Language understanding

This paper proposes that abstract concepts are represented as contextually derived structures. According to this abstract structure theory, abstract concepts are related to mostly temporal and spatial structures that underlie and can be extracted from concrete situations. Linguistic context elements of abstract concepts, such as verbs and prepositions, express these structures, and can thus aid in the acquisition of such concepts. The paper presents results from a corpus study that supports the hypothesis, and discusses implications. We propose that abstract concepts, such as faith or notion, are represented as abstract structures, which represent particular contexts in which they occur. Abstract concepts are not directly perceivable, but are often used in verbal descriptions of situations, or in utterances related to a situation. Such utterances have to unambiguously point out the entity that is referred to by the abstract noun. We argue that the relevant abstract structures can be inferred from their linguistic context, in particular, from verbs and prepositions used with the abstract nouns. Similar concepts occur in similar linguistic contexts. Miller and Charles (1991) report that contexts of similar concepts are more often classified as belonging to the same concept. For example, the sentences the patient rang for the ____ and the ____ gave the patient an injection both suggest that the word doctor or nurse would complete the sentence well. We found that a neural network could be trained to correctly select one out of seven abstract concepts based on linguistic context information in 72% of test cases (Wiemer-Hastings, 1998). An approach to learning verbs from context was described by Wiemer-Hastings, Graesser, and Wiemer-Hastings (1998; see also Hastings, 1994). This work shows a close link between context of use and concepts.

This paper proposes that abstract concepts are represented as contextually derived structures. According to this abstract structure theory, abstract concepts are related to mostly temporal and spatial structures that underlie and can be extracted from concrete situations. Linguistic context elements of abstract concepts, such as verbs and prepositions, express these structures, and can thus aid in the acquisition of such concepts. The paper presents results from a corpus study that supports the hypothesis, and discusses implications. We propose that abstract concepts, such as faith or notion, are represented as abstract structures, which represent particular contexts in which they occur. Abstract concepts are not directly perceivable, but are often used in verbal descriptions of situations, or in utterances related to a situation. Such utterances have to unambiguously point out the entity that is referred to by the abstract noun. We argue that the relevant abstract structures can be inferred from their linguistic context, in particular, from verbs and prepositions used with the abstract nouns. Similar concepts occur in similar linguistic contexts. Miller and Charles (1991) report that contexts of similar concepts are more often classified as belonging to the same concept. For example, the sentences the patient rang for the ____ and the ____ gave the patient an injection both suggest that the word doctor or nurse would complete the sentence well. We found that a neural network could be trained to correctly select one out of seven abstract concepts based on linguistic context information in 72% of test cases (Wiemer-Hastings, 1998). An approach to learning verbs from context was described by Wiemer-Hastings, Graesser, and Wiemer-Hastings (1998; see also Hastings, 1994). This work shows a close link between context of use and concepts.

The integration of cognitive neuroscience with cognitive linguistic theory represents a crucial frontier in understanding human language capacity. This paper examines how contemporary neurobiological findings support core tenets of cognitive linguistics, including embodied cognition, usage-based learning, and the integration of language with general cognitive systems. Through analysis of seminal contributions by Angela Friederici, Steven Pinker, and Evelina Fedorenko, we demonstrate that neural evidence largely validates cognitive linguistic principles while revealing additional complexity in brain-language relationships. Friederici's research on syntactic processing networks provides neural support for hierarchical organization in construction grammar, while Fedorenko's precision mapping reveals both domain-specific language networks and domain-general cognitive involvement. Pinker's evolutionary perspective, though emphasizing biological constraints, offers complementary insights into the cognitive foundations that cognitive linguistics seeks to understand. The synthesis reveals that while cognitive linguistic theory receives substantial neural validation, future models must account for the complex interplay between specialized language networks and general cognitive systems. This integration advances both fields by providing empirical grounding for theoretical claims and neurobiological constraints on cognitive models.

General purpose Vision Language Models (VLMs) have received tremendous interest in recent years, owing to their ability to learn rich vision-language correlations as well as their broad zero-shot competencies. One immensely popular line of work utilizes frozen unimodal models, by bridging vision representations to language using a trainable module called the QFormer. However, this method relies heavily on large-scale multimodal pretraining with huge computational overheads. To that end, we propose a more efficient framework for QFormer-based vision-language alignment. Our key idea relies on the observation that QFormer latents correspond more strongly to the frozen LLM's intermediate latent space. Consequently, instead of using QFormer latents as inputs to the LLM, we alter the framework by using the latents to directly condition the LLM latent space for image-to-text generation. We demonstrate the effectiveness of our approach against existing baselines in improving the efficiency of vision-language pretraining.

My deepest feelings of gratitude and indebtedness are to my advisor and mentor Mark Steedman. Mark has taught me not just linguistics and cognition, but also about thinking and behaving like a scientist. He tirelessly and carefully read through each of the many, many drafts of papers I have given him, including this thesis, providing meticulous and insightful comments; and spent countless hours explaining and debating these comments with me. I would guess that every single point I discuss in this thesis reflects his contributions. This thesis, like me, is a product of a community. The Cognitive Science community at Penn brings forth a beautifully dissonant buzz of intellectual activity. The diversity of outlooks, paradigms, methods, results and opinions about the study of the mind provides an ideal environment in which to go shopping for one's own direction. It is impossible to identify the individual members and visitors to IRCS (Institute for Research in Cognitive Science) that have shaped my perspective, so I'll just thank the two people who are responsible for enabling the tremendous growth of IRCS during my stay at Penn: Aravind Joshi and Lila Gleitman. I am very grateful to my doctoral committee: Janet Fodor, Aravind Joshi, Mitch Marcus and Ellen Prince for their helpful comments on this document, in particular to Ellen and Janet for long and fruitful discussions about the first parts of the thesis, and to Mitch for making the Penn Treebank available to me.

We describe CARDIAC, a prototype for an intelligent conversational assistant that provides health monitoring for chronic heart failure patients. CARDIAC supports user initiative through its ability to understand natural language and connect it to intention recognition. The spoken language interface allows patients to interact with CARDIAC without special training. We present speech recognition results obtained during an evaluation with fourteen chronic heart failure patients.

In this paper we present a novel algorithm for learning probabilistic subsequential transducers from a randomly drawn sample. We formalize the properties of the training data that are sufficient conditions for the learning algorithm to infer the correct machine. Finally, we report experimental evidences to backup the correctness of our proposed algorithm.

This paper addresses the problem of learning multiple spoken language understanding (SLU) tasks that have overlapping sets of slots. In such a scenario, it is possible to achieve better slot filling performance by learning multiple tasks simultaneously, as opposed to learning them independently. We focus on presenting a number of simple multi-task learning algorithms for slot filling systems based on semi-Markov CRFs, assuming the knowledge of shared slots. Furthermore, we discuss an intradomain clustering method that automatically discovers shared slots from training data. The effectiveness of our proposed approaches is demonstrated in an SLU application that involves three different yet related tasks.

In this paper, we propose to bring together the semantic web experience and statistical natural language semantic parsing modeling. The idea is that, the process for populating knowledgebases by semantically parsing structured web pages may provide very valuable implicit annotation for language understanding tasks. We mine search queries hitting to these web pages in order to semantically annotate them for building statistical unsupervised slot filling models, without even a need for a semantic annotation guideline. We present promising results demonstrating this idea for building an unsupervised slot filling model for the movies domain with some representative slots. Furthermore, we also employ unsupervised model adaptation for cases when there are some in-domain unannotated sentences available. Another key contribution of this work is using implicitly annotated natural-language-like queries for testing the performance of the models, in a totally unsupervised fashion. We believe, such an approach also ensures consistent semantic representation between the semantic parser and the backend knowledge-base.

Conversational AI has seen unprecedented growth in recent years due to which Chatbots have been made available. Conversational AI primarily focuses on text or speech inputs, identifying the intention behind them, and responding to users with relevant information. Natural Language Processing (NLP), Natural Language Understanding (NLU), Machine Learning (ML), and speech recognition offer a personalized experience that mimics human-like engagement in conversational AI systems. Conversational AI systems like Google Meena, Amazon's Alexa, Facebook's BlenderBot, and OpenAI's GPT-3 are trained using Deep Learning (DL) techniques that mimic a human brain-like structure and are trained on huge amounts of text data to provide open-domain conversations. The aim of this chapter is to highlight Conversational AI and NLP techniques behind it. The chapter focuses on DL architectures useful in building Conversational AI systems. The chapter discusses what are the recent advances in Conversational AI and how they are useful, what are the challenges, and what is the scope and future of conversational AI. This will help researchers to understand state-ofthe-art frameworks and how they are useful in building Conversational AI models.

La compréhension en lecture est une activité cognitive complexe qui met en jeu différents processus à différents niveaux de représentation, dont le but ultime est de construire une représentation cohérente de la situation décrite dans le texte. Alors que les processus liés au code ont fait l'objet de nombreuses études, la compréhension de l'oral est une composante moins bien identifiée mais qui, depuis quelques années, fait l'objet d'une attention particulière. Si certains enfants en échec face à la compréhension d'un texte lu ont des difficultés liées au traitement du code, d'autres échouent à cause de difficultés touchant la compréhension en général ; ils sont dénommés les « faibles compreneurs ». Nous exposons ici quelques recherches centrées sur le traitement des anaphores qui sont des marques linguistiques dont le traitement est crucial pour construire et intégrer une représentation mentale cohérente du texte. Toutes nos recherches confirment que l'interprétation des anaphores est une procédure qui joue un rôle important dans la compréhension de textes écrits. Les enfants faibles compreneurs ont des difficultés à organiser et à prendre en charge la double cohérence globale et locale du texte, et leurs difficultés ne sont pas spécifiques de l'écrit puisque des difficultés similaires sont observées à l'oral. Compréhension en lecture. Compréhension de l'oral. Anaphores.

We argue the case for abstract document structure as a separate descriptive level in the analysis and generation of written texts. The purpose of this representation is to mediate between the message of a text (i.e., its discourse structure) and its physical presentation (i.e., its organization into graphical constituents like sections, paragraphs, sentences, bulleted lists, figures, and footnotes). document structure can be seen as an extension of Nunberg's “text-grammar” it is also closely related to “logical” markup in languages like HTML and LaTEX. We show that by using this intermediate representation, several subtasks in language generation and language understanding can be defined more cleanly.

While the effectiveness of one-on-one human tutoring has been well established, a great deal of controversy surrounds the issue of which features of tutorial dialogue separate effective uses of dialogue in tutoring from those that are less effective. In this paper we present a formal comparison of Socratic versus Didactic style tutoring that argues in favor of the Socratic tutoring style.

Recent research in anaphora resolution has emphasized the effects of discourse structure and cohesion in determining what concepts are available as possible referents, and how discourse cohesion can aid reference resolution. Five approaches, all within this paradigm and yet all distinctly different, are presented, and their strengths and weaknesses evaluated.

Embodied theories of conceptual knowledge suggest that sensory-motor representations of actions similar to those involved in the performance of the action described are recruited during language comprehension. The extent of this recruitment, however, and the brain mechanisms supporting this process remain unknown. Using fMRI, we investigated these issues by examining how people understand sentences that convey three different degrees of physical effort and by comparing this process to action execution. To understand the effort implied by the stimulus sentences, object and action properties associated with nouns and verbs respectively needed to be integrated: pushing the piano implies more physical effort than pushing the chair. Results indicated that a pre-motor region, which was also active in action execution, was sensitive to the degree of effort implied by the language. Interestingly, the anterior inferior frontal gyrus, a region typically associated with semantic processing, was not active in action execution but was nevertheless modulated by the effort implied. Inter-region correlations also suggested that this region was strongly correlated with premotor and posterior temporal regions. Overall, results suggest that (a) language understanding elicits action representations retaining a degree of specificity that was previously unsuspected, including unique properties of interactions with objects, and (b) these representations, which result from integrating the words' semantic information, may be computed within a collaborative neural network that includes the anterior inferior frontal gyrus.

Three experiments examined the comprehension and appreciation of figurative phrases known as oxymora (e.g., intense apathy, loyal opposition). Experiment 1 showed that adjective-noun pairs in which the noun term is the hyponym of the adjective terms' antonym were judged by readers as the most poetic. Experiment 2 found that these same types of phrases were the easiest to understar,d. Experiment 3 demonstrated that interpreting o:cymora req,dres more than the simple combination of the adjective and noun terms' semantic components. These findings provide additional evidence on the influence of conceptual knowledge for both theories of complex conceptual combinations and for theories of figurative language understanding.

Working memory is important for online language processing during conversation. We use it to maintain relevant information, to inhibit or ignore irrelevant information, and to attend to conversation selectively. Working memory helps us to keep track of and actively participate in conversation, including taking turns and following the gist. This paper examines the Ease of Language Understanding model (i.e., the ELU model, Rönnberg, 2003; in light of new behavioral and neural findings concerning the role of working memory capacity (WMC) in uni-modal and bimodal language processing. The new ELU model is a meaning prediction system that depends on phonological and semantic interactions in rapid implicit and slower explicit processing mechanisms that both depend on WMC albeit in different ways. It is based on findings that address the relationship between WMC and (a) early attention processes in listening to speech, (b) signal processing in hearing aids and its effects on short-term memory, (c) inhibition of speech maskers and its effect on episodic long-term memory, (d) the effects of hearing impairment on episodic and semantic long-term memory, and finally, (e) listening effort. New predictions and clinical implications are outlined. Comparisons with other WMC and speech perception models are made.

Linguistic negation can be comprehended with the inclusion (or absence) of features and categories associated with the senses in a single step. Under this view, there is no need for explicit logical operators, as the negating word or phase is treated no differently than any other word. Negation provides additional context, whereby visualizing negation as a trajectory in a distributed, grounded perceptual simulation space can easily characterize the comprehension of negated sentences. A mousetracking experiment was conducted to ...

Classical intensional semantic frameworks, like Montague's Intensional Logic (IL), identify intensional identity with logical equivalence. This criterion of cointensionality is excessively coarse-grained, and it gives rise to several well known di culties. Theories of fine-grained intensionality have been been proposed to avoid this problem. Several of these provide a formal solution to the problem, but they do not ground this solution in a substantive account of intensional di↵erence. Applying the distinction between operational and denotational meaning, developed for the semantics of programming languages, to the interpretation of natural language expressions, o↵ers the basis for such an account. It permits us to escape some of the complications generated by the traditional modal characterization of intensions.

The lexical entry for a word must contain all the information needed to construct a semantic representation for sentences that contain the word. Because of that requirement, the formats for lexical representations must be as detailed as the semantic forms. Simple representations, such as features and frames, are adequate for resolving many syntactic ambiguities. But since those notations cannot represent all of logic, they are incapable of supporting all the function needed for semantics. Richer semantic-based approaches have been developed in both the model-theoretic tradition and the more computational AI tradition. Although superficially in conflict, these two traditions have a great deal in common at a deeper level. Both of them have developed semantic structures that are capable of representing a wide range of linguistic phenomena. This paper compares these approaches and evaluates their adequacy for various kinds of semantic information that must be stored in the lexicon. It presents conceptual graphs as a synthesis of the logicist and AI representations designed to support the requirements of both.

The galaxy system is a human-computer conversational system providing a spoken language interface for accessing on-line information. It was initially implemented for English in travel-related domains, including air travel, local city navigation, and weather. We began an effort to develop multilingual systems within the framework of galaxy several years ago. This paper describes our recent work on porting the system to Mandarin Chinese, including speech recognition, language understanding, and language generation components. Overall, the system produced reasonable responses nearly 70% of the time for spontaneous test data collected in a wizard environment.

Stochastic language models are widely used in spoken language understanding to recognize and interpret the speech signal: the speech samples are decoded into word transcriptions by means of acoustic and syntactic models and then interpreted according to a semantic model. Both for speech recognition and understanding, search algorithms use stochastic models to extract the most likely uttered sentence and its correspondent interpretation. The design of the language models has to be effective in order to mostly constrain the search algorithms and has to be efficient to comply with the storage space limits. In this work we present the Variable N-gram Stochastic Automaton (VNSA) language model that provides a unified formalism for building a wide class of language models. First, this approach allows for the use of accurate language models for large vocabulary speech recognition by using the standard search algorithm in the one-pass Viterbi decoder. Second, the unified formalism is an effective approach to incorporate different sources of information for computing the probability of word sequences. Third, the VNSAs are well suited for those applications where speech and language decoding cascades are implemented through weighted rational transductions. The VNSAs have been compared to standard bigram and trigram language models and their reduced set of parameters does not affect by any means the performances in terms of perplexity. The design of a stochastic language model through the VNSA is described and applied to word and phrase class-based language models. The effectiveness of VNSAs has been tested within the Air Travel Information System (ATIS) task to build the language model for the speech recognition and the language understanding system.

Linguistic negation can be comprehended with the inclusion (or absence) of features and categories associated with the senses in a single step. Under this view, there is no need for explicit logical operators, as the negating word or phase is treated no differently than any other word. Negation provides additional context, whereby visualizing negation as a trajectory in a distributed, grounded perceptual simulation space can easily characterize the comprehension of negated sentences. A mousetracking experiment was conducted to explore how this kind of process may be enacted in the brain and to tease apart hypotheses of logical manipulations vs. analogue signals performing this work.

This paper presents an introduction for the ESPRIT community of the recently started ESPRIT 11 project, PLUS (P5254). The goal of the project is the production of a robust natural language dialogue system integrating linguistic and non-linguistic knowledge in a principled way, based on the pragmatics theories of Grice and Searle, and using results in knowledge-base management systems and logic programming for the maintenance of dynamic contextual knowledge bases. We present the background assumptions, an ...

We think recent work in linguistics tells against the traditional claim that a string of words like (1) Every girl pushed some truck has two readings, indicated by the following formal language sentences (with restricted quantifiers): (1a) [!x:Gx]["y:Ty]Pxy (1b) ["y:Ty][!x:Gx]Pxy. In our view, (1) does not have any b-reading in which 'some truck' has widest scope. 1 The issue turns on details concerning syntactic transformations and terms like 'every'. This illustrates an important point for the study of natural language: ambiguity hypotheses are indeed hypotheses-i.e., theoretical claims to be justified in light of various considerations, not theses whose truth can be directly observed by speakers. Correlatively, strings that appear to be similar may differ with respect to scopal possibilities. While the object cannot take scope over the subject in (1), we hold that in (2) Some girl pushed every truck 'every truck' can take scope over 'some girl'. But showing that (2) is truth-conditionally ambiguous is hard, especially if natural language involves tacit quantification over events. (A string can be structurally ambiguous without having truth-conditionally distinct readings; consider [[α & β] & γ] as opposed to [α & [β & γ]].) Other cases are clearer. The subject in (3) Two girls pushed six trucks can be distributive or collective. Indeed, we use facts about plural constructions like (3) and (4) Both girls pushed six trucks to support the claim that (1) is truth-conditionally unambiguous. Finally, we argue that (5) Ralph believes that the richest man is happy has no reading in which 'the richest man' has scope over 'believes'. Our particular conclusions are less important, though, than the methodological moral. Linguistic meaning is constrained in ways that ambiguity hypotheses must respect. It can be hard to know how many readings a string of words has, especially when the scope of quantifiers is at issue. Far from being paradigms of ambiguity in natural language, (1-2) are complicated cases.

First, I want to sincerely thank my advisor James F. Allen for always being very welcoming, considerate, understanding, and encouraging. He gave me the freedom to pursue my own research, and at the same time provided me with ideas and support whenever I needed -not to mention the torture I did to him with the first draft of my documents, including this thesis, which he patiently went through and returned with insightful feedback. My special thanks to Jeffery P. Bigham for giving me the opportunity to be exposed to the Human-Computer Interaction (HCI) area, and to collaborate on exciting projects. I want to especially thank Lenhart K. Schubert for giving me constructive suggestions on various topics and for our fruitful discussions on the Timegraph (section 5.3.1

In this paper we introduce Mandarin language understanding methods developed for spoken language applications. We describe a set of strategies to improve the parsing performance for Mandarin. We also discuss two context resolution techniques adopted to handle Chinese ellipsis in a practical Mandarin spoken dialogue system. Experimental evaluation verifies the effectiveness and efficiency of our proposed parsing enhancements, in terms of both parse coverage and speed. System evaluation with human subjects also verifies the effectiveness of our proposed approaches to speech understanding and context resolution in practical conversational systems.

Virtual humans are computer‐generated characters designed to look and behave like real people. Studies have shown that virtual humans can mimic many of the social effects that one finds in human‐human interactions such as creating rapport, and people respond to virtual humans in ways that are similar to how they respond to real people. We believe that virtual humans represent a new metaphor for interacting with computers, one in which working with a computer becomes much like interacting with a person and this can bring social elements to the interaction that are not easily supported with conventional interfaces. We present two systems that embody these ideas. The first, the twins are virtual docents in the Museum of Science, Boston, designed to engage visitors and raise their awareness and knowledge of science. The second, Sim‐Coach, uses an empathetic virtual human to provide veterans and their families with information about PTSD and depression.

Latent Semantic Analysis (LSA) has been shown to perform many linguistic tasks as well as humans do, and has been put forward as a model of human linguistic competence. But LSA pays no attention to word order, much less sentence structure. Researchers in Natural Language Processing have made significant progress in quickly and accurately deriving the syntactic structure of texts. But there is little agreement on how best to represent meaning, and the representations are brittle and difficult to build. This paper evaluates a model of language understanding that combines information from rule-based syntactic processing with a vector-based semantic representation which is learned from a corpus. The model is evaluated as a cognitive model, and as a potential technique for natural language understanding.

Current dialogue systems generally operate in a pipelined, modular fashion on one complete utterance at a time. Evidence from human language understanding shows that human understanding operates incrementally and makes use of multiple sources of information during the parsing process, including traditionally “later” components such as pragmatics. In this paper we describe a spoken dialogue system that understands language incrementally, provides visual feedback on possible referents during the course of the ...

There are two significant problems in using functional neuroimaging methods to study language. Improving the state of functional brain imaging will depend on understanding how the dependent measure of brain imaging differs from behavioral dependent measures (the ''dependent measure problem'') and how the activation of the motor system may be confounded with non-motor aspects of processing in certain experimental designs (the ''motor output problem''). To address these problems, it may be necessary to shift the focus of language research from the study of linguistic competence to the understanding of language use. This will require investigations of language processing in full multi-modal and environmental context, monitoring of natural behaviors, novel experimental design, and network-based analysis. Such a combined naturalistic approach could lead to tremendous new insights into language and the brain.

This paper presents a model of language processing where word segmentation is an integral part of sentence analysis. We show that the use of a parser can enable us to achieve the best ambiguity resolution in word segmentation. The lexical component of this model resolves most of the ambiguities, but the final disambiguation takes place in the parsing process. In this model, word segmentation is a by-product of sentence analysis, where the correct segmentation is represented by the leaves of a parse tree. We also show that the complexity usually associated with the use of a parser in segmentation can be reduced dramatically by using a dictionary that contains useful information on word segmentation. With the aid of such information, the sentence analysis process is reasonably fast and does not suffer from the problems other people have encountered. The model is implemented in NLPWin, the general-purpose language understanding system developed at Microsoft Research. A demo of the syst...

Information graphics, such as bar charts and line graphs, generally have a communicative message that they are intended to convey when they appear in popular media. The communicative signals present or absent in the graphic help convey these underlying intended messages. We have designed a Bayesian system that automatically extracts communicative signals from a grouped bar chart information graphic to hypothesize its intended messages. This paper presents experiments that analyze the effect that each type of communicative evidence has on our system.

Clinical texts and documents contain a wealth of information and knowledge in the field of healthcare, and their processing, using state-of-the-art language technology, has become very important for building intelligent systems capable of supporting healthcare and providing greater social good. This processing includes creating language understanding models and translating resources into other natural languages to share domain-specific cross-lingual knowledge. In this work, we conduct investigations on clinical text machine translation by examining multilingual neural network models using deep learning methods such as Transformer-based structures. Furthermore, to address the issue of language resource imbalance, we also carry out experiments using a transfer learning methodology based on massive multilingual pre-trained language models (MMPLMs). The experimental results on three sub-tasks including 1) clinical case (CC), 2) clinical terminology (CT), and 3) ontological concept (OC) show that our models achieved toplevel performances in the ClinSpEn-2022 shared task on English-Spanish clinical domain data. Furthermore, our expert-based human evaluations demonstrate that the small-sized pre-trained language model (PLM) wins in the clinical domain fine-tuning over the other two extra-large language models by a large margin. This finding has never been previously reported in the field. Finally, the transfer learning method works well in our experimental setting using the WMT21fb model to accommodate a new Spanish language space that was not seen at the pretraining stage within WMT21fb itself-and this deserves further exploration for clinical knowledge transformation, e.g. investigation into more languages. These research findings can shed some light on domain-specific machine translation development, especially in clinical and healthcare fields. Further research projects can be carried out based on our work to improve healthcare text analytics and knowledge transformations. Our data will be openly available for research purposes at

Korece, uzerinde onemli derlem calismalari yapilmis dillerden biridir. Bu dilin, devlet kurumlarinin destegi ile Kore universite ve arastirma merkezleri tarafindan hazirlanmis oylumlu derlem calismalari bulunmaktadir. Bu yazida Korecenin genel sozvarligi ozelliklerinden hareketle bugune kadar ortaya konmus Korece derlem calismalari hakkinda bilgiler aktarilacaktir

With the recent influx of bidirectional contextualized transformer language models in the NLP, it becomes a necessity to have a systematic comparative study of these models on variety of datasets. Also, the performance of these language models has not been explored on non-GLUE datasets. The study presented in paper compares the state-of-the-art language models - BERT, ELECTRA and its derivatives which include RoBERTa, ALBERT and DistilBERT. We conducted experiments by finetuning these models for cross domain and disparate data and penned an in-depth analysis of model's performances. Moreover, an explainability of language models coherent with pretraining is presented which verifies the context capturing capabilities of these models through a model agnostic approach. The experimental results establish new state-of-the-art for Yelp 2013 rating classification task and Financial Phrasebank sentiment detection task with 69% accuracy and 88.2% accuracy respectively. Finally, the study...

We present a new temporal annotation standard, THEE-TimeML, and a corpus TheeBank enabling precise temporal information extraction (TIE) for event-based surveillance (EBS) systems in the public health domain. Current EBS must estimate the occurrence time of each event based on coarse document metadata such as document publication time. Because of the complicated language and narration style of news articles, estimated case outbreak times are often inaccurate or even erroneous. Thus, it is necessary to create annotation standards and corpora to facilitate the development of TIE systems in the public health domain to address this problem.We will discuss the adaptations that have proved necessary for this domain as we present THEE-TimeML and TheeBank. Finally, we document the corpus annotation process, and demonstrate the immediate benefit to public health applications brought by the annotations.

Dialogue systems for spatially situated tasks need to provide referential descriptions of spatially located objects and understand such descriptions from users. To construct such dialogue systems, it is useful to investigate how humans describe object locations in their immediate environment and how they ask about object locations in remote environments. In this paper, we address the semantic classification of the localization utterances found in the CReST corpus, which is a dialogue corpus of humans performing a cooperative, remote, search task. The aim is to explore the relation between specific semantic configurations and the dialogically and situationally embedded linguistic forms employed. Specifically, we first extracted different types of localization utterances from the corpus and then paired these with semantic categories provided by the linguistically motivated spatial ontology GUM. The paper concludes with a discussion of the characteristics of different types of localization expressions on the basis of spatial concepts and descriptions employed.

and to probe for possible misconceptions. The natural language understanding component uses a cascade of finite-state machines. The generation is based on lexical functional grammar. Results: Results of experiments with pretests and posttests have shown that using the system for an hour produces significant learning gains and also that even this brief use improves the student's ability to solve problems more then reading textual material on the topic. Student surveys tell us that students like the system and feel that they learn from it. The system is now in regular use in the first-year physiology course at Rush Medical College. Conclusion: We conclude that the CIRCSIM-Tutor system demonstrates that intelligent tutoring systems can implement effective natural language dialogue with current language technology.

Doherty and Perner (Metalinguistic awareness and theory of mind: just two words for the same thing? Cognitive Development, 13 (1998), 279±305) report that children's understanding of synonyms and false belief is dependent on an understanding of the representational mind. Experiment 1 extends this finding by examining children's understanding of homonyms. Children aged 3 and 4 years were asked to judge whether a puppet correctly selected the second member of a homonym pair. Performance on this task was strongly associated with performance on the false belief task even after chronological and verbal mental age had been accounted for. Experiment 2 incorporated two new tasks: a synonyms task and an adjectives task. Understanding of synonyms and homonyms significantly predicted performance on the false belief task. However, once chronological age was accounted for, only performance on the homonyms task did so. The difficulty experienced on the homonyms task was not due to a reluctance to acknowledge that the puppet can point to a different picture when the the same word label is used twice. Children had no difficulty on the adjectives task when the puppet had to point to a different picture described using the same adjective. The suggestion that the understanding of synonyms, homonyms and false belief are related by a common insight into the representational mind is therefore not supported.

In recent years, vision-language research has shifted to study tasks which require more complex reasoning, such as interactive question answering, visual common sense reasoning, and question-answer plausibility prediction. However, the datasets used for these problems fail to capture the complexity of real inputs and multimodal environments, such as ambiguous natural language requests and diverse digital domains. We introduce Mobile app Tasks with Iterative Feedback (MoTIF), a dataset with natural language commands for the greatest number of interactive environments to date. 1 MoTIF is the first to contain natural language requests for interactive environments that are not satisfiable, and we obtain follow-up questions on this subset to enable research on task uncertainty resolution. We perform initial feasibility classification experiments and only reach an F1 score of 37.3, verifying the need for richer vision-language representations and improved architectures to reason about task feasibility.

In recent years, vision-language research has shifted to study tasks which require more complex reasoning, such as interactive question answering, visual common sense reasoning, and question-answer plausibility prediction. However, the datasets used for these problems fail to capture the complexity of real inputs and multimodal environments, such as ambiguous natural language requests and diverse digital domains. We introduce Mobile app Tasks with Iterative Feedback (MoTIF), a dataset with natural language commands for the greatest number of interactive environments to date. 1 MoTIF is the first to contain natural language requests for interactive environments that are not satisfiable, and we obtain follow-up questions on this subset to enable research on task uncertainty resolution. We perform initial feasibility classification experiments and only reach an F1 score of 37.3, verifying the need for richer vision-language representations and improved architectures to reason about task feasibility.

In this paper we present a novel treebank developed to analyse marked constructions in Italian called MarkIT. The resource contains almost 1,300 sentences manually annotated with dependency relations following the Universal Dependencies paradigm. The sentences have been extracted from essays written by high-school students along several years, which accounts for the structure and the topic variability of the sentences. In this work, we detail the process to select the sentences, parse them automatically and then manually correct them. The resource covers seven types of marked constructions (839 sentences overall) plus some sentences, whose syntax can be wrongly classified as marked and which can serve as negative examples of markedness (453 sentences). We also present an evaluation of parsing performance, comparing a model trained on existing Italian treebanks with the model obtained by adding MarkIT to the training set.