On this page

Sentiment Analysis

The sentiment node scores the tone of text your pipeline has already produced. It does not look at the media itself — it takes text from an upstream node (Tika document content, OCR text, a caption, or a VLM / LLM answer) and labels it positive, neutral or negative, with a confidence score and a signed polarity value you can sort and filter on.

German and English are handled by dedicated models — German with german-sentiment-bert, English with a RoBERTa model trained on social text — rather than one generic multilingual model, because a language specialist is measurably better on its own language. Both run behind a small HTTP sidecar, so the node stays a pure client and needs no model runtime of its own.

Kind

sentiment

Applies to

Any asset (the text comes from an upstream node, not from the media)

Inputs

Text from an upstream node output — configurable, defaults to the first non-empty of tika, ocr, captioning, vlm, llm

Output keys

sentiment_label (String), sentiment_score (Number), sentiment_result (JSON String)

Languages

German and English, detected automatically; other languages fall back to a multilingual model

Requirements

A running sentiment sidecar (/v1/sentiment). CPU is sufficient; GPU optional

Persists to

asset_json_comp — one row per text source, discriminated by the source output key

What it produces

{
  "label": "NEGATIVE",
  "score": 0.87,
  "polarity": -0.81,
  "scores": { "positive": 0.06, "neutral": 0.07, "negative": 0.87 },
  "lang": "de",
  "model": "oliverguhr/german-sentiment-bert",
  "source": { "nodeId": "tika", "outputKey": "tika_content" },
  "textChars": 4211,
  "chunks": 12
}

polarity is the positive share minus the negative share, so it runs from -1 (wholly negative) to +1 (wholly positive) with neutral text near zero. It is the value to sort or threshold on when you want "the most negative documents first".

Long text is handled automatically: the sidecar splits it into sentence-aligned chunks, scores each, and combines them weighted by length — so a long complaint is not diluted by a short pleasantry.

Configuration

SentimentNodeOptions:

Option Meaning

sentimentHost / sentimentPort

Address of the sentiment sidecar (default localhost:9110)

language

de, en, or auto to detect it per text (default auto)

modelDe / modelEn

Override the sidecar’s German / English model — e.g. a finance-specific German model

textSources

Ordered nodeId:outputKey list; the first one with text wins (default tika:tika_content, ocr:ocr_text, captioning:caption_result, vlm:vlm_result, llm:llm_result)

maxChars

Upper bound on the text sent for analysis (default 200000)

Because textSources is an ordered list, one pipeline configuration covers mixed collections: a PDF is scored from its Tika text, a scanned page from its OCR text, and a photo from its caption — without separate node instances.

Use Cases

  • Review and feedback triage — surface the most negative documents in an ingest first.

  • Brand and press monitoring — track the tone of coverage across a German and English corpus.

  • Content moderation support — flag strongly negative material for human review.

  • Media library insight — combine with captioning to score the tone of what images and videos depict.

The sidecar is a small FastAPI service; see the sidecars/sentiment directory for how to run it.

Model attribution

The default English model is published under CC-BY-4.0, which permits commercial use but requires attribution. The model actually used is recorded with every result, so it can be surfaced wherever you display sentiment. If your deployment cannot carry that attribution, point modelEn at an Apache-2.0 alternative — the sidecar’s fallback multilingual model is one.