Natural Language Processing as Communication Engineering

In October 2017 Israeli police arrested a Palestinian construction worker in Beitar Illit. His Arabic caption said “good morning”. Facebook's machine translation rendered it as “attack them” in Hebrew and as “hurt them” in English. No Arabic-speaking officer read the original post. He was released after a few hours, once the error was found. Facebook's own account of it ran to one sentence: “Unfortunately, our translation systems made an error last week that misinterpreted what this individual posted.” That was Necip Fazil Ayan, an engineering manager in Facebook's language technologies group.

A support message such as “I was charged twice, but the second payment vanished” can trigger several different NLP problems. A classifier might route it. An extractor might identify two transactions. A retriever might find a policy, and a generator might draft a reply. In Beitar Illit the chain was shorter and the stakes were not. One caption became a translation, a routing decision and an arrest. No component in that chain was asked whether the sentence was a greeting. Each was asked a narrower question, and the last question was answered by a police officer rather than by a model.

The same text therefore does not define one objective. The product must specify which evidence matters, what output is needed, and what happens after the output is produced.

This map keeps language phenomena separate from the software that acts on them. The second layer is the one most often treated as neutral, and it is the hardest to keep honest. A corpus is not a recording of a language. It is the residue of the rules used to collect and clean it.

C4.EN is the 365-million-document, 156-billion-token corpus behind T5 and the Switch Transformer. It was built by discarding any page containing a word from the “List of Dirty, Naughty, Obscene or Otherwise Bad Words”. One cleaning rule, applied once. Dodge and colleagues audited what that rule had deleted and reported in 2021 that “we find that AAE and Hispanic-aligned English are removed at substantially higher rates (42% and 32%, respectively) than WAE and other English (6.2% and 7.2%, respectively)”. African American English at 42%. Hispanic-aligned English at 32%. White-aligned English at 6.2%.

The recorded-data layer had made a decision about whose language counts. Every layer above it inherited that decision without being told.

Choosing the wrong task definition can make a technically strong model useless. Each of the risks named below has been measured. The measurements are worse than the intuition.

Retrieval: in 1985 a full-text retrieval system was measured while it did real legal work. The system was IBM STAIRS, supporting the defence of an actual lawsuit; Blair and Maron ran the study. Two defence lawyers generated 51 information requests. Two paralegals turned them into queries and searched until they were satisfied they had retrieved the 75% of relevant documents the lawyers had stipulated. Recall and precision were computed for 40 of those 51 requests. Average precision was 79.0% — what came back was largely relevant. Average recall was 20.0%, with a standard deviation of 15.9. Grossman and Cormack, reviewing the study in 2011, put the finding in one line: “Although the searchers believed they had found 75% of the relevant documents, their average recall was only 20.0%.” The TREC-2006 Legal Track overview attributes the gap to “the inherent ambiguity of language”. Nothing in the search session told the searchers they were missing four documents in five.

Generation: unsupported content in summaries is not a rare tail. Three trained annotators judged 2,500 document–summary pairs — 500 randomly sampled XSum articles, each paired with summaries from four neural systems (PtGen, TConvS2S, TranS2S, BertS2S) plus the human-written BBC gold summary. Maynez and colleagues report in 2020 that “intrinsic and extrinsic hallucinations happen frequently – in more than 70% of single-sentence summaries”. Over 90% of the extrinsic hallucinations were erroneous. One sentence long, judged by humans, and unsupported more often than not.

Dialogue: MultiWOZ is the field's most-used booking-assistant benchmark, 10,000 dialogues. Re-annotating it produced MultiWOZ 2.1, whose authors report that “This correction process results in changes to over 32% of state annotations across 40% of the dialogue turns.” An independent Google team then found and fixed further state errors in 17.3% of utterances on top of 2.1, in MultiWOZ 2.2. Two rounds of correction by two teams, before the dialogue-state metric measured what it claimed to.

A human reader handles these questions almost invisibly. An NLP system needs explicit design choices. Shipped products get them wrong at rates that are easy to measure, once someone writes the tests.

Microsoft Text Analytics, Google Cloud Natural Language and Amazon Comprehend were run through their paid APIs against a behavioural test suite called CheckList. Ribeiro and colleagues report the result in 2020: “The failure rate is near 100% for all commercial models when the negation comes at the end of the sentence”. Three sentiment services sold by three of the largest vendors, defeated by where the negation sits. In the same paper's user study, practitioners given CheckList wrote twice as many tests and found almost three times as many bugs.

Use this sequence to turn a vague language feature into an auditable task. Each step exists because a measured failure above skipped it. An unnamed user decision put a police officer in front of a mistranslation. An unexamined evidence boundary left average recall at 20.0%. An output contract that rewarded fluency produced hallucinations in more than 70% of single-sentence summaries.

This path focuses on language data, linguistic structure, task design, retrieval, extraction, translation, summarization, dialogue, and NLP-specific evaluation. Neural architectures appear only when they clarify a language task.

The deeper mechanics of Transformers belong to the architecture path. Prompting, tool use, fine-tuning, and RAG product design are treated more fully in the Generative AI path.