Subword regularization, used widely in NLP, improves model performance by reducing the dependency on exact tokenizations, augmenting the training corpus, and exposing the model to more unique contexts during training. BPE and MaxMatch, two popular subword tokenization schemes, have stochastic dropout regularization variants. However, there has not been an analysis of the distributions formed by them.We show that these stochastic variants are heavily biased towards a small set of tokenizations per word. If the benefits of subword regularization are as mentioned, we hypothesize that biasedness artificially limits the effectiveness of these schemes. Thus, we propose an algorithm to uniformly sample tokenizations that we use as a drop-in replacement for the stochastic aspects of existing tokenizers, and find that it improves machine translation quality.

We explore threshold vocabulary trimming in Byte-Pair Encoding subword tokenization, a tokenization postprocessing step that replaces rare subwords with their component subwords. The technique is available in popular tokenization libraries but has not been subjected to rigorous scientific scrutiny. While the removal of rare subwords is suggested as best practice in model implementations, both as a means to reduce model size and for improving model performance through robustness, our experiments indicate that, across a large space of hyperparameter settings, vocabulary trimming fails to consistently improve model performance, and is even prone to incurring heavy degradation.

In Tokenization and the Noiseless Channel (Zouhar et al., 2023), Rényi efficiency is suggested as an intrinsic mechanism for evaluating a tokenizer: for NLP tasks, the tokenizer which leads to the highest Rényi efficiency of the unigram distribution should be chosen. The Rényi efficiency is thus treated as a predictor of downstream performance (e.g., predicting BLEU for a machine translation task), without the expensive step of training multiple models with different tokenizers. Although useful, the predictive power of this metric is not perfect, and the authors note there are additional qualities of a good tokenization scheme that Rényi efficiency alone cannot capture. We describe two variants of BPE tokenization which can arbitrarily increase Rényi efficiency while decreasing the downstream model performance. These counterexamples expose cases where Rényi efficiency fails as an intrinsic tokenization metric and thus give insight for building more accurate predictors.

Character-level language modeling has been shown empirically to perform well on highly agglutinative or morphologically rich languages while using only a small fraction of the parameters required by (sub)word models. Korean fits nicely into this framework, except that, like other CJK languages, it has a very large character vocabulary of 11,172 unique syllables. However, unlike Japanese Kanji and Chinese Hanzi, each Korean syllable can be uniquely factored into a small set of subcharacters, called jamo. We explore a “three-hot” scheme, where we exploit the decomposability of Korean characters to model at the syllable level but using only jamo-level representations. We find that our three-hot embedding and decoding scheme alleviates the two major issues with prior syllable- and jamo-level models. Namely, it requires fewer than 1% of the embedding parameters of a syllable model, and it does not require tripling the sequence length, as with jamo models. In addition, it addresses a theoretical flaw in a prior three-hot modeling scheme. Our experiments show that, even when reducing the number of embedding parameters by 99.6% (from 11.4M to just 36k), our model suffers no loss in translation quality compared to the baseline syllable model.

Probabilistic finite automata (PFAs) are com- mon statistical language model in natural lan- guage and speech processing. A typical task for PFAs is to compute the probability of all strings that match a query pattern. An impor- tant special case of this problem is computing the probability of a string appearing as a pre- fix, suffix, or infix. These problems find use in many natural language processing tasks such word prediction and text error correction. Recently, we gave the first incremental algorithm to efficiently compute the infix probabilities of each prefix of a string (Cognetta et al., 2018). We develop an asymptotic improvement of that algorithm and solve the open problem of computing the infix probabilities of PFAs from streaming data, which is crucial when process- ing queries online and is the ultimate goal of the incremental approach.

We continue the study of generating se-mantically correct regular expressions from natural language descriptions (NL). The current state-of-the-art model SemRegex produces regular expressions from NLs by rewarding the reinforced learning based on the semantic (rather than syntactic) equivalence between two regular expressions. Since the regular expression equivalence problem is PSPACE-complete, we introduce the EQ_Reg model for computing the simi-larity of two regular expressions using deep neural networks. Our EQ_Reg mod-el essentially softens the equivalence of two regular expressions when used as a reward function. We then propose a new regex generation model, SoftRegex, us-ing the EQ_Reg model, and empirically demonstrate that SoftRegex substantially reduces the training time (by a factor of at least 3.6) and produces state-of-the-art results on three benchmark datasets.

In finite-state language processing pipelines, a lexicon is often a key component. It needs to be comprehensive to ensure accuracy, reducing out-of-vocabulary misses. However, in memory-constrained environments (e.g., mobile phones), the size of the component automata must be kept small. Indeed, a delicate balance between comprehensiveness, speed, and memory must be struck to conform to device requirements while providing a good user experience.In this paper, we describe a compression scheme for lexicons when represented as finite-state transducers. We efficiently encode the graph of the transducer while storing transition labels separately. The graph encoding scheme is based on the LOUDS (Level Order Unary Degree Sequence) tree representation, which has constant time tree traversal for queries while being information-theoretically optimal in space. We find that our encoding is near the theoretical lower bound for such graphs and substantially outperforms more traditional representations in space while remaining competitive in latency benchmarks.

In natural language processing, a common task is to compute the probability of a phrase appearing in a document or to calculate the probability of all phrases matching a given pattern. For instance, one computes affix (prefix, suffix, infix, etc.) probabilities of a string or a set of strings with respect to a probability distribution of patterns. The problem of computing infix probabilities of strings when the pattern distribution is given by a probabilistic context-free grammar or by a probabilistic finite automaton is already solved, yet it was open to compute the infix probabilities in an incremental manner. The incremental computation is crucial when a new query is built from a previous query. We tackle this problem and suggest a method that computes infix probabilities incrementally for probabilistic finite automata by representing all the probabilities of matching strings as a series of transition matrix calculations. We show that the proposed approach is theoretically faster than the previous method and, using real world data, demonstrate that our approach has vastly better performance in practice.