From Logic to Language: Natural Language Generation from Logical Forms

The central theme of this thesis is the generation of natural language from a formalrepresentation of meaning. In a nutshell, the problem we want to solve is to go froma logical formula such as forall(x) man(x) implies exists(y) woman(y) and love(x,y) to the sentence“every man loves a woman”. This is achieved by employing several computer al-gorithms and statistical techniques. Moreover, not all representation formalisms areequal, some being better than others for the purpose of generation.The first chapter of this thesis starts by presenting to the reader the problem of Nat-ural Language Generation in a general way. This chapter introduces the formalrepresentation of the meaning of natural language, in particular with formalismsbased on logic, followed by the relationship between NLG and Machine Transla-tion, and the motivation behind the approach to NLG undertaken in this thesis, thatis, robustness and theoretical soundness. Finally, the research questions that drivethe work presented in the rest of the thesis are formulated.Chapter 2 contains a review of the literature in the field of Natural Language Gen-eration, with particular focus on methods and problems relevant to several aspectsof the work presented in the rest of this thesis. The chapter begins with present-ing the traditional architectural organization of NLG tasks, then presents a reviewof previous work on statistical generation, generation from knowledge bases, andprediction of surface order, that is, the order of the words in the final output. Thesecond part of the chapter shows how the representations of meaning found in theliterature do not support well approaches to generation like the one proposed in thisthesis. The chapter ends with a review of software packages for NLG.Chapter 3 introduces the plan for the architecture of a novel system that generatesnatural language expressions given formal representations of meaning as input.Two of the modules, namely the surface order prediction module and the lexical-ization module are covered in greater detail in the central chapters of the thesis. Thesurface realization module, positioned at the end of the pipeline, is also described inthis chapter. It takes the output of the previous modules and produces the completesurface form that expresses the meaning encoded in the original meaning repre-sentation. This chapter also introduces a novel formalism for the representation ofmeaning, based on formal logic. The crucial feature of this formalism, called Dis-course Representation Graphs, is that it favors the alignment between the abstractmeaning representation and the text at the level of words. Thanks to this schema ofalignment, several tasks of the NLG pipeline can be treated by supervised machinelearning approaches.Chapter 4 presents a semantically annotated resource called the Groningen MeaningBank. This resource is used to train the models introduced in the previous chapter,as well as to extract data to test the proposed approaches through experimental tri-als. A number of design choices have been made for the creation and the annotationof the resource, and several software tools were employed to automatically analyzelarge quantity of text. Moreover, crowdsourcing methods were applied to gatherlinguistic annotations from the public, through a Web interface for experts and aGame With A Purpose called Wordrobe.Chapter 5 covers the module of the system responsible for the prediction of theorder of words and phrases composing the surface form. This problem is solved byleveraging a dataset of text-aligned meaning representations and building statisticalmodels for learning to rank to predict the order of small sets of items local to eachconcept.Chapter 6 presents the other central module, that is, the module responsible forthe production of content words for the concepts contained in the original abstractmeaning representation. This module actually solves two problems: the choice ofthe correct lemma from a closed set of options, based on the semantic content toconvey, and the production of the correct morphological inflection. For the firsttask, two alternative methods are proposed: an unsupervised one and a supervisedone, trained on GMB data. For the second task, a pilot study is presented in whichthe problem is solved by a supervised model of inflectional morphology of English.The final chapter contains a series of reflections to conclude the thesis. A looka posteriori highlights the decisions that have been made for the design of the NLGsystem, thus inviting to speculate about alternative directions. Several problems arestill open and it is important to consider how they affect the performance of thesystem. While these issues need to be addressed in order to obtain better results,the approach to NLG presented in this thesis is a step forward in improving existingapproached to generation from logical forms.pas de traduction disponibl