Words Without Stories: Narrative Images in 1,212 Recorded Tamarian Utterances, 2368–2375, and a Narrative Lexicon for the Universal Translator
Abstract
For most of a century the Universal Translator rendered the speech of the Children of Tama into grammatical Federation Standard that no listener could understand. The contact at El-Adrel in 2368 showed why: Tamarians speak largely through references to figures and scenes from their own narrative tradition, and the translator converts the words while leaving the stories behind. We ask how much of Tamarian speech works this way, how it is built, and whether supplying the stories repairs translation. We coded 1,212 Tamarian utterances recorded in 14 contacts between 2368 and 2375. Of these, 1,019 (84%) were narrative images, drawing on 163 distinct images. Nearly all paired a named figure with a scene, a place, event or state, and where a figure recurred with different scenes, its meaning changed with the scene far more often than the reverse. In a crossover study, 48 Starfleet officers chose the function of 60 recorded images each, from the translator's output alone or with an entry from a narrative lexicon compiled from material the Tamarians supplied after 2368. Correct choices rose from 23%, close to the 17% expected by chance, to 71% (odds ratio 8.2, 95% CI 6.0–11.2). The translator's failure with Tamarian is not a failure of translation but of reference, and it can be repaired only by adding the stories themselves.
1. Introduction
The Universal Translator is the Federation's most successful linguistic instrument. It learns an unfamiliar language from samples of speech, infers its grammar and lexicon, and renders it into the listener's own language in real time. With the Children of Tama it failed for most of a century. Each earlier contact produced grammatical, fluent translations that made no sense: strings of personal names, place names and brief descriptions of events, offered where an answer, a proposal or a warning was expected (Wolanski, 2366).
A contact at El-Adrel IV in 2368 supplied the explanation. A Tamarian captain and the commanding officer of the USS Enterprise-D, stranded together on the planet's surface, faced a common danger, and in the course of it the Starfleet officer came to understand that the Tamarian was speaking through images drawn from Tamarian narrative. One such image, naming two figures and a place, stood for two strangers joined against a shared threat; the Tamarian captain had brought about the meeting to enact it (Starfleet Command, 2368). The captain died of his wounds, but the two crews parted with a first real exchange, and the contacts recorded since then followed from that understanding.
Word by word, the translator had been correct. What it could not render was the story each utterance pointed to, because the story was nowhere in the utterance. This paper asks how much of Tamarian speech works this way, how its images are built, and whether giving the listener the stories repairs the failure.
2. Methods
Our corpus is every Tamarian utterance in the Starfleet Contact Records for the 14 contacts between the Federation and the Children of Tama from the El-Adrel meeting in 2368 to the end of 2375, with the translator's output for each (Starfleet Command, 2376). An utterance was a single turn by a single Tamarian speaker. The records hold 1,212.
We coded each utterance as a narrative image or not. A narrative image was an utterance that referred to a scene from Tamarian narrative, usually by naming a figure together with a place, an event or a state of the figure, and occasionally by naming a place or season alone. For each image we recorded the figure and the scene separately, and, where the context of the contact made it clear, its communicative function: proposal, refusal, warning, assent, report of failure, or other. Both of us coded all 1,212 utterances; we agreed on image status in 1,187 (98%) and on function in 902 of the 1,019 images (89%), and resolved the rest by discussion.
After 2368 the Tamarians supplied Federation linguists with recited narratives, from which the Academy compiled a narrative lexicon: for each image, a summary of the story it names and the situations in which it is used (Oyelaran, 2374). The 2375 revision we used documented 131 of the 163 images in our corpus, which together account for 91% of all image uses.
In 2375 we recruited 48 Starfleet officers with no Tamarian training. Each interpreted 60 image utterances sampled at random from the 902 whose function had been coded, 30 with the translator's output alone and 30 with that output and the lexicon entry, the order of conditions balanced across officers. For each, the officer chose one of the six function categories, so chance performance is about 17%. An interpretation was correct if it matched the coded function. The outcome was binary, and we fitted a mixed-effects logistic regression with condition and order as fixed effects and random intercepts for officer and utterance.
3. Results
Of the 1,212 utterances, 1,019 (84%) were narrative images. The remaining 193 were mostly proper names used alone, numbers, and a small set of navigational and technical terms that the Tamarians appear to have adopted for contact with outsiders. The 1,019 images drew on 163 distinct images; the 20 most frequent accounted for 61% of all image uses. Of the 163, 154 named a figure with a scene and 9 named a scene alone.
Scenes carried meaning more than figures did. Forty-four figures occurred with two or more different scenes, and for 39 of them (89%) the function differed across scenes. Twenty-seven scenes occurred with two or more different figures, and for only 6 (22%) did the function differ (Fisher's exact test, p < .001). An image naming a figure at a place of meeting was a proposal of cooperation; the same figure at a moment of collapse was a report of failure.
Officers chose the correct function for 23% of utterances with the translator's output alone and 71% with the lexicon entry added. In the mixed model the odds of a correct choice were 8.2 times higher with the lexicon (95% CI 6.0–11.2; p < .001). Order made no detectable difference, and officers who met the lexicon first scored 24% without it, against 22% for those who met it second. As a check on the test, utterances whose image the lexicon did not document scored 25% in both conditions, since the officer saw the same material in each. Table 1 sets out the comparison.
4. Discussion
Tamarian speech is overwhelmingly referential. Five utterances in six point to a story, and the meaning of the utterance is the role that story plays in the present situation. A listener who knows the story understands at once; a listener who does not understands almost nothing, however well the words are translated. With the translator alone, our officers did little better than chance. This is why the Universal Translator, which works from words and grammar, produced fluent nonsense for so long. It translated what was said and could not supply what was meant.
Images in Tamarian are also built, not merely memorised. Most have a figure and a scene, and the scene carries most of the function; the contact of 2368 itself showed one figure placed in two different scenes. This suggests a productive system: a speaker can in principle combine a known figure with a new scene to say something new, provided the listener knows both parts. Our corpus contains only a few images that look newly made, and we would not claim that Tamarian speakers routinely compose them, but the structure allows it, and a translator that treats each image as a fixed idiom will miss such cases.
In practice, translating Tamarian requires the stories. A narrative lexicon supplied to the officer raised comprehension about threefold. The same material supplied to the translator itself, so that it offers the story's function alongside the words, would serve the same purpose without training every officer (Oyelaran, 2374). The undocumented images show the limit: the lexicon helps only where it holds the story, and it will need to grow with every contact.
The case has a wider lesson for xenolinguistics. The Universal Translator assumes that a language's meaning can be recovered from its samples. For most languages that assumption holds, because most languages carry their meaning in words and grammar. Tamarian carries it in a shared body of narrative outside any utterance, and no quantity of samples will recover it. The Federation should expect other languages built this way and should not read a translator's fluency as understanding (Linares-Okpara, 2371). A simple warning sign is already available: the share of an unfamiliar language's utterances that consist of proper names, which was high for Tamarian long before 2368 (T'Nira, 2370).
5. Limitations
Communicative functions were coded from the context of each contact, which is itself interpreted by Federation officers; some functions may have been misread, and these misreadings would count as errors in our comprehension test. Both coders are the authors, and one of us compiled the lexicon, which may have brought coded functions and lexicon entries closer together than an independent coder would. The corpus comes from diplomatic contacts and may over-represent formal speech. The lexicon was compiled from narratives the Tamarians chose to share, which may favour images they expected outsiders to need. Finally, our officers were untrained in Tamarian, and trained contact specialists would score higher in both conditions.
References
- Starfleet Command (2368). USS Enterprise-D contact with the Children of Tama at El-Adrel IV, logs and translator record. Starfleet Contact Records, SCR-68-117.
- Starfleet Command (2376). Contacts with the Children of Tama, 2368–2375, transcripts with translator output. Starfleet Contact Records, SCR-TAM series, contacts 1–14.
- Wolanski, C. (2366). Grammatical but opaque, the Tamarian problem in Universal Translator performance. Journal of Federation Xenolinguistics, 31(1), 7–29.
- Oyelaran, L. (2374). A narrative lexicon of Tamarian images, first edition. Journal of Federation Xenolinguistics, 39(2), supplement.
- Linares-Okpara, R. (2371). Meaning outside the utterance, limits of sample-based translation. Daystrom Institute Proceedings, 57, 88–112.
- T'Nira (2370). Referential density as a predictor of Universal Translator failure. Journal of Federation Xenolinguistics, 35(3), 201–219.
Open in Uncited Press →