For nearly three decades, language models derived from the n-gram assumption held the state of the art on the task. The key to their success lay in the application of various smoothing techniques that served to combat overfitting. However, when neural language models toppled n-gram models as the best performers, n-gram smoothing techniques became less relevant. Indeed, it would hardly be an understatement to suggest that the line of inquiry into n-gram smoothing techniques became dormant. This paper re-opens the role classical n-gram smoothing techniques may play in the age of neural language models. First, we draw a formal equivalence between label smoothing, a popular regularization technique for neural language models, and add-𝜆 smoothing. Second, we derive a generalized framework for converting any n-gram smoothing technique into a regularizer compatible with neural language models. Our empirical results find that our novel regularizers are comparable to and, indeed, sometimes outperform label smoothing on language modeling and machine translation.
Understanding memorisation in language models has practical and societal implications, e.g., studying models’ training dynamics or preventing copyright infringements.Prior work defines memorisation as the causal effect of training with an instance on the model’s ability to predict that instance. This definition relies on a counterfactual: the ability to observe what would have happened had the model not seen that instance.Existing methods struggle to provide computationally efficient and accurate estimates of this counterfactual. Further, they often estimate memorisation for a model architecture rather than for a specific model instance. This paper fills an important gap in the literature, proposing a new, principled, and efficient method to estimate memorisation based on the difference-in-differences design from econometrics. Using this method, we characterise a model’s memorisation profile–its memorisation trends across training–by only observing its behaviour on a small set of instances throughout training.In experiments with the Pythia model suite, we find that memorisation (i) is stronger and more persistent in larger models, (ii) is determined by data order and learning rate, and (iii) has stable trends across model sizes, thus making memorisation in larger models predictable from smaller ones.
Byte-Pair Encoding (BPE) is a popular algorithm used for tokenizing data in NLP, despite being devised initially as a compression method.BPE appears to be a greedy algorithm at face value, but the underlying optimization problem that BPE seeks to solve has not yet been laid down. We formalize BPE as a combinatorial optimization problem. Via submodular functions, we prove that the iterative greedy version is a 1/sigma*(1-e(-sigma))-approximation of an optimal merge sequence, where sigma is the total backward curvature with respect to the optimal merge sequence. Empirically the lower bound of the approximation is approx0.37.We provide a faster implementation of BPE which improves the runtime complexity from O(NM) to O(N log M), where N is the sequence length and M is the merge count. Finally, we optimize the brute-force algorithm for optimal BPE using memoization.
Today’s probabilistic language generators fall short when it comes to producing coherent and fluent text despite the fact that the underlying models perform well under standard metrics (e.g., perplexity). This discrepancy has puzzled the language generation community for the last few years. In this work, we posit that the abstraction of natural language generation as a discrete stochastic process—which allows for an information-theoretic analysis—can provide new insights into the behavior of probabilistic language generators, for example, why high-probability texts can be dull or repetitive. Humans use language as a means of communicating information, aiming to do so in a simultaneously efficient and error-minimizing manner; in fact, psycholinguistics research suggests humans choose each word in a string with this subconscious goal in mind. We formally define the set of strings that meet this criterion: Those for which each word has an information content close to the expected information content, namely, the conditional entropy of our model. We then propose a simple and efficient procedure for enforcing this criterion when generating from probabilistic models, which we call locally typical sampling. Automatic and human evaluations show that, in comparison to nucleus and top-k sampling, locally typical sampling offers competitive performance (in both abstractive summarization and story generation) in terms of quality while consistently reducing degenerate repetitions.
Probing has become a go-to methodology for interpreting and analyzing deep neural models in natural language processing. However, there is still a lack of understanding of the limitations and weaknesses of various types of probes. In this work, we suggest a strategy for input-level intervention on naturalistic sentences. Using our approach, we intervene on the morpho-syntactic features of a sentence, while keeping the rest of the sentence unchanged. Such an intervention allows us to causally probe pre-trained models. We apply our naturalistic causal probing framework to analyze the effects of grammatical gender and number on contextualized representations extracted from three pre-trained models in Spanish, the multilingual versions of BERT, RoBERTa, and GPT-2. Our experiments suggest that naturalistic interventions lead to stable estimates of the causal effects of various linguistic properties. Moreover, our experiments demonstrate the importance of naturalistic causal probing when analyzing pre-trained models. https://github.com/rycolab/naturalistic-causal-probing
While natural languages differ widely in both canonical word order and word order flexibility, their word orders still follow shared cross-linguistic statistical patterns, often attributed to functional pressures. In the effort to identify these pressures, prior work has compared real and counterfactual word orders. Yet one functional pressure has been overlooked in such investigations: The uniform information density (UID) hypothesis, which holds that information should be spread evenly throughout an utterance. Here, we ask whether a pressure for UID may have influenced word order patterns cross-linguistically. To this end, we use computational models to test whether real orders lead to greater information uniformity than counterfactual orders. In our empirical study of 10 typologically diverse languages, we find that: (i) among SVO languages, real word orders consistently have greater uniformity than reverse word orders, and (ii) only linguistically implausible counterfactual orders consistently exceed the uniformity of real orders. These findings are compatible with a pressure for information uniformity in the development and usage of natural languages.1
Surprisal theory posits that less-predictable words should take more time to process, with word predictability quantified as surprisal, i.e., negative log probability in context. While evidence supporting the predictions of surprisal theory has been replicated widely, much of it has focused on a very narrow slice of data: native English speakers reading English texts. Indeed, no comprehensive multilingual analysis exists. We address this gap in the current literature by investigating the relationship between surprisal and reading times in eleven different languages, distributed across five language families. Deriving estimates from language models trained on monolingual and multilingual corpora, we test three predictions associated with surprisal theory: (i) whether surprisal is predictive of reading times, (ii) whether expected surprisal, i.e., contextual entropy, is predictive of reading times, and (iii) whether the linking function between surprisal and reading times is linear. We find that all three predictions are borne out crosslinguistically. By focusing on a more diverse set of languages, we argue that these results offer the most robust link to date between information theory and incremental language processing across languages.
Zipf (1935) posited that wordforms are optimized to minimize utterances’ communicative costs. Under the assumption that cost is given by an utterance’s length, he supported this claim by showing that words’ lengths are inversely correlated with their frequencies. Communicative cost, however, can be operationalized in different ways. Piantadosi et al. (2011) claim that cost should be measured as the distance between an utterance’s information rate and channel capacity, which we dub the channel capacity hypothesis (CCH) here. Following this logic, they then proposed that a word’s length should be proportional to the expected value of its surprisal (negative log-probability in context). In this work, we show that Piantadosi et al.’s derivation does not minimize CCH’s cost, but rather a lower bound, which we term CCH-lower. We propose a novel derivation, suggesting an improved way to minimize CCH’s cost. Under this method, we find that a language’s word lengths should instead be proportional to the surprisal’s expectation plus its variance-to-mean ratio. Experimentally, we compare these three communicative cost functions: Zipf’s, CCH-lower , and CCH. Across 13 languages and several experimental settings, we find that length is better predicted by frequency than either of the other hypotheses. In fact, when surprisal’s expectation, or expectation plus variance-to-mean ratio, is estimated using better language models, it leads to worse word length predictions. We take these results as evidence that Zipf’s longstanding hypothesis holds.
Surprisal theory (Hale, 2001; Levy, 2008) posits that a word’s reading time is proportional to its surprisal (i.e., to its negative log probability given the proceeding context). Since we are unable to access a word’s ground-truth probability, surprisal theory has been empirically tested using surprisal estimates from language models (LMs). Under the premise that surprisal theory holds, we would expect that higher quality language models provide more powerful predictors of human reading behavior—a conjecture we dub the quality–power (QP) hypothesis. Unfortunately, empirical support for the QP hypothesis is mixed. Some studies in English have found correlations between LM quality and predictive power, but other studies using Japanese data, as well as using larger English LMs, find no such correlations. In this work, we conduct a systematic crosslinguistic assessment of the QP hypothesis. We train LMs from scratch on small- and medium-sized datasets from 13 languages (across five language families) and assess their ability to predict eye tracking data. We find correlations between LM quality and power in eleven of these thirteen languages, suggesting that, within the range of model classes and sizes tested, better language models are indeed better predictors of human language processing behaviors.
Sampling-based decoding strategies are widely employed for generating text from probabilistic models, yet standard ancestral sampling often results in text that is degenerate or incoherent. To alleviate this issue, various modifications to a model’s sampling distribution, such as top-p or top-k sampling, have been introduced and are now ubiquitously used in language generation systems. We propose a unified framework for understanding these techniques, which we term sampling adapters. Sampling adapters often lead to qualitatively better text, which raises the question: From a formal perspective, how are they changing the token-level distributions of language generation models? And why do these local changes lead to higher-quality text? We argue that the shift they enforce can be viewed as a trade-off between precision and recall: while the model loses its ability to produce certain strings, its precision rate on desirable text increases. While this trade-off is not reflected in standard metrics of distribution quality (such as perplexity), we find that several precision-emphasizing measures indeed indicate that sampling adapters can lead to probability distributions more aligned with the true distribution. Further, these measures correlate with higher sequence-level quality scores, specifically, Mauve.
Language modeling, a central task in natural language processing, involves estimating a probability distribution over strings. In most cases, the estimated distribution sums to 1 over all finite strings. However, in some pathological cases, probability mass can “leak” onto the set of infinite sequences. In order to characterize the notion of leakage more precisely, this paper offers a measure-theoretic treatment of language modeling. We prove that many popular language model families are in fact tight, meaning that they will not leak in this sense. We also generalize characterizations of tightness proposed in previous works.
After just a few hundred training updates, a standard probabilistic model for language generation has likely not yet learnt many semantic or syntactic rules of natural language, making it difficult to estimate the probability distribution over next tokens. Yet around this point, these models have identified a simple, loss-minimising behaviour: to output the unigram distribution of the target training corpus. The use of such a heuristic raises the question: Can we initialise our models with this behaviour and save precious compute resources and model capacity? Here we show that we can effectively endow standard neural language generation models with a separate module that reflects unigram frequency statistics as prior knowledge, simply by initialising the bias term in a model’s final linear layer with the log-unigram distribution. We use neural machine translation as a test bed for this simple technique and observe that it: (i) improves learning efficiency; (ii) achieves better overall performance; and perhaps most importantly (iii) appears to disentangle strong frequency effects by encouraging the model to specialise in non-frequency-related aspects of language.
An increasingly large percentage of natural language processing (NLP) tasks center around the generation of text from probabilistic language models. Despite this trend, techniques for improving or specifying preferences in these generated texts rely mostly on intuition-based heuristics. Further, there lacks a unified presentation of their motivations, practical implementation, successes and pitfalls. Practitioners must, therefore, choose somewhat blindly between generation algorithms—like top-p sampling or beam search—which can lead to wildly different results. At the same time, language generation research continues to criticize and improve the standard toolboxes, further adding entropy to the state of the field. In this tutorial, we will provide a centralized and cohesive discussion of critical considerations when choosing how to generate from a language model. We will cover a wide range of empirically-observed problems (like degradation, hallucination, repetition) and their corresponding proposed algorithmic solutions from recent research (like top-p sampling and its successors). We will then discuss a subset of these algorithms under a unified light; most stochastic generation strategies can be framed as locally adapting the probabilities of a model to avoid failure cases. Finally, we will then cover methods in controlled generation, that go beyond just ensuring coherence to ensure text exhibits specific desired properties. We aim for NLP practitioners and researchers to leave our tutorial with a unified framework which they can use to evaluate and contribute to the latest research in language generation.
Despite significant progress in the quality of language generated from abstractive summarization models, these models still exhibit the tendency to hallucinate, i.e., output content not supported by the source document. A number of works have tried to fix—or at least uncover the source of—the problem with limited success. In this paper, we identify a simple criterion under which models are significantly more likely to assign more probability to hallucinated content during generation: high model uncertainty. This finding offers a potential explanation for hallucinations: models default to favoring text with high marginal probability, i.e., high-frequency occurrences in the training set, when uncertain about a continuation. It also motivates possible routes for real-time intervention during decoding to prevent such hallucinations. We propose a decoding strategy that switches to optimizing for pointwise mutual information of the source and target token—rather than purely the probability of the target token—when the model exhibits uncertainty. Experiments on the dataset show that our method decreases the probability of hallucinated tokens while maintaining the Rouge and BERT-S scores of top-performing decoding strategies.
Numerous analyses of reading time (RT) data have been undertaken in the effort to learn more about the internal processes that occur during reading comprehension. However, data measured on words at the end of a sentence–or even clause–is often omitted due to the confounding factors introduced by so-called “wrap-up effects,” which manifests as a skewed distribution of RTs for these words. Consequently, the understanding of the cognitive processes that might be involved in these effects is limited. In this work, we attempt to learn more about these processes by looking for the existence–or absence–of a link between wrap-up effects and information theoretic quantities, such as word and context information content. We find that the information distribution of prior context is often predictive of sentence- and clause-final RTs (while not of sentence-medial RTs), which lends support to several prior hypotheses about the processes involved in wrap-up effects.
When generating natural language from neural probabilistic models, high probability does not always coincide with high quality: It has often been observed that mode-seeking decoding methods, i.e., those that produce high-probability text under the model, lead to unnatural language. On the other hand, the lower-probability text generated by stochastic methods is perceived as more human-like. In this note, we offer an explanation for this phenomenon by analyzing language generation through an information-theoretic lens. Specifically, we posit that human-like language should contain an amount of information (quantified as negative log-probability) that is close to the entropy of the distribution over natural strings. Further, we posit that language with substantially more (or less) information is undesirable. We provide preliminary empirical evidence in favor of this hypothesis; quality ratings of both human and machine-generated text—covering multiple tasks and common decoding strategies—suggest high-quality text has an information content significantly closer to the entropy than we would expect by chance.
Shannon entropy is often a quantity of interest to linguists studying the communicative capacity of human language. However, entropymust typically be estimated from observed data because researchers do not have access to the underlying probability distribution. While entropy estimation is a well-studied problem in other fields, there is not yet a comprehensive exploration of the efficacy of entropy estimators for use with linguistic data. In this work, we fill this void, studying the empirical effectiveness of different entropy estimators for linguistic distributions. In a replication of two recent information-theoretic linguistic studies, we find evidence that the reported effect size is over-estimated due to over-reliance on poor entropy estimators. We end this paper with a concrete recommendation for the entropy estimators that should be used in future linguistic studies.
When generating text from probabilistic models, the chosen decoding strategy has a profound effect on the resulting text. Yet the properties elicited by various decoding strategies do not always transfer across natural language generation tasks. For example, while mode-seeking methods like beam search perform remarkably well for machine translation, they have been observed to lead to incoherent and repetitive text in story generation. Despite such observations, the effectiveness of decoding strategies is often assessed on only a single task. This work—in contrast—provides a comprehensive analysis of the interaction between language generation tasks and decoding strategies. Specifically, we measure changes in attributes of generated text as a function of both decoding strategy and task using human and automatic evaluation. Our results reveal both previously observed and novel findings. For example, the nature of the diversity–quality trade-off in language generation is very task-specific; the length bias often attributed to beam search is not constant across tasks. https://github.com/gianwiher/decoding-NLG
Neural sequence-to-sequence models are currently the predominant choice for language generation tasks. Yet, on word-level tasks, exact inference of these models reveals the empty string is often the global optimum. Prior works have speculated this phenomenon is a result of the inadequacy of neural models for language generation. However, in the case of morphological inflection, we find that the empty string is almost never the most probable solution under the model. Further, greedy search often finds the global optimum. These observations suggest that the poor calibration of many neural models may stem from characteristics of a specific subset of tasks rather than general ill-suitedness of such models for language generation.
Beam search is the default decoding strategy for many sequence generation tasks in NLP. The set of approximate K-best items returned by the algorithm is a useful summary of the distribution for many applications; however, the candidates typically exhibit high overlap and may give a highly biased estimate for expectations under our model. These problems can be addressed by instead using stochastic decoding strategies. In this work, we propose a new method for turning beam search into a stochastic process: Conditional Poisson stochastic beam search. Rather than taking the maximizing set at each iteration, we sample K candidates without replacement according to the conditional Poisson sampling design. We view this as a more natural alternative to Kool et al. (2019)’s stochastic beam search (SBS). Furthermore, we show how samples generated under the CPSBS design can be used to build consistent estimators and sample diverse sets from sequence models. In our experiments, we observe CPSBS produces lower variance and more efficient estimators than SBS, even showing improvements in high entropy settings.
While there exist scores of natural languages, each with its unique features and idiosyncrasies, they all share a unifying theme: enabling human communication. We may thus reasonably predict that human cognition shapes how these languages evolve and are used. Assuming that the capacity to process information is roughly constant across human populations, we expect a surprisal–duration trade-off to arise both across and within languages. We analyse this trade-off using a corpus of 600 languages and, after controlling for several potential confounds, we find strong supporting evidence in both settings. Specifically, we find that, on average, phones are produced faster in languages where they are less surprising, and vice versa. Further, we confirm that more surprising phones are longer, on average, in 319 languages out of the 600. We thus conclude that there is strong evidence of a surprisal–duration trade-off in operation, both across and within the world’s languages.
The uniform information density (UID) hypothesis posits a preference among language users for utterances structured such that information is distributed uniformly across a signal. While its implications on language production have been well explored, the hypothesis potentially makes predictions about language comprehension and linguistic acceptability as well. Further, it is unclear how uniformity in a linguistic signal—or lack thereof—should be measured, and over which linguistic unit, e.g., the sentence or language level, this uniformity should hold. Here we investigate these facets of the UID hypothesis using reading time and acceptability data. While our reading time results are generally consistent with previous work, they are also consistent with a weakly super-linear effect of surprisal, which would be compatible with UID’s predictions. For acceptability judgments, we find clearer evidence that non-uniformity in information density is predictive of lower acceptability. We then explore multiple operationalizations of UID, motivated by different interpretations of the original hypothesis, and analyze the scope over which the pressure towards uniformity is exerted. The explanatory power of a subset of the proposed operationalizations suggests that the strongest trend may be a regression towards a mean surprisal across the language, rather than the phrase, sentence, or document—a finding that supports a typical interpretation of UID, namely that it is the byproduct of language users maximizing the use of a (hypothetical) communication channel.
Large pre-trained language models have repeatedly shown their ability to produce fluent text. Yet even when starting from a prompt, generation can continue in many plausible directions. Current decoding methods with the goal of controlling generation, e.g., to ensure specific words are included, either require additional models or fine-tuning, or work poorly when the task at hand is semantically unconstrained, e.g., story generation. In this work, we present a plug-and-play decoding method for controlled language generation that is so simple and intuitive, it can be described in a single sentence: given a topic or keyword, we add a shift to the probability distribution over our vocabulary towards semantically similar words. We show how annealing this distribution can be used to impose hard constraints on language generation, something no other plug-and-play method is currently able to do with SOTA language generators. Despite the simplicity of this approach, we see it works incredibly well in practice: decoding from GPT-2 leads to diverse and fluent sentences while guaranteeing the appearance of given guide words. We perform two user studies, revealing that (1) our method outperforms competing methods in human evaluations; and (2) forcing the guide words to appear in the generated text has no impact on the fluency of the generated text.
The uniform information density (UID) hypothesis, which posits that speakers behaving optimally tend to distribute information uniformly across a linguistic signal, has gained traction in psycholinguistics as an explanation for certain syntactic, morphological, and prosodic choices. In this work, we explore whether the UID hypothesis can be operationalized as an inductive bias for statistical language modeling. Specifically, we augment the canonical MLE objective for training language models with a regularizer that encodes UID. In experiments on ten languages spanning five language families, we find that using UID regularization consistently improves perplexity in language models, having a larger effect when training data is limited. Moreover, via an analysis of generated sequences, we find that UID-regularized language models have other desirable properties, e.g., they generate text that is more lexically diverse. Our results not only suggest that UID is a reasonable inductive bias for language modeling, but also provide an alternative validation of the UID hypothesis using modern-day NLP tools.
We propose an alternate approach to quantifying how well language models learn natural language: we ask how well they match the statistical tendencies of natural language. To answer this question, we analyze whether text generated from language models exhibits the statistical tendencies present in the human-generated text on which they were trained. We provide a framework–paired with significance tests–for evaluating the fit of language models to these trends. We find that neural language models appear to learn only a subset of the tendencies considered, but align much more closely with empirical trends than proposed theoretical distributions (when present). Further, the fit to different distributions is highly-dependent on both model architecture and generation strategy. As concrete examples, text generated under the nucleus sampling scheme adheres more closely to the type–token relationship of natural language than text produced using standard ancestral sampling; text from LSTMs reflects the natural language distributions over length, stopwords, and symbols surprisingly well.
Beam search is a go-to strategy for decoding neural sequence models. The algorithm can naturally be viewed as a subset optimization problem, albeit one where the corresponding set function does not reflect interactions between candidates. Empirically, this leads to sets often exhibiting high overlap, e.g., strings may differ by only a single word. Yet in use-cases that call for multiple solutions, a diverse or representative set is often desired. To address this issue, we propose a reformulation of beam search, which we call determinantal beam search. Determinantal beam search has a natural relationship to determinantal point processes (DPPs), models over sets that inherently encode intra-set interactions. By posing iterations in beam search as a series of subdeterminant maximization problems, we can turn the algorithm into a diverse subset selection process. In a case study, we use the string subsequence kernel to explicitly encourage n-gram coverage in text generated from a sequence model. We observe that our algorithm offers competitive performance against other diverse set generation strategies in the context of language generation, while providing a more general approach to optimizing for diversity.
Sparse attention has been claimed to increase model interpretability under the assumption that it highlights influential inputs. Yet the attention distribution is typically over representations internal to the model rather than the inputs themselves, suggesting this assumption may not have merit. We build on the recent work exploring the interpretability of attention; we design a set of experiments to help us understand how sparsity affects our ability to use attention as an explainability tool. On three text classification tasks, we verify that only a weak relationship between inputs and co-indexed intermediate representations exists—under sparse attention and otherwise. Further, we do not find any plausible mappings from sparse attention distributions to a sparse set of influential inputs through other avenues. Rather, we observe in this setting that inducing sparsity may make it less plausible that attention can be used as a tool for understanding model behavior.
This paper presents our system for the SIGMORPHON 2020 Shared Task. We build off of the baseline systems, performing exact inference on models trained on language family data. Our systems return the globally best solution under these models. Our two systems achieve 80.9% and 75.6% accuracy on the test set. We ultimately find that, in this setting, exact inference does not seem to help or hinder the performance of morphological inflection generators, which stands in contrast to its affect on Neural Machine Translation (NMT) models.
Decoding for many NLP tasks requires an effective heuristic algorithm for approximating exact search because the problem of searching the full output space is often intractable, or impractical in many settings. The default algorithm for this job is beam search—a pruned version of breadth-first search. Quite surprisingly, beam search often returns better results than exact inference due to beneficial search bias for NLP tasks. In this work, we show that the standard implementation of beam search can be made up to 10x faster in practice. Our method assumes that the scoring function is monotonic in the sequence length, which allows us to safely prune hypotheses that cannot be in the final set of hypotheses early on. We devise effective monotonic approximations to popular nonmonontic scoring functions, including length normalization and mutual information decoding. Lastly, we propose a memory-reduced variant of best-first beam search, which has a similar beneficial search bias in terms of downstream performance, but runs in a fraction of the time.
Prior work has explored directly regularizing the output distributions of probabilistic models to alleviate peaky (i.e. over-confident) predictions, a common sign of overfitting. This class of techniques, of which label smoothing is one, has a connection to entropy regularization. Despite the consistent success of label smoothing across architectures and data sets in language generation tasks, two problems remain open: (1) there is little understanding of the underlying effects entropy regularizers have on models, and (2) the full space of entropy regularization techniques is largely unexplored. We introduce a parametric family of entropy regularizers, which includes label smoothing as a special case, and use it to gain a better understanding of the relationship between the entropy of a model and its performance on language generation tasks. We also find that variance in model performance can be explained largely by the resulting entropy of the model. Lastly, we find that label smoothing provably does not allow for sparsity in an output distribution, an undesirable property for language generation models, and therefore advise the use of other entropy regularization methods in its place.
Quite surprisingly, exact maximum a posteriori (MAP) decoding of neural language generators frequently leads to low-quality results. Rather, most state-of-the-art results on language generation tasks are attained using beam search despite its overwhelmingly high search error rate. This implies that the MAP objective alone does not express the properties we desire in text, which merits the question: if beam search is the answer, what was the question? We frame beam search as the exact solution to a different decoding objective in order to gain insights into why high probability under a model alone may not indicate adequacy. We find that beam search enforces uniform information density in text, a property motivated by cognitive science. We suggest a set of decoding objectives that explicitly enforce this property and find that exact decoding with these objectives alleviates the problems encountered when decoding poorly calibrated language generation models. Additionally, we analyze the text produced using various decoding strategies and see that, in our neural machine translation experiments, the extent to which this property is adhered to strongly correlates with BLEU.