Sentiment Assignment

Sentiment Assignment: Our Approach

Sentiment analysis is one of the hardest text analytics functions to get right. Objective and subjective statements appear throughout human generated text. And correctly deciphering the intent and tone of these statements is tough enough for humans in some cases; imagine how hard it is for a machine to get right! Further, many text analytics solutions will assess the sentiment of an entire document – but don’t positive and negative sentiment often exist together in a single document? To solve this, some solutions will assess the sentiment of individual sentences in text. That’s better, but the same question persists: Within a sentence can’t there be positive and negative sentiment? And what about other emotions, like angry, sad or happy?

When we approach sentiment, Finch for Text® leverages the same context-based approach to analytics that make our entity extraction, disambiguation and enrichment so accurate. Going beyond assessing sentiment, we actually assign sentiment to individual entities. And we understand where it comes from and where it is directed. This is extremely valuable when examining transcribed customer service interactions or product reviews, for example. The following pages detail the multi-dimensional approach we take to sentiment assignment, and why we believe so strongly in the merits of our approach.

Core Sentiment Detection Using Deep Semantic Features

We go beyond just key words to understand sentiment in text. Our models were trained on a large and diverse corpus of news, conversational and narrative datasets to capture nuances in language that other products cannot. In the example at right, the inclusion of the word “doesn’t” makes this sentence – full of positive words like cleverness, wit and humor – a negative statement.  Our algorithms catch that.

Example

Movie reviewer Bob Smith said, “This movie doesn’t demonstrate cleverness, wit or any other kind of intelligent humor.”

Entity-Level Sentiment Assignment

Going a level deeper, our sentiment models understand sentiment at the entity level – rather than just at the sentence or document level. This context-based approach can be applied on large documents, rich with entities. The example at right shows the importance of understanding sentiment this way; the statement is positive about Capital One and negative about Deutsche Bank.

Example

Capital One Bank is expected to report a full-year $2B net profit on Wednesday, but the same fortune is not shared by Deutsche Bank.

Directional Sentiment Assignment Between Entities

Additionally, Finch for Text® can understand where entity-level sentiment emanates from and to. This is critically important in product reviews, reputational assessment and situational awareness contexts. In the example to the right, an observation in news about South African politics, we can see that former president Nelson Mandela was believed to have had a positive sentiment about candidate Ramaphosa and a negative sentiment toward candidate Thabo Mbeki.

Example

Mandela hoped Ramaphosa would succeed him, believing Mbeki to be too inflexible and intolerant of criticism, but the ANC elected Mbeki regardless.

Multi-Class Sentiment Models

Beyond just positive, negative or neutral sentiment, Finch for Text® can understand emotions expressed in text such as: aloofness, sympathy, anxiousness, surprise, etc. We employ a proprietary variation of a standard convolution neural network, with architectural and parameter tuning to understand deep semantic word vectors for varied classes of sentiment. In the example at right, from an employee email, it’s clear that this is an angry or aloof statement, rather than just a “negative” one

Example

I can’t believe this is our new strategy. There is so much more we should be considering. I’ll just do what I’m told and wait for it to fail like it always does.

Degrees of Sentiment

Finch for Text® can also interpret various degrees of sentiment. For example, whether someone is more positive, less positive; more aloof or less aloof, etc. These comparisons can be made from one author to another or from day to day to determine how an individual person’s sentiment has changed or evolved over time.

Example

Author 1: I think the team’s new plan is good. (positive)
Author 2: I think the team’s new plan is fantastic! (more positive)

Topic-Level Sentiment Assignment Model

We can also apply our algorithmic, entity-level approach to sentiment to specific topics discussed in a document. To do it, we examine a temporal window of textual conversation and employ our inference trained models to decipher the topics about which the author’s language is referring. In the example at right, we can see that while negative sentiment is expressed, it’s about a football game. The anxiousness detected is in response to sales numbers and that is a more valuable insight than the negative sentiment expressed in the message.

Example

Yeah, I’m so upset about the Lions’ loss. Oh well… What has me really worried are the Q3 sales figures. Have you seen them?

Supported Use Cases

Finch for Text® is being used or is under evaluation to support a number of business and mission critical use cases in the federal and commercial spaces. Among them:

  • Customer Call Center Analytics
  • Survey Analysis & Insights
  • Assessing Product Reviews
  • Media & Brand Monitoring
  • Insider Threat Detection
  • Intelligence Analysis
  • and more…

Full text of Sentiment Assignment page:

Finch for Text White Paper:

To learn more, please contact us at sales@finchcomputing.com or visit us at www.finchcomputing.com.