Text generation has a fundamental limitation almost by definition: there is no taking back tokens that have been generated, even when they are clearly problematic. In the context of language model safety, when a partial unsafe generation is produced, language models by their nature tend to happily keep on generating similarly unsafe additional text. This is in fact how safety alignment of frontier models gets circumvented in the wild, despite great efforts in improving their safety. Deviating from the paradigm of approaching safety alignment as prevention (decreasing the probability of harmful responses), we propose backtracking, a technique that allows language models to "undo" and recover from their own unsafe generation through the introduction of a special [RESET] token. Our method can be incorporated into either SFT or DPO training to optimize helpfulness and harmlessness. We show that models trained to backtrack are consistently safer than baseline models: backtracking Llama-3-8B is four times more safe than the baseline model (6.1\% $\to$ 1.5\%) in our evaluations without regression in helpfulness. Our method additionally provides protection against four adversarial attacks including an adaptive attack, despite not being trained to do so.

Dual encoders have been used for questionanswering (QA) and information retrieval (IR) tasks with good results.Previous research focuses on two major types of dual encoders, Siamese Dual Encoder (SDE), with parameters shared across two encoders, and Asymmetric Dual Encoder (ADE), with two distinctly parameterized encoders.In this work, we explore different ways in which the dual encoder can be structured, and evaluate how these differences can affect their efficacy in terms of QA retrieval tasks.By evaluating on MS MARCO, open domain NQ and the Mul-tiReQA benchmarks, we show that SDE performs significantly better than ADE.We further propose three different improved versions of ADEs by sharing or freezing parts of the architectures between two encoder towers.We find that sharing parameters in projection layers would enable ADEs to perform competitively with or outperform SDEs.We further explore and explain why parameter sharing in projection layer significantly improves the efficacy of the dual encoders, by directly probing the embedding spaces of the two encoder towers with t-SNE algorithm (van der Maaten and Hinton, 2008).

Robert L Logan IV, Andrew McCallum, Sameer Singh, Dan Bikel. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

We present our work on semi-supervised learning of discriminative language models where the negative examples for sentences in a text corpus are generated using confusion models for Turkish at various granularities, specifically, word, sub-word, syllable and phone levels. We experiment with different language models and various sampling strategies to select competing hypotheses for training with a variant of the perceptron algorithm. We find that morph-based confusion models with a sample selection strategy aiming to match the error distribution of the baseline ASR system gives the best performance. We also observe that substituting half of the supervised training examples with those obtained in a semi-supervised manner gives similar results.

This paper investigates semi-supervised methods for discriminative language modeling, whereby n-best lists are “hallucinated” for given reference text and are then used for training n-gram language models using the perceptron algorithm. We perform controlled experiments on a very strong baseline English CTS system, comparing three methods for simulating ASR output, and compare the results with training with “real” n-best list output from the baseline recognizer. We find that methods based on extracting phrasal cohorts - similar to methods from machine translation for extracting phrase tables - yielded the largest gains of our three methods, achieving over half of the WER reduction of the fully supervised methods.

Discriminative language modeling is a structured classification problem. Log-linear models have been previously used to address this problem. In this paper, the standard dot-product feature representation used in log-linear models is replaced by a non-linear function parameterized by a neural network. Embeddings are learned for each word and features are extracted automatically through the use of convolutional layers. Experimental results show that as a stand-alone model the continuous space model yields significantly lower word error rate (1% absolute), while having a much more compact parameterization (60%-90% smaller). If the baseline scores are combined, our approach performs equally well.

Automatic closed-captioning of video is a useful application of speech recognition technology but poses numerous challenges when applied to open-domain user-uploaded videos such as those on YouTube.In this work, we explore a strategy to improve decoding accuracy for video transcription by decoding each video with a language model (LM) adapted specifically to the topics that the video covers.Taxonomic topic classifiers are used to determine the topic content of videos and to build a large set of topic-specific LMs from web documents.We consider strategies for selecting and interpolating LMs in both supervised and unsupervised scenarios in a two-pass lattice rescoring framework.Experiments on a YouTube video corpus show a 3.6 absolute reduction in WER over generic single-pass transcriptions as well as a statistically significant 0.8 absolute improvement over rescoring with a very large non-adapted LM built from all the documents.

This article documents a large set of heretofore unpublished details Collins used in his parser, such that, along with Collins' (1999) thesis, this article contains all information necessary to duplicate Collins' benchmark results. Indeed, these as-yet-unpublished details account for an 11% relative increase in error from an implementation including all details to a clean-room implementation of Collins' model. We also show a cleaner and equally well-performing method for the handling of punctuation and conjunction and reveal certain other probabilistic oddities about Collins' parser. We not only analyze the effect of the unpublished details, but also reanalyze the effect of certain well-known details, revealing that bilexical dependencies are barely used by the model and that head choice is not nearly as important to overall parsing performance as once thought. Finally, we perform experiments that show that the true discriminative power of lexicalization appears to lie in the fact that unlexicalized syntactic structures are generated conditioning on the headword and its part of speech.

This paper presents the automatic construction of a Korean WordNet from pre-existing lexical resources. A set of automatic WSD techniques is described for linking Korean words collected from a bilingual MRD to English WordNet synsets. We will show how individual linking provided by each WSD method is then combined to produce a Korean WordNet for nouns.

This paper describes a first attempt at a statistical model for simultaneous syntactic parsing and generalized word-sense disambiguation. On a new data set we have constructed for the task, while we were disappointed not to find parsing improvement over a traditional parsing model, our model achieves a recall of 84.0% and a precision of 67.3% of exact synset matches on our test corpus, where the gold standard has a reported inter-annotator agreement of 78.6%.