Hugo Lewi Hammer

Sentiment classification of online political discussions: a comparison of a word-based and dependency-based method

Online political discussions have received a lot of attention over the past years. In this paper we compare two sentiment lexicon approaches to classify the sentiment of sentences from political discussions. The first approach is based on applying the number of words between the target and the sentiment words to weight the sentence sentiment score. The second approach is based on using the shortest paths between target and sentiment words in a dependency graph and linguistically motivated syntactic patterns expressed as dependency paths. The methods are tested on a corpus of sentences from online Norwegian political discussions. The results show that the method based on dependency graphs performs significantly better than the word-based approach.

Approximate forward-backward algorithm for a switching linear Gaussian model

A hidden Markov model with two hidden layers is considered. The bottom layer is a Markov chain and given this the variables in the second hidden layer are assumed conditionally independent and Gaussian distributed. The observation process is Gaussian with mean values that are linear functions of the second hidden layer. The forward-backward algorithm is not directly feasible for this model as the recursions result in a mixture of Gaussian densities where the number of terms grows exponentially with the length of the Markov chain. By dropping the less important Gaussian terms an approximate forward-backward algorithm is defined. Thereby one gets a computationally feasible algorithm that generates samples from an approximation to the conditional distribution of the unobserved layers given the data. The approximate algorithm is also used as a proposal distribution in a Metropolis-Hastings setting, and this gives high acceptance rates and good convergence and mixing properties. The model considered is related to what is known as switching linear dynamical systems. The proposed algorithm can in principle also be used for these models and the potential use of the algorithm is therefore large. In simulation examples the algorithm is used for the problem of seismic inversion. The simulations demonstrate the effectiveness and quality of the proposed approximate algorithm.

Multiplicative Update Methods for Incremental Quantile Estimation

We present a novel lightweight incremental quantile estimator which possesses far less complexity than the Tierney’s estimator and its extensions. Notably, our algorithm relies only on tuning one single parameter which is a plausible property which we could only find in the discretized quantile estimator Frugal. This makes our algorithm easy to tune for better performance. Furthermore, our algorithm is multiplicative which makes it highly suitable to handle quantile estimation in systems in which the underlying distribution varies with time. Unlike Frugal and our legacy work which are randomized algorithms, we suggest deterministic updates where the step size is adjusted in a subtle manner to ensure the convergence. The deterministic algorithm is more efficient since the estimate is updated at every iteration. The convergence of the proposed estimator is proven using the theory of stochastic learning. Extensive experimental results show that our estimator clearly outperforms legacy works.

Automatic security classification by machine learning for cross-domain information exchange

Cross-domain information exchange is necessary to obtain information superiority in the military domain, and should be based on assigning appropriate security labels to the information objects. Most of the data found in a defense network is unlabeled, and usually new unlabeled information is produced every day. Humans find that doing the security labeling of such information is labor-intensive and time consuming. At the same time there is an information explosion observed where more and more unlabeled information is generated year by year. This calls for tools that can do advanced content inspection, and automatically determine the security label of an information object correspondingly. This paper presents a machine learning approach to this problem. To the best of our knowledge, machine learning has hardly been analyzed for this problem, and the analysis on topical classification presented here provides new knowledge and a basis for further work within this area. Presented results are promising and demonstrates that machine learning can become a useful tool to assist humans in determining the appropriate security label of an information object.

Building domain specific sentiment lexicons combining information from many sentiment lexicons and a domain specific corpus

Most approaches to sentiment analysis requires a sentiment lexicon in order to automatically predict sentiment or opinion in a text. The lexicon is generated by selecting words and assigning scores to the words, and the performance the sentiment analysis depends on the quality of the assigned scores. This paper addresses an aspect of sentiment lexicon generation that has been overlooked so far; namely that the most appropriate score assigned to a word in the lexicon is dependent on the domain. The common practice, on the contrary, is that the same lexicon is used without adjustments across different domains ignoring the fact that the scores are normally highly sensitive to the domain. Consequently, the same lexicon might perform well on a single domain while performing poorly on another domain, unless some score adjustment is performed. In this paper, we advocate that a sentiment lexicon needs some further adjustments in order to perform well in a specific domain. In order to cope with these domain specific adjustments, we adopt a stochastic formulation of the sentiment score assignment problem instead of the classical deterministic formulation. Thus, viewing a sentiment score as a stochastic variable permits us to accommodate to the domain specific adjustments. Experimental results demonstrate the feasibility of our approach and its superiority to generic lexicons without domain adjustments.

Microbial specialists in below-grade foundation walls in Scandinavia

Below-grade foundation walls are often exposed to excessive moisture by water infiltration, condensation, leakage, or lack of ventilation. Microbial growth in these structures depends largely on environmental factors, elapsed time, and the type of building materials and construction setup. The ecological preferences of Actinomycetes (Actinobacteria) and the molds Ascotricha chartarum, Myxotrichum chartarum (Ascomycota), Geomyces pannorum, and Monocillium sp. (Hyphomycetes) have been addressed based on analyses of 1764 samples collected in below-grade spaces during the period of 2001-2012. Our results show a significant correlation between these taxa and moist foundation walls as ecological niches. Substrate preference was the strongest predictor of taxa distribution within the wall, but the taxas physiological needs, together with gradients of abiotic factors within the wall structure, also played a role. Our study describes for the first time how the wall environment affects microbial growth.

Detecting Threats of Violence in Online Discussions Using Bigrams of Important Words

Making violent threats towards minorities like immigrants or homosexuals is increasingly common on the Internet. We present a method to automatically detect threats of violence using machine learning. A material of 24,840 sentences from YouTube was manually annotated as violent threats or not, and was used to train and test the machine learning model. Detecting threats of violence works quit well with an error of classifying a violent sentence as not violent of about 10% when the error of classifying a non-violent sentence as violent is adjusted to 5%. The best classification performance is achieved by including features that combine specially chosen important words and the distance between those in the sentence.

Threat detection in online discussions

This paper investigates the effect of various types of linguistic features (lexical, syntactic and semantic) for training classifiers to detect threats of violence in a corpus of YouTube comments. Our results show that combinations of lexical features outperform the use of more complex syntactic and semantic features for this task.

Improving classification of tweets using word-word co-occurrence information from a large external corpus

Classifying tweets is an intrinsically hard task as tweets are short messages which makes traditional bags of words based approach inefficient. In fact, bags of words approaches ignores relationships between important terms that do not co-occur literally. In this paper we resort to word-word co-occurence information from a large corpus to expand the vocabulary of another corpus consisting of tweets. Our results show that we are able to reduce the number of erroneous classifications by 14% using co-occurence information.

Quantile estimation using the theory of stochastic learning

The goal of our research is to estimate the quantiles of a distribution from a large set of samples that arrive sequentially. Since the data set is large, the model we choose is that the data cannot be stored, but rather that estimates of the quantiles are computed in a real-time setting. In such settings, classical estimators that require storing the whole history of the data (or stream) cannot be deployed. In this paper, we present an incremental quantile estimator of a distribution, i.e., one that utilizes the previously-computed estimates and only resorts to the last sample for updating these estimates. The state-of-the-art work on obtaining incremental quantile estimators is due to Tierney [9], and is based on the theory of stochastic approximation. However, a primary shortcoming of the latter work is the requirement to incrementally build local approximations of the distribution function in the neighborhood of the quantiles. This requirement, unfortunately, increases the complexity of the algorithm. The convergence of the estimator suggested here is proven using the theory of stochastic learning. These theoretical results have been verified experimentally, and they demonstrate that our estimator outperforms the state-of-the-art estimators. In addition, it also copes with dynamic environments.