Pragmatic Congruence through Language-Specific Mappings from Semantics to Syntax
We present a study of the mappings from semantic content to a range of syntactic expressions with... more We present a study of the mappings from semantic content to a range of syntactic expressions with the aim of isolating the precise locus and role of pragmatic information in the text generation process. From a corpus of English, French, and Portuguese instructions for consumer products, we demonstrate the range of expressions of two semantic relations, Generation and Enablement [Goldman1970] in each of the three languages, and show how the available choices in each language are constrained syntactically, semantically, and pragmatically. The study reveals language-specific principles for syntactic choice, underlining the importance of empirical studies of this nature for informing multilingual NLG. A shorter version of this paper is published in the Proceedings of COLING'96. 1 Dept. English Studies, Univ. Stirling, Stirling FK9 4LA U.K. [email protected] Information Technology Research Institute Technical Report Series ITRI, Univ. of Brighton, Lewes Road, Brighton BN2 4AT, ...
INTRODUCTION The drafter system provides an environment for the author of software documentation ... more INTRODUCTION The drafter system provides an environment for the author of software documentation to construct a model of the knowledge to be conveyed in the document and to have draft manuals generated automatically in multiple languages: currently English and French. The system is relevant to both authoring and translating. However, it brings the role of the translator closer to that of author, in that it does not involve a `source' document 1 . drafter is founded on the following principles, based on the results of a user requirements analysis carried out with technical authors and translators specialising in software documentation [3]: ffl Support for knowledge reuse: Authors and translators spend much of their time acquiring the content of the document to be produced, either from the software developers, through experience with the product, or from an existing document. They often e
Proceedings of the 28th International Conference on Computational Linguistics, 2020
Knowledge bases (KBs) are essential for many downstream NLP tasks, yet their prime shortcoming is... more Knowledge bases (KBs) are essential for many downstream NLP tasks, yet their prime shortcoming is that they are often incomplete. State-of-the-art frameworks for KB completion often lack sufficient accuracy to work fully automated without human supervision. As a remedy, we propose IntKB: a novel interactive framework for KB completion from text based on a question answering pipeline. Our framework is tailored to the specific needs of a human-in-the-loop paradigm: (i) We generate facts that are aligned with text snippets and are thus immediately verifiable by humans. (ii) Our system is designed such that it continuously learns during the KB completion task and, therefore, significantly improves its performance upon initial zero-and few-shot relations over time. (iii) We only trigger human interactions when there is enough information for a correct prediction. Therefore, we train our system with negative examples and a fold-option if there is no answer. Our framework yields a favorable performance: it achieves a hit@1 ratio of 29.7% for initially unseen relations, upon which it gradually improves to 46.2%.
In this paper, we present a methodology for the contrastive analysis of comparable corpora of ins... more In this paper, we present a methodology for the contrastive analysis of comparable corpora of instructional texts in different languages. The methodology is insensitive to the fact that the texts under comparison differ widely in their semantic content, and it can be reliably applied by multiple analysts. We show the results of an empirical study of cross-linguistic variation between Portuguese, French, and English instructions which follows this methodology. Using consumer instructions for ordinary household products in the three languages, we examine expressions of the two semantic relations, generation and enablement (cf. Goldman, 1970), and their available surface syntactic expressions. We examine the role of discourse perspective, as realised by rhetorical relations such as those employed within the framework of Rhetorical Structure Theory (RST), in further narrowing down the range of choices. We demonstrate that the three languages of study tolerate different levels of ambigui...
Traditionally, Controlled Natural Languages (CNLs) are designed either to avoid ambiguity for hum... more Traditionally, Controlled Natural Languages (CNLs) are designed either to avoid ambiguity for human readers, or to facilitate automatic semantic analysis, so that texts can be transcoded to a knowledge representation language. CNLs of the second kind have recently been adapted to the requirements of knowledge formation in OWL for the Semantic Web. We suggest in this paper a variant approach based on automatic generation of texts in CNL (as opposed to automatic analysis), and argue that this provides the best of both worlds, allowing us to pursue human readability in addition to a precise mapping from texts to a formal language.
We describe techniques based on natural language generation which allow a user to author a docume... more We describe techniques based on natural language generation which allow a user to author a document in controlled language for multiple natural languages. The author is expected to be an expert in the application domain but not in the controlled language or in more than one of the supported natural languages. Because the system can produce multiple expressions of the same input in multiple languages, the author can choose among alternative expressions satisfying the constraints of the controlled language. Because the system offers only legitimate choices of wording, correction is unnecessary. Consequently, acceptance of error reports and corrections by trained authors are non-issues.
In search of a reference architecture for NLG systemsLynne
This paper addresses the question of whether Reiter's `Consensus NL Generation Architecture&#... more This paper addresses the question of whether Reiter's `Consensus NL Generation Architecture' Rei94] really exists, and if so whether it is a suitable candidate for a `reference architecture' for NLG systems. Our answer to the rst question is a tentative yes, but we are less comfortable to accept the second. In pursuit of a better understanding, we develop an alternative style of architecture model, and while this also has its shortcomings, the insights gained lead us to propose a hybrid strategy for the development of a reference architecture.
This literature review forms a deliverable in the European Network of Excellence on Semantic Inte... more This literature review forms a deliverable in the European Network of Excellence on Semantic Interoperability and Data Mining in Biomedicine. More specifically, it is part of a work package (WP27) which aims to develop and evaluate generic methods and tools for assisting patients to understand their health and healthcare by generating patient-friendly readable texts that paraphrase the content of their electronic health records. We have reviewed the literature in topics that we consider to be relevant to this work package. When appropriate, we cover variations in conditions in the four countries of the collaborating research groups (France, Germany, Sweden and the UK) and we cover corpora, tools and language technologies for the European languages of interest to these groups.
We present a natural language generator that produces a range of medical reports on the clinical ... more We present a natural language generator that produces a range of medical reports on the clinical histories of cancer patients, and discuss the problem of conceptual restatement in generating various textual views of the same conceptual content. We focus on two features of our system: the demand for “loose paraphrases” between the various reports on a given patient, with a high degree of semantic overlap but some necessary amount of distinctive content; and the requirement for paraphrasing at primarily the discourse level.
Research on paraphrase has mostly focussed on lexical or syntactic variation within individual se... more Research on paraphrase has mostly focussed on lexical or syntactic variation within individual sentences. Our concern is with larger-scale paraphrases, from multiple sentences or paragraphs to entire documents. In this paper we address the problem of generating paraphrases of large chunks of texts. We ground our discussion through a worked example of extending an existing NLG system to accept as input a source text, and to generate a range of fluent semantically-equivalent alternatives, varying not only at the lexical and syntactic levels, but also in document structure and layout.
There are two sides to natural language processing. On the one hand, work in natural language und... more There are two sides to natural language processing. On the one hand, work in natural language understanding is concerned with the mapping from some surface representation of linguistic material expressed as speech or text--to an underlying representation of the meaning carried by that surface representation. But there is also the question of how one maps from some underlying representation of meaning into text or speech: this is the domain of natural language generation. Whether our end-goal is the construction of artifacts that use natural languages intelligently, the formal characterization of phenomena in human languages, or the computational modeling of the human language processing mechanism, we cannot ignore the fact that language is both spoken (or written) and heard (or read). Both are equally large and important problems, but the literature contains much less work on natural language generation (NLG) than it does on natural language understanding (NLU). There are many reaso...
As part of the clef project, we have developed a method for allowing subject matter experts to po... more As part of the clef project, we have developed a method for allowing subject matter experts to pose complex queries to a database. The method makes use of natural language generation, whereby users compose queries through interaction with a dynamic text that is generated “on the fly”. We present here some first-step evaluations that we have conducted with our modest pool of available testers, the numbers of which are constrained for practical reasons having to do with data security. The results allow us to draw some useful conclusions about the usability of the method, and its training requirements, for subjects that are representative of our “typical” intended end-users.
Ontology authoring is a specialised task requiring amongst other things a deep knowledge of the o... more Ontology authoring is a specialised task requiring amongst other things a deep knowledge of the ontology language being used. Understanding and reusing ontologies can thus be difficult for domain experts, who tend not to be ontology experts. To address this problem, we have developed a Natural Language Generation system for transforming the axioms that form the definitions of ontology classes into Natural Language paragraphs. Our method relies on deploying ontology axioms into a top-level Rhetorical Structure Theory schema. Axioms are ordered and structured with specific rhetorical relations under rhetorical structure trees. We describe here an implementation that focuses on a sub-module of SNOMED CT. With some refinements on articles and layout, the resulting paragraphs are fluent and coherent, offering a way for subject specialists to understand an ontology's content without need to understand its logical representation.
To understand a discourse the reader needs to recover the relations between the discourse element... more To understand a discourse the reader needs to recover the relations between the discourse elements as intended by the writer Writers can and very often do help the reader along by providing explicit lexical signals of the intended discourse relations through the use of lexicalised discourse markers In this paper we present a novel approach for generating texts containing multiple embedded rhetorical relations each of which is lexically marked Our analysis inte grates syntactic factors ideas from rhetorical structure theory notions of text grammar and ndings from psycholinguistic research into a uni ed feature based account The current im plementation allows us to generate several good variants of text structures containing multiple discourse relations while blocking dispreferred variants
In the pharmaceutical industry, products have to be described by a range of document types with o... more In the pharmaceutical industry, products have to be described by a range of document types with overlapping content. Moreover, much of this documentation has to be produced in many languages. This situation is commonplace in many commercial domains, and leads to well-known problems in maintaining a set of related documents and their translations. We describe a potential solution explored in the PILLS project. All relevant knowledge about a product is entered only once, through a natural-language interface to a knowledge base. From this 'master model', specialised models for a range of document types are derived automatically; from each specialised model, documents are generated automatically in all supported languages. As an illustration of this approach, the PILLS demonstrator generates three medical document types in English, German and French.
From a generation point of view, our goal is to identify the appropriate mappings between the sem... more From a generation point of view, our goal is to identify the appropriate mappings between the semantics to be conveyed and expressions in language, in the context of multilingual instruction generation. We study this problem focusing on the identification of the realisations of the relationships of the various components of the task the reader is being instructed about. Corpus analysis to study this issue is tricky as there is a real danger of circularity, by identifying the underlying semantic relations (or styles) based on surface features the text, which renders any conclusions as to how these semantics are then expressed in text invalid. In this paper, we explain how it is necessary to go beyond the text to address this problem, and show how we have been able to apply this method in our work.
We assess the results and implications of two projects which have developed multilingual generati... more We assess the results and implications of two projects which have developed multilingual generation tools to support technical writers and translators. These projects have employed a strategy based on two ideas: first, allow technical writers or domain experts to define the content of the generated texts through a user interface specially tailored to the domain; second, restrict the scope of the system to routine passages, leaving human writers and translators to take care of the rest. We have tested our approach through evaluation studies with potential users: these studies suggest such systems could be applied commercially within a few years. 1
To produce coherent linear documents, Natural Language Generation systems have traditionally expl... more To produce coherent linear documents, Natural Language Generation systems have traditionally exploited the structuring role of textual discourse markers such as relational and referential phrases. These coherence markers of the traditional notion of text, however, do not work in non-linear documents: a new set of graphical devices is needed together with formation rules to govern their usage, supported by sound theoretical frameworks. If in linear documents graphical devices such as layout and formatting complement textual devices in the expression of discourse coherence, in non-linear documents they play a more important role. In this paper, we present our theoretical and empirical work in progress, which explores new possibilities for expressing coherence in the generation of hypertext documents. RÉSUMÉ. Dans la production de documents linéaires, les systèmes de Génération Automatique du Langage ont traditionnellement exploité le rôle structurel des marques textuelles du discours, comme les phrases relationnelles et référentielles. Pourtant, ces marques de la cohérence textuelle ne fonctionnent pas dans des documents non linéaires: des nouveaux dispositifs graphiques sont nécessaires avec des règles d'utilisation guidées par des structures théoriques. Si des dispositifs graphiques comme composition et formatage contribuent á l'expression de la cohérence dans des documents linéaires, leur rôle est beaucoup plus important dans des documents non-linéaires. Dans cet article nous explorons des nouvelles possibilités pour représenter la cohérence dans des documents hypertextuels.
Pragmatic Congruence through Language-Specific Mappings from Semantics to Syntax
We present a study of the mappings from semantic content to a range of syntactic expressions with... more We present a study of the mappings from semantic content to a range of syntactic expressions with the aim of isolating the precise locus and role of pragmatic information in the text generation process. From a corpus of English, French, and Portuguese instructions for consumer products, we demonstrate the range of expressions of two semantic relations, Generation and Enablement [Goldman1970] in each of the three languages, and show how the available choices in each language are constrained syntactically, semantically, and pragmatically. The study reveals language-specific principles for syntactic choice, underlining the importance of empirical studies of this nature for informing multilingual NLG. A shorter version of this paper is published in the Proceedings of COLING'96. 1 Dept. English Studies, Univ. Stirling, Stirling FK9 4LA U.K. [email protected] Information Technology Research Institute Technical Report Series ITRI, Univ. of Brighton, Lewes Road, Brighton BN2 4AT, ...
INTRODUCTION The drafter system provides an environment for the author of software documentation ... more INTRODUCTION The drafter system provides an environment for the author of software documentation to construct a model of the knowledge to be conveyed in the document and to have draft manuals generated automatically in multiple languages: currently English and French. The system is relevant to both authoring and translating. However, it brings the role of the translator closer to that of author, in that it does not involve a `source' document 1 . drafter is founded on the following principles, based on the results of a user requirements analysis carried out with technical authors and translators specialising in software documentation [3]: ffl Support for knowledge reuse: Authors and translators spend much of their time acquiring the content of the document to be produced, either from the software developers, through experience with the product, or from an existing document. They often e
Proceedings of the 28th International Conference on Computational Linguistics, 2020
Knowledge bases (KBs) are essential for many downstream NLP tasks, yet their prime shortcoming is... more Knowledge bases (KBs) are essential for many downstream NLP tasks, yet their prime shortcoming is that they are often incomplete. State-of-the-art frameworks for KB completion often lack sufficient accuracy to work fully automated without human supervision. As a remedy, we propose IntKB: a novel interactive framework for KB completion from text based on a question answering pipeline. Our framework is tailored to the specific needs of a human-in-the-loop paradigm: (i) We generate facts that are aligned with text snippets and are thus immediately verifiable by humans. (ii) Our system is designed such that it continuously learns during the KB completion task and, therefore, significantly improves its performance upon initial zero-and few-shot relations over time. (iii) We only trigger human interactions when there is enough information for a correct prediction. Therefore, we train our system with negative examples and a fold-option if there is no answer. Our framework yields a favorable performance: it achieves a hit@1 ratio of 29.7% for initially unseen relations, upon which it gradually improves to 46.2%.
In this paper, we present a methodology for the contrastive analysis of comparable corpora of ins... more In this paper, we present a methodology for the contrastive analysis of comparable corpora of instructional texts in different languages. The methodology is insensitive to the fact that the texts under comparison differ widely in their semantic content, and it can be reliably applied by multiple analysts. We show the results of an empirical study of cross-linguistic variation between Portuguese, French, and English instructions which follows this methodology. Using consumer instructions for ordinary household products in the three languages, we examine expressions of the two semantic relations, generation and enablement (cf. Goldman, 1970), and their available surface syntactic expressions. We examine the role of discourse perspective, as realised by rhetorical relations such as those employed within the framework of Rhetorical Structure Theory (RST), in further narrowing down the range of choices. We demonstrate that the three languages of study tolerate different levels of ambigui...
Traditionally, Controlled Natural Languages (CNLs) are designed either to avoid ambiguity for hum... more Traditionally, Controlled Natural Languages (CNLs) are designed either to avoid ambiguity for human readers, or to facilitate automatic semantic analysis, so that texts can be transcoded to a knowledge representation language. CNLs of the second kind have recently been adapted to the requirements of knowledge formation in OWL for the Semantic Web. We suggest in this paper a variant approach based on automatic generation of texts in CNL (as opposed to automatic analysis), and argue that this provides the best of both worlds, allowing us to pursue human readability in addition to a precise mapping from texts to a formal language.
We describe techniques based on natural language generation which allow a user to author a docume... more We describe techniques based on natural language generation which allow a user to author a document in controlled language for multiple natural languages. The author is expected to be an expert in the application domain but not in the controlled language or in more than one of the supported natural languages. Because the system can produce multiple expressions of the same input in multiple languages, the author can choose among alternative expressions satisfying the constraints of the controlled language. Because the system offers only legitimate choices of wording, correction is unnecessary. Consequently, acceptance of error reports and corrections by trained authors are non-issues.
In search of a reference architecture for NLG systemsLynne
This paper addresses the question of whether Reiter's `Consensus NL Generation Architecture&#... more This paper addresses the question of whether Reiter's `Consensus NL Generation Architecture' Rei94] really exists, and if so whether it is a suitable candidate for a `reference architecture' for NLG systems. Our answer to the rst question is a tentative yes, but we are less comfortable to accept the second. In pursuit of a better understanding, we develop an alternative style of architecture model, and while this also has its shortcomings, the insights gained lead us to propose a hybrid strategy for the development of a reference architecture.
This literature review forms a deliverable in the European Network of Excellence on Semantic Inte... more This literature review forms a deliverable in the European Network of Excellence on Semantic Interoperability and Data Mining in Biomedicine. More specifically, it is part of a work package (WP27) which aims to develop and evaluate generic methods and tools for assisting patients to understand their health and healthcare by generating patient-friendly readable texts that paraphrase the content of their electronic health records. We have reviewed the literature in topics that we consider to be relevant to this work package. When appropriate, we cover variations in conditions in the four countries of the collaborating research groups (France, Germany, Sweden and the UK) and we cover corpora, tools and language technologies for the European languages of interest to these groups.
We present a natural language generator that produces a range of medical reports on the clinical ... more We present a natural language generator that produces a range of medical reports on the clinical histories of cancer patients, and discuss the problem of conceptual restatement in generating various textual views of the same conceptual content. We focus on two features of our system: the demand for “loose paraphrases” between the various reports on a given patient, with a high degree of semantic overlap but some necessary amount of distinctive content; and the requirement for paraphrasing at primarily the discourse level.
Research on paraphrase has mostly focussed on lexical or syntactic variation within individual se... more Research on paraphrase has mostly focussed on lexical or syntactic variation within individual sentences. Our concern is with larger-scale paraphrases, from multiple sentences or paragraphs to entire documents. In this paper we address the problem of generating paraphrases of large chunks of texts. We ground our discussion through a worked example of extending an existing NLG system to accept as input a source text, and to generate a range of fluent semantically-equivalent alternatives, varying not only at the lexical and syntactic levels, but also in document structure and layout.
There are two sides to natural language processing. On the one hand, work in natural language und... more There are two sides to natural language processing. On the one hand, work in natural language understanding is concerned with the mapping from some surface representation of linguistic material expressed as speech or text--to an underlying representation of the meaning carried by that surface representation. But there is also the question of how one maps from some underlying representation of meaning into text or speech: this is the domain of natural language generation. Whether our end-goal is the construction of artifacts that use natural languages intelligently, the formal characterization of phenomena in human languages, or the computational modeling of the human language processing mechanism, we cannot ignore the fact that language is both spoken (or written) and heard (or read). Both are equally large and important problems, but the literature contains much less work on natural language generation (NLG) than it does on natural language understanding (NLU). There are many reaso...
As part of the clef project, we have developed a method for allowing subject matter experts to po... more As part of the clef project, we have developed a method for allowing subject matter experts to pose complex queries to a database. The method makes use of natural language generation, whereby users compose queries through interaction with a dynamic text that is generated “on the fly”. We present here some first-step evaluations that we have conducted with our modest pool of available testers, the numbers of which are constrained for practical reasons having to do with data security. The results allow us to draw some useful conclusions about the usability of the method, and its training requirements, for subjects that are representative of our “typical” intended end-users.
Ontology authoring is a specialised task requiring amongst other things a deep knowledge of the o... more Ontology authoring is a specialised task requiring amongst other things a deep knowledge of the ontology language being used. Understanding and reusing ontologies can thus be difficult for domain experts, who tend not to be ontology experts. To address this problem, we have developed a Natural Language Generation system for transforming the axioms that form the definitions of ontology classes into Natural Language paragraphs. Our method relies on deploying ontology axioms into a top-level Rhetorical Structure Theory schema. Axioms are ordered and structured with specific rhetorical relations under rhetorical structure trees. We describe here an implementation that focuses on a sub-module of SNOMED CT. With some refinements on articles and layout, the resulting paragraphs are fluent and coherent, offering a way for subject specialists to understand an ontology's content without need to understand its logical representation.
To understand a discourse the reader needs to recover the relations between the discourse element... more To understand a discourse the reader needs to recover the relations between the discourse elements as intended by the writer Writers can and very often do help the reader along by providing explicit lexical signals of the intended discourse relations through the use of lexicalised discourse markers In this paper we present a novel approach for generating texts containing multiple embedded rhetorical relations each of which is lexically marked Our analysis inte grates syntactic factors ideas from rhetorical structure theory notions of text grammar and ndings from psycholinguistic research into a uni ed feature based account The current im plementation allows us to generate several good variants of text structures containing multiple discourse relations while blocking dispreferred variants
In the pharmaceutical industry, products have to be described by a range of document types with o... more In the pharmaceutical industry, products have to be described by a range of document types with overlapping content. Moreover, much of this documentation has to be produced in many languages. This situation is commonplace in many commercial domains, and leads to well-known problems in maintaining a set of related documents and their translations. We describe a potential solution explored in the PILLS project. All relevant knowledge about a product is entered only once, through a natural-language interface to a knowledge base. From this 'master model', specialised models for a range of document types are derived automatically; from each specialised model, documents are generated automatically in all supported languages. As an illustration of this approach, the PILLS demonstrator generates three medical document types in English, German and French.
From a generation point of view, our goal is to identify the appropriate mappings between the sem... more From a generation point of view, our goal is to identify the appropriate mappings between the semantics to be conveyed and expressions in language, in the context of multilingual instruction generation. We study this problem focusing on the identification of the realisations of the relationships of the various components of the task the reader is being instructed about. Corpus analysis to study this issue is tricky as there is a real danger of circularity, by identifying the underlying semantic relations (or styles) based on surface features the text, which renders any conclusions as to how these semantics are then expressed in text invalid. In this paper, we explain how it is necessary to go beyond the text to address this problem, and show how we have been able to apply this method in our work.
We assess the results and implications of two projects which have developed multilingual generati... more We assess the results and implications of two projects which have developed multilingual generation tools to support technical writers and translators. These projects have employed a strategy based on two ideas: first, allow technical writers or domain experts to define the content of the generated texts through a user interface specially tailored to the domain; second, restrict the scope of the system to routine passages, leaving human writers and translators to take care of the rest. We have tested our approach through evaluation studies with potential users: these studies suggest such systems could be applied commercially within a few years. 1
To produce coherent linear documents, Natural Language Generation systems have traditionally expl... more To produce coherent linear documents, Natural Language Generation systems have traditionally exploited the structuring role of textual discourse markers such as relational and referential phrases. These coherence markers of the traditional notion of text, however, do not work in non-linear documents: a new set of graphical devices is needed together with formation rules to govern their usage, supported by sound theoretical frameworks. If in linear documents graphical devices such as layout and formatting complement textual devices in the expression of discourse coherence, in non-linear documents they play a more important role. In this paper, we present our theoretical and empirical work in progress, which explores new possibilities for expressing coherence in the generation of hypertext documents. RÉSUMÉ. Dans la production de documents linéaires, les systèmes de Génération Automatique du Langage ont traditionnellement exploité le rôle structurel des marques textuelles du discours, comme les phrases relationnelles et référentielles. Pourtant, ces marques de la cohérence textuelle ne fonctionnent pas dans des documents non linéaires: des nouveaux dispositifs graphiques sont nécessaires avec des règles d'utilisation guidées par des structures théoriques. Si des dispositifs graphiques comme composition et formatage contribuent á l'expression de la cohérence dans des documents linéaires, leur rôle est beaucoup plus important dans des documents non-linéaires. Dans cet article nous explorons des nouvelles possibilités pour représenter la cohérence dans des documents hypertextuels.
Uploads
Papers by Donia Scott