Appendix C — Glossary · Sözlük (EN ↔︎ TR)

A running bilingual glossary, grouped by topic and alphabetized within each group. Terms marked ◆ are conventionally left in English in Turkish technical writing; a Turkish gloss is given where a usable one exists, and proper nouns and metric names are left as they are. This table is also the terminology contract between the two editions: both pages are generated from the same glossary.csv, so edit there first and translate second.

C.1 Agents

English	Türkçe	Notes
agent loop	ajan döngüsü	a while-loop around a model with tools: the model chooses an action, the runtime executes it, the result is appended, repeat until the model stops calling tools
code mode	kod kipi	the model writes code, run in a sandbox, that calls the tools instead of calling them one at a time; at the extreme a general interpreter replaces the tool list, trading a readable allowlist for composability
compaction	sıkıştırma	summarising or dropping parts of an agent’s transcript so the context stops growing; lossy, and you choose the loss before knowing what will be needed
context isolation	bağlam yalıtımı	running a read-heavy sub-task in its own window and returning only its conclusion, so the parent’s context never carries what it read — the reading is still paid for, by the sub-agent
dead-letter queue	dead-letter queue ◆	where a task that keeps failing lands so a person reads it, instead of retrying forever or vanishing
effective context length	etkin bağlam uzunluğu	the length at which a model still performs, as opposed to the length it accepts; routinely far shorter than the advertised window
guardrail	korkuluk	a runtime check that reduces accidental harm; not a security boundary, which is Chapter 36’s distinction
handoff	handoff ◆	a coordinator passes the conversation itself to another agent, which now owns it; unlike delegation, nothing is left to summarise or reject what the second agent does
idempotency	idempotency ◆	a call that can be repeated without repeating its effect; what makes a retried agent step safe
parallel tool calling	paralel araç çağırma	a model emitting several tool calls in one turn, which collapses many loop iterations into one and is often one of the largest levers on agent cost when the calls are independent
progressive disclosure	kademeli açılım	loading only the tool definitions relevant to a step instead of all of them, so a large tool surface does not tax every call
prompt injection	prompt enjeksiyonu	instructions planted in content the model reads, which it cannot distinguish from your own
step cap	adım bütçesi	the maximum number of loop iterations; not a tuning parameter but an admission that the loop’s exit is a model decision and may never be taken
tool call	araç çağrısı	a structured request the model emits — a function name and arguments — which the runtime, not the model, decides whether to execute
tool schema	araç şeması	the name, description, types and enums a model sees; prompt text rather than documentation, because the model gets one shot from it and cannot check a guess
trajectory	trajectory ◆	the sequence of decisions, tool calls, observations and retries that led an agent to its answer; two runs can share the answer and differ in the trajectory

C.2 Annotation and data quality

English	Türkçe	Notes
adjudication	uzlaştırma	a third pass that settles disagreements between two annotators
annotation guideline	etiketleme kılavuzu	the written rules that make labels reproducible by someone else
class imbalance	sınıf dengesizliği	one label dominates, so accuracy has to be read against the majority-class baseline
Cohen’s kappa	Cohen kappa ◆	agreement corrected for how much chance alone would produce
deduplication	tekilleştirme	removing near-identical documents; yinelenen ayıklama also seen
gold standard	altın standart	the labels treated as correct when scoring a model
ground truth	referans etiketler	the labels treated as correct when scoring; manufactured by people and processes, not discovered
inter-annotator agreement (IAA)	etiketleyiciler arası uyum	how often two annotators independently pick the same label
label leakage	etiket sızıntısı	a feature that encodes the answer; inflates scores until deployment
label noise	etiket gürültüsü	labels that disagree with the text they describe; concentrates at the top of a confidence-sorted error list
model card	model kartı	the short document stating what a model was trained on and where not to use it
out-of-fold prediction	out-of-fold kestirim ◆	each training row is predicted by a model that never saw it, so confident disagreements flag labels worth reading
provenance	provenance ◆	where a record came from and when: source, date, licence, the event that produced it and how confident the extractor was

C.3 Decoding and structured output

English	Türkçe	Notes
abstention	çekimserlik	refusing to answer when confidence is low, so the answers you do give are more reliable; needs a score to threshold on
constrained decoding	kısıtlı kod çözme	masking the logits at each step to the tokens a grammar allows, so invalid output becomes unreachable rather than unlikely
decoding	kod çözme	turning the model’s per-step scores into actual tokens; çözümleme would be analysis, a different operation
logprob	log olasılık	the log probability the model assigned to a token; the cheapest confidence signal available. Cleanest when the answer is one token — over several tokens the conditional log probabilities add, and sums over different lengths or tokenizations are defined but not comparable at face value
min-p	min-p ◆	a cutoff scaled to the top token’s probability rather than to a fixed mass, so it adapts to how peaked the distribution is
structured output	yapılandırılmış çıktı	output constrained to a schema rather than merely requested in prose

C.4 Evaluation

English	Türkçe	Notes
benchmark contamination	benchmark bulaşması	test items that were in a model’s training data, so the score reports memorisation rather than capability
position bias	konum yanlılığı	a judge preferring whichever slot a candidate sits in (first or second, depending on the model); measured by running the same comparison in both orders and counting the flips
statistical power	istatistiksel güç	the probability that an evaluation detects a real difference of a given size; what decides how many items an eval set needs

C.5 Evaluation and serving

English	Türkçe	Notes
average precision (AP)	ortalama kesinlik (AP) ◆	a weighted mean of precision at each threshold where recall increases; a summary of the precision-recall curve rather than its trapezoidal area, and unlike AUC it moves with the class balance, so quote the base rate beside it
base rate	taban oran	the share of a corpus that belongs to a given class; for the largest class it is also the accuracy of always guessing it
benchmark	kıyaslama ◆	benchmark dominant
bias (in a model or corpus)	yanlılık	not önyargı, which is the psychological sense
Brier score	Brier skoru ◆	mean squared error of a predicted probability; rewards a model for being right and honest about it
calibration	kalibrasyon	whether a score of 0.8 really means “right about 80% of the time”; a separate property from accuracy
cascade	cascade ◆	answer with a cheap model and escalate only when a validator rejects the answer
checksum	sağlama toplamı	check digits that make a mistyped identifier detectable; which errors are caught depends on the scheme — the TCKN checksum in the guardrails chapter catches every single-digit error, and not every scheme does
confusion matrix	karışıklık matrisi	for a binary classifier, the four counts — TP, FP, FN, TN — that the usual metrics are computed from; with K classes it is a K×K table
continuous batching	sürekli batchleme ◆	new requests join a running batch as slots free, instead of waiting for the whole batch to finish; the throughput lever in every modern server
decision threshold	karar eşiği	the cut that turns a score into a label; a free parameter you choose on held-out data, not a property of the model
drift	kayma	a metric moving while your code did not: the model behind the API changed, the inputs moved, or the surrounding system changed — three causes that look identical on a dashboard. Input drift is the second
escalation	aktarma	sending a request the cheap path could not handle to a stronger model or a person; its rate sets what a cascade costs
evaluation	değerlendirme
expected calibration error (ECE)	beklenen kalibrasyon hatası ◆	average gap between confidence and observed accuracy across bins; always say whether you binned on the top label
F1 score / macro-F1	F1 skoru / makro-F1 ◆	harmonic mean of precision and recall; macro averages over classes, so a rare class counts as much as a common one
FLOP	FLOP ◆	one floating-point operation; the unit training compute is counted in
inference	çıkarım
intrinsic / extrinsic evaluation	içsel / dışsal değerlendirme	benchmark vs. downstream task
latency	gecikme
Matthews correlation coefficient (MCC)	Matthews korelasyon katsayısı ◆	a balanced single number for a confusion matrix; gives every degenerate baseline (always predicting one class, say) exactly 0
model FLOPs utilization (MFU)	model FLOP kullanımı ◆	share of a chip’s peak throughput a real run actually achieves
precision	kesinlik ◆	of the items the model flagged, the share it got right; precision dominant in running text, glossed once per chapter
prefill / decode	prefill / decode ◆	the two halves of a request: the prompt is read in parallel, then output tokens are produced one at a time. Different costs, different bottlenecks
prompt caching	prompt önbellekleme	reusing the KV cache of a prompt prefix across requests, so a repeated system prompt is not recomputed
quantization	nicemleme
recall	duyarlılık ◆	of the items it should have flagged, the share it actually found; recall dominant in running text, glossed once per chapter
ROC curve / AUC	ROC eğrisi / AUC ◆	recall against false-positive rate; on an imbalanced corpus it looks flattering because the denominator is the large class
service level objective (SLO)	hizmet düzeyi hedefi	the latency or availability target a system is held to, stated as a percentile rather than a mean
span	span ◆	one timed unit of work in a trace; aralık would be interval, a different thing
speculative decoding	spekülatif kod çözme ◆	a small draft model proposes several tokens and the large model verifies them in one pass; same output distribution, fewer serial steps
throughput	işlem hacmi / verimlilik
time to first token (TTFT)	ilk token’a kadar geçen süre (TTFT) ◆	how long the reader waits before anything appears — queueing, prefill and the first decode step; prefill often dominates on long prompts, and it is the number streaming is meant to improve
tracing	izleme	recording each step of a request as a linked tree of spans

C.6 Language modelling

English	Türkçe	Notes
backoff	geri çekilme	when a long context has no counts, fall back to a shorter one
chain rule (of probability)	zincir kuralı	rewrites the probability of a sentence as a product of next-word probabilities; exact, not an approximation
held-out data	ayrılmış veri	text the model never trained on; the only honest place to measure
interpolation	interpolasyon	mix estimates from several context lengths using fitted weights
Markov assumption	Markov varsayımı	only the last n−1 words matter; false, but it makes counting possible
maximum likelihood estimation (MLE)	en çok olabilirlik kestirimi	the parameter values that make the observed data most probable; for a categorical distribution that works out to the raw count ratio, which gives probability zero to anything unseen
perplexity	perplexity ◆	how many words the model is effectively choosing between at each step
smoothing	düzleştirme	move probability mass onto unseen events so nothing is impossible

C.7 Machine translation

English	Türkçe	Notes
back-translation	geri çeviri	translate monolingual target text back into the source to manufacture training pairs; the resource that exists when parallel data does not
beam search	ışın araması	keep the k best partial outputs at each step; standard in translation, and actively harmful in open-ended generation
BLEU	BLEU ◆	word n-gram precision against a reference, with a brevity penalty; harsher on morphologically rich languages for a measurable reason
chrF	chrF ◆	the same idea at the character level, which gives partial credit for a word whose stem is right and whose suffix is wrong
COMET	COMET ◆	a learned metric: a model trained to predict human judgements, so it inherits whatever the training data contained
translate-test	çeviri ile test	translate incoming text into the model’s language at inference; cheap to build, and every translation error is now a model error
translate-train	çeviri ile eğitim	translate training data into the target language and train there; pays the translation cost once instead of per request

C.8 Modelling, training, and statistics

English	Türkçe	Notes
ablation	ablasyon ◆	remove one component to measure what it was contributing
adapter	adaptör	a small trainable module added to a frozen model; LoRA is the form that won
alignment	hizalama	making a model behave the way people asked for, rather than the way its training text happened to
annealing	tavlama	decaying the learning rate to zero at the end of a run, usually on better data
baseline	karşılaştırma modeli	the simple model a complex one has to beat to be worth its cost
catastrophic forgetting	yıkıcı unutma	felaket unutması also seen; training on a new task degrades tasks the model already handled, because the same weights held both
continued pretraining	sürdürülen ön eğitim	more of the pretraining objective on new text; adds knowledge and vocabulary, not the ability to follow an instruction
cross-validation	çapraz doğrulama	rotate which slice is held out, then average the scores
curse of dimensionality	boyut laneti
design matrix	tasarım matrisi	the document-term matrix is one
direct preference optimization (DPO)	doğrudan tercih optimizasyonu ◆	DPO dominant; trains on preference pairs directly, with no reward model and no reinforcement-learning loop
epoch	devir ◆	one full pass over the training data; epok also seen
feature	öznitelik	değişken is variable, a different thing; feature engineering → öznitelik mühendisliği
fine-tuning	ince ayar
gradient descent	gradyan inişi
HDBSCAN	HDBSCAN ◆	finds clusters as regions denser than their surroundings and labels the rest noise; needs no k
hyperparameter	hiperparametre	a setting you choose rather than learn; tune on dev, never on test
isotonic regression	izotonik regresyon	a non-parametric monotone calibration map fitted by the pool-adjacent-violators algorithm (PAVA); needs roughly a thousand held-out points
k-means	k-ortalamalar	partitions points into k groups around their centres; assumes spherical, similarly sized clusters and assigns every point
knowledge distillation	bilgi damıtma	training a small model on a large one’s outputs; it can inherit fluency, but not correctness the large model never had
LoRA (low-rank adaptation)	düşük ranklı uyarlama ◆	LoRA dominant; freezes the pretrained weights and trains a small pair of matrices beside them
loss function	kayıp fonksiyonu
overfitting	aşırı öğrenme / aşırı uyum
oversampling / undersampling	aşırı örnekleme / eksik örnekleme	duplicating minority rows, or discarding majority rows, to force a training set into balance
parameter-efficient fine-tuning (PEFT)	parametre-verimli ince ayar	the family LoRA belongs to: train a small fraction of the weights, leave the rest frozen
Platt scaling	Platt ölçeklemesi ◆	fits a two-parameter sigmoid to a model’s raw scores on held-out data; the small-sample calibration default
preference pair	tercih çifti	one prompt, two answers, and a judgement of which is better; easier to collect than a correct answer
pretraining	ön eğitim
regularization	düzenlileştirme
reinforcement learning from human feedback (RLHF)	insan geri bildiriminden pekiştirmeli öğrenme ◆	RLHF dominant; fits a reward model to human comparisons and then optimizes against it
reward model	ödül modeli	a model trained to score answers, standing in for the people whose judgements it was fit to
silhouette	siluet ◆	how much closer a point is to its own cluster than to the next-nearest, on a scale from -1 to 1
supervised fine-tuning (SFT)	denetimli ince ayar	one prompt, one target answer, ordinary next-token loss; the stage that turns a base model into one that answers
temperature scaling	sıcaklık ölçeklemesi	divides the logits by one fitted constant T; T > 1 softens the distribution and T < 1 sharpens it, and neither ever changes the ranking
verifiable reward	doğrulanabilir ödül	a training signal computed by checking the answer automatically rather than asking a human; only available where correctness is cheap to check
verifier	doğrulayıcı	a model trained to score candidate solutions rather than produce them; outcome-supervised ones judge the final answer, process-supervised ones judge each step

C.9 Models and architecture

English	Türkçe	Notes
attention	dikkat (mekanizması)
base model	temel model	what pretraining produces: continues text, does not yet answer
causal mask	nedensel maske	blocks every position from attending to later ones, so the model can be trained to generate left to right
context window	bağlam penceresi
feed-forward network (FFN)	ileri beslemeli ağ	the per-token sublayer of a transformer block; about two thirds of a block’s parameters at the usual d_ff = 4d, less in other designs
KV cache	KV önbelleği ◆	stored keys and values from earlier tokens, so generation does not recompute them
language model	dil modeli
large language model (LLM)	büyük dil modeli (BDM)	◆ LLM dominant
layer normalization	katman normalleştirmesi	rescales one token’s own features; not computed across the batch
mixture of experts (MoE)	uzman karışımı ◆	MoE dominant; memory follows total parameters and latency follows active ones
multi-head attention	çok başlı dikkat	several attention patterns computed in parallel; at a fixed total hidden size each head is narrower, so the cost is close to that of one wide head
open-weights model	açık ağırlıklı model	weights downloadable; training data, code and licence are separate questions
positional encoding	konumsal kodlama	the position signal added to (or applied to) token embeddings, because attention alone is order-blind
residual connection	artık bağlantı	atlama bağlantısı also seen; the shortcut that lets a sublayer learn a correction rather than a replacement
self-attention	öz-dikkat
transformer	dönüştürücü ◆	usually left as Transformer

C.10 Multimodal

English	Türkçe	Notes
contrastive pretraining	karşıtsal ön eğitim	training image and text encoders so matching pairs land close and mismatched pairs land apart; the captions are the supervision
optical character recognition (OCR)	optik karakter tanıma	turning an image of text into text; a dedicated engine tends to fail visibly and a VLM fluently, which is a difference of degree since OCR engines use language models too
patch embedding	yama gömmesi	an image cut into fixed squares, each projected to the model’s hidden size, so a picture becomes a sequence a transformer can read
vision-language model (VLM)	görme-dil modeli	a language model with a vision encoder attached, so images and text share one sequence; reads pages, charts and photographs

C.11 Prompting and generation

English	Türkçe	Notes
chain-of-thought	düşünce zinciri
chat template	sohbet şablonu	the special tokens that mark who is speaking; fine-tuning with the wrong one trains a format the server will never send
faithfulness (of reasoning)	akıl yürütme sadakati	how well a model’s stated reasoning matches the process that actually produced its answer; separate from whether either is correct
hallucination	halüsinasyon / uydurma
in-context learning	bağlam içi öğrenme	learning a task from examples placed in the prompt; no weight is updated and nothing survives the call
instruction tuning	talimat ayarı	fine-tuning on instruction-and-answer pairs so the model answers requests instead of continuing them
prompt	istem ◆	prompt dominant in practice
temperature	sıcaklık
thinking budget	düşünme bütçesi	a cap on tokens spent inside the reasoning block; hitting it does not cancel the request, it forces an answer out of an unfinished trace
verbalised confidence	sözel güven	a 0-100 number the model writes out as text rather than a probability read off its scores; it can be right on average and still separate nothing
zero-shot classification	sıfır atışlı sınıflandırma	labelling from an instruction alone, with no examples; a plain completion returns a word rather than a score, so there is no threshold to move unless you also read logprobs or ask for one

C.12 Prompting practice

English	Türkçe	Notes
demonstration	örnek gösterim	an input-output pair placed in the prompt; the unit few-shot prompting is built from
few-shot	az örnekli	putting a handful of labelled examples in the prompt; zero-shot is the no-example case
format sensitivity	biçim duyarlılığı	the score changing when the prompt’s layout changes but its meaning does not; the reason a single-prompt model comparison is unreliable
prompt template	istem şablonu	the fixed scaffolding a prompt is built from, with slots for the input; the thing to version, not the assembled string

C.13 RAG

English	Türkçe	Notes
chunk	parça	the unit you index and retrieve; a modelling decision rather than a preprocessing detail, because it sets what can be retrieved at all
grounding	dayandırma	answering from the retrieved text rather than from the model’s own memory; separate from whether the retrieval was right
lost in the middle	ortada kaybolmak	a model using the beginning and end of a long context better than its middle, so more context is not automatically more information

C.14 Representing text as vectors

English	Türkçe	Notes
analogy (task)	analoji (görevi)	king − man + woman; the arithmetic, not reasoning
anisotropy	anizotropi	the vector distribution is not the same in every direction; a shift toward a common direction (mean cosine ≠ 0) is one form of it
bag-of-words	kelime torbası	fully naturalized; no ◆
character n-gram	karakter n-gram’ı	fastText’s unit; cf. subword
co-occurrence matrix	eş-oluşum matrisi	eş-oluşum preferred over birlikte geçme
cosine similarity	kosinüs benzerliği
distributional hypothesis	dağılımsal hipotez	Harris 1954, Firth 1957
doc2vec / word2vec / GloVe / fastText	doc2vec / word2vec / GloVe / fastText ◆	model names, never translated
document frequency (df)	belge frekansı
document-term matrix	belge-terim matrisi	rows = documents, columns = terms
dot product	iç çarpım	not nokta çarpım; the linear predictor in word2vec
embedding	embedding / gömme ◆	embedding dominant; word embedding → kelime gömmesi when Turkish is used
inverse document frequency (IDF)	ters belge frekansı ◆	IDF dominant in running text
L2 normalization	L2 normalleştirme	puts documents on the unit sphere
latent semantic analysis (LSA)	örtük anlamsal analiz ◆	LSA dominant; PPMI/SVD on counts
MinHash	MinHash ◆	a fixed-size sketch of a set whose collisions estimate Jaccard similarity; the basis of near-duplicate detection at scale
nearest neighbour	en yakın komşu	komşu listesi = neighbour list
negative sampling	negatif örnekleme	replaces the softmax with binary real-or-fabricated decisions on sampled pairs; the per-example arithmetic is a logistic regression, but the score is bilinear in two learned vectors and the joint objective is non-convex
one-hot encoding	one-hot kodlama ◆	no settled Turkish gloss; one-hot takes suffixes
out-of-vocabulary (OOV)	sözlük dışı ◆	OOV oranı = OOV rate
pointwise mutual information (PMI)	noktasal karşılıklı bilgi ◆	PPMI = pozitif PMI
relatedness vs. similarity	ilişkililik / benzerlik	WordSim-353 vs. SimLex-999; do not conflate
singular value decomposition (SVD)	tekil değer ayrışımı ◆	truncated SVD → kesik SVD
skip-gram / CBOW	skip-gram / CBOW ◆	take Turkish suffixes: skip-gram’ın
sparse matrix / sparsity	seyrek matris / seyreklik
static vs. contextual embedding	statik / bağlamsal gömme	the axis chapter 3 ends on
subsampling	altörnekleme	word2vec’s frequent-word thinning; not alt örnekleme
term frequency (tf)	terim frekansı
TF-IDF	TF-IDF ◆	never translated

C.15 Retrieval

English	Türkçe	Notes
bi-encoder	çift kodlayıcı	query and document encoded separately, so documents can be indexed ahead of time; the reason dense retrieval scales
BM25	BM25 ◆	the lexical ranking function every hybrid retriever still has one half of; never translated
cross-encoder	çapraz kodlayıcı	query and document read together in one forward pass; far more accurate and impossible to precompute, so it reranks rather than retrieves
hard negative	zor negatif	a wrong document that looks right; training on these rather than random ones is most of what makes a dense retriever work
recall@k	recall@k ◆	the share of a query’s relevant documents in the top k; with one relevant document per query it reduces to did it appear at all, also called hit rate. The ceiling on what downstream grounding can use
reciprocal rank fusion (RRF)	sıra tersine dayalı birleştirme (RRF) ◆	combine ranked lists by summing 1/(k+rank); a variance-reduction trick for comparable systems, and a way to lose accuracy when they are not

C.16 Retrieval, agents, and tools

English	Türkçe	Notes
agent	ajan / etmen
chunking	parçalama
function calling	fonksiyon çağırma
reranking	yeniden sıralama
retrieval-augmented generation	erişimle zenginleştirilmiş üretim ◆	RAG dominant
tool calling	araç çağırma
vector database	vektör veritabanı

C.17 Sequence labeling and extraction

English	Türkçe	Notes
BIO tagging	BIO etiketleme	B marks a span’s first token, I its continuation, O everything else; the scheme that turns spans into per-token labels
conditional random field (CRF)	conditional random field ◆	scores a whole label sequence at once by adding a learned transition score for each adjacent label pair; models the labels given the words rather than their joint distribution, which is what lets its features overlap freely
entity linking	varlık bağlama	mapping a mention to a canonical id; distinct from recognising that the mention is an entity at all
gazetteer	varlık sözlüğü	a memorised list of known entity strings; a strong, dumb baseline and a poor generaliser
named entity recognition (NER)	adlandırılmış varlık tanıma ◆	NER dominant in practice; finding and typing mentions of people, organisations and places
nested entity	iç içe varlık	a span inside another span; BIO cannot represent it, which is why span-based decoders exist
relation extraction	ilişki çıkarma	classifying the relation that holds between two already-identified mentions
span-level F1	span düzeyinde F1	a prediction counts only if type, start and end all match; the metric token accuracy hides
token classification	token sınıflandırma	one label per token; the architecture NER, POS tagging and chunking all share
Viterbi decoding	Viterbi kod çözme ◆	picks the highest-scoring whole label sequence rather than the best label at each position independently

C.18 Summarization and generation

English	Türkçe	Notes
abstractive summarization	soyutlayıcı özetleme	the summary is generated, so it can use words the source never used – and assert facts it never contained
coverage / density	kapsam / yoğunluk	how much of a summary is copied from its source, and in how long a stretch; characterises a corpus before any model sees it
extractive summarization	çıkarımsal özetleme	the summary is a selection of source sentences, so it cannot invent wording and cannot compress below a sentence; it can still mislead by omitting the context that governs a selected sentence
faithfulness (to source)	kaynağa sadakat	whether a summary asserts only what the source supports; separate from whether it reads well, invisible to n-gram overlap, and not the same thing as faithfulness of reasoning
lead-3	lead-3 ◆	the first three sentences as a summary; a baseline that beats trained models on some corpora and not others – measure before assuming
LLM-as-judge	LLM hakemi	scoring generated text with another model; carries position, length and self-preference biases that must be measured, not assumed away
ROUGE	ROUGE ◆	n-gram overlap against a reference summary; rewards copying, blind to faithfulness, and dependent on a tokenizer that mangles Turkish

C.19 Text, tokens, and morphology

English	Türkçe	Notes
agglutinative	eklemeli	Turkish stacks suffixes on a stem; why tokenizers fragment it
byte-pair encoding (BPE)	byte-pair encoding ◆	never translated; BPE in running text
Common Crawl	Common Crawl ◆	proper noun, never translated; the monthly web crawl most corpora start from
corpus	derlem	plural derlemler
fertility	fertility ◆	of a tokenizer: tokens per word. Used to compare tokenizers on the same text, or languages under one tokenizer on matched content — and it is not the bill: languages differ in words per sentence too
hapax legomena	hapax legomena ◆	Latin, kept as-is; words occurring exactly once
lemmatization	lemmatizasyon	reducing a word to its dictionary form; see stemming — the pair is easy to swap
n-gram / bigram	n-gram ◆ / ikili n-gram	n-gram takes Turkish suffixes: n-gram’lar
normalization	normalleştirme	the preprocessing ladder; cf. L2 normalization
stemming	gövdeleme	chopping suffixes by rule; the result need not be a real root or a dictionary form (cf. lemmatization)
stop word	stop word ◆	English lists do not transfer to Turkish; etkisiz kelime is also seen, but these words are not meaningless — which to drop depends on the task
subword	alt kelime	alt kelime tokenleştirmesi = subword tokenization
token	token ◆	left as token; jeton and birim are not used
tokenization	tokenleştirme	bölütleme is also used; this book keeps the token family together
vocabulary	sözlük	also dictionary; disambiguate when both are in play
Zipf’s law	Zipf yasası

C.20 Turkish and morphology

English	Türkçe	Notes
deasciification	Türkçe karakter geri kazanımı	restoring Turkish diacritics to text typed without them; the alternative – stripping them everywhere – merges oldu with öldü. Not şapka düzeltme, which is the circumflex (â, î, û) and a different problem
dotted and dotless i	noktalı ve noktasız i	Turkish has two i letters, dotted and dotless, so İ/i and I/ı are two pairs and four case forms; str.lower() maps them wrongly in both directions and emits a combining mark that splits words
vowel harmony	ünlü uyumu	the stem’s vowels decide which form each suffix takes, which is why one grammatical ending has four or eight spellings