Paweł Kędzia

Fextor: A Feature Extraction Framework for Natural Language Processing: A Case Study in Word Sense Disambiguation, Relation Recognition and Anaphora Resolution

Feature extraction from text corpora is an important step in Natural Language Processing (NLP), especially for Machine Learning (ML) techniques. Various NLP tasks have many common steps, e.g. low level act of reading a corpus and obtaining text windows from it. Some high-level processing steps might also be shared, e.g. testing for morpho-syntactic constraints between words. An integrated feature extraction framework removes wasteful redundancy and helps in rapid prototyping.

Recognising Compositionality of Multi-Word Expressions in the Wordnet Oriented Perspective

A method for the recognition of the compositionality of Multi Word Expressions (MWEs) is proposed. First, we study associations between MWEs and the structure of wordnet lexico-semantic relations. A simple method of splitting plWordNet’s MWEs into compositional and non-compositional on the basis of the hypernymy structure is discussed. However, our main goal is to build a classifier for the recognition of compositional MWEs. We assume prior MWE detection. Several experiments with different classification algorithms were performed for the purposes of this task, namely Naive Bayes classifier, Multinomial logistic regression model with a ridge estimator and Decision Table classifier. A heterogeneous set of features is based on: t-score measure for word co-occurrences, Measure of Semantic Relatedness and lexico-syntactic structure of MWEs. MWE compositionality classification is analysed as a knowledge source for automated wordnet expansion.

Graph-Based Complex Representation in Inter-Sentence Relation Recognition in Polish Texts

Abstract This paper presents a supervised approach to the recognition of Cross-document Structure Theory (CST) relations in Polish texts. Its core is a graph-based representation constructed for sentences. Graphs are built on the basis of lexicalised syntactic-semantic relations extracted from text. Similarity between sentences is calculated as similarity between their graphs, and the values are used as features to train the classifiers. Several different configurations of graphs, as well as graph similarity methods were analysed for this task. The approach was evaluated on a large open corpus annotated manually with 17 types of selected CST relations. The configuration of experiments was similar to those known from SEMEVAL and we obtained very promising results.

Graph-Based Approach to Recognizing CST Relations in Polish Texts.

This paper presents an supervised approach to the recognition of Cross-document Structure Theory (CST) relations in Polish texts. In the proposed, graph-based representation is constructed for sentences. Graphs are built on the basis of lexicalised syntactic-semantic relation extracted from text. Similarity between sentences is calculated from graph, and the similarity values are input to classifiers trained by Logistic Model Tree. Several different configurations of graph, as well as graph similarity methods were analysed for this tasks. The approach was evaluated on a large open corpus annotated manually with 17 types of selected CST relations. The configuration of experiments was similar to those known from SEMEVAL and we obtained very promising results.

Finding the optimal number of clusters for word sense disambiguation

Ambiguity is an inherent problem for many tasks in Natural Language Processing. Unsupervised and semi-supervised approaches to ambiguity resolution are appealing as they lower the cost of manual labour. Typically, those methods struggle with estimation of number of senses without supervision. This paper shows research on using stopping functions applied to clustering algorithms for estimation of number of senses. The experiments were performed for Polish and English. We found that estimation based on PK2 stopping functions is encouraging, but only when using coarse-grained distinctions between senses.