Alignment-Augmented Consistent Translation for Multilingual Open Information Extraction
Keshav Kolluru $^{1*}$ , Mohammed Muqeeth $^{1*}$ , Shubham Mittal $^{1}$ , Soumen Chakrabarti $^{2}$ , and Mausam $^{1}$
1 Indian Institute of Technology Delhi 2 Indian Institute of Technology Bombay keshav.kolluru@gmail.com, muqeeth101@gmail.com, shubhamiitd18@gmail.com soumen@cse.iitb.ac.in, mausam@cse.iitd.ac.in
Abstract
Progress with supervised Open Information Extraction (OpenIE) has been primarily limited to English due to the scarcity of training data in other languages. In this paper, we explore techniques to automatically convert English text for training OpenIE systems in other languages. We introduce the Alignment-Augmented Consistent Translation (AACTRANS) model to translate English sentences and their corresponding extractions consistently with each other — with no changes to vocabulary or semantic meaning which may result from independent translations. Using the data generated with AACTRANS, we train a novel two-stage generative OpenIE model, which we call GEN2OIE, that outputs for each sentence: 1) relations in the first stage and 2) all extractions containing the relation in the second stage. GEN2OIE increases relation coverage using a training data transformation technique that is generalizable to multiple languages, in contrast to existing models that use an English-specific training loss. Evaluations on 5 languages — Spanish, Portuguese, Chinese, Hindi and Telugu — show that the GEN2OIE with AACTRANS data outperforms prior systems by a margin of $6 - 25%$ F1. $^{1}$
1 Introduction
Open Information Extraction (OpenIE) is the task of converting unstructured text to semi-structured tuples of the format <subject; relation; object>, where these three components are textual phrases, broadly extracted from the original text (Etzioni et al., 2011). OpenIE tuples have shown utility in various downstream tasks (Mausam, 2016) like Question Answering (Fader et al., 2013; Khot et al., 2017), Machine Reading (Poon et al.,
2010), Multi-Document Summarization (Christensen et al., 2014; Fan et al., 2019), Schema Induction (Balasubramanian et al., 2013), and Knowledge Base Construction (Gupta et al., 2019; Chandrahas and Talukdar, 2021).
With widespread adoption of Deep Learning in NLP, Open Information Extraction (OpenIE) systems have gone through a paradigm shift from using rule-based, statistical systems to supervised neural models. However, both types of systems have been limited to only a few languages - earlier systems required language-specific OpenIE insights, and current systems require annotated training corpus that pose a barrier, particularly for low-resource languages.
Related tasks such as Semantic Role Labeling face similar challenges in extending to multiple languages. X-SRL (Daza and Frank, 2020) addresses this by automatic translation of English sentences to the target language followed by label projection to infer the semantic role labels in the translated sentence. However, translating the sentence alone may be insufficient for OpenIE because the generated tuples (also referred to as extractions) can include additional words absent in the sentence or require some changes to the word morphology used in the sentence. Although less prevalent in English, these characteristics need to be addressed in other languages.
X-SRL approach may be extended such that each extraction can also be automatically translated and subject, relation, object labels projected from English extractions. However, independent translation of sentence and extraction may introduce unwanted lexical (e.g. synonyms) or semantic (e.g., change in gender) variations between the translations, as shown in Table 1. Such translation inconsistencies in the training data lead to invalid OpenIE examples.
To maintain consistency between translations of a sentence and its extractions, both the trans
| Lexical Inconsistency English Sentence English Extraction Spanish Sentence Spanish Extraction (Indp) Spanish Extraction (Const) | The shield of Athena Parthenos, sculpted by Phideas, depicts a fallen Amazon <s> The shield of Athena Parthenos </s> <r> depicts </r> <o> a fallen Amazon </o> El escudo de Atena Parthenos, sculptado por Phideas, representa un Amazonas fallecido <s> El escudo de Atena Parthenos </s> <r> representa </r> <o> un Amazonas caído </o> <s> El escudo de Atena Parthenos </s> <r> representa </r> <o> un Amazonas fallecido </o> |
| Semantic Inconsistency English Sentence English Extraction Spanish Sentence Spanish Extraction (Indp) Spanish Extraction (Const) | The discovery was remarkable as the skeleton was almost identical to a modern Kuvasz <s> skeleton </s> <r> was </r> <o> almost identical to a modern Kuvasz </o> Un descubrimiento notable porque fósil era casi identica a un Kuvasz moderno <s> skeletó </s> <r> era </r> <o> casi identica a una Kuvasz moderna </o> <s> fósil </s> <r> era </r> <o> casi identica a un Kuvasz moderno </o> |
Table 1: OpenIE examples transferred from English to Spanish, using both Independent (Indp) and Consistent (Const) translations. Independent translation results in inconsistencies which may have the same meaning (by using synonyms, fallecido vs. caído) or may change the meaning (changing gender from male to female, moderno to moderna). Consistent translation avoids these issues, resulting in better quality of training data.
lations must use same words or their morphological variants as much as possible. Hence, we propose Alignment-Augmented Consistent Translation (AACTRANS), a seq2seq model that translates the given input text in a way that is consistent with a reference translation by biasing the translation to use words similar to the reference. To ensure that translations of sentence and extractions are consistent with each other, we use AACTRANS model to translate each of them with the same reference. In Section 4.1, we describe the reference used in training and inference.
Both generation based (Kolluru et al., 2020b) and labeling based (Ro et al., 2020) architectures have shown competitive performance on English OpenIE. However, labeling based models cannot naturally introduce new words or change morphology of sentence words required in some languages. Therefore, we use a new generative model, GEN2OIE, that contains two stages: the first stage produces all the relations in the sentence and the second stage generates the extractions containing the given relation. We also use a training heuristic specific to two stage models that increases relation coverage across multiple languages.
Our major contributions are that we:
introduce a novel technique for transferring data from English to other languages using the AACTRANS model and label projection,
propose two-stage generative model, GEN2OIE, for training OpenIE system in multiple languages,
release OpenIE evaluation datasets for two Indian languages, Hindi and Telugu, and
outperform prior systems by $6 - 25%$ in F1 over five languages.
2 Related Work
Our work is in line with the recent trend of extending IE and knowledge-based NLP systems to multiple languages. Recent works have explored distantly supervised relation extraction (Rathore et al., 2022; Bhartiya et al., 2022), knowledge-base completion (Singh et al., 2021), and fact linking (Kolluru et al., 2021). Our focus is OpenIE.
Many of the prior OpenIE systems, both nonneural (OpenIE-4 (Pal and Mausam, 2016; Christensen et al., 2011), OpenIE-5 (Saha et al., 2017; Saha and Mausam, 2018), ClausIE (Del Corro and Gemulla, 2013)) and neural (RnnOIE (Stanovsky et al., 2018), OpenIE-6 (Kolluru et al., 2020a)) have been deployed for English. Moreover, OpenIE systems built for other languages often work only for a single language due to their reliance on language-specific resources. For example, Bassa et al. (2018); Rahat and Talebpour (2018); Romadhony et al. (2018); Guarasci et al. (2020); Papadopoulos et al. (2021) focus on German, Persian, Indonesian, Italian, and Greek, respectively. Claro et al. (2019) present the importance of and various challenges involved with building multilingual OpenIE systems. Neural models like Logician (Sun et al., 2018) and CrossOIE (Cabral et al., 2020) use language-specific training data. Reliance on manually-annotated data or language-specific resources makes it infeasible to develop systems for the plurality of languages in the world, due to the cost and effort involved. However, our automated data conversion method can handle even low-resource languages like Telugu.
Non-neural systems such as PredPatt (White et al., 2016) and ArgOE (Gamallo and Garcia, 2015) work for multiple languages by using CoNLL-X and Universal Dependency parses respectively, to extract predicate-argument structures. Owing to their pipelined nature, their performance is below that of neural systems like Multi $^2$ OIE (Ro et al., 2020). Multi $^2$ OIE is a two-stage labeling model that works for English, Spanish and Portuguese. GEN2OIE extends this 2-stage design to the generative paradigm which allows for better modeling of the OpenIE task. The underlying mBERT encoder in Multi $^2$ OIE allows for cross-lingual generalization across various languages even after training with only English supervised data. However, dependence on zero-shot generalization limits the performance of the model.
Two types of methods have been proposed for constraining the outputs of the machine translation systems: 1) altering the decoding algorithm (Hasler et al., 2018), or 2) modifying the training methodology (Chen et al., 2020; Dinu et al., 2019). We follow the second approach for constraining translations by AACTRANS to be consistent to that of a reference sentence. Unlike prior work which focuses on constraining translations of few words, our task requires constraining the entire translation. We make use of awesome-align (Dou and Neubig, 2021a), an unsupervised word alignment technique (Och and Ney, 2003), that outputs the alignment between words in sentences of two languages. Awesome-align is trained using only parallel set of sentences in the two languages and generates aligned target words for each source word.
Transferring linguistic annotations from source to target language has been pioneered by (David et al., 2001) and has been used in context of Semantic Role Labeling (Annesi and Basili, 2010) and PoS-tagging (Zennaki et al., 2019). After consistent translation, we make use of Crosslingual Projection (Faruqui, 2015), to transfer OpenIE tags.
3 Notation
For the transfer of OpenIE data from one language to another, we represent the source language $^2$ as $E$ and the target language as $F$ . Further, we use $\text{sent}_E$ and $\text{ext}_E$ to represent a sentence and extraction in the source language and $\text{aact-sent}_F$ and $\text{aact-ext}_F$ to represent the transferred sentence and extraction in the target language.
To aid in the translation of extractions, we create a sub-sentence from each extraction by concatenating the phrases in all the fields of the extraction. The order of concatenation is such that the formed sub-sentence is grammatically valid. We refer to this sub-sentence as an ext-sentence and represent it as $es_{L}$ , where the subscript $L$ represents its language. For most English extractions, the ext-sentence corresponds to concatenating the fields in the order of subject, relation and object. However, other languages may follow a different order or allow for multiple orders. We rely on the output of system that translates the English ext-sentence to determine the ext-sentence in other languages. Moreover, each extraction can be seen as a labeling over the words of ext-sentence with either the Subject, Relation or Object tags. Tags for each word in the ext-sentence can also be regarded as the extraction.
4 Crosslingual Data Transfer
In this section we describe the technique used to convert OpenIE training data from source language $E$ to a target language $F$ . The source sentence, $\text{sent}_E$ , and all its corresponding ext-sentences, $es_E$ , are consistently translated to language $F$ (Section 4.1), and then, for each extraction in language $E$ , $ext_E$ , the S, R or O labels are projected to the translated ext-sentence, $es_F$ , to form the extraction, $ext_F$ , in language $F$ (Section 4.2). Figure 1 describes the pipeline with the help of an example.
4.1 Consistent Translation
We introduce a new Seq2Seq-based translation model called Alignment-Augmented Consistent Translation (AACTRANS) to ensure that sentences and ext-sentences are translated consistently from languages $E$ to $F$ . We define two translations as consistent if similar phrases have same grammatical structure, vocabulary and morphology while allowing for minimal changes necessary to ensure fluency.
To ensure consistency among translations of multiple pieces of text (both the sentence and respective ext-sentences present in an English OpenIE instance), we make use of a reference text in language $F$ to guide all of their translations. By individually maintaining consistency with the reference, their respective translations end up being consistent to one another as well.
Figure 1: Crosslingual Data Transfer pipeline from English to Spanish. The sentence and ext-sentence in English are aligned with a translation of the sentence. The AACTRANS model uses the aligned text to generate the final consistent translations. Cross Lingual Projection (CLP) introduces S, R, O tags in the extraction.
To generate a translation $\mathbf{f}$ (language $F$ ) of text $\mathbf{e}$ (language $E$ ), consistent with a reference $\mathbf{r}$ (language $F$ ), we use the following procedure.
Firstly, given $\mathbf{e} = e_1e_2\ldots e_N$ and $\mathbf{r} = r_1r_2\ldots r_M$ , we find the set of aligned words $A_{e_i} = {r_j}$ for each word $e_i$ in $\mathbf{e}$ , using a word alignment model.
Secondly, the aligned text $\mathbf{e}^{\prime}$ is constructed by concatenating each of the words $e_i$ in $\mathbf{e}$ , with their aligned words $A_{e_i}$ , using $# #$ as a separator (shown as $<1>$ , $<3> \rightarrow <4>$ and $<2>$ , $<3> \rightarrow <5>$ in Figure 1). If $e_i$ is aligned to the words $r_j$ , $r_k$ ( $j < k$ ), then $\mathbf{e}^{\prime}$ contains $e_i # # r_j r_k #$ . If $e_i$ has no aligned words, then $\mathbf{e}^{\prime}$ contains $e_i #$ .
Thirdly, the AACTRANS model takes $\mathbf{e}'$ as input and produces the sequence $\mathbf{f}$ as output, which represents a translation of $\mathbf{e}$ that is biased to use the aligned reference words (shown as $<4>\rightarrow<7>$ and $<5>\rightarrow<8>$ in Figure 1).
Next we discuss the training and inference of AACTRANS model.
Training: We use parallel sentences of languages $E$ and $F$ that are available in existing translation corpora for training the AACTRANS model. For each parallel sentence pair $\mathbf{e}$ and $\mathbf{f}$ , we use the sentence $\mathbf{f}$ itself as the reference $\mathbf{r}$ . Using the alignments between the words of $\mathbf{e}$ and $\mathbf{f}$ , we form the input $\mathbf{e}'$ , as discussed. The AACTRANS Seq2Seq model is trained with $\mathbf{e}'$ as input and $\mathbf{f}$ as output. Since $\mathbf{e}'$ has words from $\mathbf{f}$ , the model learns to use them during training.
Inference: Here, we consistently translate English sentence $sent_{E}$ and each of its ext-sentences $es_{E}$ . We use an off-the-shelf translation system to translate $sent_{E}$ to language $F$ , represented as $t - sent_{F}$ . $t - sent_{F}$ is used as the common reference $\mathbf{r}$ for constructing aligned sentence $al - sent_{EF}$ and aligned
ext-sentence $al\text{-sent}{EF}$ from sentence $sent{E}$ and ext-sentence $es_{E}$ , respectively. We then apply the trained AACTRANS model on $al\text{-sent}{EF}$ and $al\text{-sent}{EF}$ to generate target sentence $aact\text{-sent}{F}$ and target ext-sentence $aact-ess{F}$ respectively.
4.2 Crosslingual Label Projection (CLP)
Each word in the target ext-sentence, $aact - es_{F}$ , must be labeled with either the Subject, Relation, or Object tag to form the completed extraction in language $F$ . The tags from the corresponding $ext_{E}$ are projected onto $aact - es_{F}$ using the Crosslingual Projection algorithm (Faruqui, 2015) (described in Appendix A), which uses word alignments between $es_{E}$ and $aact - es_{F}$ and produces as output, the tags over $aact - es_{F}$ , giving extraction $aact - ext_{F}$ . The final set of <sentence, extractions> pairs constitute the data for training OpenIE system in language $F$ .
Thus the overall flow is: 1) AACTRANS model training is done on parallel corpus, 2) AACTRANS model inference is applied on language $E$ OpenIE examples, 3) CLP projection is used to obtain the labelled extractions, and 4) the generated data is used to train OpenIE system like GEN2OIE, which is discussed next.
5 Gen2OIE Model
To train OpenIE systems in multiple languages, we use a novel GEN2OIE model that extends the 2-stage design of Multi $^2$ OIE (Ro et al., 2020) to a generative paradigm. The first stage generates all possible relations and the second stage generates all extractions that contain a given relation.
GEN2OIE can produce overlapping relations and multiple extractions containing the same rela
Figure 2: GEN2OIE model contains two Seq2Seq models. In Stage-1, it generates all relations in the sentence, separated by an [SEP] token. For each detected relation in Stage-2, it generates extractions containing the relation.
tion, thus overcoming the limitations of Multi $^2$ OIE model. Moreover, due to its generative nature, GEN2OIE can add new words or introduce changes in morphology that may be necessary for producing correct extractions, which cannot be achieved by labeling models.
Both the stages of the GEN2OIE (shown in Figure 2) use Seq2Seq models as follows:
Stage-1 Seq2Seq: The input sentence is passed to the encoder and decoder generates a string formed by concatenating the set of relations from all the extractions, separated by an [SEP] token. During training, the target relations are concatenated in the order in which they occur in the sentence. We find that a deterministic order is important for adding stability to the model training.
Stage-2 Seq2Seq: To produce extractions corresponding to each relation generated in Stage-1, the relation $r$ is concatenated with the input sentence $s$ and passed to the encoder as " $r[SEP]s$ ". The decoder is trained to generate all the extractions containing the relation $r$ . Multiple extractions are separated by an $< e>$ token and each extraction contains delimiters tokens to identify the various parts of the extraction. The surrounding $< s>...$ , $< r>...$ and $< o>...$ tokens are used to identify the subject, relation and object phrases.
Labeling models like OpenIE-6 (Kolluru et al., 2020a) have used constrained training to increase the relation coverage. However, the constraints are limited to English and specific to labeling architectures. We introduce a simple parts-of-speech based heuristic during Stage-1 training of GEN2OIE that increases the relation coverage in the generative paradigm while being applicable across languages.
Relation Coverage (RC): We observe that for generating all possible extractions, all the verbs in the sentence must be contained in some relation. However, the extractions of training data may be incomplete and not satisfy this property. Therefore, dur
ing the training phase, we modify the input to the Stage-1 model by removing the verbs in the sentence which are not present in relation of any extraction. Thus the model learns that every verb must be included in some relation and applies the same during inference as well. This heuristic does not effect Stage-2 model training.
6 Confidence Scoring
The word log probabilities assigned by the Stage-2 decoder can be summed up to be used as confidence score for the extractions generated by GEN2OIE. We experiment with using a separate model for obtaining the confidence scores. A sequence-labeling model is trained on each language's extractions with ext-sentence as input and S, R, O labels over the ext-sentence as the output. The log probabilities given by the sequence-labeling model to the labels predicted by the GEN2OIE model are summed up to get the new confidence scores.
7 Experimental Setting
We train OpenIE systems in 5 languages, Spanish (ES), Portuguese (PT), Chinese (ZH), Hindi (HI) and Telugu (TE), by using the training data transferred from English to the respective language. For training the Seq2Seq models used in the data generation pipeline and the OpenIE systems based on the GEN2OIE architecture, we choose either the mBART (Liu et al., 2020) or mT5 (Xue et al., 2020) model depending on the particular language. Both of them are pre-trained multilingual Seq2Seq models that are trained with a span denoising objective on a large corpus of text containing many languages. mBART is pre-trained on CC25 and mT5 is pre-trained on mC4 corpus which contain text in 25 and 101 languages, respectively. Since mBART does not support Portuguese and Telugu, we use mT5 for these two languages and mBART for the
remaining 3 languages. We use the default hyperparameters recommended for these models and they are reported in Appendix F.
Training Datasets: For training the AACTRANS model, we make use of parallel English, language $F$ sentences available in standard translation corpora using the method described in Section 4. For Spanish we use parallel sentences from EuroParl corpus (Koehn et al., 2005), and for Portuguese we use a subset of the ParaCrawl corpus (Banon et al., 2019), as chosen by Lopes et al. (2020). For Hindi we use the IIT-B corpus (Kunchukuttan et al., 2018), and for Telugu we use the Samanantar corpus (Ramesh et al., 2021). For Chinese we use the data released for WMT19 (Barrault et al., 2019). We list the BLEU scores of the various systems in Appendix C.
We use the OIE4 training corpus from Kolluru et al. (2020b) and transfer it to other languages for training OpenIE systems.
Evaluation Datasets and Metrics: For evaluating translation systems we use the test sets available in the respective corpora and use SacreBLEU (Post, 2018) as the metric. For evaluating different OpenIE systems we use the Optimal F1 and Area Under Curve (AUC) as computed by the CaRB (Bhardwaj et al., 2019) scoring function. For Spanish, Portuguese OpenIE we use test sets provided in Ro et al. (2020). For Chinese OpenIE, we randomly choose $10%$ of the SAOKE dataset (Sun et al., 2018).
In order to evaluate our method on medium and low resource languages, we release new OpenIE test sets in Hindi and Telugu. Human annotators who are fluent in both the language and are knowledgeable about the OpenIE task translated about 300 randomly chosen sentences and their corresponding extractions from CaRB test set. They were paid $2.5 per sentence.
Table 2 lists the number of examples in different languages used for training and evaluating translation and OpenIE systems.
8 Experiments
We perform experiments to answer the questions:
- How effective is the GEN2OIE model?
- What is the quality of data generated with the AACTRANS+CLP pipeline, assessed both by
| EN | ES | PT | ZH | HI | TE | |
| Translation | ||||||
| Train | - | 1.9M | 5M | 1M | 1.6M | 4.8M |
| Test | - | 38473 | 99,087 | 2001 | 2507 | 2390 |
| OpenIE | ||||||
| Train | 91K | 91K | 91K | 91K | 91K | 91K |
| Test | 641 | 594 | 594 | 3833 | 298 | 302 |
Table 2: Data statistics for OpenIE examples and (English, language $F$ ) parallel sentences.
| Model | EN | |
| F1 | AUC | |
| IMoJIE | 53.6 | 33.3 |
| IGL | 52.5 | 33.8 |
| CIGL | 54 | 36 |
| OpenIE6 | 52.7 | 33.7 |
| Multi2OIE | 52.5 | 31.6 |
| GENOIE | 52.1 | 30.3 |
| GEN2OIE w/o RC | 51.9 | 29.7 |
| GEN2OIE | 54.4 | 32.3 |
| label-rescore) | 54.5 | 38.9 |
Table 3: Performance of OpenIE systems in English, evaluated with the CaRB metric. GEN2OIE along with Label Rescoring produces the best performance.
the final performance of systems trained using it and with metrics defined for evaluating consistency?
- What are the roles of different components in the GEN2OIE and AACTRANS+CLP data?
8.1 Effectiveness of GEN2OIE
To study the baseline monolingual effectiveness of GEN2OIE, we first train and evaluate the system on English data. The results are shown in Table 3. We compare with previously proposed English OpenIE models such as Multi $^2$ OIE (Ro et al., 2020), OpenIE6 (Kolluru et al., 2020a) and IMoJIE (Kolluru et al., 2020b). We also consider individual components in OpenIE6, the IGL and Constrained-IGL (CIGL) architectures. CIGL achieves the highest performance among all prior models but uses of English specific constraints in training.
We find that GEN2OIE, which uses the proposed language-agnostic relation coverage (RC) outperforms CIGL by $0.4%$ F1. However, its AUC remains lower. Therefore, we rescore the generated extractions with labeling-based rescoring model (Section 6). This results in a new state of the art for English in F1 and AUC with the labeling-based rescoring resulting in a $2.9%$ AUC gain over CIGL.
| Model | Training Data | ES | PT | ZH | HI | TE | |||||
| F1 | AUC | F1 | AUC | F1 | AUC | F1 | AUC | F1 | AUC | ||
| (Faruqui, 2015) | English | 45.5 | 28.6 | 48.5 | 31.5 | 13.7 | 3.3 | 30.4 | 12.5 | 36.7 | 16.2 |
| Multi2OIE | English | 60.0 | 41.5 | 60.2 | 41.1 | 23.7 | 8.1 | 28.8 | 10.9 | 16.5 | 4.1 |
| Multi2OIE | SentTrans+CLP | 62.0 | 42.8 | 60.9 | 41.3 | 21.2 | 6.5 | 48.1 | 27.6 | 33.4 | 15.4 |
| OpenIE6 | SentTrans+CLP | 56.8 | 37.4 | 58.7 | 39.4 | 18.2 | 4.8 | 46.3 | 28 | 39 | 18.3 |
| IMoJIE | AACTRANS+CLP | 61.6 | 43.1 | 59.7 | 39.9 | 15.4 | 4.0 | 47.5 | 26.3 | 33.9 | 15.5 |
| GENOIE | SentTrans+CLP | 60.4 | 40.6 | 63.5 | 43.7 | 20.9 | 4.9 | 51.5 | 28.5 | 41.7 | 16.3 |
| SentExtTrans+CLP | 58.3 | 39.7 | 57.3 | 36.5 | 20.8 | 5.6 | 51.6 | 28.1 | 36.6 | 13.9 | |
| AACTRANS+CLP | 60.8 | 41.3 | 63.9 | 44.8 | 23.1 | 5.9 | 51.6 | 28.6 | 39.3 | 15.1 | |
| GEN2OIE | SentTrans+CLP | 64.2 | 44.6 | 65.6 | 50.0 | 29.0 | 8.9 | 52.3 | 30.8 | 40.3 | 15.6 |
| SentExtTrans+CLP | 64.7 | 46.1 | 63.7 | 45.5 | 29.3 | 10.2 | 52.5 | 31.0 | 39.8 | 15.6 | |
| AACTRANS+CLP | 65.9 | 47.2 | 66.4 | 49.2 | 29.8 | 10.3 | 52.8 | 32.0 | 41.5 | 16.6 | |
| labelled-rescore) | AACTRANS+CLP | 65.9 | 51.5 | 66.5 | 53.8 | 29.8 | 13.8 | 52.8 | 37.6 | 41.5 | 24.9 |
| GEN2OIE-mT5 | AACTRANS+CLP | 67.9 | 48.5 | 66.4 | 49.2 | 33.3 | 12.7 | 53.6 | 30.9 | 41.5 | 16.6 |
| labelled-rescore) | AACTRANS+CLP | 68.0 | 53.6 | 66.5 | 53.8 | 33.2 | 15.8 | 53.6 | 38.1 | 41.5 | 24.9 |
Table 4: F1 and AUC performance of OpenIE systems in Spanish (ES), Portuguese (PT), Chinese (ZH), Hindi (HI) and Telugu (TE). Training with AACTRANS+CLP data shows strong performance with both GENOIE and GEN2OIE models. Labeling-based rescoring improves AUC in all languages. We also report the results of training GEN2OIE model with mT5 on all languages.
To further analyze the effectiveness of our 2-stage architecture, we introduce another model called GENOIE that outputs all extractions for a sentence as a single string, separated by an $$ token. We find that using GENOIE results in (2.3, $2.0)%$ drop in F1, AUC compared to GEN2OIE which leverages RC. We also report GEN2OIE performance without using RC.
8.2 Quality of AACTRANS+CLP data
In order to test the quality of the OpenIE examples generated using the AACTRANS+CLP pipeline, we train both the GENOIE and GEN2OIE models over the data generated for different languages. In Table 4, we compare it with examples generated from two other methods, SentTrans and SentExtTrans.
SentTrans+CLP represents an adaptation of X-SRL (Daza and Frank, 2020) for OpenIE where only the sentence is translated and each extraction, which is expressed as labeling over the words in the sentence, are projected onto the translated sentence using the CLP algorithm described in Section 4.2. The projected extraction is now a labeling over the translated sentence and hence it uses the same morphology as the sentence and cannot add new words. SentExtTrans+CLP uses independent translation of English sentence and ext-sentences followed by CLP algorithm between the English and translated ext-sentences to transfer the labels. Although this allows for adding new words and changing morphology, it can result in a lack of con
sistency between the translations.
We find that both GENOIE and GEN2OIE show consistent gains with AACTRANS+CLP data across various languages, when compared with SentExt-Trans+CLP and SentTrans+CLP data.
We further use rescoring models that are trained on the same AACTRANS+CLP data. Labeling-based rescoring achieves significantly higher AUC, with as much as $8.3%$ gain in Telugu.
We experiment with two versions of Multi $^2$ OIE: 1) trained only on English OpenIE data and applied to other languages in a zero-shot manner and 2) using language-specific training data generated from SentTrans+CLP. We specifically choose SentTrans+CLP data as all the extractions can be expressed as labels over the sentence, which is a requirement for training Multi $^2$ OIE which is itself a labeling model. We find that Multi $^2$ OIE model trained with SentTrans+CLP data improves over the zero-shot setting in all languages other than Chinese (discussed below). However, it performs significantly worse than GEN2OIE by $(5.2, 3.3)%$ in (F1, AUC) on average, even on training with the same SentTrans+CLP data. This can be attributed to Multi $^2$ OIE's lack of capability to handle: 1) overlapping relations, 2) multiple extractions per relation, 3) adding auxiliary words or 4) changing inflectional forms, as shown in Table 5.
We train IMoJIE and OpenIE6 (initialized with mBERT) on AACTRANS+CLP and Sent-Trans+CLP data. We find that they underperform
| Sentence Extractions | George Bluth Sr., patriarch of the Bluth family, is the founder and former CEO of the Bluth Company. <s> George Bluth Sr. </s> <r> is patriarch of </r> <o> the Bluth family </o> <s> George Bluth Sr. </s> <r> is </r> <o> the founder and former CEO of the Bluth Company </o> <s> George Bluth Sr. </s> <r> is </r> <o> patriarch of the Bluth family </o> |
| Telugu English Extraction | ious 5 ous 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 |
Table 5: Sentence and OpenIE predictions of GEN2OIE in English, Telugu and Hindi. It is capable of generating overlapping relations (is, is patriarch of), multiple extractions per relation (is), add auxiliary words (ajarne - > jafat) or change inflection forms ((\text{串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串串
| Model (Data) | ES | ZH | HI | |||
| F1 | AUC | F1 | AUC | F1 | AUC | |
| GEN2OIE (AACTRANS+CLP) | 65.9 | 47.2 | 29.8 | 10.3 | 52.8 | 32.0 |
| GEN2OIE (AACTRANS w/o Sentence Consistency+CLP) | 64.0 | 44.3 | 29.6 | 10.3 | 51.9 | 30.8 |
| GEN2OIE w/o Relation Ordering (AACTRANS+CLP) | 65.2 | 45.6 | 29.6 | 9.8 | 52.5 | 31.8 |
| GEN2OIE w/o Relation Coverage (AACTRANS+CLP) | 60.6 | 40.3 | 23.9 | 6.6 | 52.8 | 32.3 |
GEN2OIE and Multi $^{2}$ OIE. Compared to the two-stage models, both IMoJIE and OpenIE6 generate all the extractions autoregressively, which makes them more susceptible to noise in the automatically generated training data.
We additionally compare with Faruqui (2015), where the test sentence is translated into English, extractions are generated using OpenIE6 and they are projected back onto the test sentence. We find that the system results in poor performance due to lack of language-specific training.
We observe that all systems have low performance on Chinese. We attribute this to the various artifacts present in the SAOKE test set, that include special relations such DESC, TIME, ISA, etc. Since these extractions cannot be generated in our pipeline, we observe performance of only $33.2%$ F1 and $15.8%$ AUC with our best model, when compared to training GEN2OIE with SAOKE training data, which gives $52.5%$ F1 and $32%$ AUC.
We additionally train the GEN2OIE model using mT5 on AACTRANS data for all five languages (GEN2OIE-mT5 in Table 4) and find improvements of $(2.1%, 3.5%, 0.8%)$ F1 over the mBART models used for ES, ZH and HI.
8.3 Evaluating Consistency
In order to measure the inconsistency of the generated extractions with respect to the sentence, we
Table 6: Ablations of GEN2OIE model trained with AACTRANS+CLP data on ES, ZH and HI. We analyze the effect of removing 3 components and re-training the model: 1. Sentence Consistency used in AACTRANS data generation, and 2. Relation Ordering used and 3. Relation Coverage used in Stage-1 model training.
| ES | PT | ZH | HI | TE | |
| SenExtTrans+CLP | 12.2 | 9.5 | 24.5 | 13.3 | 19.6 |
| AACTrans+CLP | 5.4 | 3.9 | 5.7 | 6.9 | 10.3 |
Table 7: Evaluating inconsistency between translated extractions and corresponding sentences.
compute the fraction of words that occur in the extraction but are absent in the sentence. In Table 7, we find that across languages, the fraction is lower for training examples generated through the consistent translation methodology (AACTRANS+CLP) when compared against independent translations (SentExtTrans+CLP). This indicates that AAC-TRANS+CLP indeed achieves better consistency.
In order to analyze the reasons for improvement in CaRB performance, we compute the fraction of words that are present in model predictions but absent in the gold extractions of the test set (denoted by AG - Absent in Gold). In Table 8, we see that GEN2OIE trained on AACTRANS+CLP achieves lower values than the same model trained on SentExtTrans+CLP data and this correlates with the increased CaRB performance. This shows that the model generates words closer to gold extractions (and hence closer to input sentence), which contributes to higher performance.
| Data | ES | PT | ZH | HI | TE | |||||
| AG↓ | F1↑ | AG↓ | F1↑ | AG↓ | F1↑ | AG↓ | F1↑ | AG↓ | F1↑ | |
| SentExtTrans+CLP | 2.74 | 64.7 | 3.51 | 63.7 | 10.55 | 29.3 | 1.78 | 52.5 | 2.36 | 39.8 |
| AACTRANS+CLP | 2.31 | 65.9 | 2.22 | 66.4 | 9.67 | 29.8 | 1.6 | 52.8 | 2.09 | 41.5 |
Table 8: Evaluating CaRB F1 and AG of GEN2OIE predictions trained on SentExtTrans+CLP and AACTrans+CLP data. We find a decreasing trend of AG with increasing F1.
8.4 Ablation Study
We choose three representative languages to conduct the ablation study — Spanish, Chinese, and Hindi. Portuguese and Telugu belong to the same language family as Spanish and Hindi, respectively. In Table 6, we show the results of individually removing components from the GEN2OIE trained on AACTRANS+CLP data.
In AACTRANS w/o Sentence Consistency, we use regular translation of sentence while using consistent translation of extraction. This leads to a drop of (1.9, 0.2, $0.9%$ in F1 for the three languages, and shows the importance of using consistent translation on both the sentence and extraction.
In GEN2OIE w/o Relation Ordering, we train Stage-1 GEN2OIE with randomly shuffled relations. This reduces the performance as our model uses auto-regressive training which benefits from following a fixed order, which we choose as the order of occurrence of the relations in the sentence.
In GEN2OIE w/o Relation Coverage, we find that performance decreases in Spanish and Chinese by $5.3%$ and $5.9%$ in F1, respectively, but remains the same in Hindi, possibly due to the smaller number of examples in the test set.
Error Analysis: We find that the AAC-TRANS+CLP suffers from: 1) missing or 2) wrong word alignments and 3) inability to label discontinuous S, R, O phrases. We show examples of these cases in Appendix B.
9 Conclusion
We develop a novel AACTRANS+CLP pipeline for consistently transferring English OpenIE examples to other languages and present a novel two-stage generative model, GEN2OIE, for training OpenIE systems in various languages. We show improvements over the existing baseline of Multi $^2$ OIE, with an average improvement of $7.2%$ in F1 and $16.1%$ in AUC. It is effective in five languages, which is the largest number of languages covered by a single OpenIE technique known to us. To encourage research in medium and low-resource lan
guages, we additionally release new OpenIE evaluation examples in Hindi and Telugu.
Acknowledgements
Keshav is supported by TCS Research Fellowship. Mausam is supported by grants from Huawei, Google, Bloomberg and IBM, and a Jai Gupta Chair Fellowship. Soumen is partly supported by a Jagadish Bose Fellowship and an AI Horizons Network grant from IBM. We thank IIT Delhi HPC facility and TFRC program for compute resources.
References
Paolo Annesi and Roberto Basili. 2010. Cross-lingual alignment of framenet annotations through hidden markov models. In International Conference on Intelligent Text Processing and Computational Linguistics.
Niranjan Balasubramanian, Stephen Soderland, Mausam, and Oren Etzioni. 2013. Generating coherent event schemas at scale. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, EMNLP 2013, 18-21 October 2013, Grand Hyatt Seattle, Seattle, Washington, USA, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 1721-1731. ACL.
Marta Banón, Pinzhen Chen, Barry Haddow, Kenneth Heafield, Hieu Hoang, Miquel Esplà-Gomis, Mikel L. Forcada, Amir Kamran, Faheem Kirefu, Philipp Koehn, Sergio Ortiz Rojas, Leopoldo Pla Sempere, Gema Ramírez-Sánchez, Elsa Sarrías, Marek Strelec, Brian Thompson, William Waites, Dion Wiggins, and Jaume Zaragoza. 2019. ParaCrawl: Web-scale acquisition of parallel corpora. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.
Loic Barrault, Ondrej Bojar, Marta R Costa-Jussa, Christian Federmann, Mark Fishel, Yvette Graham, Barry Haddow, Matthias Huck, Philipp Koehn, Shervin Malmasi, et al. 2019. Findings of the 2019 conference on machine translation (wmt19). In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1).
Akim Bassa, Mark Kröll, and Roman Kern. 2018. Gerie-an open information extraction system for the
german language. J. Univers. Comput. Sci., 24(1):2-24.
Sangnie Bhardwaj, Samarth Aggarwal, and Mausam. 2019. CaRB: A Crowdsourced Benchmark for OpenIE. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2019, pages 6263-6268.
Abhyuday Bhartiya, Kartikeya Badola, and Mausam. 2022. Dis-rex: A multilingual dataset for distantly supervised relation extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, Dublin, Ireland. Association for Computational Linguistics.
Bruno Souza Cabral, Rafael Glauber, Marlo Souza, and Daniela Barreiro Claro. 2020. Crossoie: Crosslingual classifier for open information extraction. In PROPOR, pages 368-378.
Chandrahas and Partha Talukdar. 2021. OKGIT: Open knowledge graph link prediction with implicit types. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2546-2559, Online. Association for Computational Linguistics.
Guanhua Chen, Yun Chen, Yong Wang, and V. Li. 2020. Lexical-constraint-aware neural machine translation via data augmentation. In IJCAI.
Janara Christensen, Mausam, Stephen Soderland, and Oren Etzioni. 2011. An analysis of open information extraction based on semantic role labeling. In Proceedings of the sixth international conference on Knowledge capture, pages 113-120. ACM.
Janara Christensen, Stephen Soderland, Gagan Bansal, and Mausam. 2014. Hierarchical summarization: Scaling up multi-document summarization. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, ACL 2014, June 22-27, 2014, Baltimore, MD, USA, Volume 1: Long Papers, pages 902-912. The Association for Computer Linguistics.
Daniela Barreiro Claro, Marlo Souza, Clarissa Castellã Xavier, and Leandro Oliveira. 2019. Multilingual open information extraction: Challenges and opportunities. Information.
Yarowsky David, Ngai Grace, Wicentowski Richard, et al. 2001. Inducing multilingual text analysis tools via robust projection across aligned corpora. In Proceedings of the First International Conference on Human Language Technology Research.
Angel Daza and Anette Frank. 2020. X-srl: A parallel cross-lingual semantic role labeling dataset. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).
Luciano Del Corro and Rainer Gemulla. 2013. ClausIE: clause-based open information extraction. In Proceedings of the 22nd international conference on World Wide Web (WWW), 2013, pages 355-366. ACM.
Georgiana Dinu, Prashant Mathur, Marcello Federico, and Y. Al-Onaizan. 2019. Training neural machine translation to apply terminology constraints. In ACL.
Zi-Yi Dou and Graham Neubig. 2021a. Word alignment by fine-tuning embeddings on parallel corpora. In Conference of the European Chapter of the Association for Computational Linguistics (EACL).
Zi-Yi Dou and Graham Neubig. 2021b. Word alignment by fine-tuning embeddings on parallel corpora.
Oren Etzioni, Anthony Fader, Janara Christensen, Stephen Soderland, and Mausam. 2011. Open information extraction: The second generation. In IJCAI 2011, Proceedings of the 22nd International Joint Conference on Artificial Intelligence, Barcelona, Catalonia, Spain, July 16-22, 2011, pages 3-10. IJCAI/AAAI.
Anthony Fader, Luke Zettlemoyer, and Oren Etzioni. 2013. Paraphrase-driven learning for open question answering. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics.
Angela Fan, Claire Gardent, Chloe Braud, and Antoine Bordes. 2019. Using local knowledge graph construction to scale Seq2Seq models to multi-document inputs. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China. Association for Computational Linguistics.
Manaal Faruqui. 2015. Multilingual open relation extraction using cross-lingual projection. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.
Pablo Gamallo and Marcos Garcia. 2015. Multilingual open information extraction. In Portuguese Conference on Artificial Intelligence, pages 711-722. Springer.
Raffaele Guarasci, Emanuele Damiano, Aniello Minutolo, Massimo Esposito, and Giuseppe De Pietro. 2020. Lexicon-grammar based open information extraction from natural language sentences in Italian. Expert Systems with Applications, 143:112954.
Swapnil Gupta, Sreyash Kenkre, and Partha Talukdar. 2019. CaRe: Open knowledge graph embeddings. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China. Association for Computational Linguistics.
E. Hasler, A. Gispert, Gonzalo Iglesias, and B. Byrne. 2018. Neural machine translation decoding with terminology constraints. In NAACL.
Tushar Khot, Ashish Sabharwal, and Peter Clark. 2017. Answering complex questions using open information extraction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), Vancouver, Canada. Association for Computational Linguistics.
Philipp Koehn et al. 2005. Europarl: A parallel corpus for statistical machine translation. In MT summit, volume 5, pages 79-86. Citeseer.
Keshav Kolluru, Vaibhav Adlakha, Samarth Aggarwal, Mausam, and Soumen Chakrabarti. 2020a. OpenIE6: Iterative Grid Labeling and Coordination Analysis for Open Information Extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).
Keshav Kolluru, Samarth Aggarwal, Vipul Rathore, Mausam, and Soumen Chakrabarti. 2020b. IMoJIE: Iterative Memory-Based Joint Open Information Extraction. In The 58th Annual Meeting of the Association for Computational Linguistics (ACL), Seattle, U.S.A.
Keshav Kolluru, Martin Rezk, Pat Verga, William W. Cohen, and Partha P. Talukdar. 2021. Multilingual fact linking. In 3rd Conference on Automated Knowledge Base Construction, AKBC 2021, Virtual, October 4-8, 2021.
Anoop Kunchukuttan, Pratik Mehta, and Pushpak Bhattacharyya. 2018. The IIT Bombay English-Hindi parallel corpus. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018).
Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742.
Alexandre Lopes, Rodrigo Nogueira, Roberto Lotufo, and Helio Pedrini. 2020. Lite training strategies for Portuguese-English and English-Portuguese translation. In Proceedings of the Fifth Conference on Machine Translation.
Mausam. 2016. Open information extraction systems and downstream applications. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence (IJCAI), 2016, pages 4074-4077. AAAI Press.
Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational linguistics, 29(1):19-51.
Harinder Pal and Mausam. 2016. Demonyms and compound relational nouns in nominal OpenIE. In Proceedings of the 5th Workshop on Automated Knowledge Base Construction, pages 35-39.
Dimitris Papadopoulos, Nikolaos Papadakis, and Nikolaos Matsatsinis. 2021. PENELOPIE: Enabling open information extraction for the Greek language through machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop.
Hoifung Poon, Janara Christensen, Pedro Domingos, Oren Etzioni, Raphael Hoffmann, Chloe Kiddon, Thomas Lin, Xiao Ling, Mausam, Alan Ritter, et al. 2010. Machine reading at the university of washington. In Proceedings of the NAACL HLT 2010 First International Workshop on Formalisms and Methodology for Learning by Reading, pages 87-95. Association for Computational Linguistics.
Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, Brussels, Belgium. Association for Computational Linguistics.
Mahmoud Rahat and Alireza Talebpour. 2018. Parsa: An open information extraction system for persian. Digital Scholarship in the Humanities, 33(4):874-893.
Gowtham Ramesh, Sumanth Doddapaneni, Aravinth Bheemaraj, Mayank Jobanputra, Raghavan AK, Ajitesh Sharma, Sujit Sahoo, Harshita Diddee, Mahalakshmi J, Divyanshu Kakwani, Navneet Kumar, Aswin Pradeep, Kumar Deepak, Vivek Raghavan, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh Shantadevi Khapra. 2021. Samanantar: The largest publicly available parallel corpora collection for 11 indic languages.
Vipul Rathore, Kartikeya Badola, Parag Singla, and Mausam. 2022. PARE: a simple and strong baseline for monolingual and multilingual distantly supervised relation extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, Dublin, Ireland. Association for Computational Linguistics.
Youngbin Ro, Yukyung Lee, and Pilsung Kang. 2020. Multi^2OIE: Multilingual open information extraction based on multi-head attention with BERT. In Findings of the Association for Computational Linguistics: EMNLP 2020.
Ade Romadhony, Ayu Purwarianti, and Dwi H Widyan-toro. 2018. Rule-based indonesian open information extraction. In 2018 5th International Conference on Advanced Informatics: Concept Theory and Applications (ICAICTA), pages 107-112. IEEE.
Swarnadeep Saha and Mausam. 2018. Open information extraction from conjunctive sentences. In Proceedings of the 27th International Conference on Computational Linguistics, pages 2288-2299.
Swarnadeep Saha, Harinder Pal, and Mausam. 2017. Bootstrapping for numerical OpenIE. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 317-323. Association for Computational Linguistics.
Harkanwar Singh, Soumen Chakrabarti, Prachi Jain, Shared Roy Choudhury, and Mausam. 2021. Multilingual knowledge graph completion with joint relation and entity alignment. In 3rd Conference on Automated Knowledge Base Construction, AKBC 2021, Virtual, October 4-8, 2021.
Gabriel Stanovsky, Julian Michael, Luke Zettlemoyer, and Ido Dagan. 2018. Supervised Open Information Extraction. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), Volume 1 (Long Papers), pages 885-895.
Mingming Sun, Xu Li, Xin Wang, Miao Fan, Yue Feng, and Ping Li. 2018. Logician: A unified end-to-end neural approach for open-domain information extraction. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, pages 556-564.
Aaron Steven White, Drew Reisinger, Keisuke Sakaguchi, Tim Vieira, Sheng Zhang, Rachel Rudinger, Kyle Rawlins, and Benjamin Van Durme. 2016. Universal Decompositional Semantics on Universal Dependencies. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing.
Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2020. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.
Othman Zennaki, Nasredine Seminar, and Laurent Besacier. 2019. A neural approach for inducing multilingual resources and natural language processing tools for low-resource languages.
Alignment-Augmented Consistent Translation for Multilingual Open Information Extraction (Appendix)
A Crosslingual Label Projection (CLP)
In this section, we discuss CLP algorithm for projecting labels from English extraction to other language. Consider English sentence, E: Dutil - Dumas experiment was promoted by an organization called Encounter 2001 denotes and Spanish sentence, S: Experimento Dutil - Dumas fue promovido por una organizacion llama Encounter 2001. The word alignments between these sentences are listed in Figure 3 and equivalent phrases from the phrase extract algorithm are shown in Table 9. Consider the English extraction, (Dumas experiment; was promoted; by an organization). For each phrase in the tuple, CLP algorithm looks for the highest BLEU match phrase from Table 9. The subject phrase Dumas experiment has best BLEU match to Dutil - Dumas experiment and so the corresponding Spanish phrase Experimento Dutil - Dumas will be marked as subject. Note that the phrase Dumas experiment is not present in Table 9 because its aligned phrase is not continuous in Spanish sentence as can be seen in Figure 3. Similarly for the relation phrase was promoted, we find fue promovido from Table 9. Continuing the same algorithm, we get (Experimento Dutil - Dumas; fue promovido; por una organizacion) as the final Spanish extraction.
B Error Analysis
We list three cases that decrease the quality of transferred data using the AACTRANS+CLP pipeline.
Missing word alignments: For example, English extraction, A couple of trojans have also been found orbiting with Mars translates to Alternatively se han encontrar un par de��rajas en orbita con Mars in Spanish. The verb orbiting changes to the form en orbita (in orbit) (nominalization). The word en in Spanish does not align with any word in the English extraction as can be seen in Figure 4. So, projection of (A couple of trojans; have also been found; orbiting with Mars) leads to (un par de��rajas; Alternatively se han encontrar; orbita con Mars) which is not fluent because of missing word en in the object phrase.
In languages like Spanish and Portuguese, we found alignments to be of high precision but of
ten miss some alignments, as shown above. Next, we see how wrong alignments can affect projection quality.
Wrong word alignments: Consider the following English (E) and Hindi (H) ext-sentences, E: Many organizations like the Samskrita Bharati are conducting Speak Sanskrit workshops to popularize Sanskrit and H: नंस्कूल्ता गारती जोरे कहिके संगधन् शन्दूल्ति को लोकप्रय खनाने के किलेके बोल्ति संस्कूल्ति कागर्षार्वाके अशुल्ति कर रहे कहिके. We find that the word the is wrongly aligned to the Hindi word, कहर . So, the subject phrase Many organizations like the Samskrita Bharati does not have a continuous phrase in Hindi sentence because it has many words till कहर that do not map to the subject phrase in English sentence. Therefore, the CLP algorithm matches a partial phrase Many organizations like which is the best BLEU match to the given subject phrase and its equivalent continuous phrase जोरे कहिके संगधन् सं�्कूल्ति को gets tagged as subject in Hindi. Whereas नंस्कूल्ता गारती जोरे कहिके सं�धन् सं�्कूल्ति को would be an ideal subject phrase.
Discontinuous phrases: Phrase extract in the CLP algorithm assumes continuous phrases in English map to continuous phrase in other language. This assumption would lead to incomplete extractions in the other languages. For example, consider English extraction E: (Winston Churchill; twice suggested; naming a British battleship) and its Telugu extraction sentence T: 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 4. The relation phrase twice suggested is mapped as follows in Telugu: The word twice is mapped to $\text{三} _ { \text{一} }$ and suggested is mapped to $\text{三} _ { \text{一} }$ . The equivalent phrase twice suggested is no longer continuous in Telugu language. CLP algorithm looks for best BLEU match that results in matching to the phrase twice and its equivalent $\text{三} _ { \text{一} }$ is tagged as relation. The ideal relation in this example would be $\text{三} _ { \text{一} }$
C BLEU scores
Table 10 contains the BLEU scores of both the normal as well as consistent translations. We find that the performance remains nearly the same, indicating that the improved OpenIE performance stems from the consistency in the translations.
D Effect of word alignments quality
In order to understand the effect of alignment quality, we replace the language-specific trained
Figure 3: Equivalent English and Spanish sentence with corresponding word alignments between them
Figure 4: Equivalent English and Spanish sentence with corresponding word alignments between them
| English Phrases | Spanish Phrases |
| Dutil - Dumas experiment | Experimento Dutil - Dumas |
| Dumas | Dumas |
| experiment | Experimento |
| was promoted | fue promovido |
| .... | .... |
Table 9: Mapped continuous phrases between English (E) and Spanish (S) language sentences from the phrase extract algorithm
| BLEU | ES | PT | ZH | HI | TE |
| Translation | 45.2 | 48.4 | 26.8 | 20.5 | 7.0 |
| AACTranslation | 43.7 | 47.8 | 28.2 | 20.1 | 7.5 |
Table 10: BLEU scores of translation and AAC-translation are similar showing that the performance improvement is because of the added consistency.
| Language | MA | TA |
| ES | 0.38 | 0.19 |
| HI | 0.49 | 0.20 |
aligners (TA), with a standard pre-trained mBERT model (MA). First note in Table 11 that MA has a much higher alignment perplexity (used as a measure of unsupervised alignment quality in (Dou and Neubig, 2021b)). We now perform an experiment to replace TA with MA in our methodology. Aligners are used at two places in our setup - 1. Alignment-Constrained Translation and 2. Crosslingual Label Projection. We replace each of them with an mBERT aligner (MA), and show the results in Table 12. We find that there is some performance drop by using MA, but it is quite less compared to the drop in alignment perplexity. This suggests that our model is relatively robust to the quality of alignment.
E Alternatives to CLP
Following (Zennaki et al., 2019), we experiment with a neural mBERT-based tagging model. We train the mBERT model for tagging the Subject, Relation and Object tags in English. Due to the language-agnostic features of mBERT, we can apply the model to other languages in a zero-shot manner. These tagged examples can then be used for training the OpenIE model. In Table 13, we find that this does not improve over our CLP-based tag
Table 11: Unsupervised alignment perplexity for mBERT (MA) and Trained (TA) aligners
| (AACTRANS,CLP) | HI | ES | ||
| F1 | AUC | F1 | AUC | |
| (TA,TA) | 62.1 | 38.8 | 65.9 | 47.2 |
| (TA,MA) | 58.7 | 34.4 | 64.7 | 46.2 |
| (MA,TA) | 59.4 | 37.9 | 65.6 | 46.7 |
Table 12: F1 and AUC of GEN2OIE trained with examples generated using TA and MA alignment strategies. (1, 2) corresponds to aligner 1 being used in AACTRANS and aligner 2 being used in CLP.
| AACTRANS | HI | ES | ||
| F1 | AUC | F1 | AUC | |
| CLP | 62.1 | 38.8 | 65.9 | 47.2 |
| mBERT | 43.7 | 20.5 | 65.3 | 48.1 |
Table 13: GEN2OIE performance trained on examples tagged with either CLP or mBERT model.
ging. However, combining signals from both techniques could be interesting future work. HI results in Table 12 and Table 13 use a subset of the final test set which was initially used for development purposes.
F Reproducibility
Compute Infrastructure: We use V100 (32 GB) GPU for training the mBERT models and use TPU v3-8 for training the mT5 models.
Hyper-parameters: We list the final hyperparameters used for training mBART model in Table 14 and mT5 model in Table 15. We don't conduct any grid search and use the default hyperparameters suggested in the respective systems.
Number of parameters: mBART has 610 million parameters and mT5-base has 580 million parameters.
| Hyper-parameter | Value |
| Maximum tokens per batch | 1024 |
| Learning Rate | 3e-5 |
| LR Scheduler | Polynomial Decay |
| Warmup Updates | 2500 |
| Dropout | 0.3 |
| Max Updates | 40,000 (for OpenIE) and 1,00,000 (for translation) |
Table 14: mBART hyperparameters
| Hyper-parameter | Value |
| Maximum tokens per batch | 24576 |
| Learning Rate | 0.001 |
| LR Scheduler | Constant |
| Warmup Updates | 0 |
| Dropout | 0.1 |
| Max Updates | 20,000 (for OpenIE) and 1,00,000 (for translation) |
Table 15: mT5 hyperparameters