CLLT

Learning to disambiguate by combining multiple sense representations

This talk will discuss ongoing work investigating the combination of multiple sense representation methods for word sense disambiguation (WSD). A variety of recent methods have been proposed for learning representations of semantic senses in different domains, and there is some evidence that different methods capture complementary information for WSD. We consider a simple but competitive cosine similarity-based model for WSD, and augment it by learning to produce a context-sensitive linear transformation of representations of candidate senses. In addition to transforming the input sense space, our method allows us to jointly project multiple sense representations into a single space. We find that a single learned projection matches or outperforms directly updated sense embeddings for single embedding methods, and demonstrate that combining multiple representations improves over any individual method alone. Further, by transforming and conjoining complete embedding spaces, we gain the ability to transfer model knowledge to ambiguous terms not seen during training; we are currently investigating the effectiveness of this transfer.

Evaluating state-of-the-art models of speaker commitment

When a speaker, Mary, utters “John did not discover that Bill lied”, we take Mary to be committed to Bill having lied, whereas in “John didn’t say that Bill lied”, we do not take that she is. Extracting such inferences arising from speaker commitment (aka event factuality) is crucial for information extraction and question answering. In this talk, we evaluate the state-of-the-art models for speaker commitment and natural language inference on the CommitmentBank, an English dataset of naturally occurring discourses, annotated with speaker commitment towards the content of the complement (“lied” in the example) of clause-embedding verbs (“discover”, “say”) under four entailment-canceling environment (negation, conditional, question, and modal). The CommitmentBank thus focuses on specific linguistic constructions and can be viewed as containing “adversarial” examples for speaker commitment models. We perform a detailed error analysis of the models’ outputs by breaking down items into classes according to various linguistic features. We show that these models can achieve good performance on certain classes of items, but fail to generalize to the diverse linguistic constructions that are present in natural language, highlighting directions for improvement.

Prediction is All You Need: A Large-Scale Study of the Effects of Word Frequency and Predictability in Naturalistic Reading

A number of psycholinguistic studies have factorially manipulated words’ contextual predictabilities and corpus frequencies and shown separable effects of each on measures of human sentence processing, a pattern which has been used to support distinct processing effects of prediction on the one hand and strength of memory representation on the other. This paper examines the generalizability of this finding to more realistic conditions of sentence processing by studying effects of frequency and predictability in three large-scale naturalistic reading corpora. Results show significant effects of word frequency and predictability in isolation but no effect of frequency over and above predictability, and thus do not provide evidence of distinct effects. The non-replication of separable effects in a naturalistic setting raises doubts about the existence of such a distinction in everyday sentence comprehension. Instead, these results are consistent with previous claims that apparent effects of frequency are underlyingly effects of predictability.

Improving classification of speech transcripts

Off-the-shelf speech recognition systems can yield useful results and accelerate application development, but general-purpose systems applied to specialized domains can introduce acoustically small–but semantically catastrophic–errors. Furthermore, sufficient audio data may not be available to develop custom acoustic models for niche tasks. To address these problems, we propose a concept to improve performance in text classification tasks that use speech transcripts as input, without any in-domain audio data. Our method augments available typewritten text training data with inferred phonetic information so that the classifier will learn semantically important acoustic regularities, making it more robust to transcription errors from the general purpose ASR. We successfully pilot our method in a speech-based virtual patient used for medical training, recovering up to 62% of errors incurred by feeding a small test set of speech transcripts to a classification model trained on typescript.

Exploring Mimic Loss for Robust ASR

We have recently devised a non-local criterion, called mimic loss, for training a model for speech denoising. This objective, which uses feedback from a senone classifier trained on clean speech, ensures that the denoising model produces spectral features are useful for speech recognition. We combine this knowledge transfer technique with the traditional local criterion to train the speech enhancer. We achieve a new state-of-the-art for the CHiME-2 corpus by feeding the denoised outputs to an off-the-shelf Kaldi recipe. An in-depth analysis of mimic loss reveals that this performance correlates with better reproduction of consonants with low average energy.

Tailoring “language agnostic” blackboxes to Arabic Dialects

Learning from the best: A teacher-student framework multilingual models for low-resource languages.

Automatic Speech Recognition (ASR) in low resource languages is problematic because of the absence of transcripted speech. The amount of training data for any specific language in this category does not exceed 100 hours of speech. Recently, it has been found that knowledge obtained from a huge multilingual dataset (~ 1500 hours) is advantageous for ASR systems in low resource settings, i.e. the neural speech recognition models pre-trained on this dataset and then fine-tuned on language-specific data report a gain in performance as compared to training on language-specific data only. However, it goes without saying that a lot of time and resources are required to pre-train these models, specially the ones with recurrent connections. This work investigates the effectiveness of Teacher-Student (TS) learning to transfer knowledge from a recurrent speech recognition model (TDNN-LSTM) to a non-recurrent model (TDNN) in the context of multilingual speech recognition. Our results are interesting in more than one level. First, we find that student TDNN models trained using TS learning from a recurrent model (TDNN-LSTM) perform much better than their counterparts pre-trained using supervised learning. Second, these student models are trained only with language-specific data instead of the bulky multilingual dataset. Finally, the TS architecture allows us to leverage untranscribed data (previously untouched during supervised training) resulting in further improvement in the performance of the student TDNNs.