The word “Social” has taken a whole new meaning in today’s digital era. Simply going out to enjoy is no longer the only “social” criteria. Social now is — giving a peek in your personal and professional life to your connections. Facebook, Twitter, Instagram, and other leading platforms have connected people in ways that were unimaginable 20 years ago. More than that, these platforms have become an excellent resource for businesses to reach new customers, understand the perception of their brand, seek feedback, and improve customer experience. Today, analytics has become a driving force for businesses, enabling them to use the power of data to grow user base and revenues.

在当今的数字时代，“社交”一词具有全新的含义。 单纯地出去享受不再是唯一的“社交”标准。 现在的社交活动-窥视您的个人和职业生活中的人际关系。 Facebook，Twitter，Instagram和其他领先平台以人们20年前无法想象的方式联系人们。 不仅如此，这些平台已成为企业吸引新客户，了解其品牌认知，寻求反馈并改善客户体验的绝佳资源。 如今，分析已成为企业的驱动力，使企业能够利用数据的力量来扩大用户群和​​收入。

情感分析到底是什么？ (What exactly is Sentiment Analysis?)

One such power, Sentiment Analysis — a concept popular amongst Machine Learning enthusiasts — is leveraged by companies to understand customer sentiment or emotion regarding the company’s products. Such an analysis crawls social media platforms to gather data on user sentiments on specific products by the company. It analyzes each user’s comment to classify it as positive, negative, or neutral, and to provide the overall result. The market has witnessed an exponential growth since 2016.

公司利用“ 情感分析” ( Sentiment Analysis)这种强大的功能来吸引客户对公司产品的情感或情感，这种概念在机器学习爱好者中很流行。 此类分析会爬行社交媒体平台，以收集有关公司针对特定产品的用户情绪的数据。 它分析每个用户的评论以将其分类为肯定，否定或中立 ，并提供总体结果。 自2016年以来，该市场呈指数增长。

传统方法瓦解的地方！！ (Where the traditional approach falls apart!!)

I see Sentiment Analysis as a powerful technique that is yet to be fully tap-into. The current textual analysis has many shortcomings. “Yeah, no one does it better than you!”. The software will classify it as a positive sentiment. But what if it was Sarcasm? Here is another one. “…Amazon always does it”. What should I make of this comment? Let’s take a step back in the text and see a sentence before this. “Amazon is great in delivering products on time. Amazon always does it” — makes the latter sentence positive. However, “Amazon just delivered a damaged piece. Amazon always does it”- makes the text negative. Context is very crucial in understanding a sentiment. Text analysis has been missing the context of the conversation!

我认为情绪分析是一项功能强大的技术，目前尚未充分利用。 当前的文本分析有许多缺点。 “是的，没有人比你做得更好！” 该软件会将其归类为积极情绪。 但是，如果是讽刺呢？ 这是另一个。 “……亚马逊总是这么做”。 我该如何评价？ 让我们退后一步，看看前面的句子。 “亚马逊在按时交付产品方面很棒。 亚马逊总是这样做。 但是，“亚马逊刚刚交付了一块损坏的物品。 亚马逊总是这么做”-使文字否定。 背景对于理解情绪至关重要。 文本分析一直缺少对话的内容！

Another problem is — we have been analyzing only textual data. There are more channels to explore. “iPhone 11 pro complete review” — a user searches on YouTube. From a teenager to an adult, everyone searches for a product review on the internet before buying it. “This is my review video on the new Bose’s Noise Cancellation headphones” — a passionate technology user on Twitter. Defeating word of mouth, word of such videos has become a direct influencer of people’s buying decisions. These videos are an excellent resource for businesses to seek user sentiment and feedback on their products.

另一个问题是-我们一直仅分析文本数据。 还有更多探索渠道。 “ iPhone 11专业版完整评论”-用户在YouTube上搜索。 从青少年到成年人，每个人都在购买前在互联网上搜索产品评论。 “这是我关于新型Bose降噪耳机的评论视频” – Twitter上的一位热情技术用户。 这类影片的口碑不佳，已成为人们购买决定的直接影响者。 这些视频是企业寻求用户情绪和产品反馈的绝佳资源 。

欢迎来到语境情感分析的新时代-我的方法 (Welcome to the new age of context sentiment analysis — My approach)

In this article, I suggest a new approach towards Sentiment Analysis — Context-based hierarchical analysis of videos uploaded by product users. I decided to analyze not only text but also visuals and audio by extracting important attributes — facial expressions, speech tone, and voice intensity. To capture the context, we will analyze the smallest unit of speech separated by pauses — ‘’utterance’’. We want each utterance to seek information from the previous and the next utterance. Bi-directional Long short-term memory (LSTM) serves such requirements. A side benefit of LSTM– it resolved the vanishing/exploding gradient issue that I feared the network would face while learning long term dependencies. We want to analyze not only the video frames but also the changes in consecutive frames. 3D- Convolutional Neural Network was built just for the job!

在本文中，我建议一种新的情感分析方法-对产品用户上传的视频进行基于上下文的层次分析。 我决定通过提取重要的属性(面部表情，语音和语音强度)来分析文本，还分析视觉和音频 。 为了捕获上下文，我们将分析由停顿分隔的最小语音单位-“话语”。 我们希望每个话语都从上一个和下一个话语中寻找信息。 双向长期短期记忆(LSTM)满足了此类要求。 LSTM的附带好处–解决了我担心网络在学习长期依赖关系时将面临的消失/爆炸梯度问题。 我们不仅要分析视频帧，还要分析连续帧的变化。 3D卷积神经网络原为 专为工作而建！

Tackling one problem at a time, I developed the following algorithm:

一次解决一个问题，我开发了以下算法：

1. Extract Features for Text, Audio, and Video for each utterance

1.为每种话语提取文本，音频和视频的功能

• Text features from transcripts of spoken words using Convolutional Neural Network使用卷积神经网络从口语笔录中提取文字特征
• Audio features by using an open-source tool like OpenSMile通过使用诸如OpenSMile之类的开源工具的音频功能
• Visual feature extraction using 3D-CNN使用3D-CNN进行视觉特征提取

2. For each channel (Text, Audio, and Video)

2.对于每个频道(文本，音频和视频)

• Send the extracted features through a Bidirectional Long short-term memory (bi-LSTM) Neural Network to obtain context incorporating features

  通过双向长短期记忆(bi-LSTM)神经网络发送提取的特征，以获得包含特征的上下文

3. Append context features of Text, Audio, and Video channels and feed to an LSTM network

3.附加文本，音频和视频通道的上下文功能，并提供给LSTM网络

4. Send the output to a dense layer and then SoftMax layer for classification, using categorical cross-entropy on utterance’s SoftMax output for training

4.将输出发送到一个密集层，然后发送到SoftMax层进行分类，使用话语的SoftMax输出上的分类交叉熵进行训练

5. After the training phase, pass the test set through the network to get context incorporating features

5.在训练阶段之后，使测试集通过网络以获取包含功能的上下文

6. Feed those features through SVM for classification

6.通过SVM提供这些功能以进行分类

The above approach incorporates the shortcomings of traditional text sentiment analysis and achieves a better accuracy of ~80% on MOSI data, which contains video reviews annotated by sentiment polarity.

上述方法结合了传统文本情感分析的缺点，并在MOSI数据上获得了约80％的更好准确性 ，该数据包含以情感极性注释的视频评论。

为企业创造价值 (Value creation for businesses)

A) To make a more informed decision regarding your brand and products

A) 对您的品牌和产品做出更明智的决定

Insightful sentiment analysis eliminates the guesswork involved in evaluating the performance of your products. Based on the insights, you can adjust to the current market needs, and increase your customer satisfaction. With such data in hand, precisely calculating customer-retention becomes easier. You can also use sentiment analysis to evaluate a new product concept before bringing it to life by putting the idea through concept testing and analyzing customer sentiment.

深入的情感分析消除了评估产品性能时的猜测。 根据这些见解，您可以适应当前的市场需求，并提高客户满意度。 有了这些数据，精确计算客户保留率变得更加容易。 您还可以使用情感分析来评估新产品概念，然后通过概念测试和分析客户情感将其付诸实践。

B) To gain a competitive edge in the market

B) 获得市场竞争优势

Run the tool to get sentiments on your competitor’s products. Such knowledge will act as an incentive to keep up with the market and boost the performance of your products. It can also help you realize consumer trends early and leverage the same to get the edge in the market.

运行该工具，以获取竞争对手产品的观点。 这些知识将激励您紧跟市场发展并提高产品性能。 它还可以帮助您及早实现消费者趋势，并利用它们来获得市场优势。

C) To enhance customer experience

C) 增强客户体验

Many consumers share their experiences with the internet community through their online feedback. Their tone and temperament can be identified and labeled as a positive, negative, or neutral sentiment. So, you can know what is correctly implemented in your products, and what needs further improvement.

许多消费者通过在线反馈与互联网社区分享他们的经验。 他们的语气和气质可以被识别并标记为积极，消极或中性的情绪。 因此，您可以知道在您的产品中正确实现了哪些功能以及需要进一步改进的内容。