German Legal Sentences

German Legal Sentences (GLS) is an automatically generated training dataset for semantic sentence matching and citation recommendation in the domain in german legal documents. It follows the concept of weak supervision, where imperfect labels are generated using heursitical expert functions. For this purpose we use a combination of legal citation matching and BM25 similarity. The basis of our dataset is given by Open Legal Data, which is a collection of more than 200k German legal documents.

Why German Legal Documents?

In recent years, many efforts have been made to make German court decisions freely available to the public. Although a decent number of decisions can now be found on the Internet, it can be difficult for the layperson to find relevant information in them. With this dataset, we want to accelerate research on semantic search technologies in German domain-specific language. We hope that emerging search technologies will enable the layperson to find relevant information without knowing the specific terms used by the professionals. Furthermore, we think that the lessons learned due the work with this dataset can be applied to other domains and languages.

What is in the dataset?

GLS consists of semantically related sentence pairs from German judicial decisions. Their semantic relations are derived from shared references to laws other judicial decisions. These pairs can be used to train sentence similarity models. For this purpose we also provide pseudo-negative sentences, which we extract from rankings formed by the search engine Elasticsearch.

Download and Contents

The dataset has been split along documents (from which the sentences originate) into a 90/5/5 split for training, validation and testing. We ensure that any query sentence in the validation or test split is not contained by the training set. The sentences to match with, however, are the same for all three sets.

The current version of the Dataset can be downloaded as a zip archive: GermanLegalSentences_v0.0.2.zip. This archive contains the following tab-separated value (tsv) files.

Filename	Description	Num Records	Size	Format
train.sentences.tsv	Training Sentences	1.542.499	376MB	s_id, d_id, sentence
valid.sentences.tsv	Validation Sentences	85.375	21MB	s_id, d_id, sentence
test.sentences.tsv	Test Sentences	85.405	21MB	s_id, d_id, sentence
train.pairs.tsv	Positive Sentence Pairs (Training)	1.404.271	20MB	s_id, s_id'
valid.pairs.tsv	Positive Sentence Pairs (Validation)	78.472	1.1MB	s_id, s_id'
test.pairs.tsv	Positive Sentence Pairs (Test)	76.626	20MB	s_id, s_id'
refs.tsv	All References	790.508	20MB	r_id, TYPE, reference
sent_ref_map.tsv	References used by Sentence X	1.713.279	25MB	s_id, r_id0 r_id1 ...
es_neighborts.tsv	Elasticsearch Neighbors	466.154	25MB	s_id(query), s_id0 s_id1 ...

🤗 Hugging Face Dataset

Hugging Face Datasets makes accessing scientific Datasets very fast and easy. We have uploaded German Legal Sentences to the Hugging Face Datasets Hub using the identifier lavis-nlp/german_legal_sentences.

Usage

>>> import datasets
>>> ds = datasets.load_dataset("lavis-nlp/german_legal_sentences", "pairs")
Downloading: 100%|████████████████████████████████████████████████████████
...
>>> ds
DatasetDict({
    train: Dataset({
        features: ['query.sent_id', 'query.doc_id', 'query.text', 'query.ref_ids', 'related.sent_id', 'related.doc_id', 'related.text', 'related.ref_ids'],
        num_rows: 1404271
    })
    validation: Dataset({
        features: ['query.sent_id', 'query.doc_id', 'query.text', 'query.ref_ids', 'related.sent_id', 'related.doc_id', 'related.text', 'related.ref_ids'],
        num_rows: 78472
    })
    test: Dataset({
        features: ['query.sent_id', 'query.doc_id', 'query.text', 'query.ref_ids', 'related.sent_id', 'related.doc_id', 'related.text', 'related.ref_ids'],
        num_rows: 76626
    })
})

Example

{'query.doc_id': 28860,
 'query.ref_ids': [6215, 248, 248],
 'query.sent_id': 304863,
 'query.text': 'Zudem ist zu berücksichtigen , dass die Vollverzinsung nach '
               '[REF] i. V. m. [REF] gleichermaßen zugunsten wie zulasten des '
               'Steuerpflichtigen wirkt , sodass bei einer Überzahlung durch '
               'den Steuerpflichtigen der Staat dem Steuerpflichtigen neben '
               'der Erstattung ebenfalls den entstandenen potentiellen Zins- '
               'und Liquiditätsnachteil in der pauschalierten Höhe des [REF] '
               'zu ersetzen hat , unabhängig davon , in welcher Höhe dem '
               'Berechtigten tatsächlich Zinsen entgangen sind .',
 'related.doc_id': 56348,
 'related.ref_ids': [248, 6215, 62375],
 'related.sent_id': 558646,
 'related.text': 'Ferner ist zu berücksichtigen , dass der Zinssatz des [REF] '
                 'im Rahmen des [REF] sowohl für Steuernachforderung wie auch '
                 'für Steuererstattungen und damit gleichermaßen zugunsten wie '
                 'zulasten des Steuerpflichtigen wirkt , Vgl. BVerfG , '
                 'Nichtannahmebeschluss vom [DATE] [REF] , juris , mit der '
                 'Folge , dass auch Erstattungsansprüche unabhängig davon , ob '
                 'und in welcher Höhe dem Berechtigten tatsächlich Zinsen '
                 'entgangen sind , mit monatlich 0,0 % verzinst werden .'}

Generation

The follows describes the process of transforming the Open Legal Data corpus into a dataset of semantically related sentences pairs.

Pre-Processing and Citation Parsing

The documents we take from Open Legal Data are first preprocessed by removing line breaks, enumeration characters and headings. Afterwards we parse legal citations using hand-crafted regular expressions. Each citation is split into it components and normalized, thus different variants of the same citation are matched together. For instance, "§211 Absatz 1 des Strafgesetzbuches" is normalized to "§ 211 Abs. 1 StGB". Every time we discover an unknown citation, we assign an unique id to it. We use these ids to replace parsed citations in the document text with a simple reference tag containing this id (e.g [REF321]). At the same time we parse dates and replace them with the date tag [DATE]. Both remove dots which can may be confused with the end of a sentence, which makes the next stage easier.

Sentence Tokenizing

We use SoMaJo to perform sentence tokenizing on the pre-processed documents. Each sentence that does not contain at least one legal citation is discarded. For the rest we assign sentence ids, remove reference ids from them as well as any contents in braces (braces often contain large enumerations of citations and their sources). At the same time we keep track of the corresponding document from which a sentence originates and which references occur in it.

Sentence Pairing

Based on the parsed references and their occurrences in the tokenized sentences, we create an inverted index. This allows us to quickly retrieve all sentences that share a reference with a target sentence. We assume that a substantial portion of these sentences are semantically related to the target sentence due to fact that they share at least one reference. This basically is our first expert function. Our second expert function exploits the documents and their references, from which the sentences originate. We assume that a pair of sentences whose source documents share many references, have an above-average probability of being semantically related. For this we measure the Jaccard-Similarity on the citations used in the source documents. The goal of the third expert function is to find a topic overlap in the sentence besides the reference. For this we currently use a low BM25-Similarity threshold. If one of the expert functions is not satisfied, we discard the respective sentence pair candidate.

Task and Evaluation

We support a ranking task for the evaluation of semantic similarity models. There, our 1.542.499 training sentences serve as the collection to index search in, while the 85.405 test sentences are used as queries. However, all sentences, that are not covered by the test pairs can be skipped. We propose to form rankings of size 200 and measure MRR@10, MAP and Recall as evaluation metrics. We plan to provide evaluation scripts in the future.

Baselines

For the baselines we have tuned BM25 using the implementation for bayesian optimization of Optuna. Furthermore we have used our data to train CoRT, which is a representation-focused neural retrieval model.

Method	MRR@10	MAP@200	Recall@200
BM25 - default (k1=1.2; b=0.75)	25.7	17.6	42.9
BM25 - tuned (k1=0.47; b=0.97)	26.2	18.1	43.3
CoRT	31.2	21.4	56.2
CoRT + BM25	32.1	22.1	67.1