint, takes the best timetable and uses it as an initial solution for the Simulated Annealing. Our results show that our implementation of Simulated Annealing performs better than the hybrid and magnitudes better than the Genetic Algorithm. We believe one reason for this is that the dataset used was too simple, the Genetic Algorithm might scale better as the complexity of the dataset increases.

A recommender system is a tool for recommending personalized content for users based on previous behaviour. This thesis examines the impact of considering item and user bias in matrix factorization for implementing recommender systems. Previous work have shown that user bias have an impact on the predicting power of a recommender system. In this study two different implementations of matrix factorization using stochastic gradient descent are applied to the MovieLens 10M dataset to extract latent features, one of which takes movie and user bias into consideration. The algorithms performed similarly when looking at the prediction capabilities. When examining the features extracted from the two algorithms there was a strong correlation between extracted features and movie genres. We show that each feature form a distinct category of movies where each movie is represented as a combination of the categories. We also show how features can be used to recommend similar movies.

Topological sorting is a graph problem encountered in various different areas in computer science. Many graph problems have benefited from execution on a GPU rather than a CPU due to the GPU's capability for parallelism. The purpose of this report is to determine if topological sorting may benefit from a naive implementation on the GPU compared to the CPU. This is accomplished by constructing a parallel implementation using the CUDA platform by NVIDIA for GPGPU programing. The runtime of this implementation running on several different graphs is compared to a sequential implementation in C running on the CPU. The results indicate that the GPU algorithm only works beneficially on large, shallow graphs.

This paper examines how to write code to gain high performance on modern computers as well as the importance of well planned data structures. The experiments were run on a computer with an Intel i5-5200U CPU, 8GB of RAM running Linux Mint 17.For the measurements Winograd's variant of Strassen's matrix multiplication algorithm was implemented and eventually compared to Intel's math kernel library (MKL). A quadtree data structure was implemented to ensure good cache locality. Loop unrolling and tiling was combined to improve cache performance on both L1 and L2 cache taking into regard the out of order behavior of modern CPUs. Compiler hints were partially used but a large part of the time critical code was written in pure assembler. Measurements of the speed performance of both floats and doubles were performed and substantial differences in running times were found.While there was a substantial difference between the best implementation and MKL for both doubles and floats at smaller sizes, a difference of only 1\% in execution time was achieved for floats at the size of 2^14. This was achieved without any specific tuning and could be expected to be improved if more time was spent on the implementation.

Denna rapport syftar till att undersöka om ämnesmodellering och klustring kan användas till eller underlätta arbetet med trolldetektering. De två ämnesmodellerna Latent Semantic Indexing (LSI) och Latent Dirichlet Allocation (LDA) används samt klustringsmetoden K-means. En grupp om tio användare som bedöms som troll undersöks följt av en undersökning av tio liknande användare för vart och en av de tio trollen. Resultatet visar på att en relativt god mängd av av de relaterade användarna också kunde bedömas som troll. Klustringen kunde också avslöja en del mindre grupper varav några bestod av bottar. Slutsatsen som dras är att ämnesmodellering och klustring tycks vara en god väg att gå men att ytterligare studier behövs. / This report aims to investigate if topic modeling and clustering can be used for or ease the work with troll detection. The two topic models Latent Semantic Indexing (LSI) och Latent Dirichlet Allocation (LDA) as well as the clustering algorithm K-means are selected for this investigation. For each model, each troll in a group of ten trolls has ten related users extracted and then judged on whether they are trolls or not. The results show that a relatively large amount of the related users also were judged as likely to be trolls. The clustering revealed a couple of small groups which seem to consist of a small network of bots. The drawn conclusion is that topic modeling and clustering are seemingly a good choice for aiding in the detection of trolls but further studies are required.

Recommender systems are an important research topic in todays society as the amount of data increases across the globe. In order for commercial systems to give their users good and personalized recommendations on what data may be of interest to them in an effective manner, such a system must be able to give recommendations quickly and scale well as data increases. The purpose of this study is to evaluate two such algorithms with this in mind.  The two different algorithm families tested are classified as item-based collaborative filtering but work very differently. It is therefore of interest to see how their complexities affect their performance, accuracy as well as scalability. The Slope One family is much simpler to implement and proves to be equally as efficient, if not even more efficient than the conventional item-based ones. Both families do require a precomputation stage before recommendations are possible to give, this is the stage where Slope One suffers in comparison to the conventional item-based one. The algorithms are tested using Lenskit, on data provided by GroupLens and their MovieLens project.