Unlocking Text Sentiment Analysis: Concepts, Tasks, and Huawei Cloud’s Advances

Basic Concepts

Why: With the widespread adoption of mobile internet, users routinely voice opinions online—product reviews, social‑media comments, policy feedback—creating valuable commercial insights. Companies can detect spikes in negative sentiment and respond quickly, which is a core use case of sentiment analysis.

What: Text sentiment analysis aims to extract five elements from unstructured comments—entity, aspect, sentiment, holder, and time. The entity and aspect together form the evaluation target, which can be identified and classified.

Task Types

Current research usually ignores holder and time, focusing on the remaining three elements. Simplified tasks include word‑level, sentence/paragraph‑level, and target‑level sentiment analysis.

Word‑Level Sentiment Analysis

The goal is to build a sentiment lexicon that assigns sentiment information to individual words. Two representation methods are common: discrete (e.g., {positive, negative, neutral}) and multidimensional (e.g., Valence‑Arousal‑Dominance, Evaluation‑Potency‑Activity).

Example: happy‑positive, birthday‑positive, car‑accident‑negative, disaster‑negative.

Multidimensional examples: the VAD model represents “car accident” as (2.05, 6.26, 3.76) on a scale of 1‑9.

Common Lexicon Construction Methods

Manual annotation yields high accuracy but is costly. Automated approaches start with seed words and expand labels using methods such as pointwise mutual information, label propagation on word graphs, or regression/classification models trained on seed features.

Our Progress (Word‑Level)

Using annotated seed lexicons and automated expansion, we built the industry’s largest multidimensional sentiment lexicon (≈6 million entries) based on the Valence‑Arousal model with values ranging from –1 to 1.

Sentence‑Level Sentiment Analysis

This task predicts the overall sentiment polarity of a sentence or document. It is a typical text‑classification problem, often solved with pre‑trained language models fine‑tuned on labeled data.

Our Progress (Sentence‑Level)

We have deployed sentiment models for e‑commerce, automotive, and social media domains, supporting Chinese text with binary labels (positive/negative) and confidence scores.

Target‑Level Sentiment Analysis

Unlike sentence‑level analysis, target‑level analysis distinguishes sentiment toward specific entities or entity‑aspect pairs. Three sub‑tasks are defined: TG‑ABSA (entity with multiple aspects), TN‑ABSA (entity only), and T‑ABSA (entity‑aspect pairs).

Our Progress (Target‑Level)

We propose a single‑model multi‑attribute output approach, enabling simultaneous prediction of sentiment for multiple aspects. In the automotive domain, the model predicts sentiment for eight attributes (interior, power, appearance, cost‑performance, handling, energy consumption, space, comfort) with confidence‑based filtering.

Summary

This article presented the concept of sentiment analysis, illustrated Huawei Cloud’s research and product progress across word‑level, sentence‑level, and target‑level tasks, and highlighted available services in the “EI Experience Space” mini‑program for practical use.