Deep Learning systems augmented with a memory module are starting to show impressive results in natural language processing.

Judge as of February 2015. TRUE: by 18 February 2015, work had been published or posted showing deep networks augmented with an explicit memory module (for example, Memory Networks or Neural Turing Machines) achieving notable results on NLP tasks such as question answering, and these results were regarded as impressive and emerging. MOSTLY TRUE: results existed, but mainly on synthetic or toy NLP tasks rather than on real-world benchmarks. MOSTLY FALSE: memory-augmented models existed, but had shown only marginal or non-NLP results. FALSE: no memory-augmented deep learning NLP results had appeared by that date.

Set before the research began.

1. E1 Memory Networks arXiv (rendered via ar5iv.labs.arxiv.org)

   “We investigate these models in the context of question answering (QA) where the long-term memory effectively acts as a (dynamic) knowledge base, and the output is a textual response. We evaluate them on a large-scale QA task, and a smaller, but more complex, toy task generated from a simulated world.”

2. E2 Memory Networks arXiv (rendered via ar5iv.labs.arxiv.org)

   “Some answers appear fine, whereas others are nonsensical.”

3. E3 Memory Networks arXiv (rendered via ar5iv.labs.arxiv.org)

   “Our work was submitted to arxiv just before the Neural Turing Machine work of Graves et al. 2014, which is one of the most relevant related methods.”

4. E4 Neural Turing Machines arXiv (rendered via ar5iv.labs.arxiv.org)

   “Preliminary results demonstrate that Neural Turing Machines can infer simple algorithms such as copying, sorting, and associative recall from input and output examples.”

E3 shows that the Memory Networks preprint was submitted to arXiv in 2014, just before the Graves et al. 2014 Neural Turing Machine work, which is before the claim date. It tested a memory-augmented network on question answering, using a large-scale QA task and a toy task generated from a simulated world. The evidence does not show these results being described as impressive: the paper itself (E2) says some answers appear fine while others are nonsensical. The Neural Turing Machine paper (E4) reports preliminary results on algorithmic tasks such as copying, sorting and associative recall; the quoted passage does not mention language results. Memory-augmented NLP results clearly existed by February 2015, but nothing here shows they were seen as impressive on real-world benchmarks, so the claim holds largely but not fully.