Search

This paper presents a description about our adopted approach for the information retrieval and textual entailment tasks of the COLIEE 2017 competition. We address the information retrieval task by implementing a partial string matching and a topic clustering method. For the textual entailment task, we propose a Long Short-Term Memory (LSTM) - Convolutional Neural Network (CNN) model which utilizes word embeddings trained on the Google News vectors. We evaluated our approach for both tasks on the COLIEE 2017 dataset. The results demonstrate that the topic clustering method outperformed the partial string matching method in the information retrieval task. The performance of LSTM-CNN model was competitive with other textual entailment systems.

Textual Inference is a research trend in Natural Language Processing (NLP) that has recently received a lot of attention by the sci- entific community. Textual Entailment (TE) is a specific task in Textual Inference that aims at determining whether a hypothesis is entailed by a text. This paper employs the Child-Sum Tree-LSTM for solving the chal- lenging problem of textual entailment. Our approach is simple and able to generalize well without excessive parameter optimization. Evaluation done on SNLI, SICK and other TE datasets shows the competitiveness of our approach.

The goal of this thesis is to present a multifaceted way of inducing semantic representation from legal documents as well as accessing information in a precise and timely manner. The thesis explored approaches for semantic information retrieval (IR) in the Legal context with a technique that maps specific parts of a text to the relevant concept. This technique relies on text segments, using the Latent Dirichlet Allocation (LDA), a topic modeling algorithm for performing text segmentation, expanding the concept using some Natural Language Processing techniques, and then associating the text segments to the concepts using a semi-supervised text similarity technique. This solves two problems, i.e., that of user specificity in formulating query, and information overload, for querying a large document collection with a set of concepts is more fine-grained since specific information, rather than full documents is retrieved. The second part of the thesis describes our Neural Network Relevance Model for E-Discovery Information Retrieval. Our algorithm is essentially a feature-rich Ensemble system with different component Neural Networks extracting different relevance signal. This model has been trained and evaluated on the TREC Legal track 2010 data. The performance of our models across board proves that it capture the semantics and relatedness between query and document which is important to the Legal Information Retrieval domain.

In this paper, we present a multi-featured supervised automatic keyword extraction system. We extracted salient semantic features which are descriptive of candidate keyphrases, a Random Forest classifier was used for training. The system achieved an accuracy of 58.3 % precision and has shown to outperform two top performing systems when benchmarked on a crowdsourced dataset. Furthermore, our approach achieved a personal best Precision and F-measure score of 32.7 and 25.5 respectively on the Semeval Keyphrase extraction challenge dataset. The paper describes the approaches used as well as the result obtained.1

This paper describes a system which uses entity and topic coherence for improved Text Segmentation (TS) accuracy. First, the Linear Dirichlet Allocation (LDA) algorithm was used to obtain topics for sentences in the document. We then performed entity mapping across a window in order to discover the transition of entities within sentences. We used the information obtained to support our LDA-based boundary detection for proper boundary adjustment. We report the significance of the entity coherence approach as well as the superiority of our algorithm over existing work.

Case-Based Reasoning (CBR) is a branch of AI that is employed to solving problems which emphasizes the use of previous solutions in solving similar new problems. This work presents TAMDS, a Temperament and Mood Detection system which employs Case-Based Reasoning technique. The proposed system is adapted to the field of psychology to help psychologists solve part of the problems in their complex domain. We have designed TAMDS to detect temperament and moods of individuals. A major aim of our system is to help individuals who are out of reach of a professional psychologist to manage their personality and moods because as humans, moods affect our perceptions, personal health, the way we view the world around us and the way we react to it.

Continuous miniaturization of mobile devices has greatly increased its adoption and use by people in various facets of our lives. This has also increased the popularity of face recognition and image processing. Face recognition is now being employed for security purpose opening up the need for further research in recent time. Image compression becomes useful in cases when images need to be transmitted across networks in a less costly way by increasing data volume while reducing transmission time. This work discusses our findings on image compression and its effect on face recognition systems. We studied and implemented three well known face recognition algorithms and observed their recognition accuracy when gallery / probe images were compressed and/or uncompressed as one would naturally expect. For compression purposes, we adopted the JPEG and JPEG2000 coding standard. The face recognition algorithms studied are PCA, ICA and LDA. As a form of an extensive research, experiments conducted include both in compressed and uncompressed domains where the three algorithms have been exhaustively analyzed. We statistically present the results obtained which showed no significant depreciation in the recognition accuracies.

We propose a task independent neural networks model, based on a Siamese-twin architecture. Our model specifically benefits from two forms of attention scheme, which we use to extract high level feature representation of the underlying texts, both at word level (intra-attention) as well as sentence level (inter-attention). The inter attention scheme uses one of the text to create a contextual interlock with the other text, thus paying attention to mutually important parts. We evaluate our system on three tasks, i.e. Textual Entailment, Paraphrase Detection and Answer-Sentence selection. We set a near state-of-the-art result on the textual entailment task with the SNLI corpus while obtaining strong performance across the other tasks that we evaluate our model on.

We describe in this paper, a report of our participation at COLIEE 2016 Information Retrieval (IR) and Legal Question Answering (LQA) tasks. Our solution for the IR part employs the use of a simple but effective Machine Learning (ML) procedure. Our Question Answering solution answers "YES or 'NO' to a question, i.e., 'YES' if the question is entailed by a text and 'NO' otherwise. With recent exploit of Multi-layered Neural Network systems at language modeling tasks, we presented a Deep Learning approach which uses an adaptive variant of the Long-Short Term Memory (LSTM), i.e. the Child Sum Tree LSTM (CST-LSTM) algorithm that we modified to suit our purpose. Additionally, we benchmarked this approach by handcrafting features for two popular ML algorithms, i.e., the Support Vector Machine (SVM) and the Random Forest (RF) algorithms. Even though we used some features that have performed well from similar works, we also introduced some semantic features for performance improvement. We used the results from these two algorithms as the baseline for our CST-LSTM algorithm. All evaluation was done on the COLIEE 2015 training and test sets. The overall result conforms the competitiveness of our approach.

Distributional Semantics is based on the idea of extracting semantic information from lexical information in (multilingual) corpora using statistical algorithms. This paper presents the challenging aim of the SemBurst research projectwhich applies distributional methods not only to words, but to sets of semantic information taken from existing semantic resources and associated with words in syntactic contexts. The idea is to inject semantics into vector space models to find correlations between statements (rather than between words). The proposal may have strong impact on key applications such as Word Sense Disambiguation, Textual Entailment, and others. La semantica distribuzionale si basa sull’idea di estrarre automaticamente informazione semantica attraverso algoritmi statistici applicati ad occorrenze lessicali in grandi corpora. Questo articolo presenta l’idea del progetto SemBurst che applica metodi distribuzionali non solo alle parole, ma ad insiemi di informazioni semantiche tratte da risorse semantiche disponibili e associate alle parole nei relativi contesti sintattici. Lo scopo e’ quello infatti di iniettare semantica negli spazi vettoriali per trovare correlazioni tra informazioni semantiche (piuttosto che tra elementi lessicali). Questo nuovo approccio potra’ avere un alto impatto su applicazioni chiave come Word Sense Disambiguation, Textual Entailment, e altri.

CLiC-it 2015 is held in Trento on December 3-4 2015, hosted and locally organized by Fondazione Bruno Kessler (FBK), one the most important Italian research centers for what concerns CL. The organization of the conference is the result of a fruitful conjoint effort of different research groups (Università di Torino, Università di Roma Tor Vergata and FBK) showing the nationwide spreading of CL in Italy. As in the first edition, the main aim of the event is at establishing a reference forum on CL, covering all the aspects needed to describe the multi-faceted and cross-disciplinary reality of the involved research topics and of the Italian community working in this area. Indeed the spirit of CLiC-it is inclusive, in order to build a scenario as much as possible comprehensive of the complexity of language phenomena and approaches to address them, bringing together researchers and scholars with different competences and skills and working on different aspects according to different perspectives. The large number of researchers that have decided to present their work at CLiC-it and the number of directions here investigated are proof of the maturity of our community and a promising indication of its vitality. We received a total of 64 paper submissions, out of which 52 have been accepted to appear in the Conference Proceedings, which are available online and on the OpenEdition platform. Overall, we collected 129 authors from 15 countries.

with the recent surge in acceptance of face recognition systems, more and more work is needed to perfect the existing grey areas. A major concern is the issue of illumination intensities in the images used as probe and images trained in the database. This paper presents the adoption and use of fuzzy histogram equalization in combating illumination variations in face recognition systems. The face recognition algorithm used is Principal Component Analysis, PCA. Histogram equalization was enhanced using some fuzzy rules in order to get an efficient light normalization. The algorithms were implemented and tested exhaustively with and without the application of fuzzy histogram equalization to test our approach. A good and considerable result was achieved.

Developing Countries are gradually transiting from cash to an electronic based economy by virtue of cashless policy implementation. With this development, cyber criminals and hackers who hitherto attacked businesses and individuals across the Atlantic now see this development as a new venture for their criminal acts and are thus re-directing their energies towards exploiting possible loopholes in the electronic payment system in order to perpetuate fraud. In this paper, we proposed an enhanced approach to detecting phishing attempts and preventing unauthorized online banking withdrawal and transfer. We employed the use of Semantics Content Analysis, Earth Mover Distance and Biometric Authentication with finger print to construct a model. We demonstrated the efficacy of the implemented model with the experiments conducted, a good and considerable result was achieved.

Image compression has received tremendous research input in recent times; however, only few works have been done with respect to face recognition regardless of the fact that, usually, face images need to be transmitted across networks for recognition purposes as seen in surveillance systems. This work analytically investigates the effects of image compression on recognition accuracies of three selected PCA-based algorithms. The algorithms investigated include standard PCA, MPCA and MPCALDA algorithms. As a form of an extensive research, experiments conducted include both in compressed and uncompressed domains where the three algorithms have been exhaustively analyzed. We statistically present the results obtained which shows no significant depreciation in the recognition accuracies.

This paper presents a deep learning based dialogue system which has been trained to answer user queries posed as questions during a conversation. The proposed system, though generative, takes advantage of domain specific knowledge for generating valid answers. The evaluation analysis shows that the proposed system obtained a promising result.