📄 Squared English Word: A Method of Generating Glyph to Use Super Characters for Sentiment Analysis

Summary

Super Characters classifies text by rendering it as an image and passing that image to a 2D-CNN. The method fits scripts such as Chinese, Japanese and Korean, where each character is already a square glyph. English is harder because a word is a run of letters of varying length. This paper introduces SEW (Squared English Word). Each word gets its own square cell, and its letters are packed into that cell in a small sub-grid sized to the word's length. The cells are then tiled row by row into one image for the whole sentence. Two extensions are added. The first is an attention variant that draws chosen words in larger cells. The second encodes structured user-profile fields as extra words in the same image, so a single CNN sees both text and tabular data. On DBpedia, SEW beats the original Super Characters by 2.1 points. On the small HappyDB set from the CL-AFF Shared Task, plain SEW helps only slightly, while profile features combined with attention give the best results. The work matters because it adapts an image-based text classifier to alphabetic languages and to mixed text-plus-table inputs.

Key points

• DBpedia results (14 ontology classes; 40,000 training and 5,000 test items per class; 560,000 training texts in total):
  • Super Characters: 96.2%
  • SEW: 98.3%
  • Setup: SE-net-154 backbone, 14x14 = 196 word cells.
• HappyDB / CL-AFF setup (Affective Content Analysis workshop, AAAI 2019):
  • 10,560 happy-moment sentences, split 80/20 for training and validation.
  • Two binary labels, Agency and Social.
  • Text-only SEW uses a 6x6 grid, so sentences are cut at 36 words. This truncated 1.39% of labeled sentences and 1.91% of the 170k shared-task test set.
• Attention variant:
  • The first four words are drawn at double size in the centre of an 8x8 layout.
  • The layout is a blueprint array: 0 marks a normal cell, 1 starts an enlarged word, and -1 marks a cell to skip.
• Profile features: age, country, marital status and gender were drawn as short tokens or initials. Parenthood was dropped after an XGBoost variable-importance analysis.
• HappyDB accuracies (Agency / Social):

Interesting ideas

• I1. Fixing one cell per word, with letters packed inside it, keeps word boundaries explicit and the encoding reversible, so the original sentence can be recovered from the image.
• I2. Salience is expressed by giving chosen words more space in the input rather than through hidden weights, which makes the emphasis visible and easy to inspect.
• I3. Structured metadata (age, country, gender) is turned into short symbolic tokens and placed alongside the text, so one model consumes both sources.
• I4. Candidate side features were pruned first with a cheap importance analysis (XGBoost), and only the useful ones were kept.
• I5. The truncation length was picked from the corpus length distribution (median, max, percentiles), and the share of inputs actually cut was reported.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (back): We could require every Mova intermediate representation to pass a similar reversibility check, confirming that the text → graph → text compressor recovers each token's identity and order.
• I2 (frames): Prag could mark salient constituents (e.g. topic or headline-like sentence starts) with an explicit, readable prominence attribute that later steps consult openly.
• I3 (world): Document metadata such as source, date or genre could enter the world of the text as ordinary entities and facts, so reasoning treats them like any other stated information.
• I4 (learn): Before adding features to the perceptron parser or gparse, a quick importance pass over silver data could discard weak features and keep models smaller and easier to explain.
• I5 (tree): gparse sequence limits could be set from treebank sentence-length percentiles, with the fraction of truncated sentences logged as part of each evaluation.