Graeme Hirst

Krishnapriya Vishnubhotla, Adam Hammond, Graeme Hirst · utoronto

We present the Project Dialogism Novel Corpus, or PDNC, an annotated dataset of quotations for English literary texts. PDNC contains annotations for 35,978 quotations across 22 full-length novels, and is by an order of magnitude the largest corpus of its kind. Each quotation is annotated for the speaker, addressees, type of quotation, referring expression, and character mentions within the quotation text. The annotated attributes allow for a comprehensive evaluation of models of quotation attribution and coreference for literary texts.

Krishnapriya Vishnubhotla, Frank Rudzicz, Graeme Hirst et al. · utoronto

Current models for quotation attribution in literary novels assume varying levels of available information in their training and test data, which poses a challenge for in-the-wild inference. Here, we approach quotation attribution as a set of four interconnected sub-tasks: character identification, coreference resolution, quotation identification, and speaker attribution. We benchmark state-of-the-art models on each of these sub-tasks independently, using a large dataset of annotated coreferences and quotations in literary novels (the Project Dialogism Novel Corpus). We also train and evaluate models for the speaker attribution task in particular, showing that a simple sequential prediction model achieves accuracy scores on par with state-of-the-art models.

Krishnapriya Vishnubhotla, Adam Hammond, Graeme Hirst et al.

Stories are rich in the emotions they exhibit in their narratives and evoke in the readers. The emotional journeys of the various characters within a story are central to their appeal. Computational analysis of the emotions of novels, however, has rarely examined the variation in the emotional trajectories of the different characters within them, instead considering the entire novel to represent a single story arc. In this work, we use character dialogue to distinguish between the emotion arcs of the narration and the various characters. We analyze the emotion arcs of the various characters in a dataset of English literary novels using the framework of Utterance Emotion Dynamics. Our findings show that the narration and the dialogue largely express disparate emotions through the course of a novel, and that the commonalities or differences in the emotional arcs of stories are more accurately captured by those associated with individual characters.

Jing Yi Xie, Graeme Hirst, Yang Xu

Developing moral awareness in intelligent systems has shifted from a topic of philosophical inquiry to a critical and practical issue in artificial intelligence over the past decades. However, automated inference of everyday moral situations remains an under-explored problem. We present a text-based approach that predicts people's intuitive judgment of moral vignettes. Our methodology builds on recent work in contextualized language models and textual inference of moral sentiment. We show that a contextualized representation offers a substantial advantage over alternative representations based on word embeddings and emotion sentiment in inferring human moral judgment, evaluated and reflected in three independent datasets from moral psychology. We discuss the promise and limitations of our approach toward automated textual moral reasoning.

Jing Yi Xie, Renato Ferreira Pinto, Graeme Hirst et al.

We present a text-based framework for investigating moral sentiment change of the public via longitudinal corpora. Our framework is based on the premise that language use can inform people's moral perception toward right or wrong, and we build our methodology by exploring moral biases learned from diachronic word embeddings. We demonstrate how a parameter-free model supports inference of historical shifts in moral sentiment toward concepts such as slavery and democracy over centuries at three incremental levels: moral relevance, moral polarity, and fine-grained moral dimensions. We apply this methodology to visualizing moral time courses of individual concepts and analyzing the relations between psycholinguistic variables and rates of moral sentiment change at scale. Our work offers opportunities for applying natural language processing toward characterizing moral sentiment change in society.

Kawin Ethayarajh, David Duvenaud, Graeme Hirst

Word embeddings are often criticized for capturing undesirable word associations such as gender stereotypes. However, methods for measuring and removing such biases remain poorly understood. We show that for any embedding model that implicitly does matrix factorization, debiasing vectors post hoc using subspace projection (Bolukbasi et al., 2016) is, under certain conditions, equivalent to training on an unbiased corpus. We also prove that WEAT, the most common association test for word embeddings, systematically overestimates bias. Given that the subspace projection method is provably effective, we use it to derive a new measure of association called the $\textit{relational inner product association}$ (RIPA). Experiments with RIPA reveal that, on average, skipgram with negative sampling (SGNS) does not make most words any more gendered than they are in the training corpus. However, for gender-stereotyped words, SGNS actually amplifies the gender association in the corpus.

Kawin Ethayarajh, David Duvenaud, Graeme Hirst

A surprising property of word vectors is that word analogies can often be solved with vector arithmetic. However, it is unclear why arithmetic operators correspond to non-linear embedding models such as skip-gram with negative sampling (SGNS). We provide a formal explanation of this phenomenon without making the strong assumptions that past theories have made about the vector space and word distribution. Our theory has several implications. Past work has conjectured that linear substructures exist in vector spaces because relations can be represented as ratios; we prove that this holds for SGNS. We provide novel justification for the addition of SGNS word vectors by showing that it automatically down-weights the more frequent word, as weighting schemes do ad hoc. Lastly, we offer an information theoretic interpretation of Euclidean distance in vector spaces, justifying its use in capturing word dissimilarity.

Mohamed Abdalla, Graeme Hirst

Current approaches to cross-lingual sentiment analysis try to leverage the wealth of labeled English data using bilingual lexicons, bilingual vector space embeddings, or machine translation systems. Here we show that it is possible to use a single linear transformation, with as few as 2000 word pairs, to capture fine-grained sentiment relationships between words in a cross-lingual setting. We apply these cross-lingual sentiment models to a diverse set of tasks to demonstrate their functionality in a non-English context. By effectively leveraging English sentiment knowledge without the need for accurate translation, we can analyze and extract features from other languages with scarce data at a very low cost, thus making sentiment and related analyses for many languages inexpensive.

Vanessa Q. Marinho, Graeme Hirst, Diego R. Amancio

The vast amount of data and increase of computational capacity have allowed the analysis of texts from several perspectives, including the representation of texts as complex networks. Nodes of the network represent the words, and edges represent some relationship, usually word co-occurrence. Even though networked representations have been applied to study some tasks, such approaches are not usually combined with traditional models relying upon statistical paradigms. Because networked models are able to grasp textual patterns, we devised a hybrid classifier, called labelled subgraphs, that combines the frequency of common words with small structures found in the topology of the network, known as motifs. Our approach is illustrated in two contexts, authorship attribution and translationese identification. In the former, a set of novels written by different authors is analyzed. To identify translationese, texts from the Canadian Hansard and the European parliament were classified as to original and translated instances. Our results suggest that labelled subgraphs are able to represent texts and it should be further explored in other tasks, such as the analysis of text complexity, language proficiency, and machine translation.

Vanessa Queiroz Marinho, Graeme Hirst, Diego Raphael Amancio

Concepts and methods of complex networks can be used to analyse texts at their different complexity levels. Examples of natural language processing (NLP) tasks studied via topological analysis of networks are keyword identification, automatic extractive summarization and authorship attribution. Even though a myriad of network measurements have been applied to study the authorship attribution problem, the use of motifs for text analysis has been restricted to a few works. The goal of this paper is to apply the concept of motifs, recurrent interconnection patterns, in the authorship attribution task. The absolute frequencies of all thirteen directed motifs with three nodes were extracted from the co-occurrence networks and used as classification features. The effectiveness of these features was verified with four machine learning methods. The results show that motifs are able to distinguish the writing style of different authors. In our best scenario, 57.5% of the books were correctly classified. The chance baseline for this problem is 12.5%. In addition, we have found that function words play an important role in these recurrent patterns. Taken together, our findings suggest that motifs should be further explored in other related linguistic tasks.

Vanessa Wei Feng, Graeme Hirst

Previous attempts at RST-style discourse segmentation typically adopt features centered on a single token to predict whether to insert a boundary before that token. In contrast, we develop a discourse segmenter utilizing a set of pairing features, which are centered on a pair of adjacent tokens in the sentence, by equally taking into account the information from both tokens. Moreover, we propose a novel set of global features, which encode characteristics of the segmentation as a whole, once we have an initial segmentation. We show that both the pairing and global features are useful on their own, and their combination achieved an $F_1$ of 92.6% of identifying in-sentence discourse boundaries, which is a 17.8% error-rate reduction over the state-of-the-art performance, approaching 95% of human performance. In addition, similar improvement is observed across different classification frameworks.

Saif M. Mohammad, Bonnie J. Dorr, Graeme Hirst et al.

Knowing the degree of semantic contrast between words has widespread application in natural language processing, including machine translation, information retrieval, and dialogue systems. Manually-created lexicons focus on opposites, such as {\rm hot} and {\rm cold}. Opposites are of many kinds such as antipodals, complementaries, and gradable. However, existing lexicons often do not classify opposites into the different kinds. They also do not explicitly list word pairs that are not opposites but yet have some degree of contrast in meaning, such as {\rm warm} and {\rm cold} or {\rm tropical} and {\rm freezing}. We propose an automatic method to identify contrasting word pairs that is based on the hypothesis that if a pair of words, $A$ and $B$, are contrasting, then there is a pair of opposites, $C$ and $D$, such that $A$ and $C$ are strongly related and $B$ and $D$ are strongly related. (For example, there exists the pair of opposites {\rm hot} and {\rm cold} such that {\rm tropical} is related to {\rm hot,} and {\rm freezing} is related to {\rm cold}.) We will call this the contrast hypothesis. We begin with a large crowdsourcing experiment to determine the amount of human agreement on the concept of oppositeness and its different kinds. In the process, we flesh out key features of different kinds of opposites. We then present an automatic and empirical measure of lexical contrast that relies on the contrast hypothesis, corpus statistics, and the structure of a {\it Roget}-like thesaurus. We show that the proposed measure of lexical contrast obtains high precision and large coverage, outperforming existing methods.

Saif M Mohammad, Graeme Hirst

The automatic ranking of word pairs as per their semantic relatedness and ability to mimic human notions of semantic relatedness has widespread applications. Measures that rely on raw data (distributional measures) and those that use knowledge-rich ontologies both exist. Although extensive studies have been performed to compare ontological measures with human judgment, the distributional measures have primarily been evaluated by indirect means. This paper is a detailed study of some of the major distributional measures; it lists their respective merits and limitations. New measures that overcome these drawbacks, that are more in line with the human notions of semantic relatedness, are suggested. The paper concludes with an exhaustive comparison of the distributional and ontology-based measures. Along the way, significant research problems are identified. Work on these problems may lead to a better understanding of how semantic relatedness is to be measured.

Saif M. Mohammad, Graeme Hirst

The ability to mimic human notions of semantic distance has widespread applications. Some measures rely only on raw text (distributional measures) and some rely on knowledge sources such as WordNet. Although extensive studies have been performed to compare WordNet-based measures with human judgment, the use of distributional measures as proxies to estimate semantic distance has received little attention. Even though they have traditionally performed poorly when compared to WordNet-based measures, they lay claim to certain uniquely attractive features, such as their applicability in resource-poor languages and their ability to mimic both semantic similarity and semantic relatedness. Therefore, this paper presents a detailed study of distributional measures. Particular attention is paid to flesh out the strengths and limitations of both WordNet-based and distributional measures, and how distributional measures of distance can be brought more in line with human notions of semantic distance. We conclude with a brief discussion of recent work on hybrid measures.