Motifs are distinctive, recurring, widely used idiom-like words or phrases, often originating in folklore and usually strongly anchored to a particular cultural or national group. Motifs are significant communicative devices across a wide range of media—including news, literature, and propaganda—because they can concisely imply a large set of culturally relevant associations. One difficulty of understanding motifs is that their meaning is usually implicit, so for an out-group person the meaning is inaccessible. We present the Motif Implicit Meaning Extractor (MIME), a proof-of-concept system designed to automatically identify a motif’s implicit meaning, as evidenced by textual uses of the motif across a large set data. MIME uses several sources (including motif indices, Wikipedia pages on the motifs, explicit explanations of motifs from in-group informants, and news/social media posts where the motif is used) and can generate a structured report of information about a motif understandable to an out-group person. In addition to a variety of examples and information drawn from structured sources, the report includes implicit information about a motif such as the type of reference (e.g., a person, an organization, etc.), it’s general connotation (strongly negative, slightly negative, neutral, etc.), and it’s associations (typically adjectives). We describe how MIME works and demonstrate its operation on a small set of manually curated motifs. We perform a qualitative evaluation of the output, and assess the difficulty of the problem, showing that explicit motif information provided by cultural informants is critical to high quality output, although mining motif usages in news and social media provides useful additional depth. A system such as MIME, appropriately scaled up, would potentially be quite useful to an out-group person trying to understand in-group usages of motifs, and has wide potential applications in domains such as literary criticism, cultural heritage, marketed and branding, and intelligence analysis.

pyTLEX is an implementation of the TimeLine EXtraction algorithm (TLEX; Finlayson et al.,2021) that enables users to work with TimeML annotations and perform advanced temporal analysis, offering a comprehensive suite of features. TimeML is a standardized markup language for temporal information in text. pyTLEX allows users to parse TimeML annotations, construct TimeML graphs, and execute the TLEX algorithm to effect complete timeline extraction. In contrast to previous implementations (i.e., jTLEX for Java), pyTLEX sets itself apart with a range of advanced features. It introduces a React-based visualization system, enhancing the exploration of temporal data and the comprehension of temporal connections within textual information. Furthermore, pyTLEX incorporates an algorithm for increasing connectivity in temporal graphs, which identifies graph disconnectivity and recommends links based on temporal reasoning, thus enhancing the coherence of the graph representation. Additionally, pyTLEX includes a built-in validation algorithm, ensuring compliance with TimeML annotation guidelines, which is essential for maintaining data quality and reliability. pyTLEX equips researchers and developers with an extensive toolkit for temporal analysis, and its testing across various datasets validates its accuracy and reliability.

Motifs are distinctive, recurring, widely used idiom-like words or phrases, often originating from folklore, whose meaning are anchored in a narrative. Motifs have significance as communicative devices because they concisely imply a constellation of culturally relevant information. Their broad usage suggests their cognitive importance as touchstones of cultural knowledge. We present GOLEM, the first dataset annotated for motific information. The dataset comprises 7,955 English articles (2,039,424 words). The corpus identifies 26,078 motif candidates across 34 motif types from three cultural or national groups: Jewish, Irish, and Puerto Rican. Each motif candidate is labeled with the type of usage (Motific, Referential, Eponymic, or Unrelated), resulting in 1,723 actual motific instances. Annotation was performed by individuals identifying as members of each group and achieved a Fleiss’ kappa of >0.55. We demonstrate that classification of candidate type is a challenging task for LLMs using a few-shot approach; recent models such as T5, FLAN-T5, GPT-2, and Llama 2 (7B) achieved a performance of 41% accuracy at best. These data will support development of new models and approaches for detecting (and reasoning about) motific information in text. We release the corpus, the annotation guide, and the code to support other researchers building on this work.

jTLEX is a programming library that provides a Java implementation of the TimeLine EXtraction algorithm (TLEX; Finlayson et al.,2021), along with utilities for programmatic manipulation of TimeML graphs. Timelines are useful for a number of natural language understanding tasks, such as question answering, cross-document event coreference, and summarization & visualization. jTLEX provides functionality for (1) parsing TimeML annotations into Java objects, (2) construction of TimeML graphs from scratch, (3) partitioning of TimeML graphs into temporally connected subgraphs, (4) transforming temporally connected subgraphs into point algebra (PA) graphs, (5) extracting exact timeline of TimeML graphs, (6) detecting inconsistent subgraphs, and (7) calculating indeterminate sections of the timeline. The library has been tested on the entire TimeBank corpus, and comes with a suite of unit tests. We release the software as open source with a free license for non-commercial use.

TimeML is a scheme for representing temporal information (times, events, & temporal relations) in texts. Although automatic TimeML annotation is challenging, there has been notable progress, with F1s of 0.8–0.9 for events and time detection subtasks, and F1s of 0.5–0.7 for relation extraction. Individually, these subtask results are reasonable, even good, but when combined to generate a full TimeML graph, is overall performance still acceptable? We present a novel suite of eight metrics, combined with a new graph-transformation experimental design, for holistic evaluation of TimeML graphs. We apply these metrics to four automatic TimeML annotation systems (CAEVO, TARSQI, CATENA, and ClearTK). We show that on average 1/3 of the TimeML graphs produced using these systems are inconsistent, and there is on average 1/5 more temporal indeterminacy than the gold-standard. We also show that the automatically generated graphs are on average 109 edits from the gold-standard, which is 1/3 toward complete replacement. Finally, we show that the relationship individual subtask performance and graph quality is non-linear: small errors in TimeML subtasks result in rapid degradation of final graph quality. These results suggest current automatic TimeML annotators are far from optimal and significant further improvement would be useful.

TimeML is an annotation scheme for capturing temporal information in text. The developers of TimeML built the TimeBank corpus to both validate the scheme and provide a rich dataset of events, temporal expressions, and temporal relationships for training and testing temporal analysis systems. In our own work we have been developing methods aimed at TimeML graphs for detecting (and eventually automatically correcting) temporal inconsistencies, extracting timelines, and assessing temporal indeterminacy. In the course of this investigation we identified numerous previously unrecognized issues in the TimeBank corpus, including multiple violations of TimeML annotation guide rules, incorrectly disconnected temporal graphs, as well as inconsistent, redundant, missing, or otherwise incorrect annotations. We describe our methods for detecting and correcting these problems, which include: (a) automatic guideline checking (109 violations); (b) automatic inconsistency checking (65 inconsistent files); (c) automatic disconnectivity checking (625 incorrect breakpoints); and (d) manual comparison with the output of state-of-the-art automatic annotators to identify missing annotations (317 events, 52 temporal expressions). We provide our code as well as a set of patch files that can be applied to the TimeBank corpus to produce a corrected version for use by other researchers in the field.

There have been several attempts to create an accurate and thorough emotion lexicon in English, which identifies the emotional content of words. Of the several commonly used resources, the NRC emotion lexicon (Mohammad and Turney, 2013b) has received the most attention due to its availability, size, and its choice of Plutchik’s expressive 8-class emotion model. In this paper we identify a large number of troubling entries in the NRC lexicon, where words that should in most contexts be emotionally neutral, with no affect (e.g., ‘lesbian’, ‘stone’, ‘mountain’), are associated with emotional labels that are inaccurate, nonsensical, pejorative, or, at best, highly contingent and context-dependent (e.g., ‘lesbian’ labeled as Disgust and Sadness, ‘stone’ as Anger, or ‘mountain’ as Anticipation). We describe a procedure for semi-automatically correcting these problems in the NRC, which includes disambiguating POS categories and aligning NRC entries with other emotion lexicons to infer the accuracy of labels. We demonstrate via an experimental benchmark that the quality of the resources is thus improved. We release the revised resource and our code to enable other researchers to reproduce and build upon results.

Stereotypical character roles-also known as archetypes or dramatis personae-play an important function in narratives: they facilitate efficient communication with bundles of default characteristics and associations and ease understanding of those characters’ roles in the overall narrative. We present a fully unsupervised k-means clustering approach for learning stereotypical roles given only structural plot information. We demonstrate the technique on Vladimir Propp’s structural theory of Russian folktales (captured in the extended ProppLearner corpus, with 46 tales), showing that our approach can induce six out of seven of Propp’s dramatis personae with F1 measures of up to 0.70 (0.58 average), with an additional category for minor characters. We have explored various feature sets and variations of a cluster evaluation method. The best-performing feature set comprises plot functions, unigrams, tf-idf weights, and embeddings over coreference chain heads. Roles that are mentioned more often (Hero, Villain), or have clearly distinct plot patterns (Princess) are more strongly differentiated than less frequent or distinct roles (Dispatcher, Helper, Donor). Detailed error analysis suggests that the quality of the coreference chain and plot functions annotations are critical for this task. We provide all our data and code for reproducibility.

Determining whether an event in a news article is a foreground or background event would be useful in many natural language processing tasks, for example, temporal relation extraction, summarization, or storyline generation. We introduce the task of distinguishing between foreground and background events in news articles as well as identifying the general temporal position of background events relative to the foreground period (past, present, future, and their combinations). We achieve good performance (0.73 F1 for background vs. foreground and temporal position, and 0.79 F1 for background vs. foreground only) on a dataset of news articles by leveraging discourse information in a featurized model. We release our implementation and annotated data for other researchers

One of the most fundamental elements of narrative is character: if we are to understand a narrative, we must be able to identify the characters of that narrative. Therefore, character identification is a critical task in narrative natural language understanding. Most prior work has lacked a narratologically grounded definition of character, instead relying on simplified or implicit definitions that do not capture essential distinctions between characters and other referents in narratives. In prior work we proposed a preliminary definition of character that was based in clear narratological principles: a character is an animate entity that is important to the plot. Here we flesh out this concept, demonstrate that it can be reliably annotated (0.78 Cohen’s κ), and provide annotations of 170 narrative texts, drawn from 3 different corpora, containing 1,347 character co-reference chains and 21,999 non-character chains that include 3,937 animate chains. Furthermore, we have shown that a supervised classifier using a simple set of easily computable features can effectively identify these characters (overall F1 of 0.90). A detailed error analysis shows that character identification is first and foremost affected by co-reference quality, and further, that the shorter a chain is the harder it is to effectively identify as a character. We release our code and data for the benefit of other researchers

Temporal Dependency Trees (TDTs) have emerged as an alternative to full temporal graphs for representing the temporal structure of texts, with a key advantage being that TDTs can be straightforwardly computed using adapted dependency parsers. Relative to temporal graphs, the tree form of TDTs naturally omits some fraction of temporal relationships, which intuitively should decrease the amount of temporal information available, potentially increasing temporal indeterminacy of the global ordering. We demonstrate a new method for quantifying this indeterminacy that relies on solving temporal constraint problems to extract timelines, and show that TDTs result in up to a 109% increase in temporal indeterminacy over their corresponding temporal graphs for the three corpora we examine. On average, the increase in indeterminacy is 32%, and we show that this increase is a result of the TDT representation eliminating on average only 2.4% of total temporal relations. This result suggests that small differences can have big effects in temporal graphs, and the use of TDTs must be balanced against their deficiencies, with tasks requiring an accurate global temporal ordering potentially calling for use of the full temporal graph

Cross-Document Event Coreference (CDEC) is the task of finding coreference relationships between events in separate documents, most commonly assessed using the Event Coreference Bank+ corpus (ECB+). At least two different approaches have been proposed for CDEC on ECB+ that use only event triggers, and at least four have been proposed that use both triggers and entities. Comparing these approaches is complicated by variation in the systems’ use of gold vs. computed labels, as well as variation in the document clustering pre-processing step. We present an approach that matches or slightly beats state-of-the-art performance on CDEC over ECB+ with only event trigger annotations, but with a significantly simpler framework and much smaller feature set relative to prior work. This study allows us to directly compare with prior systems and draw conclusions about the effectiveness of various strategies. Additionally, we provide the first cross-validated evaluation on the ECB+ dataset; the first explicit evaluation of the pairwise event coreference classification step; and the first quantification of the effect of document clustering on system performance. The last in particular reveals that while document clustering is a crucial pre-processing step, improvements can at most provide for a 3 point improvement in CDEC performance, though this might be attributable to ease of document clustering on ECB+.

Identifying the discourse structure of documents is an important task in understanding written text. Building on prior work, we demonstrate an improved approach to automatically identifying the discourse function of paragraphs in news articles. We start with the hierarchical theory of news discourse developed by van Dijk (1988) which proposes how paragraphs function within news articles. This discourse information is a level intermediate between phrase- or sentence-sized discourse segments and document genre, characterizing how individual paragraphs convey information about the events in the storyline of the article. Specifically, the theory categorizes the relationships between narrated events and (1) the overall storyline (such as Main Events, Background, or Consequences) as well as (2) commentary (such as Verbal Reactions and Evaluations). We trained and tested a linear chain conditional random field (CRF) with new features to model van Dijk’s labels and compared it against several machine learning models presented in previous work. Our model significantly outperformed all baselines and prior approaches, achieving an average of 0.71 F1 score which represents a 31.5% improvement over the previously best-performing support vector machine model.

Identifying emotions as expressed in text (a.k.a. text emotion recognition) has received a lot of attention over the past decade. Narratives often involve a great deal of emotional expression, and so emotion recognition on narrative text is of great interest to computational approaches to narrative understanding. Prior work by Kim et al. 2010 was the work with the highest reported emotion detection performance, on a corpus of fairy tales texts. Close inspection of that work, however, revealed significant reproducibility problems, and we were unable to reimplement Kim’s approach as described. As a consequence, we implemented a framework inspired by Kim’s approach, where we carefully evaluated the major design choices. We identify the highest-performing combination, which outperforms Kim’s reported performance by 7.6 F₁ points on average. Close inspection of the annotated data revealed numerous missing and incorrect emotion terms in the relevant lexicon, WordNetAffect (WNA; Strapparava and Valitutti, 2004), which allowed us to augment it in a useful way. More generally, this showed that numerous clearly emotive words and phrases are missing from WNA, which suggests that effort invested in augmenting or refining emotion ontologies could be useful for improving the performance of emotion recognition systems. We release our code and data to definitely enable future reproducibility of this work.

Recognizing the internal structure of events is a challenging language processing task of great importance for text understanding. We present a supervised model for automatically identifying when one event is a subevent of another. Building on prior work, we introduce several novel features, in particular discourse and narrative features, that significantly improve upon prior state-of-the-art performance. Error analysis further demonstrates the utility of these features. We evaluate our model on the only two annotated corpora with event hierarchies: HiEve and the Intelligence Community corpus. No prior system has been evaluated on both corpora. Our model outperforms previous systems on both corpora, achieving 0.74 BLANC F1 on the Intelligence Community corpus and 0.70 F1 on the HiEve corpus, respectively a 15 and 5 percentage point improvement over previous models.

Characters are a key element of narrative and so character identification plays an important role in automatic narrative understanding. Unfortunately, most prior work that incorporates character identification is not built upon a clear, theoretically grounded concept of character. They either take character identification for granted (e.g., using simple heuristics on referring expressions), or rely on simplified definitions that do not capture important distinctions between characters and other referents in the story. Prior approaches have also been rather complicated, relying, for example, on predefined case bases or ontologies. In this paper we propose a narratologically grounded definition of character for discussion at the workshop, and also demonstrate a preliminary yet straightforward supervised machine learning model with a small set of features that performs well on two corpora. The most important of the two corpora is a set of 46 Russian folktales, on which the model achieves an F1 of 0.81. Error analysis suggests that features relevant to the plot will be necessary for further improvements in performance.

Animacy is a necessary property for a referent to be an agent, and thus animacy detection is useful for a variety of natural language processing tasks, including word sense disambiguation, co-reference resolution, semantic role labeling, and others. Prior work treated animacy as a word-level property, and has developed statistical classifiers to classify words as either animate or inanimate. We discuss why this approach to the problem is ill-posed, and present a new approach based on classifying the animacy of co-reference chains. We show that simple voting approaches to inferring the animacy of a chain from its constituent words perform relatively poorly, and then present a hybrid system merging supervised machine learning (ML) and a small number of hand-built rules to compute the animacy of referring expressions and co-reference chains. This method achieves state of the art performance. The supervised ML component leverages features such as word embeddings over referring expressions, parts of speech, and grammatical and semantic roles. The rules take into consideration parts of speech and the hypernymy structure encoded in WordNet. The system achieves an F1 of 0.88 for classifying the animacy of referring expressions, which is comparable to state of the art results for classifying the animacy of words, and achieves an F1 of 0.75 for classifying the animacy of coreference chains themselves. We release our training and test dataset, which includes 142 texts (all narratives) comprising 156,154 words, 34,698 referring expressions, and 10,941 co-reference chains. We test the method on a subset of the OntoNotes dataset, showing using manual sampling that animacy classification is 90% +/- 2% accurate for coreference chains, and 92% +/- 1% for referring expressions. The data also contains 46 folktales, which present an interesting challenge because they often involve characters who are members of traditionally inanimate classes (e.g., stoves that walk, trees that talk). We show that our system is able to detect the animacy of these unusual referents with an F1 of 0.95.

Discourse structure is a key aspect of all forms of text, providing valuable information both to humans and machines. We applied the hierarchical theory of news discourse developed by van Dijk to examine how paragraphs operate as units of discourse structure within news articles—what we refer to here as document-level discourse. This document-level discourse provides a characterization of the content of each paragraph that describes its relation to the events presented in the article (such as main events, backgrounds, and consequences) as well as to other components of the story (such as commentary and evaluation). The purpose of a news discourse section is of great utility to story understanding as it affects both the importance and temporal order of items introduced in the text—therefore, if we know the news discourse purpose for different sections, we should be able to better rank events for their importance and better construct timelines. We test two hypotheses: first, that people can reliably annotate news articles with van Dijk’s theory; second, that we can reliably predict these labels using machine learning. We show that people have a high degree of agreement with each other when annotating the theory (F1 > 0.8, Cohen’s kappa > 0.6), demonstrating that it can be both learned and reliably applied by human annotators. Additionally, we demonstrate first steps toward machine learning of the theory, achieving a performance of F1 = 0.54, which is 65% of human performance. Moreover, we have generated a gold-standard, adjudicated corpus of 50 documents for document-level discourse annotation based on the ACE Phase 2 corpus.

Psychiatric evaluation reports represent a rich and still mostly-untapped source of information for developing systems for automatic diagnosis and treatment of mental health problems. These reports contain free-text structured within sections using a convention of headings. We present a model for automatically detecting the position and type of different psychiatric evaluation report sections. We developed this model using a corpus of 150 sample reports that we gathered from the Web, and used sentences as a processing unit while section headings were used as labels of section type. From these labels we generated a unified hierarchy of labels of section types, and then learned n-gram models of the language found in each section. To model conventions for section order, we integrated these n-gram models with a Hierarchical Hidden Markov Model (HHMM) representing the probabilities of observed section orders found in the corpus, and then used this HHMM n-gram model in a decoding framework to infer the most likely section boundaries and section types for documents with their section labels removed. We evaluated our model over two tasks, namely, identifying section boundaries and identifying section types and orders. Our model significantly outperformed baselines for each task with an F1 of 0.88 for identifying section types, and a 0.26 WindowDiff (Wd) and 0.20 and (Pk) scores, respectively, for identifying section boundaries.

Story detection is the task of determining whether or not a unit of text contains a story. Prior approaches achieved a maximum performance of 0.66 F1, and did not generalize well across different corpora. We present a new state-of-the-art detector that achieves a maximum performance of 0.75 F1 (a 14% improvement), with significantly greater generalizability than previous work. In particular, our detector achieves performance above 0.70 F1 across a variety of combinations of lexically different corpora for training and testing, as well as dramatic improvements (up to 4,000%) in performance when trained on a small, disfluent data set. The new detector uses two basic types of features–ones related to events, and ones related to characters–totaling 283 specific features overall; previous detectors used tens of thousands of features, and so this detector represents a significant simplification along with increased performance.

We describe the N2 (Narrative Networks) Corpus, a new language resource. The corpus is unique in three important ways. First, every text in the corpus is a story, which is in contrast to other language resources that may contain stories or story-like texts, but are not specifically curated to contain only stories. Second, the unifying theme of the corpus is material relevant to Islamist Extremists, having been produced by or often referenced by them. Third, every text in the corpus has been annotated for 14 layers of syntax and semantics, including: referring expressions and co-reference; events, time expressions, and temporal relationships; semantic roles; and word senses. In cases where analyzers were not available to do high-quality automatic annotations, layers were manually double-annotated and adjudicated by trained annotators. The corpus comprises 100 texts and 42,480 words. Most of the texts were originally in Arabic but all are provided in English translation. We explain the motivation for constructing the corpus, the process for selecting the texts, the detailed contents of the corpus itself, the rationale behind the choice of annotation layers, and the annotation procedure.