In this paper, we take a detailed look at the performance of components of an idealized question answering system on two different tasks: the TREC Question Answering task and a set of reading comprehension exams. We carry out three types of analysis: inherent properties of the data, feature analysis, and performance bounds. Based on these analyses we explain some of the performance results of the current generation of Q/A systems and make predictions on future work. In particular, we present four findings: (1) Q/A system performance is correlated with answer repetition; (2) relative overlap scores are more effective than absolute overlap scores; (3) equivalence classes on scoring functions can be used to quantify performance bounds; and (4) perfect answer typing still leaves a great deal of ambiguity for a Q/A system because sentences often contain several items of the same type.

The previous Emerging Trends article (Church et al., 2021. Natural Language Engineering27(5), 631–645.) introduced deep nets to poets. Poets is an imperfect metaphor, intended as a gesture toward inclusion. The future for deep nets will benefit by reaching out to a broad audience of potential users, including people with little or no programming skills, and little interest in training models. That paper focused on inference, the use of pre-trained models, as is, without fine-tuning. The goal of this paper is to make fine-tuning more accessible to a broader audience. Since fine-tuning is more challenging than inference, the examples in this paper will require modest programming skills, as well as access to a GPU. Fine-tuning starts with a general purpose base (foundation) model and uses a small training set of labeled data to produce a model for a specific downstream application. There are many examples of fine-tuning in natural language processing (question answering (SQuAD) and GLUE benchmark), as well as vision and speech.

We propose a novel approach to developing a tractable affective dialogue model for probabilistic frame-based dialogue systems. The affective dialogue model, based on Partially Observable Markov Decision Process (POMDP) and Dynamic Decision Network (DDN) techniques, is composed of two main parts: the slot-level dialogue manager and the global dialogue manager. It has two new features: (1) being able to deal with a large number of slots and (2) being able to take into account some aspects of the user's affective state in deriving the adaptive dialogue strategies. Our implemented prototype dialogue manager can handle hundreds of slots, where each individual slot might have hundreds of values. Our approach is illustrated through a route navigation example in the crisis management domain. We conducted various experiments to evaluate our approach and to compare it with approximate POMDP techniques and handcrafted policies. The experimental results showed that the DDN–POMDP policy outperforms three handcrafted policies when the user's action error is induced by stress as well as when the observation error increases. Further, performance of the one-step look-ahead DDN–POMDP policy after optimizing its internal reward is close to state-of-the-art approximate POMDP counterparts.

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.

Current methods for assessing the impact of authors and scientific media employ tools such as H-Index, Co-Citation and PageRank. These tools are primarily based on citation counting, which considers all citations to be equal. This type of methods can produce perverse incentives to publish controversial or incomplete papers, as mixed or negative reviews often generate larger citation counts and better indexes, regardless of whether the citations were critical or exerted minimal influence on the citing document. Passing citations that are employed to establish background, which do not have a real impact on the citing paper, are common in scientific literature. However, these citations have equal weight in impact evaluations. Notable researchers have emphasized the need to correct this situation by developing estimation methods that consider the different roles of quotations in citing papers. To accomplish this type of evaluation, a context citation analysis should be applied to determine the nature of the citations. We propose that citations should be categorized using four dimensions – FUNCTION, POLARITY, ASPECTS and INFLUENCE – as these dimensions provide adequate information that can be employed toward the generation of a qualitative method to measure the impact of a given publication in a citing paper. In this paper, we used interchangeably the words influence and impact. We present a method for obtaining this information using our proposed classification scheme and manually annotated corpus, which is marked with meaningful keywords and labels to help identify the characteristics or properties that constitute what we call ASPECTS. We develop a classification scheme which considers purpose definition shared by previous works. Our contribution is to abstract purpose classes from several other schemes and divide a complex structure in more manageable parts, to attain a simple system that combines low granularity dimensions but nevertheless produces a fine-grained classification. For annotators, the classification process is simple because in a first step, the coders distinguish only four primary classes, and in a second pass, they add the information contained in ASPECTS keyword and labels to obtain the more specific functions. This way, we gain a high granularity labeling that gives enough information about the citations to characterize and classify them, and we achieve this detailed coding with a straightforward process where the level of human error could be minimized.

This paper introduces deep neural network (DNN)–hidden Markov model (HMM)-based methods to tackle speech recognition in heterogeneous groups of speakers including children. We target three speaker groups consisting of children, adult males and adult females. Two different kind of approaches are introduced here: approaches based on DNN adaptation and approaches relying on vocal-tract length normalisation (VTLN). First, the recent approach that consists in adapting a general DNN to domain/language specific data is extended to target age/gender groups in the context of DNN–HMM. Then, VTLN is investigated by training a DNN–HMM system by using either mel frequency cepstral coefficients normalised with standard VTLN or mel frequency cepstral coefficients derived acoustic features combined with the posterior probabilities of the VTLN warping factors. In this later, novel, approach the posterior probabilities of the warping factors are obtained with a separate DNN and the decoding can be operated in a single pass when the VTLN approach requires two decoding passes. Finally, the different approaches presented here are combined to take advantage of their complementarity. The combination of several approaches is shown to improve the baseline phone error rate performance by thirty per cent to thirty-five per cent relative and the baseline word error rate performance by about ten per cent relative.

Graphs are a powerful representation formalism that can be applied to a variety of aspects related to language processing. We provide an overview of how Natural Language Processing problems have been projected into the graph framework, focusing in particular on graph construction – a crucial step in modeling the data to emphasize the phenomena targeted.

We describe an opinion mining system which classifies the polarity of Spanish texts. We propose an NLP approach that undertakes pre-processing, tokenisation and POS tagging of texts to then obtain the syntactic structure of sentences by means of a dependency parser. This structure is then used to address three of the most significant linguistic constructions for the purpose in question: intensification, subordinate adversative clauses and negation. We also propose a semi-automatic domain adaptation method to improve the accuracy of our system in specific application domains, by enriching semantic dictionaries using machine learning methods in order to adapt the semantic orientation of their words to a particular field. Experimental results are promising in both general and specific domains.

Text preprocessing is not only an essential step to prepare the corpus for modeling but also a key area that directly affects the natural language processing (NLP) application results. For instance, precise tokenization increases the accuracy of part-of-speech (POS) tagging, and retaining multiword expressions improves reasoning and machine translation. The text corpus needs to be appropriately preprocessed before it is ready to serve as the input to computer models. The preprocessing requirements depend on both the nature of the corpus and the NLP application itself, that is, what researchers would like to achieve from analyzing the data. Conventional text preprocessing practices generally suffice, but there exist situations where the text preprocessing needs to be customized for better analysis results. Hence, we discuss the pros and cons of several common text preprocessing methods: removing formatting, tokenization, text normalization, handling punctuation, removing stopwords, stemming and lemmatization, n-gramming, and identifying multiword expressions. Then, we provide examples of text datasets which require special preprocessing and how previous researchers handled the challenge. We expect this article to be a starting guideline on how to select and fine-tune text preprocessing methods.

Designing models for learning textual entailment recognizers from annotated examples is not an easy task, as it requires modeling the semantic relations and interactions involved between two pairs of text fragments. In this paper, we approach the problem by first introducing the class of pair feature spaces, which allow supervised machine learning algorithms to derive first-order rewrite rules from annotated examples. In particular, we propose syntactic and shallow semantic feature spaces, and compare them to standard ones. Extensive experiments demonstrate that our proposed spaces learn first-order derivations, while standard ones are not expressive enough to do so.

In this paper, we describe NKRL (Narrative Knowledge Representation Language), a language designed for representing, in a standardized way, the semantic content (the ‘meaning’) of complex narrative texts. After having introduced informally the four ‘components’ (specialized sub-languages) of NKRL, we will describe (some of) the data structures proper to each of them, trying to show that the NKRL coding retains the main informational elements of the original narrative expressions. We will then focus on an important subset of NKRL, the so-called AECS sub-language, showing in particular that the operators of this sub-language can be used to represent some sorts of ‘plural’ expressions.

This paper presents a very simple and effective approach to using parallel corpora for automatic bilingual lexicon acquisition. The approach, which uses the Random Indexing vector space methodology, is based on finding correlations between terms based on their distributional characteristics. The approach requires a minimum of preprocessing and linguistic knowledge, and is efficient, fast and scalable. In this paper, we explain how our approach differs from traditional cooccurrence-based word alignment algorithms, and we demonstrate how to extract bilingual lexica using the Random Indexing approach applied to aligned parallel data. The acquired lexica are evaluated by comparing them to manually compiled gold standards, and we report overlap of around 60%. We also discuss methodological problems with evaluating lexical resources of this kind.

Over the last two decades, determining the similarity between words as well as between their meanings, that is, word senses, has been proven to be of vital importance in the field of Natural Language Processing. This paper provides the reader with an introduction to the tasks of computing word and sense similarity. These consist in computing the degree of semantic likeness between words and senses, respectively. First, we distinguish between two major approaches: the knowledge-based approaches and the distributional approaches. Second, we detail the representations and measures employed for computing similarity. We then illustrate the evaluation settings available in the literature and, finally, discuss suggestions for future research.

Being able to identify which rhetorical relations (e.g., contrast or explanation) hold between spans of text is important for many natural language processing applications. Using machine learning to obtain a classifier which can distinguish between different relations typically depends on the availability of manually labelled training data, which is very time-consuming to create. However, rhetorical relations are sometimes lexically marked, i.e., signalled by discourse markers (e.g., because, but, consequently etc.), and it has been suggested (Marcu and Echihabi, 2002) that the presence of these cues in some examples can be exploited to label them automatically with the corresponding relation. The discourse markers are then removed and the automatically labelled data are used to train a classifier to determine relations even when no discourse marker is present (based on other linguistic cues such as word co-occurrences). In this paper, we investigate empirically how feasible this approach is. In particular, we test whether automatically labelled, lexically marked examples are really suitable training material for classifiers that are then applied to unmarked examples. Our results suggest that training on this type of data may not be such a good strategy, as models trained in this way do not seem to generalise very well to unmarked data. Furthermore, we found some evidence that this behaviour is largely independent of the classifiers used and seems to lie in the data itself (e.g., marked and unmarked examples may be too dissimilar linguistically and removing unambiguous markers in the automatic labelling process may lead to a meaning shift in the examples).

Advances in language engineering may be dependent on theoretical principles originating from linguistics, since both share a common object of enquiry, natural language structures. We outline an approach to term extraction that rests on theoretical claims about the structure of words. We use the structural properties of compound words to specifically elicit the sets of terms defined by type hierarchies such as hyponymy and meronymy. The theoretical claims revolve around the head-modifier principle, which determines the formation of a major class of compounds. Significantly it has been suggested that the principle operates in languages other than English. To demonstrate the extendibility of our approach beyond English, we present a case study of term extraction in Chinese, a language whose written form is the vehicle of communication for over 1.3 billion language users, and therefore has great significance for the development of language engineering technologies.

In this article we illustrate and evaluate an approach to create high quality linguistically annotated resources based on the exploitation of aligned parallel corpora. This approach is based on the assumption that if a text in one language has been annotated and its translation has not, annotations can be transferred from the source text to the target using word alignment as a bridge. The transfer approach has been tested and extensively applied for the creation of the MultiSemCor corpus, an English/Italian parallel corpus created on the basis of the English SemCor corpus. In MultiSemCor the texts are aligned at the word level and word sense annotated with a shared inventory of senses. A number of experiments have been carried out to evaluate the different steps involved in the methodology and the results suggest that the transfer approach is one promising solution to the resource bottleneck. First, it leads to the creation of a parallel corpus, which represents a crucial resource per se. Second, it allows for the exploitation of existing (mostly English) annotated resources to bootstrap the creation of annotated corpora in new (resource-poor) languages with greatly reduced human effort.

Machine Reading Comprehension (MRC) is a challenging task and hot topic in Natural Language Processing. The goal of this field is to develop systems for answering the questions regarding a given context. In this paper, we present a comprehensive survey on diverse aspects of MRC systems, including their approaches, structures, input/outputs, and research novelties. We illustrate the recent trends in this field based on a review of 241 papers published during 2016–2020. Our investigation demonstrated that the focus of research has changed in recent years from answer extraction to answer generation, from single- to multi-document reading comprehension, and from learning from scratch to using pre-trained word vectors. Moreover, we discuss the popular datasets and the evaluation metrics in this field. The paper ends with an investigation of the most-cited papers and their contributions.

This paper describes two experiments: one exploring the amount of information relevant to sense disambiguation contained in the part-of-speech field of entries in a Machine Readable Dictionary (MRD); the other, more practical, experiment attempts sense disambiguation of all content words in a text assigning MRD homographs as sense tags using only part-of-speech information. We have implemented a simple sense tagger which successfully tags 94% of words using this method. A plan to extend this work and implement an improved sense tagger is included.

Crowd-sourced assessments of machine translation quality allow evaluations to be carried out cheaply and on a large scale. It is essential, however, that the crowd's work be filtered to avoid contamination of results through the inclusion of false assessments. One method is to filter via agreement with experts, but even amongst experts agreement levels may not be high. In this paper, we present a new methodology for crowd-sourcing human assessments of translation quality, which allows individual workers to develop their own individual assessment strategy. Agreement with experts is no longer required, and a worker is deemed reliable if they are consistent relative to their own previous work. Individual translations are assessed in isolation from all others in the form of direct estimates of translation quality. This allows more meaningful statistics to be computed for systems and enables significance to be determined on smaller sets of assessments. We demonstrate the methodology's feasibility in large-scale human evaluation through replication of the human evaluation component of Workshop on Statistical Machine Translation shared translation task for two language pairs, Spanish-to-English and English-to-Spanish. Results for measurement based solely on crowd-sourced assessments show system rankings in line with those of the original evaluation. Comparison of results produced by the relative preference approach and the direct estimate method described here demonstrate that the direct estimate method has a substantially increased ability to identify significant differences between translation systems.

This paper describes the Kestrel text normalization system, a component of the Google text-to-speech synthesis (TTS) system. At the core of Kestrel are text-normalization grammars that are compiled into libraries of weighted finite-state transducers (WFSTs). While the use of WFSTs for text normalization is itself not new, Kestrel differs from previous systems in its separation of the initial tokenization and classification phase of analysis from verbalization. Input text is first tokenized and different tokens classified using WFSTs. As part of the classification, detected semiotic classes – expressions such as currency amounts, dates, times, measure phases, are parsed into protocol buffers (https://code.google.com/p/protobuf/). The protocol buffers are then verbalized, with possible reordering of the elements, again using WFSTs. This paper describes the architecture of Kestrel, the protocol buffer representations of semiotic classes, and presents some examples of grammars for various languages. We also discuss applications and deployments of Kestrel as part of the Google TTS system, which runs on both server and client side on multiple devices, and is used daily by millions of people in nineteen languages and counting.