Advances in Artificial Intelligence

With the explosive growth of social communities, users of social Web systems have experienced considerable difficulty with discovering communities relevant to their interests. In this paper we address the problem of recommending communities (or groups) to individual users. We regard this problem as tag-based personalized searches. Based on social tags used by members of communities, we first represent communities in a low-dimensional space, the so-called latent semantic space, by using Latent Semantic Analysis. Then, for recommending communities to a given user, we capture how each community is relevant to both that user’s personal tag usage and other community members’ tagging patterns in the latent space. Our evaluation on the CiteULike dataset shows that our approach can significantly improve the recommendation quality.

In AI planning, macro learning is the task of finding sub-sequences of operators that can be added to the planning domain to improve planner performance. Typically, a single set is added to the domain for all problem instances. A number of techniques have been developed to generate such a macro set based on offline analysis of problem instances. We build on recent work on instance-specific and fixed-set macros, and recast the macro generation problem as parameter configuration: the macros in a domain are viewed as parameters of the planning problem. We then apply an existing parameter configuration system to reconfigure a domain either once or per problem instance. Our empirical results demonstrate that our approach outperforms existing macro acquisition and filtering tools. For instance-specific macros, our approach almost always achieves equal or better performance than a complete evaluation approach, while often being an order of magnitude faster offline.

The pairwise classification approach tends to perform better than other well-known approaches when dealing with multiclass classification problems. In the pairwise approach, however, the nuisance votes of many irrelevant classifiers may result in a wrong prediction class. To overcome this problem, a novel method, Local Crossing Off (LCO), is presented and evaluated in this paper. The proposed LCO system takes advantage of nearest neighbor classification algorithm because of its simplicity and speed, as well as the strength of other two powerful binary classifiers to discriminate between two classes. This paper provides a set of experimental results on 20 datasets using two base learners: Neural Networks and Support Vector Machines. The results show that the proposed technique not only achieves better classification accuracy, but also is computationally more efficient for tackling classification problems which have a relatively large number of target classes.

We present results of sentiment analysis in Twitter messages that disclose personal health information. In these messages (tweets), users discuss ailment, treatment, medications, etc. We use the author-centric annotation model to label tweets as positive sentiments, negative sentiments or neutral. The results of the agreement among three raters are reported and discussed. We then use Machine Learning methods on multi-class and binary classification of sentiments. The obtained results are comparable with previous results in the subjectivity analysis of user-written Web content.

As machine translation (MT) tools have become mainstream, machine translated text has increasingly appeared on multilingual websites. Trustworthy multilingual websites are used as training corpora for statistical machine translation tools; large amounts of MT text in training data may make such products less effective. We performed three experiments to determine whether a support vector machine (SVM) could distinguish machine translated text from human written text (both original text and human translations). Machine translated versions of the Canadian Hansard were detected with an F-measure of 0.999. Machine translated versions of six Government of Canada web sites were detected with an F-measure of 0.98. We validated these results with a decision tree classifier. An experiment to find MT text on Government of Ontario web sites using Government of Canada training data was unfruitful, with a high rate of false positives. Machine translated text appears to be learnable and detectable when using a similar training corpus.

Mining association rules is a fundamental data mining task. However, depending on the choice of the parameters (the minimum confidence and minimum support), current algorithms can become very slow and generate an extremely large amount of results or generate too few results, omitting valuable information.This is a serious problem because in practice users have limited resources for analyzing the results and thus are often only interested in discovering a certain amount of results, and fine tuning the parameters is time-consuming.To address this problem, we propose an algorithm to mine the top-

k

association rules, where

k

is the number of association rules to be found and is set by the user. The algorithm utilizes a new approach for generating association rules named rule expansions and includes several optimizations. Experimental results show that the algorithm has excellent performance and scalability, and that it is an advantageous alternative to classical association rule mining algorithms when the user want to control the number of rules generated.

In some real-world applications, it is time-consuming or expensive to collect much labeled data, while unlabeled data is easier to obtain. Many semi-supervised learning methods have been proposed to deal with this problem by utilizing the unlabeled data. On the other hand, on some datasets, misclassifying different classes causes different costs, which challenges the common assumption in classification that classes have the same misclassification cost. For example, misclassifying a fraud as a legitimate transaction could be more serious than misclassifying a legitimate transaction as fraudulent. In this paper, we propose a cost-sensitive self-training method (CS-ST) to improve the performance of Naive Bayes when labeled instances are scarce and different misclassification errors are associated with different costs. CS-ST incorporates the misclassification costs into the learning process of self-training, and approximately estimates the misclassification error to help select unlabeled instances. Experiments on 13 UCI datasets and three text datasets show that, in terms of the total misclassification cost and the number of correctly classified instances with higher costs, CS-ST has better performance than the self-training method and the base classifier learned from the original labeled data only.

Weighted Partial MaxSAT

(WPMS) is a well-known optimization variant of Boolean Satisfiability (SAT) that finds a wide range of practical applications. WPMS divides the formula in two sets of clauses: The

hard

clauses that must be satisfied and the

soft

clauses that can be unsatisfied with a penalty given by their associated

weight

. However, some applications may require each constraint to be modeled as a

set

or

group

of clauses. The resulting formalism is referred to as

Group MaxSAT

. This paper overviews

Group maxSAT

, and shows how several optimization problems can be modeled as Group MaxSAT. Several

encodings

from Group MaxSAT to standard MaxSAT are formalized and refined. A comprehensive empirical study compares the performance of several MaxSAT solvers with the proposed encodings. The results indicate that, depending on the underlying MaxSAT solver and problem domain, the solver may perform better with a given encoding than with the others.

An agent may have preferences over states and an agent may have preferences over actions. In this paper, we explore the connection between these distinct forms of preference, in the context where action effects are given by a transition system. We illustrate that preferences over actions can not always be reduced to preferences over states, even under very general conditions. It is possible, however, to define a natural notion of consistency between the two forms of preference. Moreover, it is possible to precisely specify which preferences over actions can be expressed in terms of preferences over states. We encode preferences over actions in a logic programming framework that allows us to automatically determine when preferences over actions can be reduced to preferences over states. Our framework facilitates the high-level analysis of preferences by making conflicts explicit. We conclude with a general discussion of conflicting preferences, and we suggest some topics for future work.

In this paper, we explore the application of partial weighted MaxSAT techniques for preference-based planning (PBP). To this end, we develop a compact partial weighted MaxSAT encoding for PBP based on the popular SAS

 + 

planning formalism. Our encoding extends a SAS

 + 

based encoding for SAT-based planning, SASE, to allow for the specification of simple preferences. To the best of our knowledge, the SAS

 + 

formalism has never been exploited in the context of PBP. Our MaxSAT-based PBP planner, MSP

lan

, significantly outperformed the state-of-the-art STRIPS-based MaxSAT approach for PBP with respect to running time, solving more problems in a few cases. Interestingly, when compared to three state-of-the-art heuristic search planners for PBP, MSP

lan

consistently generated plans with comparable quality, slightly outperforming at least one of these three planners in almost every case. Our results illustrate the effectiveness of our SASE based encoding and suggests that MaxSAT-based PBP is a promising area of research.

News is not simply a straight re-telling of events, but rather an interpretation of those events by a reporter, whose feelings and opinions can often become part of the story itself. Research on automatic summarization of news articles has thus far focused on facts rather than emotions, but perhaps emotions can be significant in news stories too. This article describes research done at the University of Ottawa to create an emotion-aware summarization system, which participated in the Text Analysis Conference last year. We have established that increasing the number of emotional words could help ranking sentences to be selected for the summary, but there was no overall improvement in the final system. Although this experiment did not improve news summarization as evaluated by a variety of standard scoring techniques, it was successful at generating summaries with more emotional words while maintaining the overall quality of the summary.

It is well-known that the observation of a variable in a Bay-esian network can affect the effective connectivity of the network, which in turn affects the efficiency of inference. Unfortunately, the observed variables may not be known until runtime, which limits the amount of compile-time optimization that can be done in this regard. In this paper, we consider how to improve inference when we know the likelihood of a variable being observed. We show how these

probabilities of observation

can be exploited to improve existing heuristics for choosing elimination orderings for inference. Empirical tests over a set of benchmark networks using the Variable Elimination algorithm show reductions of up to 50%, 70%, and 55% in multiplications, summations, and runtime, respectively.

Mobile ad hoc networks (MANETs) are formed by a set of mobile nodes without relying on a preexisting infrastructure. In MANETs, ad hoc nodes should count on intermediate nodes to relay messages between two distant nodes. However, due to the inherent dynamicity of such networks, secure and reliable packet forwarding is difficult to achieve. Besides, considering the scarcity of nodes’ power and computational resources, different nodes might find it economically rational to act selfishly in order to maximize their own welfare during their lifecycle. Therefore, in order to keep the network functional we need a mechanism to enforce nodes’ contribution to network operations despite their internal characteristics and willingness.

In this paper we propose a reputation-based schema to foster cooperation amongst ad hoc nodes in the packet forwarding game and isolate selfish behaviours within the MANETs. In the proposed approach, each node is equipped with a security mechanism to estimate the credibility of its neighbouring nodes as well as adaptive measures to compute the cooperation and defection payoffs in order to strategically choose the most profitable behaviour toward its partners. Experimental results indicate the efficacy of our approach in identifying selfish nodes and promote the adoption of a strategic behaviour determination mechanism in a dynamic network in which different nodes with conflicting tendencies participate.

Statistical Machine Translation (SMT) is currently used in real-time and commercial settings to quickly produce initial translations for a document which can later be edited by a human. The SMT models specialized for one domain often perform poorly when applied to other domains. The typical assumption that both training and testing data are drawn from the same distribution no longer applies. This paper evaluates domain adaptation techniques for SMT systems in the context of end-user feedback in a real world application. We present our experiments using two adaptive techniques, one relying on

log-linear models

and the other using

mixture models

. We describe our experimental results on legal and government data, and present the human evaluation effort for post-editing in addition to traditional automated scoring techniques (BLEU scores). The human effort is based primarily on the amount of time and number of edits required by a professional post-editor to improve the quality of machine-generated translations to meet industry standards. The experimental results in this paper show that the domain adaptation techniques can yield a significant increase in BLEU score (up to four points) and a significant reduction in post-editing time of about one second per word.

Basic extreme learning machines apply least square solution to calculate the neural network’s output weights. In the presence of outliers and multi-collinearity, the least square solution becomes invalid. In order to fix this problem, a new kind of extreme learning machine is proposed. An outlier detection technique is introduced to locate outliers and avoid their interference. The least square solution is replaced by regularization for output weights calculation during which the number of hidden nodes is also automatically chosen. Simulation results show that the proposed model has good prediction performance on both normal datasets and datasets contaminated by outliers.

Monitoring the levels of radioxenon isotopes in the atmosphere has been proposed as a means of verifying the Comprehensive Nuclear-Test-Ban Treaty (CTBT). This translates into a classification problem, whereby the measured concentrations either belong to an explosion class or a background class. Instances drawn from the explosions class are extremely rare, if not non-existent. Therefore, the resulting dataset is extremely imbalanced, and inherently suited for one-class classification. Further exacerbating the problem is the fact that the background distribution can be extremely complex, and thus, modelling it using one-class learning is difficult. In order to improve upon the previous classification results, we investigate the augmentation of one-class learning methods with clustering. The purpose of clustering is to convert a complex distribution into simpler distributions, the clusters, over which more effective models can be built. The resulting model, built from one-class learners trained over the clusters, performs more effectively than a model that is built over the original distribution. This thesis is empirically tested on three different data domains; in particular, a number of artificial datasets, datasets from the UCI repository, and data modelled after the extremely challenging CTBT. The results offer credence to the fact that there is an improvement in performance when clustering is used with one-class classification on complex distributions.

Prior work has reported the benefit of transfer learning on domains of single output tasks for classification or prediction of a scalar. We investigate the use of transfer learning on a domain of tasks where each task has multiple outputs (

ie.

output is a vector). Multiple Task Learning (MTL) and

Context-sensitive

Multiple Task Learning (

cs

MTL) neural networks are considered for a domain of image transformation tasks. Models are developed to transform images of neutral human faces to that of corresponding images of angry, happy and sad faces. The MTL approach proves problematic because the size of the network grows as a multiplicative function of the number of outputs and number of tasks. Empirical results show that

cs

MTL neural networks are capable of developing superior models to single task learning models when beneficial transfer occurs from one or more secondary tasks.

Using Situation Calculus, this paper provides a formal account of the relationship between the possible truth problem and the necessary truth problem in dynamical systems. In addition, advantages of applying this formal method to the domain are demonstrated through examples.

Given the breadth of constraint satisfaction problems (CSPs) and the wide variety of CSP solvers, it can be difficult to determine

a priori

which solving method is best suited to a problem. We explore the use of machine learning to predict which solving method will be most effective for a given problem. Our investigation studies the problem of attribute selection for CSPs, and supervised learning to classify CSP instances drawn from four distinct CSP classes. We limit our study to the choice of two well-known, but simple, CSP solvers. We show that the average performance of the resulting solver is very close to the average performance of a CSP solver based on an oracle.

In this paper, we evaluate several factors that influence the performance of n-gram-based music similarity algorithms. Those algorithms are derived from textual information retrieval and adapted to operate on music data. The influence of n-gram length, applied feature extraction method, term weighting approach and similarity measure to the final performance of the similarity measure has been analyzed. MIREX 2005 data and MIREX 2011 evaluation framework for symbolic music similarity task have been used to measure the impact of each of the factors. The paper concludes that the choice of a proper feature extraction method and n-gram length are more important than the applied similarity measure or term weighting technique.

We consider multiagent cooperative decision in stochastic environments, and focus on online decision during which agents communicate. We generalize partial evaluation from a specific application to a class of collaborative decision networks (CDNs), and propose a distributed decision algorithm based on partial evaluation. We show that when agents have private decision variables, the new algorithm can significantly speed up decision in comparison with the earlier CDN algorithm.

The development of location-based social networking (LBSN) services is growing rapidly these days. Users of LBSN services are more interested in sharing tips and experiences of their visits to various locations. In this paper, we aim at a study of recommending locations to users on LBSNs by collaborative filtering (CF) recommenders based only on users’ check-in data. We first design and develop a distributed crawler to collect a large amount of check-in data from Gowalla. Then, we use three ways to utilize the check-in data, namely, the binary utilization, the

FIF

utilization, and the probability utilization. According to different utilizations, we introduce different CF recommenders, namely, user-based, item-based and probabilistic latent semantic analysis (PLSA), to do the location recommendation. Finally, we conduct a set of experiments to compare the performances of different recommenders using different check-in utilizations. The experimental results show that PLSA recommender with the probability utilization outperforms other combinations of recommenders and utilizations for recommending locations to users on LBSN.

Role assignment is an important task in a multi-agent system. There are many constraints in the process of role assignment. This paper formally identifies and defines the role assignment problems with conflicting agent constraints; proposes useful algorithms; conducts experiments by randomly creating agent qualifications and constraints; compares the two algorithms’ performances based on experimental results, and indicates future research activities. The contribution of this work includes: formally defining the problem of role assignment with conflicting agents, and proposing a satisfactory algorithm.

The SmartWheeler project aims at developing an intelligent wheelchair for handicapped people. In this paper, we model the dialogue manager of SmartWheeler in MDP and POMDP frameworks using its collected dialogues. First, we learn the model components of the dialogue MDP based on our previous works. Then, we extend the dialogue MDP to a dialogue POMDP, by proposing two observation models learned from dialogues: one based on learned keywords and the other based on learned intentions. The subsequent keyword POMDP and intention POMDP are compared based on accumulated mean reward in simulation runs. Our experimental results show that the quality of the intention model is significantly higher than the keyword one.

There are several methods that lead to the emergence of specialization in agent societies. Two such methods are the Genetic Threshold Model (GTM) and the Social Inhibition Model (SIM). Based on the premises of these models, such as the availability of social networks, or the presence of genetic thresholds, it is difficult to compare results across these models. We present a model that can mimic both these models, while aiming to increase the effect of agent skill on task choice when agents possess different aptitudes for tasks. Using a metric that quantifies the quality of work performed, we are able to see meaningful increases in work quality, but with a side effect of reduced levels of specialization.

This paper presents recent results in extending the well known Machine Learning ensemble method, boosting. The main idea is to vary the “weak” base classifier with each step of the method, using a classifier which performs “best” on the data presented in that iteration. We show that the solution is sensitive to the loss function used, and that the exponential loss function is detrimental to the performance of this kind of boosting. An approach which uses a logistic loss function performs better, but tends to overfit with a growing number of iterations. We show that this drawback can be overcome with the use of resampling technique, taken from the research on learning from imbalanced data.

We present an algorithm for mining sequential rules common to several sequences, such that rules have to appear within a maximum time span. Experimental results with real-life datasets show that the algorithm can reduce the execution time, memory usage and the number of rules generated by several orders of magnitude compared to previous algorithms.

Falls among the elderly are very common and have a great impact on the health services and the community, as well as on individuals. Many medical studies have focused on the possible risk factors associated with falling in the elderly population, but predicting who is at risk for falling is still an open research question. In this paper, we investigate the use of supervised learning methods for predicting falls in individuals based on the administrative data on their medication use. The data is obtained from a cohort of elderly people in the province of Quebec, and our preliminary empirical investigation yields promising results.

The purpose of this paper is to propose an unsupervised approach for measuring the similarity of texts that can compete with supervised approaches. Finding the inherent properties of similarity between texts using a corpus in the form of a word

n

-gram data set is competitive with other text similarity techniques in terms of performance and practicality. Experimental results on a standard data set show that the proposed unsupervised method outperforms the state-of-the-art supervised method and the improvement achieved is statistically significant at 0.05 level. The approach is language-independent; it can be applied to other languages as long as

n

-grams are available.

This paper proposes a method for extracting the hidden characteristics of machine learning domains. It does so by evaluating the performance of various classifiers on these domains as well as on artificial data whose characteristics are visible since they were purposely included in the generation process. The results obtained on both the real and artificial data are analyzed simultaneously using a classical visualization tool for hierarchical clustering called a dendogram. The idea is to map the real-world domains to the artificial ones according to how well they are learnt by a variety of classifiers and, through this relationship, extract their characteristics. The method is able to determine how difficult it is to classify a specific domain and whether this difficulty stems from the complexity of the concept it embodies, the amount of overlap between each class, the dearth of training data or its dimensionality. This is an important contribution as it allows researchers to understand the underlying nature of their data, and, thus converge quickly toward novel, well-adapted solutions to their particular problems.

Humans and animals acquire their wide repertoire of motor skills through an incremental learning process, during which progressively more complex skills are acquired and subsequently integrated with prior abilities. Inspired by this general idea, we develop an approach for learning motor skills based on a two-level curriculum. At the high level, the curriculum specifies an order in which different skills should be learned. At the low level, the curriculum defines a process for learning within a skill. We develop a set of integrated motor skills for a planar articulated figure capable of doing parameterized hops, flips, rolls, and acrobatic sequences. The same curriculum can be applied to yield individualized motor skill sets for articulated figures of varying proportions.

One-Class Classifiers build classification models in the absence of negative examples, which makes it harder to estimate the class boundary. The predictive accuracy of one-class classifiers can be exacerbated by the presence of missing data in the positive class. In this paper, we propose two approaches based on Bayesian Multiple Imputation (BMI) for imputing missing data in the one-class classification framework called Averaged BMI and Ensemble BMI. We test and compare our approaches against the common method of Mean imputation and Expectation Maximization on several datasets. Our preliminary experiments suggest that as the missingness in the data increases, our proposed imputation approaches can do better on some data sets.

We present a three level Cognitive architecture for the simulation of human behaviour based on Stanovich’s tripartite framework (2010), which provides an explanation of how reflective and adaptive human behaviour emerges from the interaction of three distinct cognitive levels. The architecture is a symbolic dynamical system implemented in a multiagent system. We describe our methodology and a first validation: a classical and semantic Stroop task simulation.

Anomaly detection using One-Class Support Vector Machine (OCSVM) have attracted wide attention in practical applications. Recent research focuses on enhancing OCSVM using either ensemble learning techniques or Multiple Kernel Learning (MKL) since single kernels such as the Gaussian Radial-Based Function (GRBF) kernel might not be flexible enough to construct a proper feature space. In this paper, we develop a new kernel, called centralized GRBF. Further, the two GRBF and centralized GRBF are combined by using a new ensemble kernel technique, called Coupled Ensemble-Kernels (CEK), to improve OCSVM for anomaly detection. Therefore, the final classification model is itself a large-margin classifier while it is actually an ensemble classifier coined with two sub-large-margin models. We show that the proposed CEK outperforms previous approaches using traditional ensemble learning methods and MKL for anomaly detection.

This paper presents a novel method for answering complex temporal ordering questions in the context of an event and query-based text summarization. This task is accomplished by precisely mapping the problem of “query-based summarization of temporal ordering questions” in the field of Natural Language Processing to “verifying a finite state model against a temporal formula” in the realm of Model Checking. This mapping requires specific definitions, structures, and procedures. The output of this new approach is promisingly a readable and informative summary satisfying the user’s needs.

Knowledge compilation is a compelling technique for dealing with the intractability of propositional reasoning. One particularly effective target language is Deterministic Decomposable Negation Normal Form (d-DNNF). We exploit recent advances in #SAT solving in order to produce a new state-of-the-art CNF → d-DNNF compiler: D

sharp

. Empirical results demonstrate that D

sharp

is generally an order of magnitude faster than

c

2

d

, the de facto standard for compiling to d-DNNF, while yielding a representation of comparable size.

In this article, a multiagent system is proposed to solve Job Shop Scheduling Problems. In the proposed system, a number of autonomous agents cooperate in a Multi-Population Cultural Algorithm (MP-CA) framework. The proposed multiagent system consists of a number of groups of agents called sub-populations. The agents in each sub-population are co-evolving using a local CA. The local CAs are working in parallel and communicating to each other to exchange their extracted knowledge. The knowledge is migrated in the form of structured belief which is defined as a statistical records of an agent or a group of agents. Experiments show that our method outperforms some existing methods by offering better solutions as well as a better convergence rate.

The well-known AGM framework provides an intuitively plausible model of nonmonotonic belief revision, but it has the drawback that it is not computational. A computational variant has been proposed by Hansson, and subsequently Hansson and Wassermann have identified a notion of local belief change and discussed how this can modeled in an adaptation of Hansson framework. Briefly, the belief set is compartmentalized in such a way that consistency may be preserved in one compartment, while inconsistency may be entertained in another compartment without the entire belief system degenerating to the trivial case where all propositions are believed. An alternative to the AGM framework is the Dynamic Reasoning System (DRS), which models reasoning explicitly as a temporal activity. The objective in this paper is to show how the phenomenon of local belief change studied by Hansson and Wassermann can be modeled in the DRS framework.

Recent question answering (QA) research has started to incorporate deep natural language processing (NLP) such as syntactic and semantic parsing in order to enhance the capability of selecting the most relevant answers to a given question. However, current NLP technology involves intensive computing and thus hard to meet the real-time demand of synchronous QA. To improve e-learning we introduce NLP into a QA system that specifically exploits the communication latency between student and instructor. We present how the system will fit for educational environment, and how semantic similarity matching between a question and its candidate answers can be improved by semantic roles. The designed system and its running results show the perspective and potential of this research.

The one-to-many agent system is a typical multi-agent system that involves interaction between agents through negotiation. The one-to-many negotiation form is a complicated problem especially when the negotiation is about distinct negotiation objects characterized by multiple negotiation issues. The complexity of the problem comes from the existence of many variables in the negotiation process. For example, the number of agents, the number of objects and the number of negotiation issues contribute to the problem complexity. Few existing works address some aspects of the coordination problem in the one-to-many negotiation form. However, most works address simple negotiation scenarios such as negotiation over a single object characterized by a single issue or multiple issues. The aim of this research is to investigate possible coordination problems in the one-to-many negotiation form and propose effective and robust solutions for a number negotiation scenarios. We test our solutions by evaluating some negotiation objective criteria such as utility gain, agreement rate etc.

Our work addresses the problem of generalizing a plan and representing it for efficient execution. A key area of automated planning is the study of how to generate a plan for an agent to execute. The plan itself may take on many forms: a sequence of actions, a partial ordering over a set of actions, or a procedure-like description of what the agent should do. Once a plan is found, the question remains as to how the agent should execute the plan. For simple forms of representation (e.g., a sequence of actions), the answer to this question is straightforward. However, when the plan representation is more expressive (e.g., a GOLOG program [4]), or the agent is acting in an uncertain world, execution can be considerably more challenging. We focus on the problem of how to generalize various plan representations into a form that an agent can use for efficient and robust online execution.

Srivistava et al. propose a definition of a generalized plan as an algorithm that maps problem instances to a sequence of actions that solves the instance [7]. Our work fits nicely into this formalism, and in Section 3 we describe how a problem (i.e., a state of the world and goal) is mapped to a sequence of actions (i.e., what the agent should do).

Software Requirements Specifications (SRS) documents are important artifacts in the software industry. A SRS contains all the requirements specifications for a software system, either as functional requirements (FR) or non-functional requirements (NFR). FRs are the features of the system-to-be, whereas NFRs define its quality attributes. NFRs impact the system as a whole and interact both with each other and with the functional requirements. SRS documents are typically written in informal natural language [1], which impedes their automated analysis. The goal of this work is to support software engineers with semantic analysis methods that can automatically extract and analyze requirements written in natural language texts, in order to (i) make SRS documents machine-processable by transforming them into an ontological representation; (ii) apply quality assurance (QA) methods on the extracted requirements, in order to detect defects, like ambiguities or omissions; and (iii) attempt to build traceability links between NFRs and the FRs impacted by them, in order to aid effort estimation models.

Research applying logic and reasoning to natural language processing (NLP) requires large costly hand built knowledge resources. Time, effort, and the cost of producing these resources usually leads to limited coverage. There is a need to extract knowledge from textual knowledge resources, such as dictionaries, encyclopedias, and textbooks and then convert it into a format usable by reasoning NLP systems. My goal is to begin this task with extracting knowledge from dictionaries and demonstrating its utility in NLP tasks.