Preprocessing
- Split on
;{}and newlines - Merge tiny fragments (< 3 tokens)
- CamelCase / snake_case identifier splitting
- Java keyword + English stopword filtering
NLP · CODE SUMMARIZATION
Four ways to summarize Java source code
Upload a .java file and watch corpus-fitted extractive baselines,
a semantic embedding model, and a fine-tuned transformer each generate a
java code summary, compared live side by side.
4models
liveinference
01 — Architecture
How a file becomes four summaries
Extractive models summarize the whole file from split statements. CodeT5 runs once per Java method, the same setup used in the CodeXGLUE evaluation.
Java upload
A single .java file
Preprocess
Split statements; CodeT5 splits by method
TF
TF-IDF
Term scoring
LR
LexRank
Graph centrality
ST
Sentence-T
Embeddings
T5
CodeT5
Generation
Summary
Four summaries
Corpus fitting
- TF-IDF & LexRank IDF from CodeXGLUE Java train + validation
- Weights cached to
cache/idf_weights_train_val.pkl - Neural models use frozen pretrained checkpoints
- One-time load, then served from memory
Output
- Extractive models return top-N statements from the whole file
- CodeT5 generates one English sentence per method (evaluation setup)
- Per-model latency tracked for each run
- Results compared in a single view
02 — Models
The four summarizers
Each model represents a different tier of prior knowledge. Click a card to expand its step-by-step algorithm, strengths, and limitations.
Picks statements packed with rare, high-signal terms.
Scores each code statement by TF-IDF using IDF weights fitted on the CodeXGLUE Java train + validation corpus.
How it works
- 1Fit inverse-document-frequency (IDF) weights over the Java train + validation corpus.
- 2Tokenize each statement: split identifiers, lowercase, drop stopwords.
- 3Score every statement as the sum of term-frequency x IDF.
- 4Return the top-N highest-scoring statements as the summary.
Strengths
- Fast and fully offline
- Interpretable scores
- No GPU required
Limitations
- Output is code-like, not prose
- Ignores word order and context
- Input
- Statement fragments
Selects statements most representative of the whole file.
Builds a similarity graph over statements and runs PageRank to pick the most central fragments.
How it works
- 1Build a TF-IDF vector for each statement using shared corpus IDF.
- 2Compute pairwise cosine similarity to form a statement graph.
- 3Threshold weak edges, then run PageRank over the graph.
- 4Return the most central (highest-ranked) statements.
Strengths
- Captures redundancy / centrality
- Offline and interpretable
- Robust on longer files
Limitations
- Needs several statements to rank
- Still extractive, not generative
- Input
- Statement fragments
Uses semantic meaning to find the most central statements.
Encodes statements with all-MiniLM-L6-v2 and selects those closest to the centroid embedding.
How it works
- 1Embed each statement with the all-MiniLM-L6-v2 transformer.
- 2Average the embeddings into a single centroid vector.
- 3Rank statements by cosine similarity to the centroid.
- 4Return the statements closest to the semantic center.
Strengths
- Understands English semantics
- Order-aware encoder
- No corpus fitting needed
Limitations
- Pretrained on prose, not code
- Heavier than TF-IDF/LexRank
- Input
- Statement fragments
Writes a fresh English sentence describing the code.
Generates natural-language summaries from raw Java source using a CodeT5 checkpoint fine-tuned on CodeXGLUE.
How it works
- 1Split the file into individual Java methods.
- 2Byte-level BPE tokenize each method (first 256 tokens).
- 3Decode with beam search — one English sentence per method, same as evaluation.
- 4Show each method summary separately in the results view.
Strengths
- True natural-language output
- Fine-tuned on Java code-comment pairs
- Best quality summaries
Limitations
- Slow on CPU
- 256-token input limit
- Can hallucinate details
- Input
- Per-method Java source (256-token window each)
03 — Try it
Summarize your Java file
Upload a .java file. Extractive models use the whole file; CodeT5 summarizes each method separately.