tapir/internal/reconcile/disputes.go

package reconcile

import "strings"

// Dispute is one place where the two transcripts disagree. BaseStart:BaseEnd and
// AltStart:AltEnd are token ranges into the two Tokenize results; an empty range
// means that engine heard nothing there. Base and Alt are the two readings as each
// engine wrote them, so a replaced span takes the alternative's own spelling.
type Dispute struct {
	BaseStart, BaseEnd int
	AltStart, AltEnd   int
	Base, Alt          string
}

// FindDisputes aligns base against alt word by word and returns the disagreements.
// Differences separated by a single shared word are merged into one dispute: for a
// reader that is one decision, and asking about "b" and "d" separately in "b c d"
// hides that "c" sits between them. Two shared words keep disputes apart.
func FindDisputes(base, alt []Token) []Dispute {
	a, b := keysOf(base), keysOf(alt)
	ops := Diff(a, b)
	var out []Dispute
	for i := 0; i < len(ops); i++ {
		if ops[i].Kind == Equal {
			continue
		}
		cur := ops[i]
		for i+2 < len(ops) && ops[i+1].Kind == Equal && ops[i+1].AEnd-ops[i+1].AStart == 1 && ops[i+2].Kind != Equal {
			cur.AEnd, cur.BEnd = ops[i+2].AEnd, ops[i+2].BEnd
			i += 2
		}
		out = append(out, Dispute{
			BaseStart: cur.AStart, BaseEnd: cur.AEnd,
			AltStart: cur.BStart, AltEnd: cur.BEnd,
			Base: join(base[cur.AStart:cur.AEnd]),
			Alt:  join(alt[cur.BStart:cur.BEnd]),
		})
	}
	return out
}

// fillers are Swedish discourse particles. A dispute that is only one or two of them
// on one side changes no meaning, so it is resolved without asking a model. The list
// is deliberately short and leaves out words that can be an answer ("nej") or that
// carry content.
var fillers = map[string]bool{
	"ju": true, "då": true, "liksom": true, "alltså": true, "så": true, "ja": true,
	"eh": true, "ehm": true, "öh": true, "mm": true, "hm": true, "typ": true, "va": true, "asså": true,
}

// maxFillerRun is how many filler words may be waved through together.
const maxFillerRun = 2

// AutoResolve reports whether a dispute needs no model, in which case the base
// reading is kept: either the two sides are the same letters once spaces are removed
// (a compound written as one word or two), or one side is empty and the other is a
// short run of filler words.
func AutoResolve(d Dispute) bool {
	baseKeys, altKeys := keysOfText(d.Base), keysOfText(d.Alt)
	if strings.Join(baseKeys, "") == strings.Join(altKeys, "") {
		return true
	}
	return (len(baseKeys) == 0 && onlyFillers(altKeys)) || (len(altKeys) == 0 && onlyFillers(baseKeys))
}

func onlyFillers(keys []string) bool {
	if len(keys) == 0 || len(keys) > maxFillerRun {
		return false
	}
	for _, k := range keys {
		if !fillers[k] {
			return false
		}
	}
	return true
}

// ContextAround returns up to window words of the base before and after the dispute,
// as the base wrote them.
func ContextAround(base []Token, d Dispute, window int) (before, after string) {
	return join(base[max(0, d.BaseStart-window):d.BaseStart]), join(base[d.BaseEnd:min(len(base), d.BaseEnd+window)])
}

func keysOf(ts []Token) []string {
	out := make([]string, len(ts))
	for i, t := range ts {
		out[i] = t.Key
	}
	return out
}

func keysOfText(s string) []string { return keysOf(Tokenize(s)) }

func join(ts []Token) string {
	parts := make([]string, len(ts))
	for i, t := range ts {
		parts[i] = t.Text
	}
	return strings.Join(parts, " ")
}