Elisabetta Fersini

Sentiment analysis

The huge amount of textual data on the Web has grown in the last few years rapidly creating unique contents of massive dimension. In a decision making context, one of the most relevant tasks is polarity classification of a text source, which is usually performed through supervised learning methods. Most of the existing approaches select the best classification model leading to over-confident decisions that do not take into account the inherent uncertainty of the natural language. In this paper, we pursue the paradigm of ensemble learning to reduce the noise sensitivity related to language ambiguity and therefore to provide a more accurate prediction of polarity. The proposed ensemble method is based on Bayesian Model Averaging, where both uncertainty and reliability of each single model are taken into account. We address the classifier selection problem by proposing a greedy approach that evaluates the contribution of each model with respect to the ensemble. Experimental results on gold standard datasets show that the proposed approach outperforms both traditional classification and ensemble methods. A novel ensemble learning methodology is proposed for polarity classification task.A selection strategy is studied to reduce the search space of candidate ensembles.The proposed model has been shown to be effective and efficient in several domains.

A hierarchical document clustering environment based on the induced bisecting k-means

The steady increase of information on WWW, digital library, portal, database and local intranet, gave rise to the development of several methods to help user in Information Retrieval, information organization and browsing. Clustering algorithms are of crucial importance when there are no labels associated to textual information or documents. The aim of clustering algorithms, in the text mining domain, is to group documents concerning with the same topic into the same cluster, producing a flat or hierarchical structure of clusters. In this paper we present a Knowledge Discovery System for document processing and clustering. The clustering algorithm implemented in this system, called Induced Bisecting k-Means, outperforms the Standard Bisecting k-Means and is particularly suitable for on line applications when computational efficiency is a crucial aspect.

Expressive signals in social media languages to improve polarity detection

To capture the sentiment of messages, several expressive forms are investigated.Expressive signals enrich the feature space of baseline and ensemble classifiers.Only adjectives play a fundamental role as expressive signal.Pragmatic particles and expressive lengthening could lead to the de finition of erratic polarity classifiers. Social media represents an emerging challenging sector where the natural language expressions of people can be easily reported through blogs and short text messages. This is rapidly creating unique contents of massive dimensions that need to be efficiently and effectively analyzed to create actionable knowledge for decision making processes. A key information that can be grasped from social environments relates to the polarity of text messages. To better capture the sentiment orientation of the messages, several valuable expressive forms could be taken into account. In this paper, three expressive signals - typically used in microblogs - have been explored: (1) adjectives, (2) emoticon, emphatic and onomatopoeic expressions and (3) expressive lengthening. Once a text message has been normalized to better conform social media posts to a canonical language, the considered expressive signals have been used to enrich the feature space and train several baseline and ensemble classifiers aimed at polarity classification. The experimental results show that adjectives are more discriminative and impacting than the other considered expressive signals.

Enhance User-Level Sentiment Analysis on Microblogs with Approval Relations

Sentiment Analysis for polarity classification on microblogs is generally based on the assumption that texts are independent and identically distributed (i.i.d). Although these methods are aimed at handling the complex characteristics of natural language, usually they do not consider microblogs as networked data. Early approaches for overcoming this limitation consist in exploiting friendship relationships, since connected users may be more likely to hold similar opinions (Homophily and Social Influence). However, the assumption about the friendship relations does not reflect the real world, where two connected users could have different opinions about the same topic. In order to overcome these shortcomings, we propose a semi-supervised framework that estimates user polarities about a given topic by combining post contents and weighted approval relations, which are intended to better represent the contagion on social networks. The experimental investigation reveals that incorporating approval relations can lead to statistically significant improvements over the performance of complex supervised classifiers based only on textual features.

A probabilistic relational approach for web document clustering

The exponential growth of information available on the World Wide Web, and retrievable by search engines, has implied the necessity to develop efficient and effective methods for organizing relevant contents. In this field document clustering plays an important role and remains an interesting and challenging problem in the field of web computing. In this paper we present a document clustering method, which takes into account both contents information and hyperlink structure of web page collection, where a document is viewed as a set of semantic units. We exploit this representation to determine the strength of a relation between two linked pages and to define a relational clustering algorithm based on a probabilistic graph representation. The experimental results show that the proposed approach, called RED-clustering, outperforms two of the most well known clustering algorithm as k-Means and Expectation Maximization.

Emotional states in judicial courtrooms: An experimental investigation

Thanks to the recent progress in the judicial proceedings management, especially related to the introduction of audio/video recording facilities, the challenge of identification of emotional states can be tackled. Discovering affective states embedded into speech signals could help in semantic retrieval of multimedia clips, and therefore in a deep understanding of mechanisms behind courtroom debates and judges/jurors decision making processes. In this paper two main contributions are given: (1) the collection of real-world human emotions coming from courtroom audio recordings; (2) the investigation of a hierarchical classification system, based on a risk minimization method, able to recognize emotional states from speech signatures. The accuracy of the proposed classification approach - named Multilayer Support Vector Machines - has been evaluated by comparing its performance with traditional machine learning approaches, by using both benchmark datasets and real courtroom recordings. Results in recognition obtained by the proposed technique outperform the prediction power achieved by traditional approaches like SVM, k-Nearest Neighbors, Naive Bayes, Decision Trees and Bayesian Networks.

Detecting irony and sarcasm in microblogs: The role of expressive signals and ensemble classifiers

The automatic detection of sarcasm and irony in user generated contents is one of the most challenging task of Natural Language Processing. In this paper we address this problem by introducing Bayesian Model Averaging (BMA), an ensemble approach to take into account several classifiers according to their reliabilities and their marginal probability predictions. The impact of the most used expressive signals (pragmatic particles and POS tags) have been evaluated in baseline models (traditional classifiers and majority voting) as well as in the proposed BMA approach. Experimental results highlight two main findings: (1) not all the features are equally able to characterize sarcasm and irony and (2) BMA not only outperforms traditional state of the art models, but is also able to ensure notable generalization capabilities both on ironic and sarcastic text.

Audio-Based Emotion Recognition in Judicial Domain: A Multilayer Support Vector Machines Approach

Thanks to the recent progresses in judicial proceedings management, especially related to the introduction of audio/video recording systems, semantic retrieval is a key challenge. In this context emotion recognition engine, through the analysis of vocal signature of actors involved in judicial proceedings, could provide useful annotations for semantic retrieval of multimedia clips. With respect to the generation of semantic emotional tag in judicial domain, two main contributions are given: (1) the construction of an Italian emotional database for Italian proceedings annotation; (2) the investigation of a hierarchical classification system, based on risk minimization method, able to recognize emotional states from vocal signatures. In order to estimate the degree of affection we compared the proposed classification method with SVM, K-Nearest Neighbors and Naive Bayes, highlighting in terms of classification accuracy, the improvements given by a hierarchical learning approach.

Soft-constrained inference for Named Entity Recognition

Much of the valuable information in supporting decision making processes originates in text-based documents. Although these documents can be effectively searched and ranked by modern search engines, actionable knowledge need to be extracted and transformed in a structured form before being used in a decision process. In this paper we describe how the discovery of semantic information embedded in natural language documents can be viewed as an optimization problem aimed at assigning a sequence of labels (hidden states) to a set of interdependent variables (textual tokens). Dependencies among variables are efficiently modeled through Conditional Random Fields, an indirected graphical model able to represent the distribution of labels given a set of observations. The Markov property of these models prevent them to take into account long-range dependencies among variables, which are indeed relevant in Natural Language Processing. In order to overcome this limitation we propose an inference method based on Integer Programming formulation of the problem, where long distance dependencies are included through non-deterministic soft constraints.

Bayesian Model Averaging and Model Selection for Polarity Classification

One of the most relevant task in Sentiment Analysis is Polarity Classification. In this paper, we discuss how to explore the potential of ensembles of classifiers and propose a voting mechanism based on Bayesian Model Averaging (BMA). An important issue to be addressed when using ensemble classification is the model selection strategy. In order to help in selecting the best ensemble composition, we propose an heuristic aimed at evaluating the a priori contribution of each model to the classification task. Experimental results on different datasets show that Bayesian Model Averaging, together with the proposed heuristic, outperforms traditional classification methods and the well known Majority Voting mechanism.