Comparing Generated Responses to Reference Data

Suppose you need to judge whether a generated response is good. A tempting shortcut is to hold it up against a reference response, a trusted answer written in advance, and ask how much the two texts look alike on the surface. In evaluation work this shortcut has a name: lexical similarity. It is one of the most common ways to score generated text, and it is easy to apply. It is also easy to misread.

The belief many learners start with is this: if a generated response shares more words with the reference, it must be the better response. The reasoning sounds sensible. Overlap seems like evidence of correctness, and a score is a number that can be ranked. The trouble is that the measurement only sees matching surface material. It does not see whether the meaning is right, whether the answer is useful, or whether a different but correct wording was simply phrased differently from the reference.

This article follows one reference sentence through several measurements, and it ends by showing where those measurements break. Keep the reference sentence in mind: My cats scare the mice.

Lexical similarity measures how much two texts overlap. The first step is to break each text into smaller units called tokens. In the simplest form of the measure, you count how many tokens two texts have in common, and you express that count relative to the reference. In the source's example, each token is a word, so the reference My cats scare the mice has 5 tokens.

Overlap is the number of tokens that appear in both the generated response and the reference, divided by the number of tokens in the reference. The reference is the denominator because it is the yardstick the response is measured against. Shared tokens are the numerator. The result is usually written as a percentage.

The diagram uses a generated response, Cats frighten the mice with my friends, to show the split. Only scare is missing from the response's match. Here the response has the words my, cats, the, and mice in common with the reference, so the overlap count comes from those four shared tokens. Any token that appears only in the response, like frighten, adds nothing to the count, and any reference token the response lacks, like scare, reduces it.

Counting whole shared words is coarse. A response that spells a word slightly differently gets no credit for that word at all. Approximate string matching, sometimes called fuzzy matching, handles this by measuring similarity through the number of edits needed to turn one text into another. That number is called the edit distance. A smaller edit distance means the texts are more similar.

The three usual edit operations are deletion, which removes a character, insertion, which adds a character, and substitution, which replaces one character with another. Each operation changes the text by exactly one character. Edit distance counts how many of these operations are needed to get from the first text to the second.

N-gram similarity changes the unit of comparison. Instead of checking single tokens, it checks sequences of tokens, called n-grams. A 1-gram, also called a unigram, is a single token. A 2-gram, or bigram, is a set of two consecutive tokens. Larger n-grams follow the same pattern with longer windows. Because the unit is a sequence, n-gram similarity notices when words appear in the same order, which single-word counting does not.

To find the bigrams of a sentence, slide a two-token window across it one position at a time. Each position produces one bigram. The reference sentence My cats scare the mice has five tokens, so the window produces four bigrams in order.

The measure works by checking what percentage of the n-grams in the reference also appear in the generated response. For the reference sentence My cats scare the mice, the four bigrams are my cats, cats scare, scare the, and the mice. Setting the single-word units beside these two-word units shows the difference in granularity: five unigrams against four bigrams, with each bigram carrying information about which word came next.

This is the main gain from bigrams. The scrambled response would still share all five unigrams with the reference, but it shares few bigrams. Bigram overlap therefore rewards phrases that match the reference's local word order. That is useful, but it still counts surface sequences rather than meaning.

Several named metrics apply these ideas in different ways, including BLEU, ROUGE, METEOR++, TER, and CIDEr. They differ in exactly how the overlapping is calculated. BLEU in particular is a common lexical similarity metric, especially for translation tasks.

Before foundation models became prevalent, BLEU, ROUGE, and their relatives were common. Since then, fewer benchmarks use lexical similarity, though some still do, such as WMT, COCO Captions, and GEMv2.

The misconception breaks down for two separate reasons, and both come from the reference data rather than from the generated text. Lexical similarity can only reward resemblance to the references that were written. If a correct answer uses different words, it has nothing to match against.