The main RANLP conference will be preceeded by two days tutorials delivered by distinguished lecturers. We plan 4 half-day tutorials, each with duration of 220 minutes, distributed as follows: 60 min talk + 20 min break + 60 min talk + 20 min break + 60 min talk.
|9:30 – 13:10||14:30 – 18:10|
|September 7||1. Preslav Nakov &
Vivi Nastase &
Diarmuid Ó Séaghdha &
|2. Iryna Gurevych &
|September 8||3. Violeta Seretan||4. Dekai Wu|
Preslav Nakov (Qatar Computing Research Institute, Qatar Foundation)
Vivi Nastase (Fondazione Bruno Kessler)
Diarmuid Ó Séaghdha (Cambridge University)
Stan Szpakowicz (University of Ottawa)
"Learning Semantic Relations from Text"
Summary: Every non-trivial text describes interactions and relations between people, institutions, activities, events and so on. What we know about the world consists in large part of such relations, and that knowledge contributes to the understanding of what texts refer to. Newly found relations can in turn become part of the knowledge stored for future use.
To grasp a text's semantic content, an automatic system must be able to recognize relations in texts and reason about them. This may be done by applying and updating previously acquired knowledge. We focus here in particular on semantic relations which describe the interactions among nouns and compact noun phrases, and we present such relations from both a theoretical and a practical perspective. The theoretical exploration sketches the historical path which has brought us to the contemporary view and interpretation of semantic relations. We discuss a wide range of relation inventories proposed by linguists and by language processing people. Such inventories vary by domain, granularity and suitability for downstream applications.
On the practical side, we investigate the recognition and acquisition of relations from texts. In a look at supervised learning methods, we present available datasets, the variety of features which can describe relation instances, and learning algorithms found appropriate for the task. Next, we present weakly supervised and unsupervised learning methods of acquiring relations from large corpora with little or no previously annotated data. We show how enduring the bootstrapping algorithm based on seed examples or patterns has proved to be, and how it has been adapted to tackle Web-scale text collections. We also show a few machine learning techniques which can perform fast and reliable relation extraction by taking advantage of data redundancy and variability.
Iryna Gurevych (Technical University Darmstadt)
Judith Eckle-Kohler (Technical University Darmstadt)
"The Practitioner's Cookbook for Linked Lexical Resources"
Summary: There is a big demand for lexical-semantic resources in Natural Language Processing to improve the performance of language processing, such as word sense disambiguation and Information Extraction. This demand is not met by single resources, neither by expert-constructed resources, such as WordNet and FrameNet, nor by collaboratively constructed resources, such as Wikipedia and Wiktionary. Previously, this bottleneck has been addressed by projects on linking lexical resources at the word sense level in order to increase their coverage and enrich sense representations. More recently, methods for the automatic linking of lexical resources at the sense level (as opposed to manual linking) are being developed. This has to go hand in hand with a harmonization or even standardization of the lexical resources at the representation level.
This tutorial is intended to convey a fundamental understanding of the emerging field of linked lexical resources. The first part of the tutorial will consider linking of lexical resources and the foundations of collaboratively constructed resources. It will start by characterizing the task of linking lexical resources at the word sense level. Then, we will introduce the collaboratively constructed lexical-semantic resources based on Wikipedia and Wiktionary. Thirdly, a method for automatic linking of lexical resources will be presented and illustrated by the automatic word sense alignment of Wiktionary and WordNet.
The second part will focus on a large-scale example of a linked lexical resource and its use in language processing, UBY (http://www.ukp.tu-darmstadt.de/uby). First, we will show how the standardized format of UBY can be used for the automatic linking of verb senses. Then, we will demonstrate the practical use of UBY for language processing by considering knowledge-based word sense disambiguation as a use case. Detailed examples of accessing different resources and information types will be presented. In particular, we will explain how to exploit the sense links between resources for word sense disambiguation.
Violeta Seretan (University of Geneva)
"The Analytics of Word Sociology"
Summary: This tutorial provides a walk-through introduction to the methods developed in corpus linguistics for the analysis of word combinatorics. Taking as objective the identification of words typically co-occurring in a corpus (a task known as “collocation extraction”), I will review the most influential work in the field, providing a comparison at multiple levels: syntactic configuration (which kind of combinatorial phenomena were targeted, e.g., verb-noun, verb-particle, noun-noun); linguistic pre-processing (which kind of technology was used to pre-process the source corpus, from lemmatisation to syntactic parsing or beyond); and association measures (which methods were used to distinguish between genuinely interesting combinations and combinations that are less interesting).
I will start with a theoretical introduction to the phenomenon of collocation, highlighting the main features of this (still ill-defined) concept. After the literature review, I will discuss to what extent the practical work responds to theoretical stipulations. I will discuss open challenges and situate the work within a more global context, relating it to work on identification of non-compositional items and other types of multi-word units. Finally, recent trends will be identified, and the exploitation of multi-word units in other natural language processing applications will be discussed.
Dekai Wu (Hong Kong University of Science & Technology)
"Deeply Integrated Semantic Statistical Machine Translation"
Summary: Semantic SMT is rapidly attracting interest at a time when statistical MT is increasingly in danger of becoming trapped in a plateau, due to over-reliance on architectures that do not attack fundamental problems in machine learning of the right cross-lingual abstractions. Systems that emphasize memorization of ever-larger parallel corpora fail to induce meaningful generalizations, and require far more computational and data resources than should be necessary. Meanwhile, it is difficult to `do the right thing' in the face of complex `spaghetti architectures' involving long chains of often-mismatched heuristic modules.
We will survey central issues and techniques in `deeply theoretically integrated' models of semantic, syntactic, and structured SMT (rather than hybrid approaches such as superficial statistical aggregation or system combination of outputs produced by traditional symbolic components). On one hand, we will explore the trade-offs for semantic SMT between learnability and representational expressiveness. After looking at new studies quantifying how stochastic transduction grammars are ideal encodings of cross-lingual semantic frames, we will examine very recent approaches to automatic unsupervised induction of various classes of such transduction grammars. We will show why stochastic linear transduction grammars (LTGs, LITGs, and PLITGs) are proving to be particularly promising models for the bootstrapping of inducing full-fledged stochastic inversion transduction grammars (ITGs). On the other hand, we will explore the trade-offs for SMT involved in applying various lexical semantics models. We will first examine word sense disambiguation, and discuss why traditional WSD models that are not deeply integrated within the SMT model tend, surprisingly, to fail - whereas in contrast, a deeply embedded phrase sense disambiguation (PSD) approach succeeds. We will also look at semantic role labeling, and discuss the challenges of applying SRL models to SMT. Finally, on semantic MT evaluation, we will explore the latest human and fully-automatic metrics based on semantic frame agreement. By keeping the metrics deeply grounded within the theoretical framework of semantic frames, the new HMEANT and MEANT metrics can significantly outperform even the state-of-the-art expensive HTER and TER metrics used by DARPA, while at the same time maintaining the desirable characteristics of simplicity, inexpensiveness, and representational transparency. We will discuss how deeply embedding the semantic frame criteria into the objective function for SMT tuning robustly increases translation accuracy.