Ángel Iglesias

SyMSS: A syntax-based measure for short-text semantic similarity

Sentence and short-text semantic similarity measures are becoming an important part of many natural language processing tasks, such as text summarization and conversational agents. This paper presents SyMSS, a new method for computing short-text and sentence semantic similarity. The method is based on the notion that the meaning of a sentence is made up of not only the meanings of its individual words, but also the structural way the words are combined. Thus, SyMSS captures and combines syntactic and semantic information to compute the semantic similarity of two sentences. Semantic information is obtained from a lexical database. Syntactic information is obtained through a deep parsing process that finds the phrases in each sentence. With this information, the proposed method measures the semantic similarity between concepts that play the same syntactic role. Psychological plausibility is added to the method by using previous findings about how humans weight different syntactic roles when computing semantic similarity. The results show that SyMSS outperforms state-of-the-art methods in terms of rank correlation with human intuition, thus proving the importance of syntactic information in sentence semantic similarity computation.

Dealing with written language semantics by a connectionist model of cognitive reading

Although machines perform much better than human beings in most of the tasks, it is not the case of natural language processing. Computational linguistic systems usually rely on mathematical and statistical formalisms, which are efficient and useful but far from human procedures and therefore not so skilled. This paper proposes a computational model of natural language reading, called Cognitive Reading Indexing Model (CRIM), inspired by some aspects of human cognition, that tries to become as more psychologically plausible as possible. The model relies on a semantic neural network and it produces not vectors but nets of activated concepts as text representations. Based on these representations, measures of semantic similarity are also defined. Human comparison results show that the system is suitable to model human reading. Additional results also point out that the system could be used in real applications concerning natural language processing tasks.

Detecting phishing e-mails by heterogeneous classification

This paper presents a system for classifying e-mails into two categories, legitimate and fraudulent. This classifier system is based on the serial application of three filters: a Bayesian filter that classifies the textual content of e-mails, a rule- based filter that classifies the non grammatical content of e-mails and, finally, a filter based on an emulator of fictitious accesses which classifies the responses from websites referenced by links contained in e-mails. This system is based on an approach that is hybrid, because it uses different classification methods, and also integrated, because it takes into account all kind of data and information contained in e-mails. This approach aims to provide an effective and efficient classification. The system first applies fast and reliable classification methods, and only when the resulting classification decision is imprecise does the system apply more complex analysis and classification methods.

A Comparison of Hybrid Decision Making Methods for Emergency Support

A situation consisting in evaluating and choosing among alternative actions can be managed from the point of view of Decision Making (DM). This paper presents an approach to design and develop Decision Support Systems (DSS) to be applied in emergency situations. In these situations the decision maker is under heavy stress because each different decision implies different important outcomes related with human and economic losses. First of all, a domain knowledge base has to be built from both the properties of emergency situations and the actions devoted to counteract them. From this knowledge, three different DM methods, based on the Probability Theory and the Possibility Theory, process the incoming emergency information and choose the best action for putting out the emergency situation. The resulting decisions of each method over a set of plausible emergency situations can be evaluated by a domain expert and the method with the best average performance can be built in the DSS. This DSS can help a decision maker find out an optimal decision in a short period of time maximizing security and minimizing stress.

An integrated approach to filtering phishing e-mails

This paper presents a system for classifying e-mails into two categories, legitimate and fraudulent. This classifier system is based on the serial application of three filters: a Bayesian filter that classifies the textual content of e-mails, a rule based filter that classifies the nongrammatical content of e-mails and, finally, a filter based on an emulator of fictitious accesses which classifies the responses from websites referenced by links contained in e-mails. The approach of this system is hybrid, because it uses different classification methods, and also integrated, because it takes into account all kind of data and information contained in e-mails.

A computational knowledge-based model for emulating human performance in the Iowa Gambling Task

A new computational knowledge-based model for emulating human performance in decision making tasks is proposed. This model is mainly based on the knowledge acquired through past experience, the knowledge extracted from the environment and the relationships between the concepts that represent these two kinds of knowledge. The proposed model divides the decision making process into two phases. The first phase lies in the estimation of the decision outcomes using a net of concepts. In the second phase, the proposed model uses a value function to score each possible alternative. The design of the model focuses on some psychological and neurophysiological evidence from current research. In order to validate the model, it is compared with other widely used models that implement different theories of decision making under risk and uncertainty. The model comparison is centered on a well defined task, the Iowa Gambling Task, used in several psychological experiments. The comparison applies an evaluation method based on the optimization of each model in order to emulate human performance individually starting both the participant and the model from the same environmentally available information. The results show that the performance of the proposed model is quantitatively better than the other compared models. Besides, using relevant concepts extracted from interviews with the participants increases the performance of the proposed model.

SMS normalization: combining phonetics, morphology and semantics

The language used in electronic communications such as emails, chats and SMS texts presents special phenomena and important deviations from natural language. Typical machine translation approaches are difficult to adapt to SMS language due to the many irregularities this kind of language shows. This paper presents a new approach for SMS normalization that combines lexical and phonological translation techniques with disambiguation algorithms at two different levels: lexical and semantic. The results obtained by the system outperform some of the existing methods of SMS normalization despite the fact that the corpus created has some features that complicates the normalization task.

Assessing aspects of reading by a connectionist model

Anthropocentrism of computational systems is totally justified when the task concerns to natural language. Computational linguistics systems usually rely on mathematical and statistical formalisms, which are efficient and useful but far from human procedures and therefore not so skilled. The presented work proposes a computational model of natural language reading, called cognitive reading indexing model (CRIM), inspired by some aspects of human cognition, trying to become as psychologically plausible as possible. The model relies on a semantic neural network and it produces nets of activated concepts as text representations. The experimental evaluation shows that the system is suitable to model human reading, and it provides a framework to validate and assess hypothesis concerning reading from other cognitive science fields.

Characterizing Individual Interest by a Computational Model of Reading

Individual topic interest seems to be strongly related to reading attention and depth of processing. Although there exists a lot of research about the study of the relationship between the latter aspects, there is little work in the quantification of this interaction during reading. This paper presents a way for the detection and quantification of the influence of individual interest in the reading process by using several parameters that characterize a computational model of reading called CRIM. Experiment results have pointed out a certain dependency between the implicated issues, motivating further and deeper research.

A methodology for the characterization and diagnosis of cognitive impairments-Application to specific language impairment

OBJECTIVES The diagnosis of mental disorders is in most cases very difficult because of the high heterogeneity and overlap between associated cognitive impairments. Furthermore, early and individualized diagnosis is crucial. In this paper, we propose a methodology to support the individualized characterization and diagnosis of cognitive impairments. The methodology can also be used as a test platform for existing theories on the causes of the impairments. We use computational cognitive modeling to gather information on the cognitive mechanisms underlying normal and impaired behavior. We then use this information to feed machine-learning algorithms to individually characterize the impairment and to differentiate between normal and impaired behavior. We apply the methodology to the particular case of specific language impairment (SLI) in Spanish-speaking children. METHODS AND MATERIALS The proposed methodology begins by defining a task in which normal and individuals with impairment present behavioral differences. Next we build a computational cognitive model of that task and individualize it: we build a cognitive model for each participant and optimize its parameter values to fit the behavior of each participant. Finally, we use the optimized parameter values to feed different machine learning algorithms. The methodology was applied to an existing database of 48 Spanish-speaking children (24 normal and 24 SLI children) using clustering techniques for the characterization, and different classifier techniques for the diagnosis. RESULTS The characterization results show three well-differentiated groups that can be associated with the three main theories on SLI. Using a leave-one-subject-out testing methodology, all the classifiers except the DT produced sensitivity, specificity and area under curve values above 90%, reaching 100% in some cases. CONCLUSIONS The results show that our methodology is able to find relevant information on the underlying cognitive mechanisms and to use it appropriately to provide better diagnosis than existing techniques. It is also worth noting that the individualized characterization obtained using our methodology could be extremely helpful in designing individualized therapies. Moreover, the proposed methodology could be easily extended to other languages and even to other cognitive impairments not necessarily related to language.