README.md

metadata

language:
  - tr
license:
  - cc-by-sa-4.0
multilinguality:
  - monolingual
config_names:
  - BOUN
  - IMST
dataset_info:
  - config_name: BOUN
    features:
      - name: id
        dtype: string
      - name: text
        dtype: string
      - name: tokens
        list:
          dtype: string
      - name: upos
        list:
          dtype: string
      - name: heads
        list:
          dtype: int32
      - name: rels
        list:
          dtype: string
      - name: feats
        list:
          dtype: string
      - name: feats_dict_json
        list:
          dtype: string
  - config_name: IMST
    features:
      - name: id
        dtype: string
      - name: text
        dtype: string
      - name: tokens
        list:
          dtype: string
      - name: upos
        list:
          dtype: string
      - name: heads
        list:
          dtype: int32
      - name: rels
        list:
          dtype: string
      - name: feats
        list:
          dtype: string
      - name: feats_dict_json
        list:
          dtype: string
    splits:
      - name: train
        num_bytes: 116892
        num_examples: 3435
      - name: validation
        num_bytes: 116892
        num_examples: 1100
      - name: test
        num_bytes: 116892
        num_examples: 1100
configs:
  - config_name: BOUN
    data_files:
      - split: train
        path: BOUN/train.jsonl
      - split: test
        path: BOUN/test.jsonl
      - split: validation
        path: BOUN/dev.jsonl
  - config_name: IMST
    data_files:
      - split: train
        path: IMST/train.jsonl
      - split: test
        path: IMST/test.jsonl
      - split: validation
        path: IMST/dev.jsonl

Turkish Treebank Benchmarking

This is the repo for Turkish treebank benchmarking, namely evaluating Tranformer models on POS-Dep-Morph task. For the data, we used two treebank, IMST and BOUN. We converted conllu format to json lines for being compatible to HF dataset formats.

Here are treebank sizes at a glance:

Dataset	train lines	dev lines	test lines
BOUN	7803	979	979
IMST	3435	1100	1100

A typical instance from the dataset looks like:

{
  "id": "ins_1267",
  "tokens": [
    "Rüzgâr",
    "yine",
    "güçlü",
    "esiyor",
    "du",
    "."
  ],
  "upos": [
    "NOUN",
    "ADV",
    "ADV",
    "VERB",
    "AUX",
    "PUNCT"
  ],
  "heads": [
    4,
    4,
    4,
    0,
    4,
    4
  ],
  "rels": [
    "nsubj",
    "advmod",
    "advmod",
    "root",
    "cop",
    "punct"
  ],
  "feats": [
    "Case=Nom|Number=Sing|Person=3",
    "_",
    "_",
    "Aspect=Imp|Polarity=Pos|VerbForm=Part",
    "Aspect=Perf|Evident=Fh|Number=Sing|Person=3|Tense=Past",
    "_"
  ],
  "text": "Rüzgâr yine güçlü esiyor du .",
  "feats_dict_json": [
    "{\"Case\":\"Nom\",\"Number\":\"Sing\",\"Person\":\"3\"}",
    "{}",
    "{}",
    "{\"Aspect\":\"Imp\",\"Polarity\":\"Pos\",\"VerbForm\":\"Part\"}",
    "{\"Aspect\":\"Perf\",\"Evident\":\"Fh\",\"Number\":\"Sing\",\"Person\":\"3\",\"Tense\":\"Past\"}",
    "{}"
  ]
}

Benchmarking

Benchmarking is done by scripts on accompanying Github repo. Please proceed to this repo for running the experiments. Here are the benchmarking results for BERTurk with our scripts:

Metric	BOUN	IMST
pos_acc	0.9263	0.9377
uas	0.8151	0.7680
las	0.7459	0.6960
morph_Abbr_acc	0.4657	0.6705
morph_Aspect_acc	0.1141	0.1152
morph_Case_acc	0.1196	0.0586
morph_Echo_acc	0.4261	0.4875
morph_Evident_acc	0.3072	0.3953
morph_Mood_acc	0.0654	0.0651
morph_NumType_acc	0.2694	0.2991
morph_Number_acc	0.3986	0.4782
morph_Number[psor]_acc	0.4348	0.2333
morph_Person_acc	0.4021	0.4726
morph_Person[psor]_acc	0.2490	0.0671
morph_Polarity_acc	0.3350	0.1674
morph_PronType_acc	0.1535	0.2680
morph_Reflex_acc	0.5620	0.7051
morph_Tense_acc	0.2149	0.1241
morph_Typo_acc	0.5081	—
morph_VerbForm_acc	0.4912	0.2364
morph_Voice_acc	0.0201	0.2602
morph_Polite_acc	—	0.1436
morph_micro_acc	0.3076	0.2915

Notes:

• — means that metric wasn’t present in that dataset’s reported results (e.g., morph_Typo_acc only in BOUN; morph_Polite_acc only in IMST).

Acknowledgments

This research was supported with Cloud TPUs from Google's TPU Research Cloud (TRC), like most of our projects. Many thanks to TRC team once again.