Coreference Relations Research Articles

A graph propagation model with rich event structures for joint event relation extraction

The task of event relation extraction (ERE) aims to organize multiple events and their relations as a directed graph. However, existing ERE methods exhibit two limitations: (1) Events in a document are typically expressed with merely a trigger word or phrase neglecting the rich semantic structure of event arguments and roles. (2) Various event relations such as coreference, temporal, causal, and subevent relations are correlated with each other, while little work has explored the benefits of extracting all relations in a joint manner. In this work, we investigate an event-relation graph propagation network and a novel document-level semantically-enriched event representation for joint ERE. First, we address the lack of event argument annotations by extracting cross-sentence implicit arguments based on the AMR graph. Then triggers and argument roles are aggregated with a structure-aware encoder. Second, based on the rich event information, event relation interactions are incorporated with a subgraph propagation and aggregation mechanism. During training, we further develop a novel triadic contrastive loss to capture high-order event pair relationships. We conduct experiments based on the public MAVEN-ERE benchmark and the results show that our model achieves 60.7%, 37.4%, 32.9% F1-scores on temporal, causal and subevent relations, and 86.1% MUC-score on coreferences, outperforming the current joint models by a large margin. Further in-depth analysis shows the effectiveness of our model in capturing event-event dependencies in document context. The proposed model can be used for event graph construction and storyline understanding.

BBScore: A Brownian Bridge Based Metric for Assessing Text Coherence

Measuring the coherence of text is a vital aspect of evaluating the quality of written content. Recent advancements in neural coherence modeling have demonstrated their efficacy in capturing entity coreference and discourse relations, thereby enhancing coherence evaluation. However, many existing methods heavily depend on static embeddings or focus narrowly on nearby context, constraining their capacity to measure the overarching coherence of long texts. In this paper, we posit that coherent texts inherently manifest a sequential and cohesive interplay among sentences, effectively conveying the central theme, purpose, or standpoint. To explore this abstract relationship, we introduce the "BB Score," a novel reference-free metric grounded in Brownian bridge theory for assessing text coherence. Our findings showcase that when synergized with a simple additional classification component, this metric attains a performance level comparable to state-of-the-art techniques on standard artificial discrimination tasks. We also establish in downstream tasks that this metric effectively differentiates between human-written documents and text generated by large language models within specific domains. Furthermore, we illustrate the efficacy of this approach in detecting written styles attributed to various large language models, underscoring its potential for generalizability. In summary, we present a novel Brownian bridge coherence metric capable of measuring both local and global text coherence, while circumventing the need for end-to-end model training. This flexibility allows for its application in various downstream tasks.

No Head Left Behind – Multi-Head Alignment Distillation for Transformers

Knowledge distillation aims at reducing model size without compromising much performance. Recent work has applied it to large vision-language (VL) Transformers, and has shown that attention maps in the multi-head attention modules of vision-language Transformers contain extensive intra-modal and cross-modal co-reference relations to be distilled. The standard approach is to apply a one-to-one attention map distillation loss, i.e. the Teacher's first attention head instructs the Student's first head, the second teaches the second, and so forth, but this only works when the numbers of attention heads in the Teacher and Student are the same. To remove this constraint, we propose a new Attention Map Alignment Distillation (AMAD) method for Transformers with multi-head attention, which works for a Teacher and a Student with different numbers of attention heads. Specifically, we soft-align different heads in Teacher and Student attention maps using a cosine similarity weighting. The Teacher head contributes more to the Student heads for which it has a higher similarity weight. Each Teacher head contributes to all the Student heads by minimizing the divergence between the attention activation distributions for the soft-aligned heads. No head is left behind. This distillation approach operates like cross-attention. We experiment on distilling VL-T5 and BLIP, and apply AMAD loss on their T5, BERT, and ViT sub-modules. We show, under vision-language setting, that AMAD outperforms conventional distillation methods on VQA-2.0, COCO captioning, and Multi30K translation datasets. We further show that even without VL pre-training, the distilled VL-T5 models outperform corresponding VL pre-trained VL-T5 models that are further fine-tuned by ground-truth signals, and that fine-tuning distillation can also compensate to some degree for the absence of VL pre-training for BLIP models.

Online reflexive resolution and interference

ABSTRACT A reflexive referentially depends on another NP. Long-standing inquiries in sentence processing literature revolve around whether and how forming this coreference relation during online sentence processing is susceptible to similarity-based interference. To address these questions, we conducted four experiments using L-maze tasks. The results revealed (delayed) interference in ungrammatical sentences, indicating facilitatory interference. We also found that structural information associated with a distractor and related to clause finiteness may influence memory retrieval. Interference effects were not observed in grammatical sentences across all experiments. These results suggest that online reflexive resolution may be susceptible to interference only when the reflexive mismatches its structurally accessible antecedent in features, and that structural information may influence how a distractor interferes with memory retrieval during online sentence processing.

The lexical semantics of finite control: A view from Japanese

In this paper, we propose a semantic analysis of control verbs in Japanese that take finite clauses marked by the nominalizer koto. We argue for an analysis in which the invisible subject of the embedded clause is a run-of-the-mill zero pronoun and where the obligatory coreference relation between the controller and the embedded subject is mediated by a primarily semantic factor. At the heart of our analysis lies the idea that there is a common underlying meaning shared across apparently heterogeneous classes of koto-taking control verbs, which consists of a causal relation between a volitional action inherent in the meaning of the verb and a de se attitude denoted by the embedded clause. The semantic analysis we offer not only explains language-internal properties of Japanese control verbs with respect to tense morpheme distribution that have been attributed to syntactic factors in the previous literature, but it also suggests a hitherto unnoticed possible cross-linguistic generalization about finite control and embedded tense interpretation which we dub ‘hypothesis of relative tense in finite control.’

Code Book for the Annotation of Diverse Cross-Document Coreference of Entities in News Articles

This paper presents a scheme for annotating coreference across news articles, extending beyond traditional identity relations by also considering near-identity and bridging relations. It includes a precise description of how to set up Inception, a respective annotation tool, how to annotate entities in news articles, connect them with diverse coreferential relations, and link them across documents to Wiki data’s global knowledge graph. This multi-layered annotation approach is discussed in the context of the problem of media bias. Our main contribution lies in providing a methodology for creating a diverse cross-document coreference corpus which can be applied to the analysis of media bias by word-choice and labelling

Chinese EFL learners different from English natives in cataphora resolution: Evidence from eye-tracking studies.

Previous studies on English natives have shown that encountering an English cataphoric pronoun triggers an active search for its antecedent and this searching process is modulated by syntactic constraints. It remains unknown whether the conclusion is universal to EFL (English as a Foreign Language) learners, particularly those with distinct L1 like Chinese in linguistic typology. Therefore, this study used two eye-tracking experiments to investigate how Chinese EFL learners resolve English cataphora. The experiments adopted the gender-mismatch paradigm. Experiment 1 investigated whether Chinese EFL learners with different proficiency would adopt the similar processing pattern to English natives and found that gender congruency elicited longer reading times than gender incongruency between the first potential antecedent and the cataphoric pronoun, the effect early observed in high-proficiency relative to low-proficiency learners. Experiment 2 explored whether the cataphora resolution process was modulated by Binding Principle B and revealed that longer first fixation durations and first pass reading times were observed in gender-mismatch than in gender-match conditions no matter the antecedents are binding-accessible or not while longer regression path durations occurred in gender-mismatch than in gender-match conditions only as the antecedents are binding-accessible. Taken together, these results indicate that Chinese EFL learners also adopt an active search mechanism to resolve cataphoric pronouns, yet along a processing path distinct from English natives'. Specifically, Chinese EFL learners predictively link a cataphoric pronoun to the first potential antecedent in the sentence but only a gender-matching antecedent can prompt them to engage in deep processing of the antecedent. Moreover, the processing time varies with the learners' English proficiency. Furthermore, unlike native English speakers' early application of syntactic constraints in their cataphora resolution, Chinese EFL learners try to establish co-reference relations between cataphoric pronouns and antecedents regardless of following or flouting Binding Principle B in early processing stages whereas they exclusively link the cataphoric pronouns to the binding-accessible antecedents in late processing stages. This study adds evidence to the Shallow Structure Hypothesis whereby L2 learners resort to lexical prior to syntactic cues to process sentences in general, which is just opposite to the fashion adopted by the natives.

A deep neural network model for coreference resolution in geological domain

Coreference resolution of geological entities is an important task in geological information mining. Although the existing generic coreference resolution models can handle geological texts, a dramatic decline in their performance can occur without sufficient domain knowledge. Due to the high diversity of geological terminology, coreference is intricately governed by the semantic and expressive structure of geological terms. In this paper, a framework CorefRoCNN based on RoBERTa and convolutional neural network (CNN) for end-to-end coreference resolution of geological entities is proposed. Firstly, the fine-tuned RoBERTa language model is used to transform words into dynamic vector representations with contextual semantic information. Second, a CNN-based multi-scale structure feature extraction module for geological terms is designed to capture the invariance of geological terms in length, internal structure, and distribution. Thirdly, we incorporate the structural feature and word embedding for further determinations of coreference relations. In addition, attention mechanisms are used to improve the ability of the model to capture valid information in geological texts with long sentence lengths. To validate the effectiveness of the model, we compared it with several state-of-the-art models on the constructed dataset. The results show that our model has the optimal performance with an average F1 value of 79.78%, which is a 1.22% improvement compared to the second-ranked method.

Information extraction pipelines for knowledge graphs

In the last decade, a large number of knowledge graph (KG) completion approaches were proposed. Albeit effective, these efforts are disjoint, and their collective strengths and weaknesses in effective KG completion have not been studied in the literature. We extend Plumber, a framework that brings together the research community’s disjoint efforts on KG completion. We include more components into the architecture of Plumber to comprise 40 reusable components for various KG completion subtasks, such as coreference resolution, entity linking, and relation extraction. Using these components, Plumber dynamically generates suitable knowledge extraction pipelines and offers overall 432 distinct pipelines. We study the optimization problem of choosing optimal pipelines based on input sentences. To do so, we train a transformer-based classification model that extracts contextual embeddings from the input and finds an appropriate pipeline. We study the efficacy of Plumber for extracting the KG triples using standard datasets over three KGs: DBpedia, Wikidata, and Open Research Knowledge Graph. Our results demonstrate the effectiveness of Plumber in dynamically generating KG completion pipelines, outperforming all baselines agnostic of the underlying KG. Furthermore, we provide an analysis of collective failure cases, study the similarities and synergies among integrated components and discuss their limitations.

The Text Anonymization Benchmark (TAB): A Dedicated Corpus and Evaluation Framework for Text Anonymization

Abstract We present a novel benchmark and associated evaluation metrics for assessing the performance of text anonymization methods. Text anonymization, defined as the task of editing a text document to prevent the disclosure of personal information, currently suffers from a shortage of privacy-oriented annotated text resources, making it difficult to properly evaluate the level of privacy protection offered by various anonymization methods. This paper presents TAB (Text Anonymization Benchmark), a new, open-source annotated corpus developed to address this shortage. The corpus comprises 1,268 English-language court cases from the European Court of Human Rights (ECHR) enriched with comprehensive annotations about the personal information appearing in each document, including their semantic category, identifier type, confidential attributes, and co-reference relations. Compared with previous work, the TAB corpus is designed to go beyond traditional de-identification (which is limited to the detection of predefined semantic categories), and explicitly marks which text spans ought to be masked in order to conceal the identity of the person to be protected. Along with presenting the corpus and its annotation layers, we also propose a set of evaluation metrics that are specifically tailored toward measuring the performance of text anonymization, both in terms of privacy protection and utility preservation. We illustrate the use of the benchmark and the proposed metrics by assessing the empirical performance of several baseline text anonymization models. The full corpus along with its privacy-oriented annotation guidelines, evaluation scripts, and baseline models are available on: https://github.com/NorskRegnesentral/text-anonymization-benchmark.

Multi-aware coreference relation network for visual dialog

Multi-aware coreference relation network for visual dialog

Hard-style Selective Context Utilization for dialogue generation based on what user just said

Dialogue is a process of information exchanging, where global background is stable while local focuses are transiting. Thus, at the ongoing dialogue turn, there are both relevant and irrelevant semantics existing in dialogue contexts. How to filter out noises and selectively utilize context can pave the way to successful dialogue generation. Current work on dialogue context utilization either processes contexts as vanilla monologue text ignoring dynamic conversation flows, or depends on weighted strategies to fuse all contexts where irrelevant utterances cannot be filter out even may overwhelm relevant ones. To deal with this, this paper proposes a Hard-style Selective Context Utilization method (HardSCU). We first define and measure the information density of the last utterance (query) of a dialogue, marking it as “strong” or “weak”. For a dialogue with strong query, HardSCU directly inputs the query into a RNN-based or T5-based encoder–decoder framework to generate a response; for a dialogue with weak query, HardSCU conducts a selective context utilization for dialogue generation, where a semantic interaction module introduces relevant semantics of context to enrich the query and the co-reference relations existing in dialogue are extracted to promote the learning process of response decoder. Extensive experiments on two benchmark conversation corpora verify that our HardSCU method can outperform competitive baselines on generating appropriate responses for chit-chat-bots with yielding strong robustness to the variations of dialogue lengths.

CDCAT: A multi-language cross-document entity and event coreference annotation tool

A tool for the manual annotation of cross-document entity and event coreferences that helps annotators to label mention coreference relations in text is essential for the annotation of coreference corpora. To the best of our knowledge, CROss-document Main Events and entities Recognition (CROMER) is the only open-source manual annotation tool available for cross-document entity and event coreferences. However, CROMER lacks multi-language support and extensibility. Moreover, to label cross-document mention coreference relations, CROMER requires the support of another intra-document coreference annotation tool known as Content Annotation Tool, which is now unavailable. To address these problems, we introduce Cross-Document Coreference Annotation Tool (CDCAT), a new multi-language open-source manual annotation tool for cross-document entity and event coreference, which can handle different input/output formats, preprocessing functions, languages, and annotation systems. Using this new tool, annotators can label a reference relation with only two mouse clicks. Best practice analyses reveal that annotators can reach an annotation speed of 0.025 coreference relations per second on a corpus with a coreference density of 0.076 coreference relations per word. As the first multi-language open-source cross-document entity and event coreference annotation tool, CDCAT can theoretically achieve higher annotation efficiency than CROMER.

Analysis of the Full-Size Russian Corpus of Internet Drug Reviews with Complex NER Labeling Using Deep Learning Neural Networks and Language Models

The paper presents the full-size Russian corpus of Internet users’ reviews on medicines with complex named entity recognition (NER) labeling of pharmaceutically relevant entities. We evaluate the accuracy levels reached on this corpus by a set of advanced deep learning neural networks for extracting mentions of these entities. The corpus markup includes mentions of the following entities: medication (33,005 mentions), adverse drug reaction (1778), disease (17,403), and note (4490). Two of them—medication and disease—include a set of attributes. A part of the corpus has a coreference annotation with 1560 coreference chains in 300 documents. A multi-label model based on a language model and a set of features has been developed for recognizing entities of the presented corpus. We analyze how the choice of different model components affects the entity recognition accuracy. Those components include methods for vector representation of words, types of language models pre-trained for the Russian language, ways of text normalization, and other pre-processing methods. The sufficient size of our corpus allows us to study the effects of particularities of annotation and entity balancing. We compare our corpus to existing ones by the occurrences of entities of different types and show that balancing the corpus by the number of texts with and without adverse drug event (ADR) mentions improves the ADR recognition accuracy with no notable decline in the accuracy of detecting entities of other types. As a result, the state of the art for the pharmacological entity extraction task for the Russian language is established on a full-size labeled corpus. For the ADR entity type, the accuracy achieved is 61.1% by the F1-exact metric, which is on par with the accuracy level for other language corpora with similar characteristics and ADR representativeness. The accuracy of the coreference relation extraction evaluated on our corpus is 71%, which is higher than the results achieved on the other Russian-language corpora.

КОРЕФЕРЕНТНІСТЬ У ПОЕТИЧНИХ ТВОРАХ ЛЕСІ УКРАЇНКИ

The category of coreference plays an important role in the creation of texts, as it is based on fundamental principles of text organization. The linearity of a text and the non-linearity of a situation described in the text are closely related to the mention of elements of a situation with another nomination of this object. Therefore, chains of coreferential names appear in the text and, having a binary relation, partly cause the emergence of a secondary nomination, which is an anaphoric relation with the primary nomination. The purpose of this article is to study the category of coreference in the poetic speech of Lesia Ukrainka based on the corpus of the Ukrainian language represented on mova.info. The object of the research is the idiostyle of Lesia Ukrainka’s poetry, and the subject is the functioning of the category of coreference. The categorical essence of coreference is characterized, the types of representatives of real objects, their functional and stylistic varieties and typical objects of representation in a poetic text are determined, and the most frequent structures are considered. Software for working with coreference has been developed and implemented. It has a user-friendly interface, which allows searching, sort, and doing the quantitative processing of the collected information according to the needs of researchers. On the material of 153 texts, 1520 referentially identical pairs are established. The coreferentiality index and the index of coreference coverage of the text proposed in this article helped to quantitatively assess the saturation of the text with the category of coreference. In addition, the classification of coreferential relations is composed: 1) identity; 2) collection gap; 3) part-whole; 4) predicative identity; 5) metonymy; 6) association. A dependency tree of a sentence made it possible to determine the location of coreferential pairs in the sentence, the position relative to each other, the belonging of the segment with the coreferent or the referent to the complicators of the sentence, i.e., a subordinate clause or phrases. This representation also helped to establish the type of syntactic relationship between the coreferent and the referent, the type of syntactic-semantic relationship between them, the presence of referent-dependent words, the intersection of the coreferent and the referent in gender and number. Automatic extraction of coreferential pairs from dependency trees allowed building intersentential coreference.

Processing of coreferential relations and their relationship with the working memory: the influence of syntactic distance on anaphora processing in Brazilian Portuguese

The aim of the present work was to investigate the processing of coreferential relations, focusing on their relationship with the working memory. In a reading process, it is essential that readers continuously perform mental operations that involve the working memory, such as storing, retrieving, and manipulating information. For this reason, it is understood that the processing of coreference resolution is closely related to the working memory. In this relationship between memory and anaphora resolution, the distance factor between the referent and anaphor can play a significant role as distance increases the processing costs when reading sentences. It is known that the working memory is limited; yet there is still no consensus among researchers regarding the nature of storable items. Moreover, the syntactic node, here taken as a representative of the structure of a phrase, was observed as possible storage and recoverable unit in the working memory. Thus, the objective of the study was to verify whether the syntactic distance factor, embeddings with 1 and 2 intervening syntactic nodes, is a storable unit in the working memory, investigating its influence in the processing of different anaphora resolution. The self-paced reading experimental technique was used in an experiment. Seventy participants, all university students, were tested individually and underwent a training session before the task. The results suggest that the syntactic distance is a relevant factor in the working memory with significant differences between conditions with syntactic distance concerning those with linear distance, indicating the formation of chunks. There was an increase in reading times depending on the number of intervening embeddings, and different processing costs for each type of anaphora resolution investigated . The aim of the present work was to investigate the processing of coreferential relations, focusing on their relationship with the working memory. In a reading process, it is essential that readers continuously perform mental operations that involve the working memory, such as storing, retrieving, and manipulating information. For this reason, it is understood that the processing of coreference resolution is closely related to the working memory. In this relationship between memory and anaphora resolution, the distance factor between the referent and anaphor can play a significant role as distance increases the processing costs when reading sentences. It is known that the working memory is limited; yet there is still no consensus among researchers regarding the nature of storable items. Moreover, the syntactic node, here taken as a representative of the structure of a phrase, was observed as possible storage and recoverable unit in the working memory. Thus, the objective of the study was to verify whether the syntactic distance factor, embeddings with 1 and 2 intervening syntactic nodes, is a storable unit in the working memory, investigating its influence in the processing of different anaphora resolution. The self-paced reading experimental technique was used in an experiment. Seventy participants, all university students, were tested individually and underwent a training session before the task. The results suggest that the syntactic distance is a relevant factor in the working memory with significant differences between conditions with syntactic distance concerning those with linear distance, indicating the formation of chunks. There was an increase in reading times depending on the number of intervening embeddings, and different processing costs for each type of anaphora resolution investigated .

Heterogeneous Graph Based Extractive Summarization Considering Discourse and Coreference Relations

Modeling the relations between text spans in a document is a crucial yet challenging problem for extractive summarization. Various kinds of relations exist among text spans of different granularity, such as discourse relations between elementary discourse units and coreference relations between phrase mentions. In this paper, we utilize heterogeneous graphs that contain multiple edge/node types to model the input document as well as the various relations among text spans in it. Also, we propose a heterogeneous graph based model for extractive summarization that considers the heterogeneity of the document graph. Experimental results on a benchmark summarization dataset verify the effectiveness of our proposed method.

Building Graph for Events and Time in Natural Language Text

Events and time are two major key terms in natural language processing due to the various event-oriented tasks these are become an essential terms in information extraction. In natural language processing and information extraction or retrieval event and time leads to several applications like text summaries, documents summaries, and question answering systems. In this paper, we present events-time graph as a new way of construction for event-time based information from text. In this event-time graph nodes are events, whereas edges represent the temporal and co-reference relations between events. In many of the previous researches of natural language processing mainly individually focused on extraction tasks and in domain-specific way but in this work we present extraction and representation of the relationship between events- time by representing with event time graph construction. Our overall system construction is in three-step process that performs event extraction, time extraction, and representing relation extraction. Each step is at a performance level comparable with the state of the art. We present Event extraction on MUC data corpus annotated with events mentions on which we train and evaluate our model. Next, we present time extraction the model of times tested for several news articles from Wikipedia corpus. Next is to represent event time relation by representation by next constructing event time graphs. Finally, we evaluate the overall quality of event graphs with the evaluation metrics and conclude the observations of the entire work.

A coesão referencial em textos traduzidos do inglês para o português: a questão dos sujeitos nulos

Understanding and writing a text arise from several factors. Among them, coreference relations, which allow for the identification of the referents of linguistic expressions, are of particular importance, since they ensure referential cohesion and, thus, contribute to text cohesion. One context in which recognizing referential relations is crucial involves reference chains with anaphoric pronouns. Although all languages exhibit coreference relations, the linguistic means to establish them may vary. Thus, in the context of translation, maintaining reference chains often implies the use of adaptation strategies: when working with non-null subject source languages, such as English, and null subject target languages, such as Portuguese, translators should recognize the contexts in which, in the latter, they may or must omit the subject, without affecting interpretation nor creating vagueness, conflict of interpretations or referential ambiguity. In this work, some data about reference chains in the context of English to Portuguese translation is discussed. We identify the main mismatches in translations done by university students and present clues to help train students in translation, promoting their mastering of the grammatical and textual conditions that determine the omission vs. the realization of the subject. The results show that the main problems result either from calque of the source language properties or from the overgeneralization of null subject contexts in the target language.

Anaphoric Connectives and Long-Distance Discourse Relations in Czech

This paper is a linguistic as well as technical survey for the development of a shallow discourse parser for Czech. It focuses on long-distance discourse relations signalled by (mostly) anaphoric discourse connectives. Proceeding from the division of connectives on “structural” and “anaphoric” according to their (in)ability to accept distant (non-adjacent) text segments as their left-sided arguments, and taking into account results of related analyses on English data in the framework of the Penn Discourse Treebank, we analyze a large amount of language data in Czech. We benefit from the multilayer manual annotation of various language aspects from morphology to discourse, coreference and bridging relations in the Prague Dependency Treebank 3.0. We describe the linguistic parameters of long-distance discourse relations in Czechin connection with their anchoring connective, and suggest possible ways of their detection. Our empirical research also outlines some theoretical consequences for the underlying assumptions in discourse analysis and parsing, e.g. the risk of relying too much on different (language-specific?) part-of-speech categorizations of connectives or the different perspectives in shallow and global discourse analyses (the minimality principle vs. higher text structure).