791 research outputs found
Language Dataset Documentation Design: Learning from Deaf and Indigenous Communities
No full textThesis (Ph.D.)--University of Washington, 2023This dissertation investigates how engaging with stakeholder groups, namely natural language processing (NLP) practitioners and language communities, can contribute to the development of documentation toolkits that are more responsive to the needs of these groups. The development process follows value sensitive design in conducting a series of investigations to learn what are the needs of these groups and how iterative improvements to technology can help address those needs. Building from the data statements for NLP Version 1 schema proposed in Bender and Friedman (2018), Dr. Emily M. Bender, Dr. Batya Friedman, and I conduct an empirical investigation and a technical investigation to develop the data statementsVersion 2 schema by engaging with natural language processing professionals. To learn about the needs of indigenous and deaf communities with respect to collaborating with researchers, in a retrospective technical investigation I analyze ethical guidelines and licenses for the values frequently expressed in these communities’ stated expectations for research collaborations. I then conduct a technical investigation to meld the data statements Version 2 schema, aspects of datasheets for datasets (Gebru et al., 2021), and the results of the retrospective technical investigation into a single toolkit. Rather than documenting existing datasets, the Collaborative Discussions for the Documentation and Design of Linguistic Archival Resources (C3DAR) toolkit is designed to facilitate collaborative partnerships between communities and researchers working to develop language datasets. I conclude with possible future investigations, focusing on community researchers as key
stakeholders, and considerations for uptake
Recurrent One-Hop Predictions for Reasoning over Knowledge Graphs
No full textLarge scale knowledge graphs (KGs) such as Freebase are generally incomplete. Reasoning over multi-hop (mh) KG paths is thus an important capability that is needed for question answering or other NLP tasks that require knowledge about the world. mh-KG reasoning includes diverse scenarios, e.g., given a head entity and a relation path, predict the tail entity; or given two enti- ties connected by some relation paths, predict the unknown relation between them. We present ROPs, recurrent one-hop predictors, that predict entities at each step of mh-KB paths by using recurrent neural networks and vector representations of entities and relations, with two benefits: (i) modeling mh-paths of arbitrary lengths while updating the entity and relation representations by the training signal at each step; (ii) handling different types of mh-KG reasoning in a unified framework. Our models show state-of-the-art for two important multi-hop KG reasoning tasks: Knowledge Base Completion and Path Query Answerin
Implementing the syntax of japanese numeral classifiers
No full textWhile the sortal constraints associated with Japanese numeral classifiers are wellstudied, less attention has been paid to the details of their syntax. We describe an analysis implemented within a broadcoverage HPSG that handles an intricate set of numeral classifier construction types and compositionally relates each to an appropriate semantic representation, using Minimal Recursion Semantics
A Grammar Library for Information Structure
No full textThesis (Ph.D.)--University of Washington, 2014This dissertation makes substantial contributions to both the theoretical and computational treatment of information structure, with an eye toward creating natural language processing applications such as multilingual machine translation systems. The aim of the present dissertation is to create a grammar library of information structure for the LinGO Grammar Matrix system (Bender et al. 2010b). Information structure consists of focus, topic, contrast, and background, and refers to how speakers package semantic content they wish to convey to listeners. The information structure of individual sentences is crucial to understanding the cohesiveness of larger segments of text. Despite the crucial role information structure plays in conveying meaning, there is insufficient research on how computational language models might successfully incorporate information structure marking particularly from a multilingual perspective. Part I introduces the current study, and gives some background information. Part II provides cross-linguistic findings about information structure meanings and markings. Part III exploits a naturally occurring text in four languages (e.g. English, Spanish, Russian, and Korean) to formulate a cross-linguistic generalization about distributional properties of information structure. Drawing from these cross-linguistic findings, Part IV shows how information structure can be represented within the HPSG/MRS framework (Pollard and Sag, 1994; Copestake et al., 2005). Part V explores the construction of a grammar library for creating customized grammars incorporating information structure and shows how the information structure-based model improves performance of transfer-based machine translation
Efficient deep processing of japanese
No full textWe present a broad coverage Japanese grammar written in the HPSG formalism with MRS semantics. The grammar is created for use in real world applications, such that robustness and performance issues play an important role. It is connected to a POS tagging and word segmentation tool. This grammar is being developed in a multilingual context, requiring MRS structures that are easily comparable across languages
Adjectives in the LinGO Grammar Matrix
No full textThesis (Master's)--University of Washington, 2014The LinGO Grammar Matrix (Bender et al. 2002, 2010) provides a system for user-linguists to jump start the creation of starter Head-driven Phrase Structure Grammar precision grammars (Pollard and Sag 1994), with semantic representations in Minimal Recursion Semantics (Copestake et al. 2005). The Grammar Matrix provides an online questionnaire for users to describe their target language in a user-friendly and typologically motivated fashion. This description is utilized to produce customized, language-specific rule definitions extending a core, near universal set of types available to any grammar. I propose and implement a new library for intersective adjectives cross-linguistically, considering both attributive and predicative constructions, editing and extending the core grammar while adding additional capabilities to the online customization system to analyze adjectives in target languages and generate language-specific customized grammars with analyses of adjectives. I present a broad typological review the behavior of adjectives, including the morphology and syntax of adjectives, along with an overview of the literature on the semantics of adjectives. I also present a review of the adjectives in several large implemented deep linguistic HPSG grammars in the DELPH-IN formalism. I develop a cross-linguistic analysis of adjectives, adapting previous DELPH-IN analyses to cover significant amounts of new data. The analysis relies not only on definitions in the lexicon, but also on defining the syntactic behavior of adjectives in the morphology. I present a computational implementation of this analysis as an extension to the Grammar Matrix. Finally, I present an evaluation of this extension, showing that the extension achieves 100% coverage of development language test suites and 100% coverage of held out test language test suites, with minimal spurious ambiguity
Author Profiling for Abuse Detection
No full textThe rapid growth of social media in recent years has fed into some highly undesirable phenomena such as proliferation of hateful and offensive language on the Internet. Previous research suggests that such abusive content tends to come from users who share a set of common stereotypes and form communities around them. The current state-of-the-art approaches to abuse detection are oblivious to user and community information and rely entirely on textual (i.e., lexical and semantic) cues. In this paper, we propose a novel approach to this problem that incorporates community-based profiling features of Twitter users. Experimenting with a dataset of 16k tweets, we show that our methods significantly outperform the current state of the art in abuse detection. Further, we conduct a qualitative analysis of model characteristics. We release our code, pre-trained models and all the resources used in the public domain
Author Profiling for Abuse Detection
No full textThe rapid growth of social media in recent years has fed into some highly undesirable phenomena such as proliferation of hateful and offensive language on the Internet. Previous research suggests that such abusive content tends to come from users who share a set of common stereotypes and form communities around them. The current state-of-the-art approaches to abuse detection are oblivious to user and community information and rely entirely on textual (i.e., lexical and semantic) cues. In this paper, we propose a novel approach to this problem that incorporates community-based profiling features of Twitter users. Experimenting with a dataset of 16k tweets, we show that our methods significantly outperform the current state of the art in abuse detection. Further, we conduct a qualitative analysis of model characteristics. We release our code, pre-trained models and all the resources used in the public domain
Recommended from our members
Unsafe AI for Education: A Conversation on Stochastic Parrots and Other Learning Metaphors
No full textThis interview article discusses the impact on popular and educational discourses of the metaphor for a Large Language Model of a “stochastic parrot”. The metaphor comes from the title of an influential paper on the harms of large language models from 2021. Here we present a perspective on the short but influential history of the metaphor through an interview with one of its creators, Professor Emily M. Bender. Using the broad lens of metaphor as a way to shape and frame discourse, the editors interviewed Professor Bender and asked her a series of questions to spark discussion around AI in Education. A variety of topics were covered, including: on how metaphors and anthropomorphisation when carelessly used can elide harms and obviate responsibilities; the role of BigTech, data theft and metaphors of colonisation; Whether AI is unsafe for education and if so to which learners; Techno-solutionism; Positionality and AI “voice”; and whether EdTech is a key driver of AI bullshit and enshittification. This article aims to give readers an accessible insight into how a particular metaphor may be enacted in discourse and to contribute to wider critical debates about how GenAI operates in the context of datafication and educational harms
Implementation for discovery: A bipartite lexicon to support morphological and syntactic analysis 1
No full textThe purpose of this paper is to present and justify aspects of the Montage model of morphology. Montage (Bender et al. 2004) is a long-term project with the goal of building a suite of software tools to assist linguists in the documentation of underdescribed languages by allowing them to make use of techniques from gramma
- …
