Many research topics in natural language processing (NLP), such as explanation generation, dialog modeling, or machine translation, require evaluation that goes beyond standard metrics like accuracy or F1 score toward a more human-centered approach. Therefore, understanding how to design user studies becomes increasingly important. However, few comprehensive resources exist on planning, conducting, and evaluating user studies for NLP, making it hard to get started for researchers without prior experience in the field of human evaluation. In this paper, we summarize the most important aspects of user studies and their design and evaluation, providing direct links to NLP tasks and NLP-specific challenges where appropriate. We (i) outline general study design, ethical considerations, and factors to consider for crowdsourcing, (ii) discuss the particularities of user studies in NLP, and provide starting points to select questionnaires, experimental designs, and evaluation methods that are tailored to the specific NLP tasks. Additionally, we offer examples with accompanying statistical evaluation code, to bridge the gap between theoretical guidelines and practical applications.

Paraphrase corpora are an essential but scarce resource in Natural Language Processing. In this paper, we present the Wikipedia-based Relational Paraphrase Acquisition (WRPA) method, which extracts relational paraphrases from Wikipedia, and the derived WRPA paraphrase corpus. The WRPA corpus currently covers person-related and authorship relations in English and Spanish, respectively, suggesting that, given adequate Wikipedia coverage, our method is independent of the language and the relation addressed. WRPA extracts entity pairs from structured information in Wikipedia applying distant learning and, based on the distributional hypothesis, uses them as anchor points for candidate paraphrase extraction from the free text in the body of Wikipedia articles. Focussing on relational paraphrasing and taking advantage of Wikipedia-structured information allows for an automatic and consistent evaluation of the results. The WRPA corpus characteristics distinguish it from other types of corpora that rely on string similarity or transformation operations. WRPA relies on distributional similarity and is the result of the free use of language outside any reformulation framework. Validation results show a high precision for the corpus.

Named Entity Recognition is a basic task in Information Extraction that aims at identifying entities of interest within full text documents. The patterns used to recognize entities can be rule based, as in the popular JAPE system. However, hand-crafting effective patterns is often difficult, and yet there is little research devoted to methods capable of learning human-readable patterns, possibly with arbitrary sets of features. In this paper, we present a semi-automatic method to generate both regular expressions and a subset of the JAPE language. It does not need a corpus annotated beforehand. Instead, it employs active learning and combines clustering with an algorithm that finds alignments between symbols present in the entities discovered during the learning process. The method currently supports a fixed set of character features and an arbitrary set of token features, but it can incorporate other kinds of features as well. Through several experiments with an English corpus, we show the ability of the method to generate effective patterns at a low annotation cost, and how it can successfully help in the annotation of brand new corpora.

Finding translations for technical terms is an important problem in machine translation. In particular, in highly specialized domains such as biology or medicine, it is difficult to find bilingual experts to annotate sufficient cross-lingual texts in order to train machine translation systems. Moreover, new terms are constantly being generated in the biomedical community, which makes it difficult to keep the translation dictionaries up to date for all language pairs of interest. Given a biomedical term in one language (source language), we propose a method for detecting its translations in a different language (target language). Specifically, we train a binary classifier to determine whether two biomedical terms written in two languages are translations. Training such a classifier is often complicated due to the lack of common features between the source and target languages. We propose several feature space concatenation methods to successfully overcome this problem. Moreover, we study the effectiveness of contextual and character n-gram features for detecting term translations. Experiments conducted using a standard dataset for biomedical term translation show that the proposed method outperforms several competitive baseline methods in terms of mean average precision and top-k translation accuracy.

Generic software architectures aim to support re-use of components, focusing of research and development effort, and evaluation and comparison of approaches. In the field of natural language processing, generic frameworks for understanding have been successfully deployed to meet all of these aims, but nothing comparable yet exists for generation. The nature of the task itself, and the current methodologies available to research it, seem to make it more difficult to reach the necessary level of consensus to support generic proposals. Recent work has made progress towards establishing a generic framework for generation at the functional level, but left open the issue of actual implementation. In this paper, we discuss the requirements for such an implementation layer for generation systems, drawing on two initial attempts to implement it. We argue that it is possible and useful to distinguish “functional architecture” from “implementation architecture” for generation systems.

This paper addresses the feasibility of cross-lingual parsing with Universal Dependencies (UD) between Romance languages, analyzing its performance when compared to the use of manually annotated resources of the target languages. Several experiments take into account factors such as the lexical distance between the source and target varieties, the impact of delexicalization, the combination of different source treebanks or the adaptation of resources to the target language, among others. The results of these evaluations show that the direct application of a parser from one Romance language to another reaches similar labeled attachment score (LAS) values to those obtained with a manual annotation of about 3,000 tokens in the target language, and unlabeled attachment score (UAS) results equivalent to the use of around 7,000 tokens, depending on the case. These numbers can noticeably increase by performing a focused selection of the source treebanks. Furthermore, the removal of the words in the training corpus (delexicalization) is not useful in most cases of cross-lingual parsing of Romance languages. The lessons learned with the performed experiments were used to build a new UD treebank for Galician, with 1,000 sentences manually corrected after an automatic cross-lingual annotation. Several evaluations in this new resource show that a cross-lingual parser built with the best combination and adaptation of the source treebanks performs better (77 percent LAS and 82 percent UAS) than using more than 16,000 (for LAS results) and more than 20,000 (UAS) manually labeled tokens of Galician.

The paper presents experiments on part-of-speech and full morphological tagging of the Slavic minority language Rusyn. The proposed approach relies on transfer learning and uses only annotated resources from related Slavic languages, namely Russian, Ukrainian, Slovak, Polish, and Czech. It does not require any annotated Rusyn training data, nor parallel data or bilingual dictionaries involving Rusyn. Compared to earlier work, we improve tagging performance by using a neural network tagger and larger training data from the neighboring Slavic languages. We experiment with various data preprocessing and sampling strategies and evaluate the impact of multitask learning strategies and of pretrained word embeddings. Overall, while genre discrepancies between training and test data have a negative impact, we improve full morphological tagging by 9% absolute micro-averaged F1 as compared to previous research.

Policy learning is an active topic in dialogue systems research, but it has not been explored in relation to interactive question answering (IQA). We take a first step in learning adaptive interaction policies for question answering : we address the question of how to acquire enough reliable query constraints, how many database results to present to the user and when to present them, given the competing trade-offs between the length of the answer list, the length of the interaction, the type of database and the noise in the communication channel. The operating conditions are reflected in an objective function which we use to derive a hand-coded threshold-based policy and rewards to train a reinforcement learning policy. The same objective function is used for evaluation. We show that we can learn strategies for this complex trade-off problem which perform significantly better than a variety of hand-coded policies, for a wide range of noise conditions, user types, types of DB and turn-penalties. Our policy learning framework thus covers a wide spectrum of operating conditions. The learned policies produce an average relative increase in reward of 86.78% over the hand-coded policies. In 93% of the cases the learned policies perform significantly better than the hand-coded ones (p < .001). Furthermore we show that the type of database has a significant effect on learning and we give qualitative descriptions of the learned IQA policies.

Recently, most part-of-speech tagging approaches, such as rule-based, probabilistic and neural network approaches, have shown very promising results. In this paper, we are particularly interested in probabilistic approaches, which usually require lots of training data to get reliable probabilities. We alleviate such a restriction of probabilistic approaches by introducing a fuzzy network model to provide a method for estimating more reliable parameters of a model under a small amount of training data. Experiments with the Brown corpus show that the performance of the fuzzy network model is much better than that of the hidden Markov model under a limited amount of training data.

The wide usage of multiple spoken Arabic dialects on social networking sites stimulates increasing interest in Natural Language Processing (NLP) for dialectal Arabic (DA). Arabic dialects represent true linguistic diversity and differ from modern standard Arabic (MSA). In fact, the complexity and variety of these dialects make it insufficient to build one NLP system that is suitable for all of them. In comparison with MSA, the available datasets for various dialects are generally limited in terms of size, genre and scope. In this article, we present a novel approach that automatically develops an annotated country-level dialectal Arabic corpus and builds lists of words that encompass 15 Arabic dialects. The algorithm uses an iterative procedure consisting of two main components: automatic creation of lists for dialectal words and automatic creation of annotated Arabic dialect identification corpus. To our knowledge, our study is the first of its kind to examine and analyse the poor performance of the MSA part-of-speech tagger on dialectal Arabic contents and to exploit that in order to extract the dialectal words. The pointwise mutual information association measure and the geographical frequency of word occurrence online are used to classify dialectal words. The annotated dialectal Arabic corpus (Twt15DA), built using our algorithm, is collected from Twitter and consists of 311,785 tweets containing 3,858,459 words in total. We randomly selected a sample of 75 tweets per country, 1125 tweets in total, and conducted a manual dialect identification task by native speakers. The results show an average inter-annotator agreement score equal to 64%, which reflects satisfactory agreement considering the overlapping features of the 15 Arabic dialects.

We present the RAGS (Reference Architecture for Generation Systems) framework, a specification of an abstract Natural Language Generation (NLG) system architecture to support sharing, re-use, comparison and evaluation of NLG technologies. We argue that the evidence from a survey of actual NLG systems calls for a different emphasis in a reference proposal from that seen in similar initiatives in information extraction and multimedia interfaces. We introduce the framework itself, in particular the two-level data model that allows us to support the complex data requirements of NLG systems in a flexible and coherent fashion, and describe our efforts to validate the framework through a range of implementations.

This paper describes the key aspects of a parser developed at the University of Pennsylvania from 1958 to 1959. The parser is essentially a cascade of finite state transducers. To the best of our knowledge, this is the first application of finite state transducers to parsing. This parser was recently faithfully reconstructed from the original documentation. Many aspects of this program have a close relationship to some of the recent work on finite state transducers.

In training deep learning networks, the optimizer and related learning rate are often used without much thought or with minimal tuning, even though it is crucial in ensuring a fast convergence to a good quality minimum of the loss function that can also generalize well on the test dataset. Drawing inspiration from the successful application of cyclical learning rate policy to computer vision tasks, we explore how cyclical learning rate can be applied to train transformer-based neural networks for neural machine translation. From our carefully designed experiments, we show that the choice of optimizers and the associated cyclical learning rate policy can have a significant impact on the performance. In addition, we establish guidelines when applying cyclical learning rates to neural machine translation tasks.

Alleviating pain is good and abandoning hope is bad. We instinctively understand how words like alleviate and abandon affect the polarity of a phrase, inverting or weakening it. When these words are content words, such as verbs, nouns, and adjectives, we refer to them as polarity shifters. Shifters are a frequent occurrence in human language and an important part of successfully modeling negation in sentiment analysis; yet research on negation modeling has focused almost exclusively on a small handful of closed-class negation words, such as not, no, and without. A major reason for this is that shifters are far more lexically diverse than negation words, but no resources exist to help identify them. We seek to remedy this lack of shifter resources by introducing a large lexicon of polarity shifters that covers English verbs, nouns, and adjectives. Creating the lexicon entirely by hand would be prohibitively expensive. Instead, we develop a bootstrapping approach that combines automatic classification with human verification to ensure the high quality of our lexicon while reducing annotation costs by over 70%. Our approach leverages a number of linguistic insights; while some features are based on textual patterns, others use semantic resources or syntactic relatedness. The created lexicon is evaluated both on a polarity shifter gold standard and on a polarity classification task.

Case systems abound in natural language processing. Almost any attempt to recognize and uniformly represent relationships within a clause – a unit at the centre of any linguistic system that goes beyond word level statistics – must be based on semantic roles drawn from a small, closed set. The set of roles describing relationships between a verb and its arguments within a clause is a case system. What is required of such a case system? How does a natural language practitioner build a system that is complete and detailed yet practical and natural? This paper chronicles the construction of a case system from its origin in English marker words to its successful application in the analysis of English text.

In this paper, we discuss the results of a new unsupervised and computationally lightweight scoring of how two words are morphologically related to each other. This measure is meant to be an alternative to stemming, radicals (root) extraction, and morphological analysis in a wide range of applications; especially information extraction related ones. Compared to light stemming, which seems to be the most convenient approach for systems with efficiency concerns, our measure does not neglect unconditionally a prefix or a suffix as the light stemming does. Instead, our measure takes into account all letters of the word but with different weights. This prevents the missing of a significant letter. Compared to heavy stemming, morphological analysis, or radicals extraction, which rely on dictionaries and compatibility databases, our measure does not rely on any language-specific morphology knowledge. This makes our approach unsupervised and theoretically language independent and computationally much lighter. Our tests targeted Arabic: a Semitic language recognized to have a complex morphology due to its highly inflectional lexicon.

‘Deep-syntactic’ dependency structures that capture the argumentative, attributive and coordinative relations between full words of a sentence have a great potential for a number of NLP-applications. The abstraction degree of these structures is in between the output of a syntactic dependency parser (connected trees defined over all words of a sentence and language-specific grammatical functions) and the output of a semantic parser (forests of trees defined over individual lexemes or phrasal chunks and abstract semantic role labels which capture the frame structures of predicative elements and drop all attributive and coordinative dependencies). We propose a parser that provides deep-syntactic structures. The parser has been tested on Spanish, English and Chinese.

Statistical parsers often require careful parameter tuning and feature selection. This is a nontrivial task for application developers who are not interested in parsing for its own sake, and it can be time-consuming even for experienced researchers. In this paper we present MaltOptimizer, a tool developed to automatically explore parameters and features for MaltParser, a transition-based dependency parsing system that can be used to train parser's given treebank data. MaltParser provides a wide range of parameters for optimization, including nine different parsing algorithms, an expressive feature specification language that can be used to define arbitrarily rich feature models, and two machine learning libraries, each with their own parameters. MaltOptimizer is an interactive system that performs parser optimization in three stages. First, it performs an analysis of the training set in order to select a suitable starting point for optimization. Second, it selects the best parsing algorithm and tunes the parameters of this algorithm. Finally, it performs feature selection and tunes machine learning parameters. Experiments on a wide range of data sets show that MaltOptimizer quickly produces models that consistently outperform default settings and often approach the accuracy achieved through careful manual optimization.

In this paper, we present a new method based on co-occurrence graphs for performing Cross-Lingual Word Sense Disambiguation (CLWSD). The proposed approach comprises the automatic generation of bilingual dictionaries, and a new technique for the construction of a co-occurrence graph used to select the most suitable translations from the dictionary. Different algorithms that combine both the dictionary and the co-occurrence graph are then used for performing this selection of the final translations: techniques based on sub-graphs (communities) containing clusters of words with related meanings, based on distances between nodes representing words, and based on the relative importance of each node in the whole graph. The initial output of the system is enhanced with translation probabilities, provided by a statistical bilingual dictionary. The system is evaluated using datasets from two competitions: task 3 of SemEval 2010, and task 10 of SemEval 2013. Results obtained by the different disambiguation techniques are analysed and compared to those obtained by the systems participating in the competitions. Our system offers the best results in comparison with other unsupervised systems in most of the experiments, and even overcomes supervised systems in some cases.

In machine-learning applications, data selection is of crucial importance if good runtime performance is to be achieved. In a scenario where the test set is accessible when the model is being built, training instances can be selected so they are the most relevant for the test set. Feature Decay Algorithms (FDA) are a technique for data selection that has demonstrated excellent performance in a number of tasks. This method maximizes the diversity of the n-grams in the training set by devaluing those ones that have already been included. We focus on this method to undertake deeper research on how to select better training data instances. We give an overview of FDA and propose improvements in terms of speed and quality. Using German-to-English parallel data, first we create a novel approach that decreases the execution time of FDA when multiple computation units are available. In addition, we obtain improvements on translation quality by extending FDA using information from the parallel corpus that is generally ignored.