Integrating and evaluating neural word embeddings in information retrieval

Recent advances in neural language models have contributed new methods for learning distributed vector representations of words (also called word embeddings). Two such methods are the continuous bag-of-words model and the skipgram model. These methods have been shown to produce embeddings that capture higher order relationships between words that are highly effective in natural language processing tasks involving the use of word similarity and word analogy. Despite these promising results, there has been little analysis of the use of these word embeddings for retrieval. Motivated by these observations, in this paper, we set out to determine how these word embeddings can be used within a retrieval model and what the benefit might be. To this aim, we use neural word embeddings within the well known translation language model for information retrieval. This language model captures implicit semantic relations between the words in queries and those in relevant documents, thus producing more accurate estimations of document relevance. The word embeddings used to estimate neural language models produce translations that differ from previous translation language model approaches; differences that deliver improvements in retrieval effectiveness. The models are robust to choices made in building word embeddings and, even more so, our results show that embeddings do not even need to be produced from the same corpus being used for retrieval.

Automatic ICD-10 classification of cancers from free-text death certificates

Objective Death certificates provide an invaluable source for cancer mortality statistics; however, this value can only be realised if accurate, quantitative data can be extracted from certificates – an aim hampered by both the volume and variable nature of certificates written in natural language. This paper proposes an automatic classification system for identifying cancer related causes of death from death certificates. Methods Detailed features, including terms, n-grams and SNOMED CT concepts were extracted from a collection of 447,336 death certificates. These features were used to train Support Vector Machine classifiers (one classifier for each cancer type). The classifiers were deployed in a cascaded architecture: the first level identified the presence of cancer (i.e., binary cancer/nocancer) and the second level identified the type of cancer (according to the ICD-10 classification system). A held-out test set was used to evaluate the effectiveness of the classifiers according to precision, recall and F-measure. In addition, detailed feature analysis was performed to reveal the characteristics of a successful cancer classification model. Results The system was highly effective at identifying cancer as the underlying cause of death (F-measure 0.94). The system was also effective at determining the type of cancer for common cancers (F-measure 0.7). Rare cancers, for which there was little training data, were difficult to classify accurately (F-measure 0.12). Factors influencing performance were the amount of training data and certain ambiguous cancers (e.g., those in the stomach region). The feature analysis revealed a combination of features were important for cancer type classification, with SNOMED CT concept and oncology specific morphology features proving the most valuable. Conclusion The system proposed in this study provides automatic identification and characterisation of cancers from large collections of free-text death certificates. This allows organisations such as Cancer Registries to monitor and report on cancer mortality in a timely and accurate manner. In addition, the methods and findings are generally applicable beyond cancer classification and to other sources of medical text besides death certificates.

Idealähtöistä koulualgebraa : IDEAA-opetusmallin kehittäminen algebran opetukseen peruskoulun 7. luokalla

This research is based on the problems in secondary school algebra I have noticed in my own work as a teacher of mathematics. Algebra does not touch the pupil, it remains knowledge that is not used or tested. Furthermore the performance level in algebra is quite low. This study presents a model for 7th grade algebra instruction in order to make algebra more natural and useful to students. I refer to the instruction model as the Idea-based Algebra (IDEAA). The basic ideas of this IDEAA model are 1) to combine children's own informal mathematics with scientific mathematics ("math math") and 2) to structure algebra content as a "map of big ideas", not as a traditional sequence of powers, polynomials, equations, and word problems. This research project is a kind of design process or design research. As such, this project has three, intertwined goals: research, design and pedagogical practice. I also assume three roles. As a researcher, I want to learn about learning and school algebra, its problems and possibilities. As a designer, I use research in the intervention to develop a shared artefact, the instruction model. In addition, I want to improve the practice through intervention and research. A design research like this is quite challenging. Its goals and means are intertwined and change in the research process. Theory emerges from the inquiry; it is not given a priori. The aim to improve instruction is normative, as one should take into account what "good" means in school algebra. An important part of my study is to work out these paradigmatic questions. The result of the study is threefold. The main result is the instruction model designed in the study. The second result is the theory that is developed of the teaching, learning and algebra. The third result is knowledge of the design process. The instruction model (IDEAA) is connected to four main features of good algebra education: 1) the situationality of learning, 2) learning as knowledge building, in which natural language and intuitive thinking work as "intermediaries", 3) the emergence and diversity of algebra, and 4) the development of high performance skills at any stage of instruction.

Complete and interpretable conformance checking of business processes

This article presents a method for checking the conformance between an event log capturing the actual execution of a business process, and a model capturing its expected or normative execution. Given a business process model and an event log, the method returns a set of statements in natural language describing the behavior allowed by the process model but not observed in the log and vice versa. The method relies on a unified representation of process models and event logs based on a well-known model of concurrency, namely event structures. Specifically, the problem of conformance checking is approached by folding the input event log into an event structure, unfolding the process model into another event structure, and comparing the two event structures via an error-correcting synchronized product. Each behavioral difference detected in the synchronized product is then verbalized as a natural language statement. An empirical evaluation shows that the proposed method scales up to real-life datasets while producing more concise and higher-level difference descriptions than state-of-the-art conformance checking methods.

Towards automated content analysis of discussion transcripts: A cognitive presence case

In this paper, we present the results of an exploratory study that examined the problem of automating content analysis of student online discussion transcripts. We looked at the problem of coding discussion transcripts for the levels of cognitive presence, one of the three main constructs in the Community of Inquiry (CoI) model of distance education. Using Coh-Metrix and LIWC features, together with a set of custom features developed to capture discussion context, we developed a random forest classification system that achieved 70.3% classification accuracy and 0.63 Cohen's kappa, which is significantly higher than values reported in the previous studies. Besides improvement in classification accuracy, the developed system is also less sensitive to overfitting as it uses only 205 classification features, which is around 100 times less features than in similar systems based on bag-of-words features. We also provide an overview of the classification features most indicative of the different phases of cognitive presence that gives an additional insights into the nature of cognitive presence learning cycle. Overall, our results show great potential of the proposed approach, with an added benefit of providing further characterization of the cognitive presence coding scheme.

Semantic Classes in Topic Detection and Tracking

Topic detection and tracking (TDT) is an area of information retrieval research the focus of which revolves around news events. The problems TDT deals with relate to segmenting news text into cohesive stories, detecting something new, previously unreported, tracking the development of a previously reported event, and grouping together news that discuss the same event. The performance of the traditional information retrieval techniques based on full-text similarity has remained inadequate for online production systems. It has been difficult to make the distinction between same and similar events. In this work, we explore ways of representing and comparing news documents in order to detect new events and track their development. First, however, we put forward a conceptual analysis of the notions of topic and event. The purpose is to clarify the terminology and align it with the process of news-making and the tradition of story-telling. Second, we present a framework for document similarity that is based on semantic classes, i.e., groups of words with similar meaning. We adopt people, organizations, and locations as semantic classes in addition to general terms. As each semantic class can be assigned its own similarity measure, document similarity can make use of ontologies, e.g., geographical taxonomies. The documents are compared class-wise, and the outcome is a weighted combination of class-wise similarities. Third, we incorporate temporal information into document similarity. We formalize the natural language temporal expressions occurring in the text, and use them to anchor the rest of the terms onto the time-line. Upon comparing documents for event-based similarity, we look not only at matching terms, but also how near their anchors are on the time-line. Fourth, we experiment with an adaptive variant of the semantic class similarity system. The news reflect changes in the real world, and in order to keep up, the system has to change its behavior based on the contents of the news stream. We put forward two strategies for rebuilding the topic representations and report experiment results. We run experiments with three annotated TDT corpora. The use of semantic classes increased the effectiveness of topic tracking by 10-30\% depending on the experimental setup. The gain in spotting new events remained lower, around 3-4\%. The anchoring the text to a time-line based on the temporal expressions gave a further 10\% increase the effectiveness of topic tracking. The gains in detecting new events, again, remained smaller. The adaptive systems did not improve the tracking results.

Methods for Answer Extraction in Textual Question Answering

In this thesis we present and evaluate two pattern matching based methods for answer extraction in textual question answering systems. A textual question answering system is a system that seeks answers to natural language questions from unstructured text. Textual question answering systems are an important research problem because as the amount of natural language text in digital format grows all the time, the need for novel methods for pinpointing important knowledge from the vast textual databases becomes more and more urgent. We concentrate on developing methods for the automatic creation of answer extraction patterns. A new type of extraction pattern is developed also. The pattern matching based approach chosen is interesting because of its language and application independence. The answer extraction methods are developed in the framework of our own question answering system. Publicly available datasets in English are used as training and evaluation data for the methods. The techniques developed are based on the well known methods of sequence alignment and hierarchical clustering. The similarity metric used is based on edit distance. The main conclusions of the research are that answer extraction patterns consisting of the most important words of the question and of the following information extracted from the answer context: plain words, part-of-speech tags, punctuation marks and capitalization patterns, can be used in the answer extraction module of a question answering system. This type of patterns and the two new methods for generating answer extraction patterns provide average results when compared to those produced by other systems using the same dataset. However, most answer extraction methods in the question answering systems tested with the same dataset are both hand crafted and based on a system-specific and fine-grained question classification. The the new methods developed in this thesis require no manual creation of answer extraction patterns. As a source of knowledge, they require a dataset of sample questions and answers, as well as a set of text documents that contain answers to most of the questions. The question classification used in the training data is a standard one and provided already in the publicly available data.

vSpeak: Edge detection based feature extraction for sign to text conversion

This paper presents 'vSpeak', the first initiative taken in Pakistan for ICT enabled conversion of dynamic Sign Urdu gestures into natural language sentences. To realize this, vSpeak has adopted a novel approach for feature extraction using edge detection and image compression which gives input to the Artificial Neural Network that recognizes the gesture. This technique caters for the blurred images as well. The training and testing is currently being performed on a dataset of 200 patterns of 20 words from Sign Urdu with target accuracy of 90% and above.

iPSL: Enabling rehabilitation of deaf community in Pakistan

This paper presents an initiative taken in Pakistan for the rehabilitation of the deaf community, enabled by the use of technology. iPSL is a system that primarily aims at facilitating communication between the hearing and the deaf community in Pakistan. There is a twofold approach to achieve this. The first dimension is to implement a system that can translate signs made by deaf into natural language sentences. The second dimension is to implement tools that enable hearing people to understand and learn sign language by converting natural language sentences into sign language. This paper presents the progress made in the project so far in terms of design, implementation and evaluation. © ACM 2009.

Mining topical relevant patterns for multi-document summarization

Multi-document summarization addressing the problem of information overload has been widely utilized in the various real-world applications. Most of existing approaches adopt term-based representation for documents which limit the performance of multi-document summarization systems. In this paper, we proposed a novel pattern-based topic model (PBTMSum) for the task of the multi-document summarization. PBTMSum combining pattern mining techniques with LDA topic modelling could generate discriminative and semantic rich representations for topics and documents so that the most representative and non-redundant sentences can be selected to form a succinct and informative summary. Extensive experiments are conducted on the data of document understanding conference (DUC) 2007. The results prove the effectiveness and efficiency of our proposed approach.

Effective 20 newsgroups dataset cleaning

The rapid increase in the number of text documents available on the Internet has created pressure to use effective cleaning techniques. Cleaning techniques are needed for converting these documents to structured documents. Text cleaning techniques are one of the key mechanisms in typical text mining application frameworks. In this paper, we explore the role of text cleaning in the 20 newsgroups dataset, and report on experimental results.

Overview of the 2015 ALTA shared task: Identifying French cognates in English text

This paper presents an overview of the 6th ALTA shared task that ran in 2015. The task was to identify in English texts all the potential cognates from the perspective of the French language. In other words, identify all the words in the English text that would acceptably translate into a similar word in French. We present the motivations for the task, the description of the data and the results of the 4 participating teams. We discuss the results against a baseline and prior work.

Perspektiv på språkprocessering

Automatisk språkprocessering har efter mer än ett halvt sekel av forskning blivit ett mycket viktigt område inom datavetenskapen. Flera vetenskapligt viktiga problem har lösts och praktiska applikationer har nått programvarumarknaden. Disambiguering av ord innebär att hitta rätt betydelse för ett mångtydigt ord. Sammanhanget, de omkringliggande orden och kunskap om ämnesområdet är faktorer som kan användas för att disambiguera ett ord. Automatisk sammanfattning innebär att förkorta en text utan att den relevanta informationen går förlorad. Relevanta meningar kan plockas ur texten, eller så kan en ny, kortare text genereras på basen av fakta i den ursprungliga texten. Avhandlingen ger en allmän översikt och kort historik av språkprocesseringen och jämför några metoder för disambiguering av ord och automatisk sammanfattning. Problemområdenas likheter och skillnader lyfts fram och metodernas ställning inom datavetenskapen belyses.

Linguistic grammars with very low complexity

We have presented an overview of the FSIG approach and related FSIG gram- mars to issues of very low complexity and parsing strategy. We ended up with serious optimism according to which most FSIG grammars could be decom- posed in a reasonable way and then processed efficiently.

Do wordnets also improve human performance on NLP tasks?

FinnWordNet is a wordnet for Finnish that complies with the format of the Princeton WordNet (PWN) (Fellbaum, 1998). It was built by translating the PrincetonWordNet 3.0 synsets into Finnish by human translators. It is open source and contains 117000 synsets. The Finnish translations were inserted into the PWN structure resulting in a bilingual lexical database. In natural language processing (NLP), wordnets have been used for infusing computers with semantic knowledge assuming that humans already have a sufficient amount of this knowledge. In this paper we present a case study of using wordnets as an electronic dictionary. We tested whether native Finnish speakers benefit from using a wordnet while completing English sentence completion tasks. We found that using either an English wordnet or a bilingual English Finnish wordnet significantly improves performance in the task. This should be taken into account when setting standards and comparing human and computer performance on these tasks.