BIO Tagging Scheme

Every token in a sentence receives exactly one tag that encodes its position relative to a semantic argument:

• B (Beginning): Marks the first token of an argument span. A new argument always starts with a B tag.
• I (Inside): Marks any token that continues an argument after the initial B token. Multi-word arguments use consecutive I tags.
• O (Outside): Marks any token that is not part of any argument for the current predicate.

This scheme ensures that even when two arguments of the same type appear adjacent, the B prefix signals the boundary unambiguously.

A complete BIO tag concatenates the positional prefix with the semantic role label:

• B-ARG0: The first token of the Agent argument
• I-ARG0: Continuation tokens of the Agent argument
• B-ARGM-LOC: The first token of a Locative modifier
• O: Token not participating in the predicate's frame

Example sentence: "[ARG0 John] [V sold] [ARG1 the car] [ARGM-TMP yesterday]"

• John → B-ARG0
• sold → O (or B-V for predicate identification)
• the → B-ARG1
• car → I-ARG1
• yesterday → B-ARGM-TMP

A naive Inside-Outside (IO) scheme without the B prefix fails when two arguments of the same type are adjacent:

Problem case: "[ARG1 the book] [ARG1 the movie]"

• IO scheme: the/I-ARG1 book/I-ARG1 the/I-ARG1 movie/I-ARG1 — indistinguishable from a single argument
• BIO scheme: the/B-ARG1 book/I-ARG1 the/B-ARG1 movie/I-ARG1 — two arguments clearly separated

The B prefix acts as a boundary delimiter, making the encoding lossless and enabling the model to learn argument segmentation directly from token labels without relying on external syntactic parsers.

Extended tagging schemes add finer-grained boundary information:

• BIOES: Adds E (End) and S (Singleton) tags

  • B-ARG1: beginning of multi-token argument
  • I-ARG1: inside multi-token argument
  • E-ARG1: end of multi-token argument
  • S-ARG1: single-token argument
  • O: outside

• BILOU: Adds L (Last) and U (Unit) tags with equivalent semantics

These extensions provide explicit boundary signals that can improve span detection accuracy, particularly for longer arguments, at the cost of a larger output label space.

BIO tagging converts SRL into a standard sequence labeling task solvable with token-level classifiers:

• Input: A sequence of tokens with a marked predicate position
• Encoder: A Transformer (e.g., BERT) produces contextualized embeddings for each token
• Classifier: A linear layer + softmax predicts one BIO tag per token from the label vocabulary
• Decoding: Contiguous B-I sequences are collapsed into argument spans with their role labels

This architecture eliminates the need for span enumeration or syntactic parse features, enabling end-to-end neural SRL that learns both boundaries and roles jointly from annotated data.

The BIO tag vocabulary size depends on the role inventory:

• PropBank roles: ARG0 through ARG5, plus ARGM-* modifiers (LOC, TMP, MNR, CAU, etc.)
• Typical label count: 50-100 distinct BIO tags per predicate
• O tag dominance: The O tag accounts for 70-85% of all tokens, creating a severe class imbalance

Training robust BIO taggers requires:

• Large annotated corpora like OntoNotes (1.7M+ words)
• Weighted loss functions or focal loss to counteract O-tag dominance
• Predicate-aware input representations that condition token embeddings on the target predicate