Warped Language Models for Noise Robust Language Understanding

Detection of automatic speech recognition (ASR) errors is crucial to preventing their further propagation through statistical machine translation (SMT) in conversational spoken language translation (CSLT) systems. In this paper, we venture beyond traditional features obtained from the ASR decoder and hypothesized word sequence, and explore additional information streams provided by an error-robust CSLT system, including SMT confidence estimates and posteriors from named entity detection (NED). Another significant novelty of this work is the use of an automated word boundary detector based on acoustic-prosodic features to verify the existence of ASR-hypothesized word boundaries, which further improves ASR error detection. Offline evaluation on a test set designed to invoke ASR errors showed that at 10% false alarm rate, the proposed features provide 2.8% absolute (4.2% relative) improvement in detection rate over a state-of-the-art baseline error detector that uses a rich set of features traditionally employed in the existing literature.

A significant source of errors in Automatic Speech Recognition (ASR) systems is due to pronunciation variations which occur in spontaneous and conversational speech. Usually ASR systems use a finite lexicon that provides one or more pronunciations for each word. In this paper, we focus on learning a similarity function between two pronunciations. The pronunciations can be the canonical and the surface pronunciations of the same word or they can be two surface pronunciations of different words. This task generalizes problems such as lexical access (the problem of learning the mapping between words and their possible pronunciations), and defining word neighborhoods. It can also be used to dynamically increase the size of the pronunciation lexicon, or in predicting ASR errors. We propose two methods, which are based on recurrent neural networks, to learn the similarity function. The first is based on binary classification, and the second is based on learning the ranking of the pronunciations. We demonstrate the efficiency of our approach on the task of lexical access using a subset of the Switchboard conversational speech corpus. Results suggest that on this task our methods are superior to previous methods which are based on graphical Bayesian methods.

Pretrained masked language models (MLMs) require finetuning for most NLP tasks. Instead, we evaluate MLMs out of the box via their pseudo-log-likelihood scores (PLLs), which are computed by masking tokens one by one. We show that PLLs outperform scores from autoregressive language models like GPT-2 in a variety of tasks. By rescoring ASR and NMT hypotheses, RoBERTa reduces an end-to-end LibriSpeech model's WER by 30% relative and adds up to +1.7 BLEU on state-of-the-art baselines for low-resource translation pairs, with further gains from domain adaptation. We attribute this success to PLL's unsupervised expression of linguistic acceptability without a left-to-right bias, greatly improving on scores from GPT-2 (+10 points on island effects, NPI licensing in BLiMP). One can finetune MLMs to give scores without masking, enabling computation in a single inference pass. In all, PLLs and their associated pseudo-perplexities (PPPLs) enable plug-and-play use of the growing number of pretrained MLMs; e.g., we use a single cross-lingual model to rescore translations in multiple languages. We release our library for language model scoring at https://github.com/awslabs/mlm-scoring.

The remarkable success of pretrained language models has motivated the study of what kinds of knowledge these models learn during pretraining. Reformulating tasks as fill-in-the-blanks problems (e.g., cloze tests) is a natural approach for gauging such knowledge, however, its usage is limited by the manual effort and guesswork required to write suitable prompts. To address this, we develop AutoPrompt, an automated method to create prompts for a diverse set of tasks, based on a gradient-guided search. Using AutoPrompt, we show that masked language models (MLMs) have an inherent capability to perform sentiment analysis and natural language inference without additional parameters or finetuning, sometimes achieving performance on par with recent state-of-the-art supervised models. We also show that our prompts elicit more accurate factual knowledge from MLMs than the manually created prompts on the LAMA benchmark, and that MLMs can be used as relation extractors more effectively than supervised relation extraction models. These results demonstrate that automatically generated prompts are a viable parameter-free alternative to existing probing methods, and as pretrained LMs become more sophisticated and capable, potentially a replacement for finetuning.

Today, most of the large-scale conversational AI agents such as Alexa, Siri, or Google Assistant are built using manually annotated data to train the different components of the system including automatic speech recognition (ASR), natural language understanding (NLU), and entity resolution (ER). Typically, the accuracy of the machine learning models in these components are improved by manually transcribing and annotating data. As the scope of these systems increase to cover more scenarios and domains, manual annotation to improve the accuracy of these components becomes prohibitively costly and time con-suming. In this paper, we propose a system that leverages customer/system interaction feedback signals to automate learning without any manual annotation. Users of these systems tend to modify a previous query in hopes of fixing an error in the previous turn to get the right results. These reformulations, which are often preceded by defective experiences caused by either errors in ASR, NLU, ER, or the application. In some cases, users may not properly formulate their requests (e.g., providing partial title of a song), but gleaning across a wider pool of users and sessions reveals the underlying recurrent patterns. Our proposed self-learning system automatically detects the errors, generates reformulations, and deploys fixes to the runtime system to correct different types of errors occurring in different components of the system. In particular, we propose leveraging an absorbing Markov Chain model as a collaborative filtering mechanism in a novel attempt to mine these patterns, and coupling it with a guardrail rewrite selection mechanism that reactively evaluates these fixes using feedback friction data. We show that our approach is highly scalable, and able to learn reformulations that reduce Alexa-user errors by pooling anonymized data across millions of customers. The proposed self-learning system achieves a win-loss ratio of 11.8 and effectively reduces the defect rate by more than 30 percent on utterance level reformulations in our production A/B tests. To the best of our knowledge, this is the first self-learning large-scale conversational AI system in production.

Sequential Recommender Systems (SRSs) predict items of interest for users based on their historical interactions. Two key popular paradigms for SRs are unidirectional Next Item Prediction (NIP) and bidirectional Masked Language Modeling (MLM). NIP performs well in recommendation tasks but is constrained by its reliance on prior information and a rigid temporal order assumption, limiting its ability to capture dynamic personalized preferences. MLM, on the other hand, enhances sequence representation and personalization by leveraging bidirectional context, but its objective is misaligned with recommendation tasks. To this end, we develop MetaSR, a meta-learning-based SR approach that ingeniously combines both paradigms into a unified framework to enhance personalized sequential recommendation. Specifically, we treat the mask prediction task in MLM as the inner-loop task (i.e., the support set), adjusting mask items dynamically to generate sequences that match individual behavioral patterns to capture personalized user preferences. In the outer loop (i.e., the query set), we adopt the NIP paradigm to directly train MetaSR toward the SR target. The joint training bridges the gap between MLM and SRs tasks, also overcoming NIP’s limitations in capturing dynamic personalized preferences. Furthermore, to improve the overall efficiency, we develop a reinforcement learning-based Adaptive Masked Sequence Selection (AMSS) mechanism, automatically selecting the optimal masking prediction tasks within the meta-learning support set to accelerate the meta-learning and achieve better personalization. Experiments on multiple public benchmark datasets show that MetaSR significantly outperforms existing SRs models. The dataset and codes are available at MetaSR .

Nowadays, automatic speech recognizers have become quite goodin recognizing well prepared fluent speech (e.g. news readings).However, the recognition of spontaneous speech is still problem-atic. Some important reasons for this are that spontaneous speechis usually less articulated and contains a lot of disfluencies. In thispaper, a new methodology for coping with disfluencies is presentedand evaluated. The basic idea is to detect disfluencies and to deter-mine the nature of these disfluencies prior to the recognition, and touse that information to control/modify the search. At present, themethodology has been elaborated for filled pauses (FP) and wordrepetitions (WR). It enables us to eliminate about one associated normal word error per disfluency without introducing a significantaugmentation of the computational load. 1. Introduction The most frequently occurring disfluencies in spontaneous speechare filled pauses (FPs). If one has a sufficiently large spontaneousspeech corpus with FP transcriptions available, one can create ded-icated acoustic FP models to supplement the traditional triphonemodel set. One can also introduce lexical FP models (pronuncia-tion variants in the lexicon). In [7], a combination of acoustic andlexical FP models yielded a reduction of the word error rate (WER)of 7.8 % relative for medical dictation (with respect to a baselinesystem with one FP model in the lexicon).Even when good FP models are available, FPs may still con-fuse the recognizer because they lead to word sequences (-grams)that are not observed in the kind of written material that is normallyused for language model (LM) development. Extending this mate-rialwithspontaneous speech transcriptionswould beanother meansof improving a spontaneous speech recognizer. However, the nec-essary resources to accomplish this for a general task are usuallynot available (and certainly not for Dutch). For a limited domainlike medical dictation, the situation is different, and experiments byPakhomov [4]showed thatthetechnique can work : the WERcouldbe reduced from 32.2 to 28.5 %.A simple way of dealing with FPs at the level of the LM is toignore the FP in the LM context of the next word. However, Stol-cke and Shriberg [8] come to the surprising conclusion that discard-ing FPs from the trigram context actually increases the perplexity.However, they looked at the speech stretches that were supplied totherecognizer and thatweredelimitedonthebasisofacoustic crite-ria. The FPs occurring at sentence boundaries often appeared in themiddle of such stretches. By only discarding sentence internal FPsthe perplexity did decrease, as expected. In [5] the speech stretchescorresponded to sentences and all FPs were sentence internal. Forthis material, the discarding strategy resulted in a 4 % decrease ofthe overall perplexity and a 30 % decrease of the perplexity of theword after an FP.The problems with the discarding of FPs in the LM are that thenotion of a sentence is not always clear in spontaneous speech, andthat even when it is, the sentence boundaries only become availableafter the recognition (with FPs) has been accomplished.Apart from FPs, there are other important disfluency types inspontaneous speech, such as the word repetition (WR). It is definedas an event consisting of a completed (not interrupted) word (W)and a repetition (R) of that word. Clearly, WRs do not require anynew acoustic or lexical models. However, just like FPs, they maydisturb the recognizer at the level of the language model.In this paper we investigate adaptive search methodologies thataim at improving the recognition of spontaneous speech in thevicinity of FPsand WRs without affecting the performance in otherregions. In section 2 we briefly outline the basic principles of thenew adaptive approach and we further elaborate and evaluate it insections 3 and 4.

Cross-lingual Spoken Sentence Retrieval (CLSSR) remains a challenge, especially for queries including OOV words such as person names. This paper proposes a simple method of fuzzy matching between query names and phones of candidate audio segments. This approach has the advantage of avoiding some word decoding errors in Automatic Speech Recognition (ASR). Experiments on Mandarin-English CLSSR show that phone-based searching and conventional translation-based searching are complementary. Adding phone matching achieved 26.29% improvement on F-measure over searching on state-of-the-art Machine Translation (MT) output and 8.83% over Entity Translation (ET) output.

This paper investigates how to effectively incorporate a pre-trained masked language model (MLM), such as BERT, into an encoder-decoder (EncDec) model for grammatical error correction (GEC). The answer to this question is not as straightforward as one might expect because the previous common methods for incorporating a MLM into an EncDec model have potential drawbacks when applied to GEC. For example, the distribution of the inputs to a GEC model can be considerably different (erroneous, clumsy, etc.) from that of the corpora used for pre-training MLMs; however, this issue is not addressed in the previous methods. Our experiments show that our proposed method, where we first fine-tune a MLM with a given GEC corpus and then use the output of the fine-tuned MLM as additional features in the GEC model, maximizes the benefit of the MLM. The best-performing model achieves state-of-the-art performances on the BEA-2019 and CoNLL-2014 benchmarks. Our code is publicly available at: https://github.com/kanekomasahiro/bert-gec.

Machine reading comprehension (MRC) is a challenging natural language processing (NLP) task. It has a wide application potential in the fields of question answering robots, human‐computer interactions in mobile virtual reality systems, etc. Recently, the emergence of pretrained models (PTMs) has brought this research field into a new era, in which the training objective plays a key role. The masked language model (MLM) is a self‐supervised training objective widely used in various PTMs. With the development of training objectives, many variants of MLM have been proposed, such as whole word masking, entity masking, phrase masking, and span masking. In different MLMs, the length of the masked tokens is different. Similarly, in different machine reading comprehension tasks, the length of the answer is also different, and the answer is often a word, phrase, or sentence. Thus, in MRC tasks with different answer lengths, whether the length of MLM is related to performance is a question worth studying. If this hypothesis is true, it can guide us on how to pretrain the MLM with a relatively suitable mask length distribution for MRC tasks. In this paper, we try to uncover how much of MLM’s success in the machine reading comprehension tasks comes from the correlation between masking length distribution and answer length in the MRC dataset. In order to address this issue, herein, (1) we propose four MRC tasks with different answer length distributions, namely, the short span extraction task, long span extraction task, short multiple‐choice cloze task, and long multiple‐choice cloze task; (2) four Chinese MRC datasets are created for these tasks; (3) we also have pretrained four masked language models according to the answer length distributions of these datasets; and (4) ablation experiments are conducted on the datasets to verify our hypothesis. The experimental results demonstrate that our hypothesis is true. On four different machine reading comprehension datasets, the performance of the model with correlation length distribution surpasses the model without correlation.

Voice search is the technology that enables users to access information using spoken queries. Automatic speech recognizer (ASR) is one of the key modules for voice search systems. However, the high error rate of the state-of-the-art large vocabulary continuous speech recognition (LVCSR) is the bottleneck for most voice search systems. In this paper, we first build a baseline system using language model (LM) with domain-specific information. To improve our system, we propose a forward-backward LVCSR system combination method to decrease the search errors in speech recognition. This also helps to improve the spoken language understanding (SLU) performance. Experiment results show that our proposed method improves the performance of speech recognition by 5.7% relative CER reduction and increases the F1-measure of SLU by 1.5% absolute on our test set.KeywordsVoice searchLVCSRSLUforward-backwardsystem combination

How do masked language models (MLMs) such as BERT learn contextual representations? In this work, we analyze the learning dynamics of MLMs. We find that MLMs adopt sampled embeddings as anchors to estimate and inject contextual semantics to representations, which limits the efficiency and effectiveness of MLMs. To address these issues, we propose TACO, a simple yet effective representation learning approach to directly model global semantics. TACO extracts and aligns contextual semantics hidden in contextualized representations to encourage models to attend global semantics when generating contextualized representations. Experiments on the GLUE benchmark show that TACO achieves up to 5x speedup and up to 1.2 points average improvement over existing MLMs. The code is available at https://github.com/FUZHIYI/TACO.

Knowledge probing assesses to which degree a language model (LM) has successfully learned relational knowledge during pre-training.Probing is an inexpensive way to compare LMs of different sizes and training configurations.However, previous approaches rely on the objective function used in pre-training LMs and are thus applicable only to masked or causal LMs.As a result, comparing different types of LMs becomes impossible.To address this, we propose an approach that uses an LM's inherent ability to estimate the log-likelihood of any given textual statement.We carefully design an evaluation dataset of 7,731 instances (40,916 in a larger variant) from which we produce alternative statements for each relational fact, one of which is correct.We then evaluate whether an LM correctly assigns the highest log-likelihood to the correct statement.Our experimental evaluation of 22 common LMs shows that our proposed framework, BEAR, can effectively probe for knowledge across different LM types.We release the BEAR datasets and an open-source framework that implements the probing approach to the research community to facilitate the evaluation and development of LMs.

In this paper, a pipelining empirical mode decomposition is presented to reduce the computing time of the emotionalized spontaneous speaker or speech recognition processing. This is a novel approach for integrating the pipelining technique into the standard empirical mode decomposition of the Hilbert-Huang transform. In addition, there is reduced about 45% of the computing time when the emotionalized spoken signal through our segmentation and pipelining processes. Based on the designed processing of emotionalized spontaneous speaker or speech recognition, the segmented and processed voice signals are recomposed back for constructing the speech and speaker models, or to identify which existed model is the most similar one. In the final part of this paper, a comparison of the speech recognized rate between standard and pipelining empirical mode decompositions are presented, and an equivalent effect in the recognition will be found. In practice, speaker or speech recognitions in an emotionalized spontaneous speech are very difficult. The existing speech recognition methods often fail to capture inherent voiceprint features from an emotionalized speech, such as the voice with a passionate intonation. And some of the existed methods to extract the pure voiceprint from an emotionalized spoken signal are very expensive in computation and time, so that technique is impossible to use in a real-time environment like smart houses. But, this paper presents a solution to improve the emotionalized spontaneous speaker or speech recognition processing to fit the real-time request.

This paper presents two different approaches utilizing sta-tisticallanguagemodel(SLM)andsupportvectormachines(SVM) for sentence boundary detection of spontaneousJapanese. In the SLM-based approach, linguistic likeli-hoods and occurrence of pause are used to determine sen-tence boundaries. To suppress false alarms, heuristic pat-terns of end-of-sentence expressions are also incorporated.On the other hand, SVM is adopted to realize robust classi-fication against a wide variety of expressions and speechrecognition errors. Detection is performed by an SVM-based text chunker using lexical and pause information asfeatures. We evaluated these approaches on manual and au-tomatic transcription of spontaneous lectures and speeches,and achieved F-measures of 0.85 and 0.78, respectively. Index Terms : sentence boundary detection, spontaneousspeech, statistical language model, support vector ma-chines. 1. Introduction Recent advance of automatic speech recognition (ASR)technology, especially for spontaneous speech, enables var-ious applications such as spoken document archiving andretrieval, speech summarization and speech translation. Toorganizeaspokendocumentinastructuredformandtogiveuseful indices, transcriptions should be segmented into ap-propriate units like sentences. Moreover, these applicationsare usually built by combining an ASR system with naturallanguage processing (NLP) systems such as a parser and amachine translator, which often assume that input text is asentence. However,sentencesinspontaneousspeechareill-formed, and sentence boundaries are indistinct. Output textby ASR systems is just a sequence of words and has no ex-plicitsentenceboundaries, sothefurtherstepofsegmentingthe ASR output is required for these applications.Automatic boundary detection of spoken sentences hasbeen explored mainly on broadcast news (BN)tasks[1,2,3]and conversational telephone speech (CTS) tasks[3, 4, 5] inEnglish. As features for detection, pause, prosodic and lin-guistic information is often used. Most popular approach isa combination of prosodic and linguistic information[2, 3],which realizes high performance on BN and CTS tasks.Prosody-based approaches[1, 5] have also been investi-gated. Meanwhile, linguistic information is not used by it-self, since most of these works were performed on Englishdata, where cue words or expressions of sentence bound-aries are not easily defined.In Japanese, cue expressions are typically observed atthe end of sentences and expected to be useful for bound-ary detection. However, variety of such expressions is solarge in spoken Japanese, that it is hard to collect sufficientamountofdatafortrainingstatisticalmodelssuchasamax-imum entropy (ME) model. Moreover, many of cue expres-sions consist of particles, which are apparently difficult tobedetectedinASR.Thus,robustnessforASRerrorsshouldalso be investigated.In this paper, we address two approaches of sen-tence boundary detection for spontaneous Japanese. Asframeworks of detection, we adopt and compare statisti-cal language model (SLM) and support vector machines(SVM). The proposed approaches are evaluated with reallectures and speeches included in the Corpus of Sponta-neous Japanese (CSJ).