QwenCloudのQwen3.8-2.4T-A95Bが100万トークンまで無料と聞いたので、7月28日の記事を横目に、多言語Few-Shot係り受け解析に挑戦してみた。Google Colaboratoryだと、こんな感じ。
!pip install openai deplacy
msg=[{"role":"system","content":"Try dependency-parsing with UPOS and DEPREL."},
{"role":"user","content":"You don't know what love is"},
{"role":"assistant","content":"You_PRON_nsubj_know|do_AUX_aux_know|n't_PART_advmod_know|know_VERB_root_know|what_PRON_comp_know|love_NOUN_nsubj_what|is_AUX_cop_what"},
{"role":"user","content":"Du weißt nicht was Liebe ist"},
{"role":"assistant","content":"Du_PRON_nsubj_weißt|weißt_VERB_root_weißt|nicht_PART_advmod_weißt|was_PRON_ccomp_weißt|Liebe_NOUN_nsubj_was|ist_AUX_cop_was"},
{"role":"user","content":"Bei mir bist du schön means that you're grand"}]
url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
key="sk-your-api-key"
mdl="qwen3.8-2.4t-a95b"
if url==None:
from transformers import pipeline
nlp=pipeline("text-generation",mdl,temperature=0.00001,token=key)
res=nlp(msg)
out=res[0]["generated_text"][-1]["content"].strip().replace("¥n","|")
else:
from openai import OpenAI
cli=OpenAI(base_url=url,api_key=key)
res=cli.chat.completions.create(model=mdl,temperature=0,messages=msg)
out=res.choices[0].message.content.strip().replace("¥n","|")
clu="# output = "+out+"\n"
c=[s.split("_")+["_","_","_"] for s in out.split("|")]
a,b=[t[0] for t in c],[]
for i,t in enumerate(c):
if t[2]=="root":
h="0"
else:
try:
k=b.index(t[3])
except:
k=-1
try:
j=a[i+1:].index(t[3])
h=str(i+j+2) if k<0 or j<k else str(i-k)
except:
h="_" if k<0 else str(i-k)
clu+="\t".join([str(i+1),t[0],"_",t[1],"_","_",h,t[2],"_","_"])+"\n"
b.insert(0,t[0])
import deplacy
deplacy.serve(clu,port=None)
「Bei mir bist du schön means that you're grand」をFew-Shot係り受け解析してみたところ、私(安岡孝一)の手元では以下の結果が出力された。
# output = Bei_ADP_case_mir|mir_PRON_obl_schön|bist_AUX_cop_schön|du_PRON_nsubj_schön|schön_ADJ_csubj_means|means_VERB_root_means|that_SCONJ_mark_grand|you_PRON_nsubj_grand|'re_AUX_cop_grand|grand_ADJ_ccomp_means
1 Bei _ ADP _ _ 2 case _ _
2 mir _ PRON _ _ 5 obl _ _
3 bist _ AUX _ _ 5 cop _ _
4 du _ PRON _ _ 5 nsubj _ _
5 schön _ ADJ _ _ 6 csubj _ _
6 means _ VERB _ _ 0 root _ _
7 that _ SCONJ _ _ 10 mark _ _
8 you _ PRON _ _ 10 nsubj _ _
9 're _ AUX _ _ 10 cop _ _
10 grand _ ADJ _ _ 6 ccomp _ _
素晴らしい、完璧だ。この出力に7828トークンほど使ったのだが、さて、他の言語にも適用できるかな。