Touch Feedback as Index Modulation for Interactive Translation

When a person fixes a machine translation by tapping rather than typing, every tap is a message sent back to the translation system. The system can learn from which spots were tapped, how many were tapped, and which highlighted spots were deliberately left alone. Wireless engineers have a mature theory for exactly this kind of signal. Index modulation is a family of transmission schemes that carry information not only in the waveform but in which antennas or frequency slots are switched on, and in how many are on. This proposal treats the menu of tappable choices shown to a translator as a codebook the system designs. The system builds it to get the most correction per tap, and it reads the translator's whole selection pattern, silences included, as the transmitted message. The decoder that consumes that message is a constrained translation model fast enough to redraft after every tap.

The motivation is a specific failure of the best existing no-typing interface. Touch-Based Pre-Post-Editing of Machine Translation Output had users tap correct word sequences in a draft and redecoded with the tapped material encouraged and the untapped material discouraged. On WMT'11 French–English news this raised a Moses baseline from 28.6 to 49.7 BLEU over five rounds without a typed character. The user, however, can only approve words already in the draft. A missing word or a wrong inflection cannot be introduced, a limitation that matters most for morphologically rich and low-resource targets. Typing restores that power, but typing is the fatiguing step: with weak drafts it can be slower than translating from scratch (The Impact of Machine Translation Quality on Human Post-Editing). The opening is an interface that lets taps introduce new words while keeping each tap highly informative.

Core Claim

The claim is that a small, per-sentence menu of alternatives, chosen by the system's own uncertainty, lets taps carry substantially more correction than tapping draft spans. The menu offers a few highlighted positions in the draft, each with two or three replacement words or inflections drawn from the model's own top candidates. The gain should be largest when the menu has three properties. First, it is chosen jointly rather than position by position, because one replacement changes what the redraft does elsewhere. Second, its size varies with how uncertain the model is about that sentence. Third, the redraft treats untapped highlighted alternatives and unhighlighted positions as weak evidence rather than as no evidence.

Each of these properties corresponds to a wireless result. Variable menu size is the analogue of subcarrier number modulation. In Enhanced Orthogonal Frequency-Division Multiplexing with Subcarrier Number Modulation, the transmitter signals with the count of active frequency slots and activates only the strongest slots given current channel measurements. That paper also delivers a warning: the scheme shifts error curves by roughly 2 dB for four slots and 4 dB for eight, but it does not change their slope. Placing effort where the channel is good helps, but it cannot rescue positions where everything is bad. Our analogue predicts that adaptive menus will help sentences with a few uncertain spots and will not fix drafts that are wrong throughout. We will test that prediction directly. Reading silences as signal is the analogue of sparse vector transmission. In Sparse Vector Transmission: An Idea Whose Time Has Come, the positions of zero-valued entries carry information, which mirrors the negative language models of touch post-editing.

We use the wireless theory for two concrete things. It supplies the design objective for the menu, and it supplies a way to account for feedback in bits. A single tap among $T$ highlighted positions with $K$ alternatives each, or no tap, carries at most $\log_2(TK+1)$ bits. That gives a ceiling against which the measured quality gain per tap can be compared.

Research Program

The system has three parts. The first is a fast redrafting model. It must accept positive constraints (tapped alternatives that must appear), negative constraints (offered alternatives that were declined), and soft acceptance of untouched text. Lexically Constrained Neural Machine Translation with Levenshtein Transformer supplies the method. It is an edit-based model that inserts and deletes tokens in parallel. It reaches a 100% term usage rate and gains up to 1.16 BLEU on WMT'17 English–German at about 260 sentences per second, fast enough to redraft between taps. We extend it so that negative constraints become penalties on the insertion and deletion decisions, which that paper does not support.

Penalties are deliberately soft. Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation is the efficient constrained beam search we use as a second redrafter, and it reports that hard forcing pushes models into low-probability "garbage" while they wait to satisfy a constraint. That result constrains the design: only taps become hard constraints, and silences are never forced.

The second part is the menu designer. Let $y$ be the current draft and $C$ a candidate menu of (position, alternative) pairs. Let $r$ be the translator's response, meaning the set of tapped pairs, possibly empty. Let $D(y,r)$ be the redraft, $Q$ a quality estimate, and $\lambda$ the price of a tap. The designer picks

$$C^\star = \arg\max_{C} \; \mathbb{E}_{r \sim p(r \mid C, y)}\big[\,Q(D(y,r))\,\big] - \lambda\,\mathbb{E}[\,|r|\,].$$

The expectation over responses is estimated by sampling plausible correct translations from the model itself and simulating what an accurate tapper would select. The number of highlighted positions $T$ is part of the maximization, so it varies by sentence.

The third part is a model of the human as a noisy channel. Translators miss errors. Beyond Human-Only: Evaluating Human-Machine Collaboration for Collecting High-Quality Translation Data found that editors overlook errors in fluent-looking text, the familiarity bias. Silence at a highlighted position is therefore weighted by an estimated probability that the position was actually inspected. Interaction telemetry, namely dwell and scroll events logged in OpenTIPE: An Open-source Translation Framework for Interactive Post-Editing Research, supplies that estimate in the human study. In simulation we sweep it.

Baselines

The naive reference point is the unedited machine draft, which tells us how much any interaction adds. The first interactive baseline is touch-based span selection from Touch-Based Pre-Post-Editing of Machine Translation Output. It is reimplemented on the same neural redrafting model, so the comparison isolates the interface rather than the decoder. Its reported result is 28.6 to 49.7 BLEU after five rounds on WMT'11 French–English with a Moses system. We will report it in its own effort measure, the tap-cost-weighted edit rate $TER_{PPE}$, which prices each tap at a cost $\alpha$.

The strongest competitor on our question is greedy uncertainty menus. This baseline highlights the $T$ positions where the model is least confident, using a fixed $T$, and offers the top-$K$ model alternatives at each one. It uses the same redrafter, the same candidate pool, the same tap budget and the same hyperparameter search budget as our method. It is the strongest competitor because it already contains most of the intuition, namely spending the user's attention where the model is uncertain, and it requires no wireless theory at all.

Two ablations isolate the proposal's ingredients. One removes adaptive cardinality by fixing $T$. The other removes silence-as-signal, so that only taps constrain the redraft.

For typed editing the reference point is human post-editing of a strong system. On the English–German corpus of Beyond Human-Only, this reaches 2.98 MQM, a fine-grained error score where lower is better, against 3.79 for translation from scratch. We do not expect taps to beat typing on quality. Beating this baseline means reaching comparable MQM in less logged effort time.

Our method counts as beating span selection if it is higher on the quality-versus-tap curve across the whole $\alpha$ sweep. It counts as beating greedy menus if it does so at matched tap budgets, with a statistically reliable gap.

Why not just…?

Why not just highlight uncertain words and show alternatives, without any of this machinery? This is the real threat, and it is why greedy uncertainty menus are the headline baseline rather than a straw man. The proposal's case rests on two effects that greedy selection ignores. One is coupling: fixing one position changes the redraft elsewhere, so the two least-confident positions are often redundant. The other is that the right menu size differs from sentence to sentence. If greedy menus with a tuned fixed size match the full objective, the honest conclusion is that the index-modulation framing adds bookkeeping, through the bits-per-tap ceiling, but no method. We will report that outcome as such.

Why not just prompt a large language model with the draft and the translator's typed comments? That keeps typing, which is the effort we are trying to remove. It also adds seconds of latency per redraft against sub-second edit-based decoding. Beyond Human-Only further shows that automatic refinement does not replace human correction for top quality. We include a language-model rewrite step on top of the final tapped output as an optional arm, to check whether it absorbs our gains.

Why not just let translators type, as computer-assisted translation tools already do? For good drafts, typing is likely the best option, and we expect our method's advantage to concentrate on mid-quality drafts and mobile settings. The human study measures effort time directly against typed post-editing in the same OpenTIPE interface, so the regime where taps help, if any, will be visible.

Proposed Experiments

The simulated-user experiments run on WMT'11 news and WMT'14 medical French–English, for direct comparison with span selection, and on WMT'17 English–German with its terminology constraints. They also cover FLORES-101 English into morphologically rich, lower-resource targets. The FLORES-101 Evaluation Benchmark reports long-tail directions often below 3 spBLEU, which is where the inability to introduce new words should hurt span selection most. Simulated users tap the alternative closest to the reference, under a sweep of miss rates. Metrics are BLEU, spBLEU and an edit rate after one to five rounds, plotted against taps, along with redraft latency.

A human study with professional and bilingual annotators in OpenTIPE then compares typed editing, span selection, greedy menus and our method on matched sentence sets. It is scored by MQM with logged active time.

Three checks decide the question. The first is whether the full menu objective beats greedy uncertainty menus at matched tap budgets. The second is whether the gain over span selection is largest on FLORES targets, where new inflections are needed. The third tests the subcarrier-number warning: whether improvement concentrates on drafts with few uncertain positions and vanishes on uniformly poor ones. A failure on the first check would reduce the contribution to a better span-selection interface. A failure on the third would mean the wireless analogy predicts nothing, and we would drop it from the explanation.