Decomposing Sentiment Analysis into Multi Stage Pipelines

48 concepts42 papers1 proposal

When humans evaluate an opinion, we naturally perform a series of cognitive eliminations. Before deciding whether a sentence is a scathing critique or a rave review, we first determine if it contains an opinion at all. If we are tracking opinions on a specific topic, we ignore sentences that are off-topic, regardless of how emotional their language might be.

Early machine learning approaches, however, often treated sentiment analysis as a "flat" classification problem—feeding raw text into a model to directly output a polarity label (positive, negative, neutral) or a fine-grained rating (1 to 5 stars). This single-stage approach forces a model to solve two fundamentally different problems simultaneously: relevance detection (Is there an opinion here, and what is it about?) and polarity classification (Is it good or bad?).

To overcome this, a prominent research direction decomposes these tasks into multi-stage hierarchical pipelines. By decoupling the problem, these architectures optimize specialized features, handle complex mixed sentiments, and dramatically reduce noise.


The Gatekeeper Principle: Subjectivity Filtering

The most common multi-stage architecture introduces a "gatekeeper" stage to separate objective facts from subjective opinions. In raw text corpora, neutral or factual sentences vastly outnumber opinionated ones. A flat classifier easily gets overwhelmed by this class imbalance, frequently misclassifying subtle sentiments as neutral, or vice versa.

By implementing a two-stage pipeline, the system first filters out objective text. As Muhammad Abdul-Mageed et al. demonstrate in Subjectivity and Sentiment Analysis of Modern Standard Arabic, separating fact-opinion discrimination from sentiment direction significantly improves overall classification accuracy. This cascading logic ensures that the polarity model is only triggered for text that actually bears an opinion, a mechanism also leveraged by Eric Breck et al. in Cornell System Description for the NTCIR-6 Opinion Task to prevent the misclassification of neutral text.

[ Raw Text ] ---> [ Stage 1: Subjectivity Filter ] ---> Objective (Discard/Route)
                               |
                               v (Subjective Subset Only)
                  [ Stage 2: Polarity Classifier ] ---> Positive / Negative

This filtering is especially vital in targeted sentiment tasks where background noise is exceptionally high. For instance, Yufang Hou et al. show in The Cool Cucumber System at the 2017 TAC KBP BeSt Evaluation that filtering for subjectivity before determining polarity effectively manages the high volume of neutral data, preventing the sentiment classifier from being overwhelmed by the vast majority of objective entity pairs found in raw text. Similar pipeline structures are utilized across diverse domains, from analyzing bilingual social media text (Understanding Language Preference for Expression of Opinion and Sentiment: What do Hindi-English Speakers do on Twitter?) to building large-scale business intelligence platforms (Web Information Mining and Decision Support Platform for the Modern Service Industry).


Isolating What from How: Aspect and Topic Relevance

In real-world text, sentiment is rarely global; it is directed at specific aspects, topics, or underlying values. A single sentence might praise a restaurant's food but lambast its service. Flat models struggle to map these coexisting, often conflicting sentiments.

Multi-stage pipelines solve this by first identifying the presence of a topic or aspect, and then evaluating the sentiment towards that specific target. In the context of aspect-based sentiment analysis (ABSA), Huimin Chen et al. deployed a two-stage transformer pipeline in Country Image in COVID-19 Pandemic: A Case Study of China to first perform multi-label classification for aspect detection and then execute sentiment classification for each identified aspect. This decoupled approach allows for granular analysis of complex text where multiple topics coexist. Similarly, Lois Rink et al. show in Aspect-Based Sentiment Analysis for Open-Ended HR Survey Responses that using the identified aspects as explicit features alongside the original text significantly sharpens downstream sentiment accuracy.

This separation of "theme" from "attitude" extends naturally to stance detection and ideological profiling. Songtao Liu et al. note in Ideology Takes Multiple Looks: A High-Quality Dataset for Multifaceted Ideology Detection that structuring the process as a sequential pipeline—relevance classification followed by stance identification—ensures that ideological labels are only predicted for facets explicitly addressed within the input text, reducing noise from irrelevant dimensions. The same intuition applies when mapping human values: Nadav Borenstein et al. in Investigating Human Values in Online Communities use a multi-label model to detect the presence of specific human values before a secondary model determines the polarity of the expressed sentiment.


Feature Optimization and Intermediate Representations

Why does this separation inherently perform better? The core reason lies in feature optimization. The linguistic cues that signal whether someone is expressing an opinion or discussing a topic are often entirely different from the cues that signal which direction that opinion goes.

By decomposing the pipeline, researchers can optimize distinct feature sets or models for each subtask. Soo-Min Kim et al. explicitly leverage this in Automatic Identification of Pro and Con Reasons in Online Reviews, showing that decomposing the task into a two-stage pipeline allows for different feature sets to be optimized for relevance detection versus polarity classification.

Furthermore, the first stage can generate explicit intermediate representations—such as opinion snippets or extracted segments—that make the second stage's job much easier. For example, Yuncong Li et al. demonstrate in Multi-Instance Multi-Label Learning Networks for Aspect-Category Sentiment Analysis that recognizing opinion snippets as intermediate representations improves the accuracy of assigning sentiments to their corresponding categories, rather than trying to directly associate categories with raw words.


Compositional and Multi-Level Aggregation

Multi-stage architectures do not just filter top-down; they also aggregate bottom-up, moving from local analysis to global understanding.

To predict precise positions on a multi-point rating scale (like a 1-to-5 star rating), Bo Pang et al. show in Seeing Stars: Exploiting Class Relationships for Sentiment Categorization with Respect to Rating Scales that decomposing the task into local sentence-level sentiment identification followed by a global meta-analysis yields a much more precise position on a rating scale than a flat multi-class model.

[ Sentence 1 ] ---> Local Sentiment
[ Sentence 2 ] ---> Local Sentiment  ===> [ Global Aggregator ] ---> Precise Rating Scale (1-5 Stars)
[ Sentence 3 ] ---> Local Sentiment

Similarly, Qi Zhang et al. show in FDU at TREC 2007: Opinion Retrieval of Blog Track that combining sentence-level subjectivity classification with document-level sentiment scoring improves the accuracy of identifying opinionated content. This multi-level approach allows systems to distinguish between objective reporting and subjective commentary by aggregating local linguistic features and global document context.


Disentangling Intent and Context

Finally, decoupling classification stages allows models to handle complex pragmatic nuances, such as distinguishing between an action and the reporting of an action.

Consider the challenge of detecting sensitive or toxic content. A user might quote or report a sexist remark to condemn it, rather than to endorse it. A flat classifier would flag the text based purely on the presence of toxic words. To solve this, Patricia Chiril et al. design a hierarchical pipeline in He said "who's gonna take care of your children when you are at ACL?": Reported Sexist Acts are Not Sexist that first identifies the presence of sensitive content and then applies a secondary classifier to determine the discursive intent. This two-stage process allows the system to distinguish between the original performance of a speech act and the reporting of that act for documentation or testimony.

By breaking the problem into sequential, specialized questions, multi-stage architectures transform sentiment analysis from a brittle pattern-matching exercise into a structured, context-aware reasoning process.

Key Foundational Work

Recognizing Contextual Polarity in Phrase-Level Sentiment Analysis (2005) — contributed 2 concept elements to this direction.
Automatic Identification of Pro and Con Reasons in Online Reviews (2006) — contributed 1 concept elements to this direction.
Seeing Stars: Exploiting Class Relationships for Sentiment Categorization with Respect to Rating Scales (2005) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hierarchical_sentiment_classification_4cc5e785,
  title = {hierarchical sentiment classification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-sentiment-classification/rdr_4cc5e78548544de58c80cd28c662fe55},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_hierarchical_sentiment_classification_4cc5e785_v1,
  title = {hierarchical sentiment classification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-sentiment-classification/rdr_4cc5e78548544de58c80cd28c662fe55/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake