Subwords have become very popular, but the BERTa and ERNIEb tokenizers often produce surprising results. Byte pair encoding (BPE) trains a dictionary with a simple information theoretic criterion that sidesteps the need for special treatment of unknown words. BPE is more about training (populating a dictionary of word pieces) than inference (parsing an unknown word into word pieces). The parse at inference time can be ambiguous. Which parse should we use? For example, “electroneutral” can be parsed as electron-eu-tral or electro-neutral, and “bidirectional” can be parsed as bid-ire-ction-al and bi-directional. BERT and ERNIE tend to favor the parse with more word pieces. We propose minimizing the number of word pieces. To justify our proposal, a number of criteria will be considered: sound, meaning, etc. The prefix, bi-, has the desired vowel (unlike bid) and the desired meaning (bi is Latin for two, unlike bid, which is Germanic for offer).

The objective of this work is to set a corpus-driven methodology to quantify automatically diachronic language distance between chronological periods of several languages. We apply a perplexity-based measure to written text representing different historical periods of three languages: European English, European Portuguese, and European Spanish. For this purpose, we have built historical corpora for each period, which have been compiled from different open corpus sources containing texts as close as possible to its original spelling. The results of our experiments show that a diachronic language distance based on perplexity detects the linguistic evolution that had already been explained by the historians of the three languages. It is remarkable to underline that it is an unsupervised multilingual method which only needs a raw corpora organized by periods.

We compare the word sense disambiguation systems submitted for the English-all-words task in SENSEVAL-2. We give several performance measures for the systems, and analyze correlations between system performance and word features. A decision tree learning algorithm is employed to discover the situations in which systems perform particularly well, and the resulting decision tree is examined. We investigate using a decision tree based on the SENSEVAL systems to (i) filter out senses unlikely to be correct, and to (ii) combine WSD systems. Some combinations created in this way outperform the best SENSEVAL system.

In source reordering the order of the source words is permuted to minimize word order differences with the target sentence and then fed to a translation model. Earlier work highlights the benefits of resolving long-distance reorderings as a pre-processing step to standard phrase-based models. However, the potential performance improvement of source reordering and its impact on the components of the subsequent translation model remain unexplored. In this paper we study both aspects of source reordering. We set up idealized source reordering (oracle) models with/without syntax and present our own syntax-driven model of source reordering. The latter is a statistical model of inversion transduction grammar (ITG)-like tree transductions manipulating a syntactic parse and working with novel conditional reordering parameters. Having set up the models, we report translation experiments showing significant improvement on three language pairs, and contribute an extensive analysis of the impact of source reordering (both oracle and model) on the translation model regarding the quality of its input, phrase-table, and output. Our experiments show that oracle source reordering has untapped potential in improving translation system output. Besides solving difficult reorderings, we find that source reordering creates more monotone parallel training data at the back-end, leading to significantly larger phrase tables with higher coverage of phrase types in unseen data. Unfortunately, this nice property does not carry over to tree-constrained source reordering. Our analysis shows that, from the string-level perspective, tree-constrained reordering might selectively permute word order, leading to larger phrase tables but without increase in phrase coverage in unseen data.

Sentence embeddings are the suitable input vectors for the neural learning of a number of inferences about content and meaning. Similarity estimation, classification, emotional characterization of sentences as well as pragmatic tasks, such as question answering or dialogue, have largely demonstrated the effectiveness of vector embeddings to model semantics. Unfortunately, most of the above decisions are epistemologically opaque as for the limited interpretability of the acquired neural models based on the involved embeddings. We think that any effective approach to meaning representation should be at least epistemologically coherent. In this paper, we concentrate on the readability of neural models, as a core property of any embedding technique consistent and effective in representing sentence meaning. In this perspective, this paper discusses a novel embedding technique (the Nyström methodology) that corresponds to the reconstruction of a sentence in a kernel space, inspired by rich semantic similarity metrics (a semantic kernel) rather than by a language model. In addition to being based on a kernel that captures grammatical and lexical semantic information, the proposed embedding can be used as the input vector of an effective neural learning architecture, called Kernel-based deep architectures (KDA). Finally, it also characterizes by design the KDA explanatory capability, as the proposed embedding is derived from examples that are both human readable and labeled. This property is obtained by the integration of KDAs with an explanation methodology, called layer-wise relevance propagation (LRP), already proposed in image processing. The Nyström embeddings support here the automatic compilation of argumentations in favor or against a KDA inference, in form of an explanation: each decision can in fact be linked through LRP back to the real examples, that is, the landmarks linguistically related to the input instance. The KDA network output is explained via the analogy with the activated landmarks. Quantitative evaluation of the explanations shows that richer explanations based on semantic and syntagmatic structures characterize convincing arguments, as they effectively help the user in assessing whether or not to trust the machine decisions in different tasks, for example, Question Classification or Semantic Role Labeling. This confirms the epistemological benefit that Nyström embeddings may bring, as linguistically rich and meaningful representations for a variety of inference tasks.

Large-scale transformer-based language models (LMs) demonstrate impressive capabilities in open-text generation. However, controlling the generated text’s properties such as the topic, style, and sentiment is challenging and often requires significant changes to the model architecture or retraining and fine-tuning the model on new supervised data. This paper presents a novel approach for topical language generation (TLG) by combining a pre-trained LM with topic modeling information. We cast the problem using Bayesian probability formulation with topic probabilities as a prior, LM probabilities as the likelihood, and TLG probability as the posterior. In learning the model, we derive the topic probability distribution from the user-provided document’s natural structure. Furthermore, we extend our model by introducing new parameters and functions to influence the quantity of the topical features presented in the generated text. This feature would allow us to easily control the topical properties of the generated text. Our experimental results demonstrate that our model outperforms the state-of-the-art results on coherency, diversity, and fluency while being faster in decoding.

We present a lexical platform that has been developed for the Spanish language. It achieves portability between different computer systems and efficiency, in terms of speed and lexical coverage. A model for the full treatment of Spanish inflectional morphology for verbs, nouns and adjectives is presented. This model permits word formation based solely on morpheme concatenation, driven by a feature-based unification grammar. The run-time lexicon is a collection of allomorphs for both stems and endings. Although not tested, it should be suitable also for other Romance and highly inflected languages. A formalism is also described for encoding a lemma-based lexical source, well suited for expressing linguistic generalizations: inheritance classes, lemma encoding, morpho-graphemic allomorphy rules and limited type-checking. From this source base, we can automatically generate an allomorph indexed dictionary adequate for efficient retrieval and processing. A set of software tools has been implemented around this formalism: lexical base augmenting aids, lexical compilers to build run-time dictionaries and access libraries for them, feature manipulation libraries, unification and pseudo-unification modules, morphological processors, a parsing system, etc. Software interfaces among the different modules and tools are cleanly defined to ease software integration and tool combination in a flexible way. Directions for accessing our e-mail and web demonstration prototypes are also provided. Some figures are given, showing the lexical coverage of our platform compared to some popular spelling checkers.

Language tools that help people with their writing are now usually included in today's word processors. Although these various tools provide increasing support to native speakers of a language, they are much less useful to non-native speakers who are writing in their second language (e.g. French speakers writing in English). Real errors may go undetected and potential errors or non-errors that are flagged by the system may be taken to be genuine errors by the non-native speaker. In this paper, we present the prototype of an English writing tool which is aimed at helping speakers of French write in English. We first discuss the kind of problems non-native speakers have when writing in a second language. We then explain how we collected a corpus of errors which we used to build a typology of errors needed in the various stages of the project. This is followed by an overview of the prototype which contains a number of writing aids (dictionaries, on-line grammar helps, verb conjugator, etc.) and two checking tools: a problem word highlighter which lists all the potentially difficult words that cannot be dealt with correctly by the system (false friends, confusions, etc.) and a grammar checker which detects and corrects morphological and syntactic errors. We describe in detail the automata formalism we use to extract linguistic information, test syntactic environments and detect and correct errors. Finally, we present a first evaluation of the correction capacity of our grammar checker as compared to that of commercially available systems.

We present research aiming to build tools for the normalization of User-Generated Content (UGC). We argue that processing this type of text requires the revisiting of the initial steps of Natural Language Processing, since UGC (micro-blog, blog, and, generally, Web 2.0 user-generated texts) presents a number of nonstandard communicative and linguistic characteristics – often closer to oral and colloquial language than to edited text. We present a corpus of UGC text in Spanish from three different sources: Twitter, consumer reviews, and blogs, and describe its main characteristics. We motivate the need for UGC text normalization by analyzing the problems found when processing this type of text through a conventional language processing pipeline, particularly in the tasks of lemmatization and morphosyntactic tagging. Our aim with this paper is to seize the power of already existing spell and grammar correction engines and endow them with automatic normalization capabilities in order to pave the way for the application of standard Natural Language Processing tools to typical UGC text. Particularly, we propose a strategy for automatically normalizing UGC by adding a module on top of a pre-existing spell-checker that selects the most plausible correction from an unranked list of candidates provided by the spell-checker. To build this selector module we train four language models, each one containing a different type of linguistic information in a trade-off with its generalization capabilities. Our experiments show that the models trained on truecase and lowercase word forms are more discriminative than the others at selecting the best candidate. We have also experimented with a parametrized combination of the models by both optimizing directly on the selection task and doing a linear interpolation of the models. The resulting parametrized combinations obtain results close to the best performing model but do not improve on those results, as measured on the test set. The precision of the selector module in ranking number one the expected correction proposal on the test corpora reaches 82.5% for Twitter text (baseline 57%) and 88% for non-Twitter text (baseline 64%).

Neural Networks applied to Machine Translation need a finite vocabulary to express textual information as a sequence of discrete tokens. The currently dominant subword vocabularies exploit statistically-discovered common parts of words to achieve the flexibility of character-based vocabularies without delegating the whole learning of word formation to the neural network. However, they trade this for the inability to apply word-level token associations, which limits their use in semantically-rich areas and prevents some transfer learning approaches e.g. cross-lingual pretrained embeddings, and reduces their interpretability. In this work, we propose new hybrid linguistically-grounded vocabulary definition strategies that keep both the advantages of subword vocabularies and the word-level associations, enabling neural networks to profit from the derived benefits. We test the proposed approaches in both morphologically rich and poor languages, showing that, for the former, the quality in the translation of out-of-domain texts is improved with respect to a strong subword baseline.

We investigate the use of instance-based ranking methods for surface realization in natural language generation. Our approach to instance-based natural language generation (IBNLG) employs two components: a rule system that ‘overgenerates’ a number of realization candidates from a meaning representation and an instance-based ranker that scores the candidates according to their similarity to examples taken from a training corpus. We develop an efficient search technique for identifying the optimal candidate based on a novel extension of the A* algorithm. The rule system is produced automatically from a semantically annotated fragment of the Penn Treebank II containing management succession texts. We detail the annotation scheme and grammar induction algorithm and evaluate the efficiency and output of the generator. We also discuss issues such as input coverage (completeness) and fluency that are relevant to surface generation in general.

We describe a system that automatically learns effective and engaging dialogue strategies, generated from a library of dialogue content, using reinforcement learning from user feedback. Besides the more usual clarification and verification components of dialogue, this library contains various social elements like greetings, apologies, small talk, relational questions and jokes. We tested the method through an experimental dialogue system that encourages take-up of exercise and shows that the learned dialogue policy performs as well as one built by human experts for this system.

Comparable corpora are key translation resources for both languages and domains with limited linguistic resources. The existing approaches for building comparable corpora are mostly based on ranking candidate documents in the target language for each source document using a cross-lingual retrieval model. These approaches also exploit other evidence of document similarity, such as proper names and publication dates, to build more reliable alignments. However, the importance of each evidence in the scores of candidate target documents is determined heuristically. In this paper, we employ a learning to rank method for ranking candidate target documents with respect to each source document. The ranking model is constructed by defining each evidence for similarity of bilingual documents as a feature whose weight is learned automatically. Learning feature weights can significantly improve the quality of alignments, because the reliability of features depends on the characteristics of both source and target languages of a comparable corpus. We also propose a method to generate appropriate training data for the task of building comparable corpora. We employed the proposed learning-based approach to build a multi-domain English–Persian comparable corpus which covers twelve different domains obtained from Open Directory Project. Experimental results show that the created alignments have high degrees of comparability. Comparison with existing approaches for building comparable corpora shows that our learning-based approach improves both quality and coverage of alignments.

Statistical, linguistic, and heuristic clues can be used for the alignment of words and multi-word units in parallel texts. This article describes the clue alignment approach and the optimization of its parameters using a genetic algorithm. Word alignment clues can come from various sources such as statistical alignment models, co-occurrence tests, string similarity scores and static dictionaries. A genetic algorithm implementing an evolutionary procedure can be used to optimize the parameters necessary for combining available clues. Experiments on English/Swedish bitext show a significant improvement of about 6% in F-scores compared to the baseline produced by statistical word alignment.Most of the work described in this paper was carried out at the Department of Linguistics and Philology at Uppsala University. I would like to acknowledge technical and scientific support by people at the department in Uppsala.

Most of the morphological properties of derivational Arabic words are encapsulated in their corresponding morphological patterns. The morphological pattern is a template that shows how the word should be decomposed into its constituent morphemes (prefix + stem + suffix), and at the same time, marks the positions of the radicals comprising the root of the word. The number of morphological patterns in Arabic is finite and is well below 1000. Due to these properties, most of the current analysis algorithms concentrate on discovering the morphological pattern of the input word as a major step in recognizing the type and category of the word. Unfortunately, this process is non-determinitic in the sense that the underlying search process may sometimes associate more than one morphological pattern with the given word, all of them satisfying the major lexical constraints. One solution to this problem is to use a collection of connectionist pattern associaters that uniquely associate each word with its corresponding morphological pattern. This paper describes an LVQ-based learning pattern association system that uniquely maps a given Arabic word to its corresponding morphological pattern, and therefore deduces its morphological properties. The system consists of a collection of hetroassociative models that are trained using the LVQ algorithm plus a collection of autoassociative models that have been trained using backpropagation. Experimental results have shown that the system is fairly accurate and very easy to train. The LVQ algorithm has been chosen because it is very easy to train and the implied training time is very small compared to that of backpropagation.

Text style transfer (TST) aims at automatically changing a text’s stylistic features, such as formality, sentiment, authorial style, humor, and complexity, while still trying to preserve its content. Although the scientific community has investigated TST since the 1980s, it has recently regained attention by adopting deep unsupervised strategies to address the challenge of training without parallel data. In this manuscript, we investigate how relying on sequence-to-sequence pretraining models affects the performance of TST when the pretraining step leverages pairs of paraphrase data. Furthermore, we propose a new technique to enhance the sequence-to-sequence model by distilling knowledge from masked language models. We evaluate our proposals on three unsupervised style transfer tasks with widely used benchmarks: author imitation, formality transfer, and polarity swap. The evaluation relies on quantitative and qualitative analyses and comparisons with the results of state-of-the-art models. For the author imitation and the formality transfer task, we show that using the proposed techniques improves all measured metrics and leads to state-of-the-art (SOTA) results in content preservation and an overall score in the author imitation domain. In the formality transfer domain, we paired with the SOTA method in the style control metric. Regarding the polarity swap domain, we show that the knowledge distillation component improves all measured metrics. The paraphrase pretraining increases content preservation at the expense of harming style control. Based on the results reached in these domains, we also discuss in the manuscript if the tasks we address have the same nature and should be equally treated as TST tasks.

ANVIL is an information retrieval system using natural language processing techniques, intended for retrieval of captioned images. It extracts dependency structures from the image captions and user queries, and then applies a high accuracy matching algorithm which recursively explores the dependency structures to determine their similarity. A further algorithm allows additional contextual information to be extracted following a successful match, with the intention of helping users understand and organise the retrieval results. ANVIL was developed to high engineering standards, and as well as looking at the research aspects of the system, we also look at some of the design and development issues. English and Japanese versions of the system have been developed.

Morphological analysis and disambiguation are crucial stages in a variety of natural language processing applications, especially when languages with complex morphology are concerned. We present a system which disambiguates the output of a morphological analyzer for Hebrew. It consists of several simple classifiers and a module that combines them under the constraints imposed by the analyzer. We explore several approaches to classifier combination, as well as a back-off mechanism that relies on a large unannotated corpus. Our best result, around 83 percent accuracy, compares favorably with the state of the art on this task.

This paper presents two novel ideas of improving the Machine Translation (MT) quality by applying the word-level quality prediction for the second pass of decoding. In this manner, the word scores estimated by word confidence estimation systems help to reconsider the MT hypotheses for selecting a better candidate rather than accepting the current sub-optimal one. In the first attempt, the selection scope is limited to the MT N-best list, in which our proposed re-ranking features are combined with those of the decoder for re-scoring. Then, the search space is enlarged over the entire search graph, storing many more hypotheses generated during the first pass of decoding. Over all paths containing words of the N-best list, we propose an algorithm to strengthen or weaken them depending on the estimated word quality. In both methods, the highest score candidate after the search becomes the official translation. The results obtained show that both approaches advance the MT quality over the one-pass baseline, and the search graph re-decoding achieves more gains (in BLEU score) than N-best List Re-ranking method.