Named Entity Extraction

This paper gives an overview of the history of prosopographical projects at KU Leuven, starting with the Prosopographia Ptolemaica in the interbellum, its successor Trismegistos People, and Trismegistos' newest feature, the Names in the Ancient World portal. It presents the evolving database models that were adapted to accommodate new types of data, touches upon the question of the prosopogaphical nature of these resources, and explains Trismegistos' stance on the feasibility of a single prosopography of the ancient world.

Este libro refleja el trabajo realizado bajo investigación entre docentes investigadores con el afán de que sea útil al lector, el uso de predicciones al momento de entrenar un algoritmo clasificado de texto en procesamiento de lenguaje natural (PLN) basado en machine learning. Conformado de 4 capítulos con la utilidad para el inicio al mundo de la IA de la rama de procesamiento de lenguaje natural con Python en machine learning. El Capítulo 1 menciona conceptos y la evolución de las diferentes ramas de conocimiento que abarca la inteligencia artificial (AI), el entendimiento del NLP, machine learning, tipos de aprendizaje para resolver problemas como el supervisado, no supervisado y refuerzo. Capítulo 2 se profundiza el NLP conociendo los contenidos básicos de clasificación como: Las técnicas y diseño de LSTM, tokeización, stopword, lematización, bag of Word (part of speech tagging). Capítulo 3 es la estructuración de este capítulo el conocer las definiciones de los modelos de aprendizaje supervisado que son útiles en NLP orientado a la clasificación de texto. Capítulo 4 un caso de predicción o grado de asertividad del modelamiento de un algoritmo, la intención es demostrar la utilización de un modelo y varias técnicas aplicando NLP basado en machine learning.

The availability of large multi-parallel corpora offers an enormous wealth of material to contrastive corpus linguists, translators and language learners, if we can exploit the data properly. Necessary preparation steps include sentence and word alignment across multiple languages. Additionally, linguistic annotation such as part-of-speech tagging, lemmatisation, chunking, and dependency parsing facilitate precise querying of linguistic properties and can be used to extend word alignment to sub-sentential groups. Such highly inter-connected data is stored in a relational database to allow for efficient retrieval and linguistic data mining, which may include the statistics-based selection of good example sentences. The varying information needs of contrastive linguists require a flexible linguistic query language for ad hoc searches. Such queries in the format of generalised treebank query languages will be automatically translated into SQL queries.

Transformer-based language models have been changing the modern Natural Language Processing (NLP) landscape for high-resource languages such as English, Chinese, Russian, etc. However, this technology does not yet exist for any Ghanaian language. In this paper, we introduce the first of such models for Twi or Akan, the most widely spoken Ghanaian language. The specific contribution of this research work is the development of several pretrained transformer language models for the Akuapem and Asante dialects of Twi, paving the way for advances in application areas such as Named Entity Recognition (NER), Neural Machine Translation (NMT), Sentiment Analysis (SA) and Part-of-Speech (POS) tagging. Specifically, we introduce four different flavours of ABENA – A BERT model Now in Akan that is fine-tuned on a set of Akan corpora, and BAKO BERT with Akan Knowledge only, which is trained from scratch. We open-source the model through the Hugging Face model hub and demonstrate its use via a sim...

A sentence aligned parallel corpus is an important prerequisite in statistical machine translation. However, manual creation of such a parallel corpus is time consuming, and requires experts fluent in both languages. Automatic creation of a sentence aligned parallel corpus using parallel text is the solution to this problem. In this paper, we present the first ever empirical evaluation carried out to identify the best method to automatically create a sentence aligned Sinhala-Tamil parallel corpus. Annual reports from Sri Lankan government institutions were used as the parallel text for aligning. Despite both Sinhala and Tamil being under-resourced languages, we were able to achieve an F-score value of 0.791 using a hybrid approach that makes use of a bilingual dictionary.

Le but de cet article est de présenter de façon synthétique le dernier état du cadre théorique mis en oeuvre dans mes travaux récents. Si certains de ses éléments n'ont pas changé depuis le début, les orientations nouvelles apparues depuis, et dont l'exposé s'est trouvé dispersé au gré des thèmes abordés, n'ont pas manqué. La notion de « multiprédicativité », quant à elle, semble avoir toujours aussi mauvaise presse chez bon nombre de linguistes d'obédiences diverses. Or, les « langues multiprédicatives », qui se définissent par le fait que d'autres parties du discours que le verbe peuvent exercer la fonction de prédicat syntaxique sans l'intervention d'une copule, verbe « être » ou autre élément copulatif, nous montrent pourtant deux choses absolument essentielles : * la première est qu'en l'absence de l'élément considéré classiquement comme voué à la prédication non verbale, la bonne question est « comment ça peut marcher ? » en l'absence même de cet élément-au lieu de supposer l'existence de quelque « zéro » ou de quelque opérateur caché. On en tire un principe heuristique et un instrument de lutte contre l'ethnocentrisme qui se révèle d'une force singulière pour peu qu'on en fasse systématiquement usage dès que le cas se présente : « lorsqu'on se trouve confronté à des langues présentant une marque, ou tout autre élément particulier, et d'autres d'où cet élément est absent (comme, par exemple, des langues ou des structures « avec » vs « sans » copule), il est de bonne méthode de partir des langues ou structures « sans » avant d'expliquer les langues ou structures « avec », * Tous nos remerciements aux fidèles du séminaire (EPHE) 2020-2021 dont les interventions non seulement ont enrichi le débat mais ont apporté vie et chaleur à ces visio-conférences qui auraient été, sans eux, encore plus 'distancielles' :

Co-reference resolution is an important part of natural language understanding and it's been affected by the current corpora lacking in diversity. This project presents the implementation of two models for masked language modeling using pre-trained BERT adjusted to work for a classification problem.

In interactive searching environments, robust linguistic techniques can provide sophisticated search assistance with a reasonable tolerance to errors, because users can easily select relevant items and dismiss the noisy bits. The general idea is that the combination of Language Engineering and Information Retrieval techniques can be used to suggest complex terms or relevant pieces of information to the user, facilitating query formulation and refinement when the information need is not completely defined a priori or when the user is not familiar with the contents and/or the terminology used in the collection. In this paper, we describe an interactive search engine that suggests Named Entities extracted automatically from the collection, and related to the initial query terms, helping users to filter and structure relevant information according to the persons, locations or other entities involved.

Resumen. La identificación automática del humor resulta una tarea compleja, ya que lo que provoca el humor aún no está completamente caracterizado. Se han presentado varios enfoques para detectar humor siendo la mayoría en inglés . Esta fue una de las razones por las que se presentó durante IberLEF2019 un reto que consistió en detectar humor en tweets en español . En esta competencia se abordó la detección de humor como una tarea de clasificación de textos. Retomando algunas partes de este reto, en este trabajo se presenta la comparación del desempeño de varias técnicas de aprendizaje supervisado para detectar humor en tweets en español utilizando herramientas disponibles en scikit-learn.

In this paper we propose an approach for identifying syntactic behaviours related to lexical items and linking them to the meanings. This approach is based on the analysis of the textual content presented in LMF normalized dictionaries by means of Definition and Context classes. The main particularity of these contents is their large availability and their semantically control due to their attachment to the meanings, which promotes the effective links between the syntactic behaviours and the meaning. In order to test the performance of the proposed approach, we tested it on an available Arabic LMF normalized dictionary. The experiment treats 9,800 verbs and allows us to evaluate the identified syntactic behaviours as well as their links to the meanings.

Information about location and geographical coordinates in particular, may be very important during a crisis event, especially for search and rescue operations – but currently geo-tagged tweets are extremely rare. Improved capabilities of capturing additional location from Twitter (up to 4 times improvement) are crucial for response efforts given a vast amount of messages exchanged during a crisis event. That is why authors have designed a tool (Text Analysis TWeet lOcator – TAT2) that relies on existing open source text analysis tools with additional services to provide additional hints about people location. Validation process, complementing experimentation and test results, included involvement of end-users (i.e. Public Protection and Disaster Relief services and citizens during a realistic crisis exercise showcase. In addition, the integration of TAT2 with external tools has also been validated.

With a dataset of 1.3 million articles from arXiv, we explore the potential of classifying research papers based solely on their abstracts and titles. We extract abstracts and titles from the arXiv dataset and fine-tune multiple pre-trained large language models to classify the correct category out of 149 distinct research categories. We report our results from five different models on four different input settings with their respective context lengths. For evaluation, we measure the top-k accuracies. Results show that different pre-trained models can successfully classify the correct category, based on either the title or title and abstract, with high accuracies on top-1 and definitely top-10 highest probabilities. The finetuned models performed competitively despite context length and large parameter size differences.

Named Entity Recognition (NER) is a crucial task in Natural Language Processing (NLP), which involves identifying and categorizing named entities in unstructured text data. In recent years, deep learning-based approaches such as Long Short-Term Memory (LSTM) and Conditional Random Fields (CRF) have shown impressive performance in NER. Furthermore, word embeddings have also become a popular method for representing words in NLP tasks. In this paper, we propose a comprehensive review of recent advances in NER using AI-NLP techniques, specifically LSTM, CRF, and word embeddings. We discuss the underlying principles of these techniques and their advantages in NER. We also present a comparative analysis of these approaches on benchmark datasets and highlight their strengths and weaknesses. Furthermore, we discuss some of the key challenges in NER, such as handling rare and unknown entities, and explore potential solutions using these techniques. Overall, this paper provides a comprehensive overview of NER using AI-NLP with LSTM, CRF, and word embeddings.

Online social networks convey rich information about geospatial facets of reality. However in most cases, geographic information is not explicit and structured, thus preventing its exploitation in real-time applications. We address this limitation by introducing a novel geoparsing and geotagging technique called Geo-Semantic-Parsing (GSP). GSP identifies location references in free text and extracts the corresponding geographic coordinates. To reach this goal, we employ a semantic annotator to identify relevant portions of the input text and to link them to the corresponding entity in a knowledge graph. Then, we devise and experiment with several efficient strategies for traversing the knowledge graph, thus expanding the available set of information for the geoparsing task. Finally, we exploit all available information for learning a regression model that selects the best entity with which to geotag the input text. We evaluate GSP on a well-known reference dataset including almost 10 k event-related tweets, achieving F1 = 0.66. We extensively compare our results with those of 2 baselines and 3 state-of-the-art geoparsing techniques, achieving the best performance. On the same dataset, competitors obtain F1 ≤ 0.55. We conclude by providing in-depth analyses of our results, showing that the overall superior performance of GSP is mainly due to a large improvement in recall, with respect to existing techniques.

Online social networks convey rich information about geospatial facets of reality. However in most cases, geographic information is not explicit and structured, thus preventing its exploitation in real-time applications. We address this limitation by introducing a novel geoparsing and geotagging technique called Geo-Semantic-Parsing (GSP). GSP identifies location references in free text and extracts the corresponding geographic coordinates. To reach this goal, we employ a semantic annotator to identify relevant portions of the input text and to link them to the corresponding entity in a knowledge graph. Then, we devise and experiment with several efficient strategies for traversing the knowledge graph, thus expanding the available set of information for the geoparsing task. Finally, we exploit all available information for learning a regression model that selects the best entity with which to geotag the input text. We evaluate GSP on a well-known reference dataset including almost 10 k event-related tweets, achieving F1 = 0.66. We extensively compare our results with those of 2 baselines and 3 state-of-the-art geoparsing techniques, achieving the best performance. On the same dataset, competitors obtain F1 ≤ 0.55. We conclude by providing in-depth analyses of our results, showing that the overall superior performance of GSP is mainly due to a large improvement in recall, with respect to existing techniques.

Cross-language information retrieval consists in providing a query in one language and searching documents in different languages. Retrieved documents are ordered by the probability of being relevant to the user's request with the highest ranked being considered the most relevant document. The LIC2M cross-language information retrieval system is a weighted Boolean search engine based on a deep linguistic analysis of the query and the documents to be indexed. This system, designed to work on Arabic, Chinese, English, French, German and Spanish, is composed of a multilingual linguistic analyzer, a statistical analyzer, a reformulator, a comparator and a search engine. The multilingual linguistic analyzer includes a morphological analyzer, a part-of-speech tagger and a syntactic analyzer. In the case of Arabic, a clitic stemmer is added to the morphological analyzer to segment the input words into proclitics, simple forms and enclitics. The linguistic analyzer processes both documents to be indexed and queries to produce a set of normalized lemmas, a set of named entities and a set of nominal compounds with their morpho-syntactic tags. The statistical analyzer computes for documents to be indexed concept weights based on concept database frequencies. The comparator computes intersections between queries and documents and provides a relevance weight for each intersection. Before this comparison, the reformulator expands queries during the search. The expansion is used to infer from the original query words other words expressing the same concepts. The expansion can be in the same language or in different languages. The search engine retrieves the ranked, relevant documents from the indexes according to the corresponding reformulated query and then merges the results obtained for each language, taking into account the original words of the query and their weights in order to score the documents. Sentence alignment consists in estimating which sentence or sentences in the source language correspond with which sentence or sentences in a target language. We present in this paper a new approach to aligning sentences from a parallel corpora based on the LIC2M cross-language information retrieval system. This approach consists in building a database of sentences of the target text and considering each sentence of the source text as a "query" to that database. The aligned bilingual parallel corpora can be used as a translation memory in a computer-aided translation tool.

La acentuación de palabras cuando se escribe un texto en español es un problema de ambigüedad, debido a que muchas palabras llevan acento o no dependiendo del contexto de la frase. El problema de la ambigüedad está relacionado con la asignación de etiquetas o categorías gramaticales a las palabras dentro de una frase, es decir, cuando se indica si se trata de un verbo, un sustantivo, etcétera. En el presente artículo se propone un modelo que ayuda a determinar en forma automática si una palabra con acento diacrítico debe llevar o no acento ortográfico, con base en la asignación de etiquetas y mediante la aplicación de métodos híbridos —algoritmos supervisado y no supervisado. Posteriormente, el método se aplica en el diseño de un programa de cómputo cuya función es de apoyo en la enseñanza de las reglas de acentuación y con el cual se realiza la acentuación automática de palabras ambiguas. Este método podría ser una herramienta en un procesador de palabras.

In financial services industry, compliance involves a series of practices and controls in order to meet key regulatory standards which aim to reduce financial risk and crime, e.g. money laundering and financing of terrorism. Faced with the growing risks, it is imperative for financial institutions to seek automated information extraction techniques for monitoring financial activities of their customers. This work describes an ontology of compliance-related concepts and relationships along with a corpus annotated according to it. The presented corpus consists of financial news articles in French and allows for training and evaluating domain-specific named entity recognition and relation extraction algorithms. We present some of our experimental results on named entity recognition and relation extraction using our annotated corpus. We aim to furthermore use the the proposed ontology towards construction of a knowledge base of financial relations.

In financial services industry, compliance involves a series of practices and controls in order to meet key regulatory standards which aim to reduce financial risk and crime, e.g. money laundering and financing of terrorism. Faced with the growing risks, it is imperative for financial institutions to seek automated information extraction techniques for monitoring financial activities of their customers. This work describes an ontology of compliance-related concepts and relationships along with a corpus annotated according to it. The presented corpus consists of financial news articles in French and allows for training and evaluating domain-specific named entity recognition and relation extraction algorithms. We present some of our experimental results on named entity recognition and relation extraction using our annotated corpus. We aim to furthermore use the the proposed ontology towards construction of a knowledge base of financial relations.

La clasificación de subjetividad es un ámbito de la minería de texto poco estudiado en el idioma español, y sin embargo sus aplicaciones son extensas. Su estudio permite comprender mejor la semántica de un texto y la intención de su autor, sin mencionar las implicaciones de su uso en la inteligencia de negocios, para identificar las necesidades de los clientes y obtener métricas valiosas a partir de sus críticas. En este trabajo se intenta aplicar técnicas conocidas de análisis de subjetividad en inglés, adaptadas al español, construyendo en el proceso una base de datos y un sistema clasificador de oraciones.Facultad de Informátic

Human beings are capable of categorizing a document based on its topic. Computers are already able to perform very well on that task. However, when translating from one language to another, the human translator will use this knowledge to adapt the writing style and vocabulary for the translation to sound as natural as possible. Statistical Machine Translation (SMT) uses Probabilistic Machine Learning methods to perform translations. However, such systems do not perform well in domains different from the ones used to train them. How can the ability to recognize the topic of a document be captured by an SMT system to perform better? Methodologies for adapting a Statistical Machine Translation System to a specific domain are explored. Two methods are examined. The one mixes translation and language models, weighting them appropriately to improve translation quality. The other uses unsupervised methods to cluster a corpus into sub-corpora, train them individually and decode on a specific trained cluster according to the genre or "domain" of the new sentence to be translated. Experimentation showed improvement in translation quality using both methods. Training on a small domain-specific corpus and a large general one, can improve the performance on translating documents in the small corpus' domain.

We will report evaluation of Automatic Named Entity Extraction feature of IR tools on Dutch, French, and English text. The aim is to analyze the competency of off-the-shelf information extraction tools in recognizing entity types including person, organization, location, vehicle, time, & currency from unstructured text. Within such an evaluation one can compare the effectiveness of different approaches for identifying named entities.

The paper presents an integral framework for multilingual lexical databases (henceforth MLLD) based on Compreno technology. It differs from the existing approaches to MLLD in the following aspects: 1) it is based on a universal semantic hierarchy (SH) of thesaurus type filled with language-specific lexicon; 2) the position in the SH generally determines semantic and syntactic model of a word; 3) this model proposes a suite of elaborate tools to determine universal and language-specific semantic and syntactic properties and deals efficiently with problems of cross-lingual lexical, semantic and syntactic asymmetry. Currently, it includes English, Russian, German, French and Chinese and proves to be a compatible MLLD for typologically different languages that can be used as a comprehensive lexical-semantic database for various NLP applications.

Named entity recognition (NER) is the process of seeking to locate atomic elements in text into predefined categories such as the names of persons, organizations and locations. Most existing NER systems are based on supervised learning. This method often requires a large amount of labelled training data, which is very time-consuming to build. To solve this problem, we introduce a semi-supervised learning method for recognizing named entities in Vietnamese text by combining proper name coreference, named-ambiguity heuristics with a powerful sequential learning model, Conditional Random Fields. Our approach inherits the idea of Liao and Veeramachaneni [6] and expands it by using proper name coreference. Starting by training the model using a small data set that is annotated manually, the learning model extracts high confident named entities and finds low confident ones by using proper name coreference rules. The low confident named entities are put in the training set to learn new context features. The F-scores of the system for extracting "Person", "Location" and "Organization" entities are 83.36%, 69.53% and 65.71% when applying heuristics proposed by Liao and Veeramachaneni. Those values when using our proposed heuristics are 93.13%, 88.15% and 79.35%, respectively. It shows that our method is good in increasing the system accuracy.

Temporal event signature mining for knowledge discovery is a difficult problem. In this paper a framework is designed to know a temporal knowledge about the large scales signature mining of longitudinal heterogeneous event data. This framework mainly deals with the mining of high order latent event structure and its relationship within single and multiple event sequences. Here, the heterogeneous event sequences maps to geometric image by encoding structured into spatial temporal shape process. A Probabilistic language modeling is used to extract high order events from large scale data. Also, presents a doubly constrained conventional sparse coding to learn interpretable and shift invariant latent temporal event signature. This can be shown by the sparsity in the data and latent factor model on the β-divergence. An optimization scheme is also used to perform large scale incremental learning of group specific temporal event signatures.

These days, the number of data sources an ordinary computer user works with every day is very large and continues to grow. With the increasing number of cloud services with specialized functionalities, the users are faced with the necessity to routinely perform manual actions to interchange data among different cloud and web services, in order to perform more complex and composite actions. These actions always require a certain amount of dedicated time from the user, who has to change the context in which he work, in order to take the actions and transfer data from one system to another. In this paper, we present a software platform, based on the concepts and technologies of the Semantic Web, which provides the users with a unified and simple composite approach to the different services they use, and crates a simple flow of information from one infrastructure to another. The system is able to automatically discover the context in which the user is working, and offer him the actions which can be used over the data within the context. In this way, the user can completely focus on his tasks in his work environment, and get relevant information and possible actions in that context. This system improvements in their productivity, information exchange and efficiency. The system is called Sky represents a platform where many cloud services are interconnected by the use of semantic web technologies.

We propose a new method of using multiple documents as evidence with decreased adding to improve the performance of a question-answering system. Sometimes, the answer to a question may be found in multiple documents. In such cases, using multiple documents for prediction would generate better answers than using a single document. Thus, our method employs information from multiple documents by adding the scores of the candidate answers extracted from the various documents. Because simply adding scores degrades the performance of questionanswering systems, we add scores with decreasing weights to reduce the negative effect of simple adding. We carried out experiments using the QAC1 test collection and confirmed the effectiveness of our method through a statistical test. Our method produced a large improvement, with values of 0.05 to 0.14 for the evaluation scores (MRR/MF). We also obtained relatively good results in the experiments using the QAC2 test collection. These results, and the fact that our method is very simple and easy to use, demonstrate its feasibility and utility for question-answering systems.

Sentiment classification has been crucial for many natural language processing (NLP) applications, such as the analysis of movie reviews, tweets, or customer feedback. A sufficiently large amount of data is required to build a robust sentiment classification system. However, such resources are not always available for all domains or for all languages. In this work, we propose employing a machine translation (MT) system to translate customer feedback into another language to investigate in which cases translated sentences can have a positive or negative impact on an automatic sentiment classifier. Furthermore, as performing a direct translation is not always possible, we explore the performance of automatic classifiers on sentences that have been translated using a pivot MT system. We conduct several experiments using the above approaches to analyse the performance of our proposed sentiment classification system and discuss the advantages and drawbacks of classifying translated sente...

We will report evaluation of Automatic Named Entity Extraction feature of IR tools on Dutch, French, and English text. The aim is to analyze the competency of off-the-shelf information extraction tools in recognizing entity types including person, organization, location, vehicle, time, & currency from unstructured text. Within such an evaluation one can compare the effectiveness of different approaches for identifying named entities.

We will report evaluation of Automatic Named Entity Extraction feature of IR tools on Dutch, French, and English text. The aim is to analyze the competency of off-the-shelf information extraction tools in recognizing entity types including person, organization, location, vehicle, time, & currency from unstructured text. Within such an evaluation one can compare the effectiveness of different approaches for identifying named entities.

Transformer-based language models have been changing the modern Natural Language Processing (NLP) landscape for high-resource languages such as English, Chinese, Russian, etc. However, this technology does not yet exist for any Ghanaian language. In this paper, we introduce the first of such models for Twi or Akan, the most widely spoken Ghanaian language. The specific contribution of this research work is the development of several pretrained transformer language models for the Akuapem and Asante dialects of Twi, paving the way for advances in application areas such as Named Entity Recognition (NER), Neural Machine Translation (NMT), Sentiment Analysis (SA) and Part-of-Speech (POS) tagging. Specifically, we introduce four different flavours of ABENA – A BERT model Now in Akan that is fine-tuned on a set of Akan corpora, and BAKO BERT with Akan Knowledge only, which is trained from scratch. We open-source the model through the Hugging Face model hub and demonstrate its use via a sim...

GeoCLEF is an evaluation initiative for testing queries with a geographic specification in large set of text documents. GeoCLEF ran a regular track for the third time within the Cross Language Evaluation Forum (CLEF) 2008. The purpose of GeoCLEF is to test and evaluate cross-language geographic information retrieval (GIR). GeoCLEF 2008 consisted of two sub tasks. A search task ran for the third time and a Wikipedia pilot task (GikiP) was organized for the first time. For the GeoCLEF 2008 search task, twenty-five search topics were defined by the organizing groups for searching English, German and Portuguese document collections. Topics were developed also for English, German and Portuguese. Many topics were geographically challenging. Eleven groups submitted 131 runs. The groups used a variety of approaches, including sample documents, named entity extraction and ontology based retrieval. The evaluation methodology and results are presented in the paper.

Development of m odern Cadastral Information Systems (CIS) requires deployment of tools for automatic estimation of real estates' value which is influenced by a number of factors. After differentiation of the factors, apropriate information on certain locations needs to be acquired. Since most up-to-date information is transmited mainly as free-text documents via online media, information extraction technology plays a key role in converting such data into valuable and structured knowledge, which faciliates automatic real-estate value estimation. This article reports on creation of a corpus of Polish free-text documents, tagged with name mentions of CIS-relevant entities, which constitutes a core resource for development and evaluation of i nformation extraction components used within a cadastre framework.

Two salient properties of user behavior make Help Desk a unique speech application different from the more general transactional kind: (a) majority of users have only vague ideas about their problem, and (b) these users are likely to contextswitch (change discourse topic) during the course of a dialog. We describe a conversational Help Desk natural language call routing application and show how the alignment of Voice User Interface (VUI), Grammar Development, and Application Architecture results in a conversational user interface that is able to guide the vague user in the most optimal way while being flexible to allow mid-discourse context switches. Usability evaluation confirms the peculiar user behavior and provides empirical evidence that user's perception of time in a speech application can be influenced by the dialog; in this case, Help Desk users' tend not to become impatient going through threefive dialog turns, as long as dialog is progressing toward problem-resolution.

A framework is proposed for enterprise automated call routing system development and large scalable natural language call routing application deployment based on IBM's speech recognition and NLU application engagement practices in recently years. To facilitate employing different call classification algorithms in an easy integration manner, this framework architecture provides a plug & play environment for evaluating promising call routing algorithms and a systematic approach to carry out a large scalable enterprise application deployment. The paradigm in this paper illustrates the complementary effort to develop an automatic call routing application for enterprise call centers and covers from call classification algorithm investigation to application programming model. Experimental results on a live data testing set collected from an enterprise call center shows that the performance of the call classification algorithm implemented in this framework is outstanding.

Two salient properties of user behavior make Help Desk a unique speech application different from the more general transactional kind: (a) majority of users have only vague ideas about their problem, and (b) these users are likely to contextswitch (change discourse topic) during the course of a dialog. We describe a conversational Help Desk natural language call routing application and show how the alignment of Voice User Interface (VUI), Grammar Development, and Application Architecture results in a conversational user interface that is able to guide the vague user in the most optimal way while being flexible to allow mid-discourse context switches. Usability evaluation confirms the peculiar user behavior and provides empirical evidence that user's perception of time in a speech application can be influenced by the dialog; in this case, Help Desk users' tend not to become impatient going through threefive dialog turns, as long as dialog is progressing toward problem-resolution.

A framework is proposed for enterprise automated call routing system development and large scalable natural language call routing application deployment based on IBM's speech recognition and NLU application engagement practices in recently years. To facilitate employing different call classification algorithms in an easy integration manner, this framework architecture provides a plug & play environment for evaluating promising call routing algorithms and a systematic approach to carry out a large scalable enterprise application deployment. The paradigm in this paper illustrates the complementary effort to develop an automatic call routing application for enterprise call centers and covers from call classification algorithm investigation to application programming model. Experimental results on a live data testing set collected from an enterprise call center shows that the performance of the call classification algorithm implemented in this framework is outstanding.

Named entity recognition (NER) is the process of seeking to locate atomic elements in text into predefined categories such as the names of persons, organizations and locations. Most existing NER systems are based on supervised learning. This method often requires a large amount of labelled training data, which is very time-consuming to build. To solve this problem, we introduce a semi-supervised learning method for recognizing named entities in Vietnamese text by combining proper name coreference, named-ambiguity heuristics with a powerful sequential learning model, Conditional Random Fields. Our approach inherits the idea of Liao and Veeramachaneni [6] and expands it by using proper name coreference. Starting by training the model using a small data set that is annotated manually, the learning model extracts high confident named entities and finds low confident ones by using proper name coreference rules. The low confident named entities are put in the training set to learn new context features. The F-scores of the system for extracting "Person", "Location" and "Organization" entities are 83.36%, 69.53% and 65.71% when applying heuristics proposed by Liao and Veeramachaneni. Those values when using our proposed heuristics are 93.13%, 88.15% and 79.35%, respectively. It shows that our method is good in increasing the system accuracy.

Named entity recognition (NER) is the process of seeking to locate atomic elements in text into predefined categories such as the names of persons, organizations and locations. Most existing NER systems are based on supervised learning. This method often requires a large amount of labelled training data, which is very time-consuming to build. To solve this problem, we introduce a semi-supervised learning method for recognizing named entities in Vietnamese text by combining proper name coreference, named-ambiguity heuristics with a powerful sequential learning model, Conditional Random Fields. Our approach inherits the idea of Liao and Veeramachaneni [6] and expands it by using proper name coreference. Starting by training the model using a small data set that is annotated manually, the learning model extracts high confident named entities and finds low confident ones by using proper name coreference rules. The low confident named entities are put in the training set to learn new context features. The F-scores of the system for extracting "Person", "Location" and "Organization" entities are 83.36%, 69.53% and 65.71% when applying heuristics proposed by Liao and Veeramachaneni. Those values when using our proposed heuristics are 93.13%, 88.15% and 79.35%, respectively. It shows that our method is good in increasing the system accuracy.

In course of a breaking news event, such as natural calamity, political uproar etc., a massive crowd sourced data is generated over social media which makes social media platforms an important source of information in such scenarios. The value of the information being propagated via social media is being increasingly realised by the news organisations and the journalists. Better tools and methodologies are needed to facilitate them in utilising this information for news production. A lot of analysis over social media, by the journalists, is performed via rigorous manual labour. However, the sheer volume of the data produced on social media is overwhelming and acts as a major obstacle for manual inspection of the streaming data for finding, aggregating and contextualising the emerging event in a short time span. This is a day-today challenge for journalists and media organisations. This paper addresses the above problem for journalist in handling the voluminous social media data, viewing it from an information retrieval perspective, by proposing an 'event detection and contextualisation' framework that processes an input stream of social media data into the clusters of likely events.

Los problemas que presentan los modelos neuronales de procesamiento del lenguaje y la representación del significado derivan de dos problemas principales: el problema del 'binding' y el problema de la composicionalidad. A su vez estos dos problemas derivan del problema de cómo representar estímulos complejos con estructura interna, como son las oraciones, mediante vectores de activación. En este artículo presentamos un modelo neuronal de procesamiento de lenguaje (ANNLP) que resuelve estos dos problemas y cuyo rendimiento es muy eficiente procesando textos reales para un amplio abanico de tareas que incluyen la desambiguación del sentido en contextos proposicionales, la desambiguación sintáctica y el parsing. Para conseguir este nivel de eficiencia hemos tenido que dotar al modelo de un conjunto de características cuyo elemento central reside en la propuesta de cómo computar estructura (estructura sintáctica y de significado) mediante vectores de activación. Se aporta evidencia a favor de que el modelo y en concreto la propuesta de cómo representar estructura no es una simple solución ingenieril a los problemas del binding y la composicionalidad, sino que es plausible psicológicamente.

In this paper, we describe our approach for Named Entity rEcognition and Linking Challenge (NEEL) at the #Microposts2016. The task is to automatically recognize entities and their types from English microposts, and link them to corresponding DBpedia 2015 entries. If the resources do not exist, we use NIL identifiers instead. The task is unique as twitter data is informal in nature with non-conformational spellings, random contractions and various other noises. For this task, we developed our system using a hybrid model. We have used various existing named entity recognition (NER) systems and combined them with our classifier to improve the results.

This paper presents an approach for constructing an ontology from a stream of documents. Named entities extracted from the documents are used as instances of the ontology. Entities and co-occurring entity pairs are represented by feature vectors based on the content of the documents where they occurred. In general, concepts and relations can be formed into an ontological structure either by clustering or by classification into an existing topic hierarchy. We propose the latter using DMoz as an existing topic hierarchy. The approach is efficient and can scale to large data sets. We propose a framework that incorporates the stream mining process into a formal definition of the ontology. We describe a software component implementing this approach, and present experiments using a large collection of news.

We propose a new method of using multiple documents as evidence with decreased adding to improve the performance of question-answering systems.Sometimes,the answer to a question may be found in multiple documents.In such cases,using multiple documents to predict answers may generate better answers than using a single document.Our method therefore uses information from multiple documents,adding the scores of candidate answers extracted from various documents.However,because simply adding the scores can degrade the performance of question-answering systems, we add the scores with progressively decreasing weights to reduce the negative effect of simple adding.We carried out experiments using the Question-Answering Challenge(QAC)test collection.The results showed that our method produced a statistically significant improvement,with the degree of improvement ranging fro 0.05 to 0.14.These results,and the fact that our method is simple and easy to use,indicate its potential feasibility and utility in question-answering systems.Experiments comparing our decreased adding method with several previously proposed methods that use multiple documents showed that our method was more effective than these other methods.

Named entity recognition (NER) is the process of seeking to locate atomic elements in text into predefined categories such as the names of persons, organizations and locations. Most existing NER systems are based on supervised learning. This method often requires a large amount of labelled training data, which is very time-consuming to build. To solve this problem, we introduce a semi-supervised learning method for recognizing named entities in Vietnamese text by combining proper name coreference, named-ambiguity heuristics with a powerful sequential learning model, Conditional Random Fields. Our approach inherits the idea of Liao and Veeramachaneni [6] and expands it by using proper name coreference. Starting by training the model using a small data set that is annotated manually, the learning model extracts high confident named entities and finds low confident ones by using proper name coreference rules. The low confident named entities are put in the training set to learn new context features. The F-scores of the system for extracting "Person", "Location" and "Organization" entities are 83.36%, 69.53% and 65.71% when applying heuristics proposed by Liao and Veeramachaneni. Those values when using our proposed heuristics are 93.13%, 88.15% and 79.35%, respectively. It shows that our method is good in increasing the system accuracy.

Classification of crisis events, such as natural disasters, terrorist attacks and pandemics, is a crucial task to create early signals and inform relevant parties for spontaneous actions to reduce overall damage. Despite the crises, such as natural disaster, that can be predicted by professional institutions, certain events are first signaled by everyday citizens, i.e., civilians, such as the recent COVID-19 pandemics. Social media platforms such as Twitter often expose firsthand signals on such crises through high volume information exchange. In the case of Twitter, this corresponds to on average over half a billion tweets posted daily. Prior works proposed various crisis embeddings and classification using conventional Machine Learning and Neural Network models. However, none of the works perform crisis embedding and classification using state of the art attention-based deep neural networks models, such as Transformers and document-level contextual embeddings. This work proposes CrisisBERT, an end-to-end transformer-based model for two crisis classification tasks, namely crisis detection and crisis recognition, which shows promising results across accuracy and f1 scores. The proposed model demonstrates superior robustness over various benchmarks, as it shows marginal performance compromise while extending from 6 to 36 events with only 51.4% additional data points. We also propose Crisis2Vec, an attention-based, document-level contextual embedding architecture for crisis embedding, which achieves better performance than conventional crisis embedding methods such as Word2Vec and GloVe. To the best of our knowledge, our works are first to propose using transformer-based crisis classification and document-level contextual crisis embedding in the literature.