Hiram Calvo

Distributional Thesaurus Versus WordNet: A Comparison of Backoff Techniques for Unsupervised PP Attachment

Prepositional Phrase (PP) attachment can be addressed by considering frequency counts of dependency triples seen in a non-annotated corpus. However, not all triples appear even in very big corpora. To solve this problem, several techniques have been used. We evaluate two different backoff methods, one based on WordNet and the other on a distributional (automatically created) thesaurus. We work on Spanish. The thesaurus is created using the dependency triples found in the same corpus used for counting the frequency of unambiguous triples. The training corpus used for both methods is an encyclopaedia. The method based on a distributional thesaurus has higher coverage but lower precision than the WordNet method.

Improving Prepositional Phrase attachment disambiguation using the Web as corpus

The problem of Prepositional Phrase (PP) attachment disambiguation consists in determining if a PP is part of a noun phrase, as in He sees the room with books, or an argument of a verb, as in He fills the room with books. Volk has proposed two variants of a method that queries an Internet search engine to find the most probable attachment variant. In this paper we apply the latest variant of Volk’s method to Spanish with several differences that allow us to attain a better performance close to that of statistical methods using treebanks.

NLP for Shallow Question Answering of Legal Documents Using Graphs

Previous work has shown that modeling relationships between articles of a regulation as vertices of a graph network works twice as better than traditional information retrieval systems for returning articles relevant to the question. In this work we experiment by using natural language techniques such as lemmatizing and using manual and automatic thesauri for improving question based document retrieval. For the construction of the graph, we follow the approach of representing the set of all the articles as a graph; the question is split in two parts, and each of them is added as part of the graph. Then several paths are constructed from part A of the question to part B, so that the shortest path contains the relevant articles to the question. We evaluate our method comparing the answers given by a traditional information retrieval system--vector space model adjusted for article retrieval, instead of document retrieval--and the answers to 21 questions given manually by the general lawyer of the National Polytechnic Institute, based on 25 different regulations (academy regulation, scholarships regulation, postgraduate studies regulation, etc.); with the answer of our system based on the same set of regulations. We found that lemmatizing increases performance in around 10%, while the use of thesaurus has a low impact.

Music composition based on linguistic approach

Music is a form of expression. Since machines have limited capabilities in this sense, our main goal is to model musical composition process, to allow machines to express themselves musically. Our model is based on a linguistic approach. It describes music as a language composed of sequences of symbols that form melodies, with lexical symbols being sounds and silences with their duration in time. We determine functions to describe the probability distribution of these sequences of musical notes and use them for automatic music generation.

DILUCT: an open-source spanish dependency parser based on rules, heuristics, and selectional preferences

A method for recognizing syntactic patterns for Spanish is presented. This method is based on dependency parsing using heuristic rules to infer dependency relationships between words, and word co-occurrence statistics (learnt in an unsupervised manner) to resolve ambiguities such as prepositional phrase attachment. If a complete parse cannot be produced, a partial structure is built with some (if not all) dependency relations identified. Evaluation shows that in spite of its simplicity, the parsers accuracy is superior to the available existing parsers for Spanish. Though certain grammar rules, as well as the lexical resources used, are specific for Spanish, the suggested approach is language-independent.

A novel recurrent neural network soft sensor via a differential evolution training algorithm for the tire contact patch

In this paper we propose a novel Recurrent Neural Network Soft Sensor designed to estimate and predict the contact area that tires of a car are making with the ground. This is one of the most critical issues regarding car modelling for improving its performance. The proposed sensor is particularly useful for an active suspension because it allows its suspension to be prepared instead of reacting to a disturbance. The recurrent neural network enables the Soft Sensor to have a correct prediction of the contact area of the tire. This sensor uses data from 11 sensors mounted on the car while the tire contact patch is obtained by means of frustrated total internal reflection phenomenon. The training process of the Recurrent Neuronal Network presents several difficulties caused by the existence of spurious valleys. For this reason, we address this problem as an optimization problem, solved by using a modified differential evolution algorithm. Our Soft Sensor performance is successfully validated by physical experiments under real operation.

A Micro Artificial Immune System

In this paper, we present a new algorithm, namely, a micro artificial immune system (Micro-AIS) based on the Clonal Selection Theory for solving numerical optimization problems. For our study, we consider the algorithm CLONALG, a widely used artificial immune system. During the process of cloning, CLONALG greatly increases the size of its population. We propose a version with reduced population. Our hypothesis is that reducing the number of individuals in a population will decrease the number of evaluations of the objective function, increasing the speed of convergence and reducing the use of data memory. Our proposal uses a population of 5 individuals (antibodies), from which only 15 clones are obtained. In the maturation stage of the clones, two simple and fast mutation operators are used in a nominal convergence that works together with a reinitialization process to preserve the diversity. To validate our algorithm, we use a set of test functions taken from the specialized literature to compare our approach with the standard version of CLONALG. The same method can be applied in many other problems, for example, in text processing.

Using Graphs for Shallow Question Answering on Legal Documents

This work describes a Shallow Question Answering System (QAS) restricted to legal documents. This system returns a set of relevant articles extracted from several regulation documents. The set of relevant articles allows inferring answers to questions posed in natural language. We take the approach of representing the set of all the articles as a graph; the question is split in two parts (called A and B), and each of them is added as part of the graph. Then several paths are constructed from part A of the question to part B, so that the shortest path contains the relevant articles to the question. We evaluate our method comparing the answers given by a traditional information retrieval system--vector space model adjusted for article retrieval, instead of document retrieval--and the answers to 21 questions given manually by the general lawyer of the National Polytechnic Institute, based on 26 different regulations (academy regulation, scholarships regulation, postgraduate studies regulation, etc.); with the answer of our system based on the same set of regulations. The results show that our system performs twice as better with regard to the traditional Information Retrieval model for Question Answering.

Extracting Semantic Categories of Nouns for Syntactic Disambiguation from Human-Oriented Explanatory Dictionaries *

Syntactic disambiguation frequently requires knowledge of the semantic categories of nouns, especially in languages with free word order. For example, in Spanish the phrases pinto un cuadro un pintor (lit. painted a picture a painter) and pinto un pintor un cuadro (lit. painted a painter a picture) mean the same: ‘a painter painted a picture’. The only way to tell the subject from the object is by knowing that pintor ‘painter’ is a causal agent and cuadro is a thing. We present a method for extracting semantic information of this kind from existing machine-readable human-oriented explanatory dictionaries. Application of this procedure to two different human-oriented Spanish dictionaries gives additional information as compared with using solely Spanish EuroWordNet. In addition, we show the results of an experiment conducted to evaluate the similarity of word classifications using this method.

Acquiring Selectional Preferences from Untagged Text for Prepositional Phrase Attachment Disambiguation

Extracting information automatically from texts for database representation requires previously well-grouped phrases so that entities can be separated adequately. This problem is known as prepositional phrase (PP) attachment disambiguation. Current PP attachment disambiguation systems require an annotated treebank or they use an Internet connection to achieve a precision of more than 90. Unfortunately, these resources are not always available. In addition, using the same techniques that use the Web as corpus may not achieve the same results when using local corpora. In this paper, we present an unsupervised method for generalizing local corpora information by means of semantic classification of nouns based on the top 25 unique beginner concepts of WordNet. Then we propose a method for using this information for PP attachment disambiguation.