Cue Phrases Research Articles

Assessing linguistic generalisation in language models: a dataset for Brazilian Portuguese

Much recent effort has been devoted to creating large-scale language models. Nowadays, the most prominent approaches are based on deep neural networks, such as BERT. However, they lack transparency and interpretability, and are often seen as black boxes. This affects not only their applicability in downstream tasks but also the comparability of different architectures or even of the same model trained using different corpora or hyperparameters. In this paper, we propose a set of intrinsic evaluation tasks that inspect the linguistic information encoded in models developed for Brazilian Portuguese. These tasks are designed to evaluate how different language models generalise information related to grammatical structures and multiword expressions (MWEs), thus allowing for an assessment of whether the model has learned different linguistic phenomena. The dataset that was developed for these tasks is composed of a series of sentences with a single masked word and a cue phrase that helps in narrowing down the context. This dataset is divided into MWEs and grammatical structures, and the latter is subdivided into 6 tasks: impersonal verbs, subject agreement, verb agreement, nominal agreement, passive and connectors. The subset for MWEs was used to test BERTimbau Large, BERTimbau Base and mBERT. For the grammatical structures, we used only BERTimbau Large, because it yielded the best results in the MWE task. In both cases, we evaluated the results considering the best candidates and the top ten candidates. The evaluation was done both automatically (for MWEs) and manually (for grammatical structures). The results obtained for MWEs show that BERTimbau Large surpassed both the other models in predicting the correct masked element. However, the average accuracy of the best model was only 52% when only the best candidates were considered for each sentence, going up to 66% when the top ten candidates were taken into account. As for the grammatical tasks, the results presented better prediction, but also varied depending on the type of morphosyntactic agreement. On the one hand, cases such as connectors and impersonal verbs, which do not require any agreement in the produced candidates, had precision of 100% and 98.78% among the best candidates. On the other hand, tasks that require morphosyntactic agreement had results consistently below 90% overall precision, with the lowest scores being reported for nominal agreement and verb agreement, both having scores below 80% in overall precision among the best candidates. Therefore, we identified that a critical and widely adopted resource for Brazilian Portuguese NLP presents issues concerning MWE vocabulary and morphosyntactic agreement, even if it is prolific in most cases. These models are a core component in many NLP systems, and our findings demonstrate the need of additional improvements in these models and the importance of widely evaluating computational representations of language.

Signalling conditional relations

Abstract We investigate how discourse relations and their subtypes are signalled, extending the set of discourse signals from connectives and lexical cue phrases to the wide range of semantic, syntactic, and orthographic signals of the RST Signalling Corpus (Das, Debopam & Maite Taboada. 2018. RST signalling corpus. Language Resources and Evaluation 52. 149–184). This extension requires re-evaluating previous predictions on discourse signalling, in particular, those of Sanders, Ted. 2005. Coherence, causality and cognitive complexity in discourse. In M. Aurnague, M. Bras, A. Le Draoulec & L. Vieu (eds.), Proceedings/Actes SEM-05, first international symposium on the exploration and modelling of meaning, 105–114. Biarritz causality-by-default hypothesis, the hypothesis of uniform information density (Frank, Austin & Florian Jaeger. 2008. Speaking rationally: Uniform information density as an optimal strategy for language production. In Proceedings of the 30th annual meeting of the Cognitive Science Society, 933–938. https://escholarship.org/uc/item/7d08h6j4 (accessed 18 May 2022)), and the hypothesis that discourse is continuous by preference (Segal, Erwin, Judith Duchan & Paula Scott. 1991. The role of interclausal connectives in narrative structuring. Discourse Processes 14. 27–54; Murray, John. 1997. Connectives and narrative text. Memory and Cognition 25. 227–236). We evaluate the predictions of these theories on the conditional relations in the RST Discourse Treebank (Carlson, Lynn, Daniel Marcu & Mary Ellen Okurowski. 2002. RST Discourse Treebank. LDC2002T07. Philadelphia: Linguistic Data Consortium), using causal relations as a control group. Informativity and continuity are operationalized in terms of semantic complexity and Givón, Talmy. 1993. English grammar: A function-based introduction, vol. 2. Amsterdam: John Benjamins dimensions of deictic shift. Our results show that the hypotheses make accurate predictions only for the relation groups in their entirety but not for the observed in-group variation, in particular, the low amount of marking for the hypothetical subtype of conditional relations. We attribute this difference to the distribution of intra- and inter-sentential occurrences across the conditional subtypes: intra-sentential relations are consistently more marked than inter-sentential ones, and hypothetical relations are special in that they appear predominantly inter-sententially.

Investigating disagreement in the scientific literature.

Disagreement is essential to scientific progress but the extent of disagreement in science, its evolution over time, and the fields in which it happens remain poorly understood. Here we report the development of an approach based on cue phrases that can identify instances of disagreement in scientific articles. These instances are sentences in an article that cite other articles. Applying this approach to a collection of more than four million English-language articles published between 2000 and 2015 period, we determine the level of disagreement in five broad fields within the scientific literature (biomedical and health sciences; life and earth sciences; mathematics and computer science; physical sciences and engineering; and social sciences and humanities) and 817 meso-level fields. Overall, the level of disagreement is highest in the social sciences and humanities, and lowest in mathematics and computer science. However, there is considerable heterogeneity across the meso-level fields, revealing the importance of local disciplinary cultures and the epistemic characteristics of disagreement. Analysis at the level of individual articles reveals notable episodes of disagreement in science, and illustrates how methodological artifacts can confound analyses of scientific texts.

Phrase Embedding and Clustering for Sub-Feature Extraction From Online Data

Abstract Recently, online user-generated data have been used as an efficient resource for customer analysis. In the product design area, various methods for analyzing customer preference for product features have been suggested. However, most of them focused on feature categories rather than product components which are crucial in practical applications. To address that limitation, this paper proposes a new methodology for extracting sub-features from online data. First, the method detects phrases in the data and filtered them using product manual documents. The filtered phrases are embedded into vectors, and then they are divided into several groups by two clustering methods. The resulting clusters are labeled by analyzing items in each cluster. Finally, cue phrases for sub-features are obtained by selecting clusters with labels representing product features. The proposed methodology was tested on smartphone review data. The result provides feature clusters containing sub-feature phrases with high accuracy. The obtained cue phrases will be used in analyzing customer preferences for sub-features and this can help product designers determine the optimal component configuration in embodiment design.

Why-Type Question to Query Reformulation for Efficient Document Retrieval

Understanding the actual need of user from a question is very crucial in non-factoid why-question answering as Why-questions are complex and involve ambiguity and redundancy in their understanding. The precise requirement is to determine the focus of question and reformulate them accordingly to retrieve expected answers to a question. The paper analyzes different types of why-questions and proposes an algorithm for each class to determine the focus and reformulate it into a query by appending focal terms and cue phrase ‘because’ with it. Further, a user interface is implemented which asks input why-question, applies different components of question , reformulates it and finally retrieve web pages by posing query to Google search engine. To measure the accuracy of the process, user feedback is taken which asks them to assign scoring from 1 to 10, on how relevant are the retrieved web pages according to their understanding. The results depict that maximum precision of 89% is achieved in Informational type why-questions and minimum of 48% in opinionated type why-questions.

Predictions of Citations of a Scholarly Paper

Abstract: Bibliometrics is a statistical analysis of written publications such as books or articles. A bibliographic citationis a reference to a book, article, web page, or other published item. Thus citations are useful for identifying the progress ofthe particular work and measuring the quality of the research article. The cited papers are downloaded using the crawler. Fromthe downloaded article, identify article relation by analyzing the citation context of the article. So first extract the citation context from the article. Citation context are classifies based on cue phrases of Simon tufel. Next, identify the relation of unlabeled article by word embedding. After labeling all articles identifythe perspective behind the citation of the article. In this project, citation relation is identified based on cue phrases of Simon tufel finally article impact is quantified based on the citation network formed from citation analysis. Index Terms: bibliometrics, citation, word embedding, article

KGGCN: Knowledge-Guided Graph Convolutional Networks for Distantly Supervised Relation Extraction

Distantly supervised relation extraction is the most popular technique for identifying semantic relation between two entities. Most prior models only focus on the supervision information present in training sentences. In addition to training sentences, external lexical resource and knowledge graphs often contain other relevant prior knowledge. However, relation extraction models usually ignore such readily available information. Moreover, previous works only utilize a selective attention mechanism over sentences to alleviate the impact of noise, they lack the consideration of the implicit interaction between sentences with relation facts. In this paper, (1) a knowledge-guided graph convolutional network is proposed based on the word-level attention mechanism to encode the sentences. It can capture the key words and cue phrases to generate expressive sentence-level features by attending to the relation indicators obtained from the external lexical resource. (2) A knowledge-guided sentence selector is proposed, which explores the semantic and structural information of triples from knowledge graph as sentence-level knowledge attention to distinguish the importance of each individual sentence. Experimental results on two widely used datasets, NYT-FB and GDS, show that our approach is able to efficiently use the prior knowledge from the external lexical resource and knowledge graph to enhance the performance of distantly supervised relation extraction.

Creating a Disaster Chain Diagram from Japanese Newspaper Articles Using Mechanical Methods

A new method for creating a chain diagram of events that occur during disasters by extracting causal knowledge from Japanese newspaper articles and designing a causal network is proposed herein. Machine learning discriminant models were created for both conventional cue phrases and succession expressions with causation to extract causal sentences. We found that causal sentences can be extracted with a certain degree of accuracy from disaster articles. We were also able to create a causal network using sentences as nodes and links. The chain diagram using our new method extracted events and causal knowledge that were unavailable in a disaster chain diagram designed using conventional methods.

Signaling of Causal Relations in Spanish: Variety, Functionality, and Specificity

While some recent studies on Spanish have shown that some causal discourse markers specialize in expressing certain types of causal relations, others have revealed that causal relations may be signaled by a variety of linguistic devices. Given that we were interested not only in specificity and variety, but also in the (poly) functionality of signals, our objective in the present study was threefold. First, to identify the variety of markers used to signal causal relations in Spanish. Second, to describe the (poly) functionality of those causal markers. Third, to determine whether there exists a relationship of specificity between markers and particular types of causal relations. We analyzed a corpus of 2,514 causal coherence relations previously annotated. 40 different linguistic devices used to signal causal relations were identified. These devices were grouped into two main classes: Discourse Markers and Cue Phrases. Regarding the (poly) functionality of the markers, we found that 8 of the most frequent markers were used to signal different relations. Regarding specificity, it was observed that various conjunctions and conjunctive adverbs specialize in signaling specific relations.

Contributions of Voice Expectations to Talker Selection in Younger and Older Adults With Normal Hearing.

Focused attention on expected voice features, such as fundamental frequency (F0) and spectral envelope, may facilitate segregation and selection of a target talker in competing talker backgrounds. Age-related declines in attention may limit these abilities in older adults, resulting in poorer speech understanding in complex environments. To test this hypothesis, younger and older adults with normal hearing listened to sentences with a single competing talker. For most trials, listener attention was directed to the target by a cue phrase that matched the target talker’s F0 and spectral envelope. For a small percentage of randomly occurring probe trials, the target’s voice unexpectedly differed from the cue phrase in terms of F0 and spectral envelope. Overall, keyword recognition for the target talker was poorer for older adults than younger adults. Keyword recognition was poorer on probe trials than standard trials for both groups, and incorrect responses on probe trials contained keywords from the single-talker masker. No interaction was observed between age-group and the decline in keyword recognition on probe trials. Thus, reduced performance by older adults overall could not be attributed to declines in attention to an expected voice. Rather, other cognitive abilities, such as speed of processing and linguistic closure, were predictive of keyword recognition for younger and older adults. Moreover, the effects of age interacted with the sex of the target talker, such that older adults had greater difficulty understanding target keywords from female talkers than male talkers.

Evaluation of Content Compaction in Assamese Language

Abstract Text summarization is the task of condensing the input text documents into a shorter version by retaining the overall meaning and information of the original document. Though plenty of text compaction research works in English and other languages have been done till date, but text compaction in Assamese language is still lagging behind due to the limitations of resources for text compaction. This paper proposes a content compaction approach for Assamese Text to generate a summary by incorporating statistical and linguistic features and makes an attempt to extract the relevant points in the summary from the input Assamese document. The statistical features applied in this approach incorporate a new modified tf-idf, sentence positioning, length of sentences and identification of numericals. Whereas the linguistic features employed include identifying the nouns, cue words and cue phrases. To contrast the summaries that are generated both automatically and by humans, the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) framework is used. Furthermore, our proposed approach is also compared with text summarization approaches by using traditional tf-idf. It is observed that this approach has yielded impressive results while summarizing Assamese documents with average F-measure of 0.633 at the compression ratio of 50%.

An attention-based neural framework for uncertainty identification on social media texts

Uncertainty identification is an important semantic processing task. It is crucial to the quality of information in terms of factuality in many applications, such as topic detection and question answering. Factuality has become a premier concern especially in social media, in which texts are written informally. However, existing approaches that rely on lexical cues suffer greatly from the casual or word-of-mouth peculiarity of social media, in which the cue phrases are often expressed in substandard form or even omitted from sentences. To tackle these problems, this paper proposes an Attention-based Neural Framework for Uncertainty identification on social media texts, named ANFU. ANFU incorporates attention-based Long Short-Term Memory (LSTM) networks to represent the semantics of words and Convolutional Neural Networks (CNNs) to capture the most important semantics. Experiments were conducted on four datasets, including 2 English benchmark datasets used in the CoNLL-2010 task of uncertainty identification and 2 Chinese datasets of Weibo and Chinese news texts. Experimental results showed that our proposed ANFU approach outperformed the-state-of-the-art on all the datasets in terms of F1 measure. More importantly, 41.37% and 13.10% improvements were achieved over the baselines on English and Chinese social media datasets, respectively, showing the particular effectiveness of ANFU on social media texts.

The congruent, the incongruent, and the unexpected: Event-related potentials unveil the processes involved in schematic encoding

Learning is most effective when new information can be related to a preexisting knowledge structure or schema. In the present study, event-related potentials (ERPs) were used to investigate the temporal dynamics of the processes by which activated schemata support the encoding of schema-congruent information. Participants learned category exemplar words that were either semantically congruent or incongruent with a preceding category cue phrase. Congruent words were composed of expected (high typicality, HT) and unexpected (low typicality, LT) category exemplars. On the next day, recognition memory for the exemplars and the category cues they were presented with was tested. Semantically related lures were used in order to ascertain that memory judgements were based on episodic memory for specific category exemplars. Generally, congruent (HT and LT) exemplars were remembered better than incongruent exemplars. ERPs recorded during the encoding of the exemplar words were compared for subsequently remembered and forgotten items. Subsequent memory effects (SME) emerged in the N400 time window at frontal electrodes and did not differ between congruent and incongruent exemplars. In the same epoch, an SME with a parietal distribution was specific for congruent exemplars, suggesting that activated schemata strengthened memory for congruent exemplars by supporting the encoding of item-specific details. Subsequently remembered LT exemplars were associated with a late frontal positivity that is assumed to reflect expectancy mismatch-related processing such as the contextual integration of an unexpected word by the suppression of strongly expected words. A correlation analysis revealed that the greater the involvement of the processes reflected by the frontal positivity, the lower the level of false positive memory responses in the test phase one day later. These results suggest that the contextual integration of schema-congruent but unexpected events involves a weakening of the representations of semantically related, but unstudied items in memory and by this benefits subsequent memory.

Subjectivity in Spanish causal connectives

AbstractRecent literature has focused on specifying the nature of causal relations. Although the identification of whether connectives prototypically convey objective or subjective relations has been carried out for languages such as French, Dutch and English, the same has not occurred in the case of Spanish. The present study examines frequently used connectives and lexical cue phrases in terms of domain, propositional attitude, and presence of a Subject of Consciousness (SoC) in order to determine the degrees of subjectivity conveyed by the connectives. The findings show thatdebido a (que)is a highly specific connective used in objective causal relations about facts which do not involve SoC. On the contrary, the connectivesporqueandya quepresent similar flexibility that result in both being used to convey subjective and objective causal relations. This study supports the assumption of subjectivity being a cognitive mechanism that shows across different languages.

Causal relation extraction and network construction of web events

The explosive increase of news data on the web has created a mass of causal knowledge, which explains the causal relation between two events that effect event will occur following the occurrence of cause event. Analysis of causal knowledge has gain lots of attentions due to its widespread applications, such as question answering, event prediction, generating future scenarios, and commonsense causal reasoning. However, few researches are based on Chinese news corpus, and no effective causal template is proposed for extracting Chinese causal relationship. Therefore, the method for extracting causal relation and building network of causal events from Chinese news corpus is proposed. First, we propose a method to obtain complete cue phrases set and present four common causal patterns to extract causal relations. And then we merge the same events by similarity calculation of causal events. At last, a network of causal events is constructed. Experiments on the datasets show the effectiveness of the proposed approach.

Causal relation extraction and network construction of web events

The explosive increase of news data on the web has created a mass of causal knowledge, which explains the causal relation between two events that effect event will occur following the occurrence of cause event. Analysis of causal knowledge has gain lots of attentions due to its widespread applications, such as question answering, event prediction, generating future scenarios, and commonsense causal reasoning. However, few researches are based on Chinese news corpus, and no effective causal template is proposed for extracting Chinese causal relationship. Therefore, the method for extracting causal relation and building network of causal events from Chinese news corpus is proposed. First, we propose a method to obtain complete cue phrases set and present four common causal patterns to extract causal relations. And then we merge the same events by similarity calculation of causal events. At last, a network of causal events is constructed. Experiments on the datasets show the effectiveness of the proposed approach.

The linguistic marking of coherence relations

AbstractConnectives and cue phrases are the most prototypical linguistic elements that signal coherence relations, but by limiting our attention to connectives, we are likely missing out on important other cues readers and listeners use when establishing coherence relations. However, defining the role of other types of linguistic elements in the signaling of coherence relations is not straightforward, and it is also not obvious why and how non-connective elements function as signals for coherence relations. In this paper, we aim to develop a systematic way of categorizing segment-internal elements as signals of coherence relations on the basis of a literature review and evidence from parallel corpora. We propose a three-way distinction betweendivision of labor, agreement, andgeneral collocationto categorize the different ways in which elements inside discourse segments interact with connectives in the marking of coherence relations. In each type of interaction, segment-internal elements can function as signals for coherence relations, but the mechanism behind it is slightly different for each type.

Adversative Discourse Markers in Kurdish Literary Texts

Discourse Markers are one of an uninvestigated aspect of language in old and modern Kurdish linguistics, that has not been given due attention, neither by native nor non-native researchers. On this ground, it is hoped that the present study sheds light on this almost entirely ignored aspect of the language and this study is meant to be a systematic treatment of this group of lexical items known as Discourse Markers (henceforth, DMs), more specifically one category of them; Adversative DMs. 
 
 DMs are words, phrases and even clauses that enhance discourse coherence and are found in all languages, as tapped on by researches and investigations. Numerous terminologies are utilized to refer to such group of markers by different researchers in English and other languages, such as ‘Discourse Particles, Cue Phrases, Small Words, Pragmatic Markers, Discourse Connectives… and even they are defined differently. 
 
 It is postulated that DMs are meaningless and lay outside the domain of sentence structure. Likewise, lexical expressions that have different grammatical functions such as ‘and, also, but, or, simultaneously, at the same moment …etc, can also function as DMs to connect the previous utterance with the upcoming discourse segment. 
 
 The current investigation endeavors to answer certain specific questions: first, the extents to which DMs are operated in literary texts; second, discourse functions DMs implement. Thirdly, the word categories DMs are derived from, and to which extent Halliday and Hassan (1976)’s framework is applicable to Kurdish DMs? 
 
 For achieving the aims, the researchers analyzed one of the contemporary novels of a famous novelist entitled ‘Xezlenûs w Bâxekâni Xejâł”. By applying Halliday and Hasan’s (1976) framework and also by taking insights from Fraser (2009), DMs are categorized into different classes. One of which is Adversative DMs, which are the concern of the present study.
 
 For obtaining the frequency of each marker, the data are scrutinized manually, since there are no corpus analysis tools that can facilitate such measurements.
 
 The study concludes that Adversative DMs are frequently used in selected Kurdish literary texts and that they are similar to those found in English in terms of derived grammatical categories, taxonomy, and they have different characteristics in terms of form, position and discourse functions. Withal, it has been arrived that Adversative DMs are of different kinds analogous to those investigated in English by Halliday and Hassan (1976).

Smart Enough to Talk With Us? Foundations and Challenges for Dialogue Capable AI Systems

Smart Enough to Talk With Us? Foundations and Challenges for Dialogue Capable AI Systems

Cognitive complexity and the linguistic marking of coherence relations: A parallel corpus study

Coherence relations can be made linguistically explicit by means of connectives (e.g., but, because) or cue phrases (e.g., on the other hand, which is why), but can also be left implicit and conveyed through the juxtaposition of two clauses or sentences. However, it seems that not all relations are equally easy to reconstruct when they are implicit. In this paper, we explore which features of coherence relations make them more, or less, likely to be conveyed implicitly. We adopt the assumption that expected relations are more often implicit than relations that are not expected, and propose to determine a relation's expectedness using the notion of cognitive complexity. We test our hypotheses by means of a parallel corpus study, in which we analyze the translations of explicit English coherence relations from the Europarl Direct corpus into four target languages: Dutch, German, French, and Spanish. We find that cognitive complexity indeed influences the linguistic marking of coherence relations, and that this does not vary between the languages in our corpus. In addition, we find that a relation's relational and syntactic dependency also influences its linguistic marking, but that these measures are not completely independent of relation type.