Controlled Text Generation (CTG) steers the generation of continuations of a given context (prompt) by a Large Language Model (LLM) towards texts possessing a given attribute (e.g., topic, sentiment). In this paper we view CTG as a Continual Learning problem: how to learn at every step to steer next-word generation, without having to wait for end-of-sentence. This continual view is useful for online applications such as CTG for speech, where end-of-sentence is often uncertain. We depart from an existing model, the Plug-and-Play language models (PPLM), which perturbs the context at each step to better predict next-words that posses the desired attribute. While PPLM is intricate and has many hyper-parameters, we provide a proof that the PPLM objective function can be reduced to a Continual Reinforcement Learning (CRL) reward function, thereby simplifying PPLM and endowing it with a better understood learning framework. Subsequently, we present, the first of its kind, CTG algorithm that is fully based on CRL and exhibit promising empirical results.

Existing syntax-enriched neural machine translation (NMT) models work either with the single most-likely unlabeled parse or the set of n-best unlabeled parses coming out of an external parser. Passing a single or n-best parses to the NMT model risks propagating parse errors. Furthermore, unlabeled parses represent only syntactic groupings without their linguistically relevant categories. In this paper we explore the question: Does passing both parser uncertainty and labeled syntactic knowledge to the Transformer improve its translation performance? This paper contributes a novel method for infusing the whole labeled dependency distributions (LDD) of the source sentence’s dependency forest into the self-attention mechanism of the encoder of the Transformer. A range of experimental results on three language pairs demonstrate that the proposed approach outperforms both the vanilla Transformer as well as the single best-parse Transformer model across several evaluation metrics.

This work exploits translation data as a source of semantically relevant learning signal for models of word representation. In particular, we exploit equivalence through translation as a form of distributional context and jointly learn how to embed and align with a deep generative model. Our EmbedAlign model embeds words in their complete observed context and learns by marginalisation of latent lexical alignments. Besides, it embeds words as posterior probability densities, rather than point estimates, which allows us to compare words in context using a measure of overlap between distributions (e.g. KL divergence). We investigate our model’s performance on a range of lexical semantics tasks achieving competitive results on several standard benchmarks including natural language inference, paraphrasing, and text similarity.

MT evaluation metrics are tested for correlation with human judgments either at the sentence- or the corpus-level. Trained metrics ignore corpus-level judgments and are trained for high sentence-level correlation only. We show that training only for one objective (sentence or corpus level), can not only harm the performance on the other objective, but it can also be suboptimal for the objective being optimized. To this end we present a metric trained for corpus-level and show empirical comparison against a metric trained for sentence-level exemplifying how their performance may vary per language pair, type and level of judgment. Subsequently we propose a model trained to optimize both objectives simultaneously and show that it is far more stable than–and on average outperforms–both models on both objectives.

We present a simple and effective approach to incorporating syntactic structure into neural attention-based encoder-decoder models for machine translation. We rely on graph-convolutional networks (GCNs), a recent class of neural networks developed for modeling graph-structured data. Our GCNs use predicted syntactic dependency trees of source sentences to produce representations of words (i.e. hidden states of the encoder) that are sensitive to their syntactic neighborhoods. GCNs take word representations as input and produce word representations as output, so they can easily be incorporated as layers into standard encoders (e.g., on top of bidirectional RNNs or convolutional neural networks). We evaluate their effectiveness with English-German and English-Czech translation experiments for different types of encoders and observe substantial improvements over their syntax-agnostic versions in all the considered setups.

Existing work on domain adaptation for statistical machine translation has consistently assumed access to a small sample from the test distribution (target domain) at training time. In practice, however, the target domain may not be known at training time or it may change to match user needs. In such situations, it is natural to push the system to make safer choices, giving higher preference to domain-invariant translations, which work well across domains, over risky domain-specific alternatives. We encode this intuition by (1) inducing latent subdomains from the training data only; (2) introducing features which measure how specialized phrases are to individual induced sub-domains; (3) estimating feature weights on out-of-domain data (rather than on the target domain). We conduct experiments on three language pairs and a number of different domains. We observe consistent improvements over a baseline which does not explicitly reward domain invariance.

Existing approaches for evaluating word order in machine translation work with metrics computed directly over a permutation of word positions in system output relative to a reference translation. However, every permutation factorizes into a permutation tree (PET) built of primal permutations, i.e., atomic units that do not factorize any further. In this paper we explore the idea that permutations factorizing into (on average) shorter primal permutations should represent simpler ordering as well. Consequently, we contribute Permutation Complexity, a class of metrics over PETs and their extension to forests, and define tight metrics, a sub-class of metrics implementing this idea. Subsequently we define example tight metrics and empirically test them in word order evaluation. Experiments on the WMT13 data sets for ten language pairs show that a tight metric is more often than not better than the baselines.

In this paper we explore the novel idea of building a single universal reordering model from English to a large number of target languages. To build this model we exploit typological features of word order for a large number of target languages together with source (English) syntactic features and we train this model on a single combined parallel corpus representing all (22) involved language pairs. We contribute experimental evidence for the usefulness of linguistically defined typological features for building such a model. When the universal reordering model is used for preordering followed by monotone translation (no reordering inside the decoder), our experiments show that this pipeline gives comparable or improved translation performance with a phrase-based baseline for a large number of language pairs (12 out of 22) from diverse language families.

PARSEVAL, the default paradigm for evaluating constituency parsers, calculates parsing success (Precision/Recall) as a function of the number of matching labeled brackets across the test set. Nodes in constituency trees, however, are connected together to reflect important linguistic relations such as predicate-argument and direct-dominance relations between categories. In this paper, we present FREVAL, a generalization of PARSEVAL, where the precision and recall are calculated not only for individual brackets, but also for co-occurring, connected brackets (i.e. fragments). FREVAL fragments precision (FLP) and recall (FLR) interpolate the match across the whole spectrum of fragment sizes ranging from those consisting of individual nodes (labeled brackets) to those consisting of full parse trees. We provide evidence that FREVAL is informative for inspecting relative parser performance by comparing a range of existing parsers.

Modern statistical parsers are trained on large annotated corpora (treebanks). These treebanks usually consist of sentences addressing different subdomains (e.g. sports, politics, music), which implies that the statistics gathered by current statistical parsers are mixtures of subdomains of language use. In this paper we present a method that exploits raw subdomain corpora gathered from the web to introduce subdomain sensitivity into a given parser. We employ statistical techniques for creating an ensemble of domain sensitive parsers, and explore methods for amalgamating their predictions. Our experiments show that introducing domain sensitivity by exploiting raw corpora can improve over a tough, state-of-the-art baseline.

Lexical mappings (word translations) between languages are an invaluable resource for multilingual processing. While the problem of extracting lexical mappings from parallel corpora is well-studied, the task is more challenging when the language samples are from non-parallel corpora. The goal of this work is to investigate one such scenario: finding lexical mappings between dialects of a diglossic language, in which people conduct their written communications in a prestigious formal dialect, but they communicate verbally in a colloquial dialect. Because the two dialects serve different socio-linguistic functions, parallel corpora do not naturally exist between them. An example of a diglossic dialect pair is Modern Standard Arabic (MSA) and Levantine Arabic. In this paper, we evaluate the applicability of a standard algorithm for inducing lexical mappings between comparable corpora (Rapp, 1999) to such diglossic corpora pairs. The focus of the paper is an in-depth error analysis, exploring the notion of relatedness in diglossic corpora and scrutinizing the effects of various dimensions of relatedness (such as mode, topic, style, and statistics) on the quality of the resulting translation lexicon.

Given a probabilistic parsing model and an evaluation metric for scoring the match between parse-trees, e.g., PARSEVAL [Black et al., 1991], this paper addresses the problem of how to select the on average best scoring parse-tree for an input sentence. Common wisdom dictates that it is optimal to select the parse with the highest probability, regardless of the evaluation metric. In contrast, the Maximizing Metrics (MM) method [Goodman, 1998, Stolcke et al., 1997] proposes that an algorithm that optimizes the evaluation metric itself constitutes the optimal choice. We study the MM method within parsing. We observe that the MM does not always hold for tree-bank models, and that optimizing weak metrics is not interesting for semantic processing. Subsequently, we state an alternative proposition: the optimal algorithm must maximize the metric that scores parse-trees according to linguistically relevant features. We present new algorithms that optimize metrics that take into account increasingly more linguistic features, and exhibit experiments in support of our claim.