We provide a comprehensive analysis of the interactions between pre-trained word embeddings, character models and POS tags in a transition-based dependency parser. While previous studies have shown POS information to be less important in the presence of character models, we show that in fact there are complex interactions between all three techniques. In isolation each produces large improvements over a baseline system using randomly initialised word embeddings only, but combining them quickly leads to diminishing returns. We categorise words by frequency, POS tag and language in order to systematically investigate how each of the techniques affects parsing quality. For many word categories, applying any two of the three techniques is almost as good as the full combined system. Character models tend to be more important for low-frequency open-class words, especially in morphologically rich languages, while POS tags can help disambiguate high-frequency function words. We also show that large character embedding sizes help even for languages with small character sets, especially in morphologically rich languages.

How to make the most of multiple heterogeneous treebanks when training a monolingual dependency parser is an open question. We start by investigating previously suggested, but little evaluated, strategies for exploiting multiple treebanks based on concatenating training sets, with or without fine-tuning. We go on to propose a new method based on treebank embeddings. We perform experiments for several languages and show that in many cases fine-tuning and treebank embeddings lead to substantial improvements over single treebanks or concatenation, with average gains of 2.0-3.5 LAS points. We argue that treebank embeddings should be preferred due to their conceptual simplicity, flexibility and extensibility.

We present the Uppsala system for the CoNLL 2018 Shared Task on universal dependency parsing. Our system is a pipeline consisting of three components: the first performs joint word and sentence segmentation; the second predicts part-of- speech tags and morphological features; the third predicts dependency trees from words and tags. Instead of training a single parsing model for each treebank, we trained models with multiple treebanks for one language or closely related languages, greatly reducing the number of models. On the official test run, we ranked 7th of 27 teams for the LAS and MLAS metrics. Our system obtained the best scores overall for word segmentation, universal POS tagging, and morphological features.

This release contains errors in several files. Please use http://hdl.handle.net/11234/1-1983 instead.

Despite many years of research on Swedish language technology, there is still no well-documented standard for Swedish word processing covering the whole spectrum from low-level tokenization to morp ...

When developing a system that attempts to choose the best output from several machine translation engines for each input sentence, development data labelled with the correct answer is required. This data consists of input sentences, candidate outputs from each of the machine translation engines and labels to specify which of the outputs is best for each sentence. Previous research used an n-gram heuristic to compare each candidate output for each sentence to a reference translation in order to determine which of them contributes to the highest overall BLEU score. Here, we present an alternative method based on local search to choose the best candidate output for each sentence in the development data, and show that it leads to a higher BLEU on the labelled data itself and, more importantly, a higher BLEU on unseen test data for a combination of systems using a classifier trained on this new data.

Universal Dependencies is a project that seeks to develop cross-linguistically consistent treebank annotation for many languages, with the goal of facilitating multilingual parser development, cross-lingual learning, and parsing research from a language typology perspective. The annotation scheme is based on (universal) Stanford dependencies (de Marneffe et al., 2006, 2008, 2014), Google universal part-of-speech tags (Petrov et al., 2012), and the Interset interlingua for morphosyntactic tagsets (Zeman, 2008).

We present a method for finding oracle BLEU translations in phrase-based statistical machine translation using exact document-level scores. Experiments are presented where the BLEU score of a candi ...

Scarcity of multiword expression data sets raises a fundamental challenge to evaluating the systems that deal with these linguistic structures. In this work we attempt to address this problem for a subclass of multiword expressions by producing a large data set annotated by experts and validated by common statistical measures. We present a set of 1048 noun-noun compounds annotated as non-compositional, compositional, conventionalized and not conventionalized. We build this data set following common trends in previous work while trying to address some of the well known issues such as small number of annotated instances, quality of the annotations, and lack of availability of true negative instances.

This thesis presents two sets of experiments carried out with Docent, a document-level decoder for phrase-based statistical machine translation (SMT). In the first half of the thesis, BLEU decoding is introduced, implemented and investigated, whereby proposed changes to a translation are only accepted if the score of the automatic evaluation metric BLEU increases. Rather than leading to high quality translations as might be expected, the results show that high BLEU scores can be achieved despite patently bad translations. These results complement previous criticisms of the BLEU metric, which have thrown doubt on the claim that it always correlates well with human judgements of translation quality. The argument in this thesis is however more subtle: even if BLEU correlates well with human judgements of translation quality for translations produced independently of this metric, it does not necessarily hold that manipulating systems to achieve high BLEU scores leads to better quality translations. The second half of the thesis focuses on feature weight tuning, the process of finding optimum weights for the various feature functions employed in SMT. Several tuning paradigms already exist for standard sentence-level SMT decoders, but so far very little research has been carried out into tuning at the document level. The experiments presented here focus on adapting the PRO technique, which aims to find weights that correctly rank competing candidate translations. The ultimate goal of this work was to find a stable baseline tuning algorithm and use BLEU decoding to improve yet further on it. Despite extensive experimentation, this optimisation process fails to yield positive results in Docent: weights could not be found that improve the performance of the decoder compared to default settings. Possible reasons for these unexpected results and directions for future investigation are discussed in detail.

Universal Dependencies is a project that seeks to develop cross-linguistically consistent treebank annotation for many languages, with the goal of facilitating multilingual parser development, cross-lingual learning, and parsing research from a language typology perspective. The annotation scheme is based on (universal) Stanford dependencies (de Marneffe et al., 2006, 2008, 2014), Google universal part-of-speech tags (Petrov et al., 2012), and the Interset interlingua for morphosyntactic tagsets (Zeman, 2008).

We present ParCor, a parallel corpus of texts in which pronoun coreference - reduced coreference in which pronouns are used as referring expressions - has been annotated. The corpus is intended to be used both as a resource from which to learn systematic differences in pronoun use between languages and ultimately for developing and testing informed Statistical Machine Translation systems aimed at addressing the problem of pronoun coreference in translation. At present, the corpus consists of a collection of parallel English-German documents from two different text genres: TED Talks (transcribed planned speech), and EU Bookshop publications (written text). All documents in the corpus have been manually annotated with respect to the type and location of each pronoun and, where relevant, its antecedent. We provide details of the texts that we selected, the guidelines and tools used to support annotation and some corpus statistics. The texts in the corpus have already been translated into many languages, and we plan to expand the corpus into these other languages, as well as other genres, in the future.

The automatic extraction of verb-particle constructions (VPCs) is of particular interest to the NLP community. Previous studies have shown that word alignment methods can be used with parallel corpora to successfully extract a range of multi-word expressions (MWEs). In this paper the technique is applied to a new type of corpus, made up of a collection of subtitles of movies and television series, which is parallel in English and Spanish. Building on previous research, it is shown that a precision level of 94±4.7% can be achieved in English VPC extraction. This high level of precision is achieved despite the difficulties of aligning and tagging subtitles data. Moreover, many of the extracted VPCs are not present in online lexical resources, highlighting the benefits of using this unique corpus type, which contains a large number of slang and other informal expressions. An added benefit of using the word alignment process is that translations are also automatically extracted for each VPC. A precision rate of 75±8.5% is found for the translations of English VPCs into Spanish. This study thus shows that VPCs are a particularly good subset of the MWE spectrum to attack using word alignment methods, and that subtitles data provide a range of interesting expressions that do not exist in other corpus types.

We describe the Uppsala University systems for WMT14. We look at the integration of a model for translating pronominal anaphora and a syntactic dependency projection model for English‐French. Furthermore, we investigate post-ordering and tunable POS distortion models for English‐ German.

The visceral endoderm (VE) is a simple epithelium that forms the outer layer of the egg-cylinder stage mouse embryo. The anterior visceral endoderm (AVE), a specialised subset of VE cells, is responsible for specifying anterior pattern. AVE cells show a stereotypic migratory behaviour within the VE, which is responsible for correctly orientating the anterior-posterior axis. The epithelial integrity of the VE is maintained during the course of AVE migration, which takes place by intercalation of AVE and other VE cells. Though a continuous epithelial sheet, the VE is characterised by two regions of dramatically different behaviour, one showing robust cell movement and intercalation (in which the AVE migrates) and one that is static, with relatively little cell movement and mixing. Little is known about the cellular rearrangements that accommodate and influence the sustained directional movement of subsets of cells (such as the AVE) within epithelia like the VE. This study uses an interdisciplinary approach to further our understanding of cell movement in epithelia. Using both wild-type embryos as well as mutants in which AVE migration is abnormal or arrested, we show that AVE migration is specifically linked to changes in cell packing in the VE and an increase in multi-cellular rosette arrangements (five or more cells meeting at a point). To probe the role of rosettes during AVE migration, we develop a mathematical model of cell movement in the VE. To do this, we use a vertex-based model, implemented on an ellipsoidal surface to represent a realistic geometry for the mouse egg-cylinder. The potential for rosette formation is included, along with various junctional rearrangements. Simulations suggest that while rosettes are not essential for AVE migration, they are crucial for the orderliness of this migration observed in embryos. Our simulations are similar to results from transgenic embryos in which Planar Cell Polarity (PCP) signalling is disrupted. Such embryos have significantly reduced rosette numbers, altered epithelial packing, and show abnormalities in AVE migration. Our results show that the formation of multi-cellular rosettes in the mouse VE is dependent on normal PCP signalling. Taken together, our model and experimental observations suggest that rosettes in the VE epithelium do not form passively in response to AVE migration. Instead, they are a PCP-dependent arrangement of cells that acts to buffer the disequilibrium in cell packing generated in the VE by AVE migration, enabling AVE cells to migrate in an orderly manner.