We at Zetane are all about democratizing AI, but getting to the laudable goal of empowering more people within AI innovation requires many steps along the way. One beginner’s step on our journey involves peeling away the abstract nature of AI like the skins of an onion. We peel these layers of abstraction away by presenting fundamental components of machine learning in ways technical and non-technical professionals can appreciate. This will be a repeated theme of ours on the Zetane blog, which we call:

我们在Zetane都是关于民主化AI，但得到授权向AI创新中更多的人值得称道的目标要求沿途许多步骤。 初学者迈出的第一步包括剥开AI的抽象本质，例如洋葱皮。 通过以技术人员和非技术人员可以欣赏的方式展示机器学习的基本组成部分，我们将这些抽象层剥离掉了。 这将是Zetane博客上我们主题的重复主题，我们称之为：

This is what AI looks like.

这就是AI的样子。

To start us off we present here the Bidirectional Attention Flow model, or BiDAF.

首先，我们在这里介绍双向注意力流模型或BiDAF 。

At first glance, this model looks to be a practical attempt to understand elements of unstructured text. Its development in 2016 marked a major step forward in Natural Language Processing (NLP) for Question and Answer tasks. That may at first sound straightforward. This basic task, however, has huge implications in business as a tool to automate important tasks like customer support inquiries or common requests for information. It thus comes as no surprise that this model attracts fervent investment and research. Aside from the model’s commercial applications, BiDAF marks an important milestone in opening the doors to a not-too-distant future of appreciable Machine Comprehension. Building on knowledge gained from BiDAF, the machine learning community has since brought forth the superior NLP models of BERT and ELMo.

乍一看，该模型似乎是理解非结构化文本元素的实际尝试。 它在2016年的发展标志着自然语言处理(NLP)在问答任务方面迈出了重要的一步。 乍一看听起来很简单。 但是，此基本任务作为使重要任务(例如客户支持查询或常见的信息请求)自动化的工具在业务中具有重大意义。 因此，这种模式吸引了无数的投资和研究也就不足为奇了。 除了该模型的商业应用之外，BiDAF还标志着一个重要的里程碑，它为不久的将来可观的机器理解打开了大门。 机器学习社区以从BiDAF获得的知识为基础，提出了BERT和ELMo的高级NLP模型。

For the time being, let’s take a closer look at BiDAF and some of its complexities, beginning with an image of the full neural network.

目前，让我们从完整的神经网络的图像开始，仔细研究BiDAF及其一些复杂性。

The dimensions of the BiDAF model should grab one’s attention. When viewing the model end-to-end, it is difficult to make out the appreciable complexity of the neural network. Let’s zoom in for a closer look, focusing on the left-side input layers of the network.

BiDAF模型的尺寸应引起人们的注意。 端到端查看模型时，很难确定神经网络的明显复杂性。 让我们放大一下以集中查看网络的左侧输入层。

The above image shows a typical ‘source’, being text with a jumble of information. Upon querying the basic question ‘What color is the fox?’, the model predicts the best answer is ‘brown’, being spot-on. Of particular interest here are the four input layers of the neural network, represented by the mid-green squares. Machine-learning models often have one input, but the specifics of the question-and-answer task requires a bit more work in terms of structuring and analysing the data as text. Let’s inspect further to better understand why that is the case.

上图显示了一个典型的“来源”，即带有混杂信息的文本。 在询问基本问题“狐狸是什么颜色？”时，该模型预测最佳答案是“棕色”，即亮点。 这里特别有趣的是神经网络的四个输入层，由中间的绿色方块表示。 机器学习模型通常只有一个输入，但是问答结构的细节在以文本形式构造和分析数据方面需要更多的工作。 让我们进一步检查以更好地理解为什么会这样。

The first input layers are the context_char and query_char, being abbreviations for the inputs of character-level representations of the text and the question (“query”). The latter two inputs of contex_word and query_word are where word-level representations of the text and question become incorporated in the analysis by the neural network. This architecture shows there is an initial processing of the text and then a subsequent assessment of the words in order to achieve an appreciable level of language comprehension. Overall, the BiDAF model aims to identify a context and then keywords integral to the context, where the layered design of the neural network makes it easier for the model to tease out the answer keyword from the general context. It’s interesting to note that the model does this by converting all text into math, namely vectors, meaning its ability to understand text has little to do with text per se and is in fact a probability-based prediction of pairing words.

第一个输入层是context_char和query_char，它们是文本和问题(“查询”)的字符级表示的输入的缩写。 contex_word和query_word的后两个输入是文本和问题的单词级表示形式在神经网络分析中的合并位置。 此体系结构显示了对文本的初始处理，然后对单词进行了后续评估，以实现可理解的语言理解水平。 总体而言，BiDAF模型旨在识别上下文，然后识别上下文中不可或缺的关键字，其中神经网络的分层设计使该模型更容易从通用上下文中挑出答案关键字。 有趣的是，该模型通过将所有文本转换为数学(即向量)来实现此目的，这意味着其理解文本的能力与文本本身无关，实际上是基于概率的配对词预测。

To conclude our overview of BiDAF we present a short video capture of the model in our environment for machine learning projects. Here is a display of the breakdown and analysis of the vectorized text information as it passes through nodes of the neural network. We intermittently show the internal metrics of the information at the nodes, called tensors. The take-home message here is that whether data is in the form of text, images, video or whatnot, a neural network transforms that information into a universal language of advanced — and arguably aesthetic — statistics.

总结我们对BiDAF的概述，我们将在我们的环境中为机器学习项目展示该模型的简短视频。 这是矢量化文本信息通过神经网络节点时的细目分类和分析的显示。 我们间歇地显示节点上信息的内部度量，称为张量。 这里要传达的信息是，无论数据是文本，图像，视频还是其他形式，神经网络都会将该信息转换为高级的统计数据(可以说是美学)。

This is what AI looks like.

这就是AI的样子。

You can read more about the foundational research for BiDAF right below (or follow links to the .pdf articles here & here).

您可以在下面阅读有关BiDAF基础研究的更多信息(或在此处和此处，链接到.pdf文章)。

补充阅读 (Additional reading)

We appreciate this article by Meraldo Antonio that also provides a visual representation of the BiDAF model.

我们赞赏Meraldo Antonio的这篇文章，该文章还提供了BiDAF模型的直观表示。

致谢 (Acknowledgement)

We thank Semih Cantürk for providing the images and video for this post.

我们感谢SemihCantürk为这篇文章提供的图像和视频。