(English description of the programme follows Japanese)

python_asa

python版日本語意味役割付与システム（ASA）

動作環境

• Python 3.5 以上
• cabochaをpythonで使えるようにしておく
• cabochaのダウンロード https://taku910.github.io/cabocha/

インストール

git clone {url}

pip install -e python_asa

使用方法

cd asapyでディレクトリを移動し、 python main.py で起動して標準入力を受け付けます。

変更点

• 2020/1/12 バグの修正．辞書構造をJsonに変更．moodなど動詞ムードなどの判定変更．semroleの場所判定の修正．願望(DESIDERATIVE)を追加．

Argument Structure Analyzer (ASA) and ASA python

ASA extracts predicate-argument structures in Japanese sentences based on the predicate frame file Predicate Thesaurus (PT), which is constructed under Project of Constructing Japanese Thesaurus of Predicate-Argument Structure. Thus, ASA may fail to detect some predicate-argument structures in the cases which the predicates are not registered in PT. Although PT currently contains more than about 10,000 predicates, there may remains some bugs since the thesaurus is under construction.

ASA detects a set of predicate-arguments based on chunk dependency. Since the dependency unit employed in this analyzer is not phrasal based, you need to follow the dependency links if a full phrase is necessary.

When you input a Japanese sentence, 「太郎の本を健が捨てた」Ken threw Taro's book away, for instance, ASA punctuates it into several chunks, as shown below. The square brackets [ ] signify the boundaries of a chunk. The chunk IDs and head chunk IDs are also assigned in the brackets.

[0 1 太郎の] [1 3 本を] [2 3 健が] [3 -1 捨てた]

[0 1 Taro-no] [1 3 hon-wo] [2 3 Ken-ga] [3 -1 suteta]

[0 1 Taro-ɢᴇɴ] [1 3 book-ᴀᴄᴄ] [2 3 Ken-ɴᴏᴍ] [3 -1 throw.away]

ASA returns the predicate-arguments of the sentence as:

 [arg0 健が] [arg1 本を] [v 捨てた]  

Here, the dative chunk [太郎の] in the noun phrase 太郎の本 is ignored. In order to get a complete phrase of arg1, i.e., [arg1 太郎の本を], you have to follow the chunk dependency links and extract the chunk [太郎の].

Accuracy of detecting semantic role labels

Accuracy of semantic role labeling is about 60% evaluated on BCCWJ-PT corpus. The accuracy is not high enough, as the system utilizes a simple, rule-based approach.

How to get a set of predicate-arguments

Input sentence: 太郎は6時に次郎を追いかけた。

Output: ['追いかける', ['太郎は', '対象'], ['6時に', '場所（時）（点）'], ['次郎を', '']]

getpas.py

from ASA import ASA

def out_predarg(result_chunks):
    for chunk in result_chunks:
        if chunk.get('semantic') != None:
            # get a predicate-argument
            out_st = []
            out_st.append(chunk['main'])
            record_arg_chunk_ids = []
            for arg_chunk in chunk['frames']:
                arg_chunk_id = arg_chunk['id']
                if arg_chunk_id in record_arg_chunk_ids:
                    continue
                arg_surface = result_chunks[arg_chunk_id]['surface']
                arg_role = arg_chunk['semrole']
                #arg_nrole = result_chunks[arg_chunk_id]['surface']
                out_st.append([arg_surface,arg_role])
                record_arg_chunk_ids.append(arg_chunk_id)
            print(out_st)

if __name__ == '__main__':
    asa = ASA()
    sentences = ["太郎は6時に次郎を追いかけた", "昨日，太郎に会った", "太郎が買った本を売った"]
    for sent in sentences:
        asa.parse(sent) #Class resut.Result
        result_json = asa.dumpJson()
        print (result_json['surface']) # input sentence
        result_chunks = result_json['chunks']
        out_predarg(result_chunks)