| { | |
| "paper_id": "2021", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T10:24:07.509142Z" | |
| }, | |
| "title": "Yseop at FinSim-3 Shared Task 2021: Specializing Financial Domain Learning with Phrase Representations", | |
| "authors": [ | |
| { | |
| "first": "Hanna", | |
| "middle": [ | |
| "Abi" | |
| ], | |
| "last": "Akl", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Dominique", | |
| "middle": [], | |
| "last": "Mariko", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Hugues", | |
| "middle": [], | |
| "last": "De Mazancourt", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "In this paper, we present our approaches for the FinSim-3 Shared Task 2021: Learning Semantic Similarities for the Financial Domain. The aim of this shared task is to correctly classify a list of given terms from the financial domain into the most relevant hypernym (or top-level) concept in an external ontology. For our system submission, we evaluate two methods: a Sentence-RoBERTa (SRoBERTa) embeddings model pre-trained on a custom corpus, and a dual word-sentence embeddings model that builds on the first method by improving the proposed baseline word embeddings construction using the FastText model to boost the classification performance. Our system ranks 2 nd overall on both metrics, scoring 0.917 on Average Accuracy and 1.141 on Mean Rank.", | |
| "pdf_parse": { | |
| "paper_id": "2021", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "In this paper, we present our approaches for the FinSim-3 Shared Task 2021: Learning Semantic Similarities for the Financial Domain. The aim of this shared task is to correctly classify a list of given terms from the financial domain into the most relevant hypernym (or top-level) concept in an external ontology. For our system submission, we evaluate two methods: a Sentence-RoBERTa (SRoBERTa) embeddings model pre-trained on a custom corpus, and a dual word-sentence embeddings model that builds on the first method by improving the proposed baseline word embeddings construction using the FastText model to boost the classification performance. Our system ranks 2 nd overall on both metrics, scoring 0.917 on Average Accuracy and 1.141 on Mean Rank.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "A hypernym or hyperonym is a concept which is superordinate to another one. In computer science, it is often represented as an IS-A relationship. For example, animal is a hypernym of cat and equity index is a hypernym of S&P 500 Index [Murphy, 2003] . Hypernymy, i.e. the capability to relate generic terms or classes to their specific instances, lies at the core of human cognition [Camacho-Collados et al., 2018] . Hypernymy modeling has been widely studied in natural language processing (NLP) for decades. Particularly, results based on embeddings methods [Henderson, 2017; Nguyen et al., 2017; Wang and He, 2020; Yu et al., 2015] show promise but the challenge remains in specializing these embeddings in particular areas such as the financial domain because of different aspects of language such as precise terms (e.g. abbreviations) and specific semantics that are badly or not covered at all by general-purpose models. The FinSim 2020 shared task [Maarouf et al., 2020] was the first task that attempts to combine hypernym classification methods in the financial domain. The FinSim-3 Shared Task 2021: Learning Semantic Similarities for the Financial Domain iterates on the previous editions by proposing an extended dataset with more diversified financial concepts.", | |
| "cite_spans": [ | |
| { | |
| "start": 235, | |
| "end": 249, | |
| "text": "[Murphy, 2003]", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 383, | |
| "end": 414, | |
| "text": "[Camacho-Collados et al., 2018]", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 560, | |
| "end": 577, | |
| "text": "[Henderson, 2017;", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 578, | |
| "end": 598, | |
| "text": "Nguyen et al., 2017;", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 599, | |
| "end": 617, | |
| "text": "Wang and He, 2020;", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 618, | |
| "end": 634, | |
| "text": "Yu et al., 2015]", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 955, | |
| "end": 977, | |
| "text": "[Maarouf et al., 2020]", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In this paper, we present our approaches which focus on domain-specific learning embeddings using as little data as possible. Although the shared task permits the use of external sources, we limit our training to the Financial Industry Business Ontology (FIBO) 1 data as well as the set of prospectuses in English curated and made available by the organizers. The corpus size for the latter set is estimated to about 10 million tokens. We explore two methods: the first is based on a custom sentence-level embeddings training using SRoBERTa [Reimers and Gurevych, 2019] and a term-definition dataset compiled from the FIBO website, and the second is a concatenated sentence-word embeddings model combining the custom SRoBERTa embeddings with a FastText 2 word embeddings model trained on the prospectuses set and the constructed FIBO dataset.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "We also explore and compare empirically the performance of several classifiers. Our experimental results demonstrate that while the domain-specific custom embeddings enhance the classification performance, class imbalances still hinder the recognition of under-represented classes. We analyze these results based on the number of labels provided in the training dataset as well as those extracted from the FIBO website.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The rest of this paper is organized as follows. Section 2 introduces the technical details of our proposed approaches. Section 3 empirically evaluates the performances of our methods and presents our results. Section 4 provides the conclusions of our work.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "We make use of custom corpus and exploit sentence-level and word-level embeddings in the context of phrase representation learning. We also test several classifiers in our term classification approaches. The general framework is shown in Figure 1 . This framework consists of customized corpus collections, sentence and word representation learning methods and term classification strategies. We will elaborate on each component below.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 238, | |
| "end": 246, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Proposed Approaches", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "(FinNLP@IJCAI 2021), pages 52-57, Online, August 19, 2021. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proceedings of the Third Workshop on Financial Technology and Natural Language Processing", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "General word embeddings are trained on domainindependent corpus. However, different domains have their proper semantics. In order to learn domain-specific representations for financial data, we base our work on collected customized corpus. We use the set of English prospectuses provided by the shared task organizers that contains 203 documents which roughly amounts to an estimated size of 10 million tokens. We augment this set with an extracted corpus from the FIBO website that we also use to train the sentence embeddings. Sentence embeddings already contain contextual information. However, they suffer from the same domain specialization problem as word embeddings. We choose to work with a specialized corpus to generate our sentence embeddings and use the FIBO website provided by the shared task organizers. Starting from the predefined tags (Bonds, Forward, Funds, Future, MMIs, Option, Stocks, Swap, Equity Index, Credit Index, Securities restrictions, Parametric schedules, Debt pricing and yields, Credit Events, Stock Corporation, Central Securities Depository, Regulatory Agency), we mine their corresponding FIBO web pages for the following properties:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Customized Corpus Collection", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "\u2022 Definition \u2022 Explanatory Note \u2022 Generated Description \u2022 Synonym(s) We also iterate over their children (n+1) instances found under the \"Direct subclasses\" web page section and collect their associated definitions. We do the same for the grandchildren (n+2) of the predefined tags at which point we stop the recursion. From the collected definitions we create a corpus of definition/tag pairs whereby each definition is associated to its corresponding tag. Children and grandchildren definitions are associated to one of the parent tags we started with. We stop the recursion at (n+2) because iterating further causes an overlap between concept definitions related to one or more of the predefined tags resulting in imprecise tag associations depending on the order of the recursion. Limiting the recursion at the (n+2) stage effectively prevents noise addition caused by such overlaps. The final compiled dictionary contains a total of 2015 definitions. These definitions are used to train domain-specific sentence representations.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Customized Corpus Collection", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "In this component, we combine two representational techniques: word embeddings and sentence embeddings. By concatenating both word and sentence vectors for a phrase (i.e. the group of words that make up a term), we hope to capture the syntactic and semantic properties of the financial domain while trying to reduce the ambiguity that comes with domainspecific representational learning. To achieve this practically, we pad the word embeddings vector of dimension 300 with zeroes to obtain a new word vector of size 768 identical to the size of the sentence embeddings vector without loss of information stored by the word vector. Then we concatenate both vectors by performing a term-by-term addition operation. Our final vector model is of size 768 and combines the information captured by both the sentence and word embeddings vectors.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Phrase Representation Learning", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "For both models, The training is performed on an NVIDIA GeForce RTX 2070 with Max-Q Design 8GB GPU machine. The construction of each embeddings model is detailed in the following subsections. We use the version of SRoBERTa provided by Hugging Face 3 and train it from scratch by adopting the method for training any BERT-like model on the STSBenchmark 4 for the semantic similarity task 5 . We split our corpus into 70% train set, 10% dev set and 20% test set. We also adopt the same model parameters as the STSBenchmark method for our training:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Phrase Representation Learning", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "\u2022 Training Batch Size: 8", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sentence Representation Learning", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u2022 Number of Epochs: 4 To specialize our model, we use the extracted FIBO corpus of term definitions described earlier. In terms of preprocessing for each definition, we transform the text to lowercase and segment it into sentences based on newline and punctuation delimiters. The SentenceTransformer model has the following architecture depicted in Figure 2 . The depicted architecture consists of one RoBERTa layer and a pooling layer. We feed the input sentence or text into the RoBERTa transformer network. RoBERTa produces contextualized word embeddings for all input tokens in our text. Since we want a fixed-sized output representation (vector u), we need a pooling layer. Different pooling options are available, the most basic one being mean-pooling: we simply average all contextualized word embeddings RoBERTa produces. This gives us a fixed 768 dimensional output vector independently of how long our input text is.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 349, | |
| "end": 357, | |
| "text": "Figure 2", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Sentence Representation Learning", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For our training set, each definition is duplicated to match the total number of predefined tags we have. Each duplicate is then passed as an input sentence to the SentenceTransformer model along with a label indicating the semantic similarity with each of the tags. We use a label of 0.8 to indicate a positive example (i.e. corresponding to a matching definition-tag pair) and 0.3 for negative examples (i.e. all other duplicate instances of the definition with the remaining mismatched tags). The labels are chosen to be sufficiently far apart in value to discriminate especially for ambiguous terms. We feed the model a total of 317101 definition-label pairs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sentence Representation Learning", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We augment the corpus compiled from the FIBO web pages with the English prospectuses set and use a FastText [Bojanowski et al., 2017] model to generate custom domainspecific word embeddings. Between the two versions of custom word2vec models with dimensions 100 and 300 provided by the shared task organizers, the model with dimension 300 outperforms the smaller model. We use this as our starting point to generate two custom embeddings models, the first based on word2vec and the second on FastText, both of dimension 300, using our extracted corpus and compare their performance in the classification task. The results are detailed in Section 3.", | |
| "cite_spans": [ | |
| { | |
| "start": 108, | |
| "end": 133, | |
| "text": "[Bojanowski et al., 2017]", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Word Representation Learning", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Sentence and word representations are used as features to train classifiers for term classification. A term is represented by the sum of phrase (sentence + word) embeddings for each word contained in the term. To find the best classifiers, we test two widely used classification methods: Logistic Regression and Random Forest. Experimental studies will be discussed in Section 3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Classification Methods", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "In this section we describe the data provided by the shared task organizers. We then provide details on the empirical experiments we performed and present our final results.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The training data provided by the task organizers contains a total of 1050 entries where each entry consists of a term and its corresponding label. A label can be one of the 17 predefined tags. For the test data, there are 326 entries of terms to be correctly classified into the correct tag. The main difficulty in this classification task lies in the tag distribution: the chosen labels are not at the same ontological level as Figure 3 demonstrates. The hierarchy shows that while some labels like Forward, Future, Option and Swap are on the same level, they are not aligned with other labels like Bond. The same case can be made for Central Securities Depository and Regulatory Agency in the FIBO ontology. The issue indicates that labels cannot be learned from a simple IS-A relationship.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 430, | |
| "end": 439, | |
| "text": "Figure 3", | |
| "ref_id": "FIGREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data Description", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "To tackle this problem, we enlarge the scope of our mining while collecting data from the FIBO web pages. Instead of limiting our collections to direct subclasses of the predefined tags, we search for \"Instances\" under \"Ontological characteristic\" which allows us to enrich our corpus both vertically and horizontally and expand term relations as much as possible by capturing their semantic connections.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Description", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "We design our experiments in order to determine the best model for each of the components in our proposed framework approach.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Results and Analysis", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "For the sentence embeddings module, we pit our SRoBERTa model against other well-performing models that we pre-train using the same setup described in Section 2.2. The selection of models is done based on our computational limitations as well as the grid 6 proposed in the official Sen-tenceTransformers Documentation. We select paraphrasempnet-base-v2, paraphrase-MiniLM-L6-v2 and paraphrasedistilroberta-base-v2 as the main competitors to SRoBERTa. To measure the performance of each custom trained model, we treat the classification problem like a semantic similarity task and use cosine similarity to find the best label for each term embedding. We evaluate model performance based on the metrics proposed by the shared task organizers. The results are shown in Table 1 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 766, | |
| "end": 773, | |
| "text": "Table 1", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Results and Analysis", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "Accuracy The empirical experiment is consistent with the choice of model since SRoBERTa is specialized in tasks like clustering or semantic search.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For the word embeddings model selection, we adopt the same baseline component proposed by the task organizers and composed of an embeddings module used as a feature vector and a logistic regression term classifier. The classifier is fixed in this experiment. The type of word embeddings is the only variable. The word2vec-100 and word2vec-300 models are the ones proposed by the task organizers and trained on the English prospectuses set (which we'll call Base). The c-word2vec-300 and c-fasttext-300 models are the models trained on our custom corpus (which we'll call Custom) comprised of the FIBO term definitions and the English prospectuses. Note that each result is the average of 5 runs and the train/test ratio is 80%/20%. The results are presented in While training on the custom corpus enhances model performance, the results validate our choice of FastText as it outperforms word2vec due to the model's capability to retain subword information which results in better learning.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Finally, in order to improve term classification, we empirically study the performance of the classifier component by reversing the conditions of our previous experiment: we fix the feature vector to our best word embeddings model (c-fasttext-300) and vary the classifier. We keep the train/test split at 80%/20% and perform 5 runs. The results are displayed in Table 3 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 362, | |
| "end": 369, | |
| "text": "Table 3", | |
| "ref_id": "TABREF4" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Accuracy Mean Rank Random Forest 0.80 1.45 Logistic Regression 0.82 1.33 From this experimental study, we find that complicated classifiers like Random Forest achieve worse performances than linear classifiers, so we select Logistic Regression as the classifier in our submitted systems. This observation shows that models that learn linear boundaries tend to perform better for this type of task.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In our submitted systems, we use the SRoBERTa model trained on the extracted corpus from the FIBO web pages. We submit a first system composed only of SRoBERTa and a classifier to study the performance of specialized sentence representations on this type of task. We use the constructed vector resulting from the concatenation of both the sentence model and the c-fasttext-300 word model as feature vector to the classifier in our second submission to study the effect of combining sentence and word information in what we refer to as phrase representation learning. Logistic regression is used in both submissions as the classifier. The final results are reported in Table 4 . ACC is short for Accuracy and MR is short for Mean Rank.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 668, | |
| "end": 675, | |
| "text": "Table 4", | |
| "ref_id": "TABREF5" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "System Submissions", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "yseop 1 In this submission, we combine SRoBERTa as a feature vector with a Logistic Regression classifier. The dimension of representation for each term is 768.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "System Submissions", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "yseop 2 In this submission, we concatenate the SRoBERTa model with the padded c-fasttext-300 to produce a feature vector of size 768. We feed the resulting feature vector to a Logistic Regression classifier.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "System Submissions", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "From our submissions, yseop 2 performs best and ranks 2 nd overall on both Average Accuracy and Mean Rank metrics in the shared task.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "System Submissions", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "Test Data System ACC MR ACC MR yseop 1 0.871 1.275 0.883 1.236 yseop 2 0.883 1.234 0.917 1.141 ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Train Data", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Another issue in this shared task is the problem of data distribution. By examining our system results, we observe that our framework performs consistently better for some labels than others. We investigate the reason for the poor performance on some labels by averaging the accuracy of matched labels, i.e. labels that were correctly classified as the best choice for an entry, over 5 runs for our best system. The analysis yields:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Imbalance", | |
| "sec_num": "3.4" | |
| }, | |
| { | |
| "text": "\u2022 The distribution shows a discrepancy in label expression that may explain the over-prediction of certain labels whenever the model makes a wrong prediction. However, the train data is one source of our learning and is active at the classification component of our framework. The other main source of representation is the extracted corpus collected from FIBO. We propose to analyze the distribution of label occurrences in the corpus based on the definitions collected:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Imbalance", | |
| "sec_num": "3.4" | |
| }, | |
| { | |
| "text": "\u2022 The last set of results shows the number of times a label is expressed is not sufficient to guarantee a good model performance. Some labels that are well expressed are underrepresented in terms of definitions with respect to others. This effectively splits the data distribution problem in two ways: the first is balancing uner-represented labels using techniques such as SMOTE and the second enriching definitions for some labels to improve predictions using external sources. Both methods merit further exploration.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Imbalance", | |
| "sec_num": "3.4" | |
| }, | |
| { | |
| "text": "In this paper, we studied the task of hypernym identification in the financial domain. We trained a a phrase representation learning model by specializing and combining a SRoBERTa sentence embeddings model and a FastText word embeddings model on a relatively small data set. We also enriched the provided data by collecting term definitions and term relations for the proposed hypernyms. Our approach shows that it is possible to specialize a domain-specific model by combining sentence and word models with a linear classifier on a relatively small corpus. It would be interesting to explore future possibilities by exploiting other domain resources or enhancing under-represented labels and studying their impact on domain-specific learning.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "https://spec.edmcouncil.org/fibo/ 2 https://radimrehurek.com/gensim/models/fasttext.html", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "https://huggingface.co/sentence-transformers/nli-roberta-base-v2 4 http://ixa2.si.ehu.eus/stswiki/index.php/STSbenchmark 5 https://github.com/UKPLab/sentencetransformers/blob/master/examples/training/sts/training stsbenchmark.py", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "https://www.sbert.net/docs/pretrained models.html", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Enriching word vectors with subword information", | |
| "authors": [ | |
| { | |
| "first": "[", | |
| "middle": [], | |
| "last": "References", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Bojanowski", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "References [Bojanowski et al., 2017] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vec- tors with subword information, 2017.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Vered Shwartz, Roberto Navigli, and Horacio Saggion. SemEval-2018 task 9: Hypernym discovery", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Camacho-Collados", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of The 12th International Workshop on Semantic Evaluation", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "712--724", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Camacho-Collados et al., 2018] Jose Camacho-Collados, Claudio Delli Bovi, Luis Espinosa-Anke, Sergio Oramas, Tommaso Pasini, Enrico Santus, Vered Shwartz, Roberto Navigli, and Horacio Saggion. SemEval-2018 task 9: Hypernym discovery. In Proceedings of The 12th Interna- tional Workshop on Semantic Evaluation, pages 712-724, New Orleans, Louisiana, June 2018. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "The FinSim 2020 shared task: Learning semantic representations for the financial domain", | |
| "authors": [ | |
| { | |
| "first": "Henderson ; James Henderson ; Ismail El", | |
| "middle": [], | |
| "last": "Maarouf", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Youness", | |
| "middle": [], | |
| "last": "Mansar", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Virginie", | |
| "middle": [], | |
| "last": "Mouilleron", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dialekti", | |
| "middle": [], | |
| "last": "Valsamou-Stanislawski", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of the Second Workshop on Financial Technology and Natural Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "81--86", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Henderson, 2017] James Henderson. Learning word em- beddings for hyponymy with entailment-based distribu- tional semantics, 2017. [Maarouf et al., 2020] Ismail El Maarouf, Youness Mansar, Virginie Mouilleron, and Dialekti Valsamou-Stanislawski. The FinSim 2020 shared task: Learning semantic repre- sentations for the financial domain. In Proceedings of the Second Workshop on Financial Technology and Natural Language Processing, pages 81-86, Kyoto, Japan, 5 Jan- uary 2020. -.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Hierarchical embeddings for hypernymy detection and directionality", | |
| "authors": [ | |
| { | |
| "first": "; M. Lynne", | |
| "middle": [], | |
| "last": "Murphy", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Murphy ; Nguyen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "233--243", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Murphy, 2003] M. Lynne Murphy. Semantic Relations and the Lexicon: Antonymy, Synonymy and other Paradigms. Cambridge University Press, 2003. [Nguyen et al., 2017] Kim Anh Nguyen, Maximilian K\u00f6per, Sabine Schulte im Walde, and Ngoc Thang Vu. Hierarchi- cal embeddings for hypernymy detection and directional- ity. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 233-243, Copenhagen, Denmark, September 2017. Association for Computational Linguistics. [Reimers and Gurevych, 2019] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks, 2019.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "BiRRE: Learning bidirectional residual relation embeddings for supervised hypernymy detection", | |
| "authors": [ | |
| { | |
| "first": "Chengyu", | |
| "middle": [], | |
| "last": "He", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Xiaofeng", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "He", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3630--3640", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "and He, 2020] Chengyu Wang and Xiaofeng He. BiRRE: Learning bidirectional residual relation embed- dings for supervised hypernymy detection. In Proceedings of the 58th Annual Meeting of the Association for Compu- tational Linguistics, pages 3630-3640, Online, July 2020. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Learning term embeddings for hypernymy identification", | |
| "authors": [], | |
| "year": 2015, | |
| "venue": "Proceedings of the 24th International Conference on Artificial Intelligence, IJCAI'15", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1390--1397", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "et al., 2015] Zheng Yu, Haixun Wang, Xuemin Lin, and Min Wang. Learning term embeddings for hyper- nymy identification. In Proceedings of the 24th Inter- national Conference on Artificial Intelligence, IJCAI'15, page 1390-1397. AAAI Press, 2015.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "uris": null, | |
| "text": "Framework of our proposed approach", | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF1": { | |
| "uris": null, | |
| "text": "Basic SentenceTransformer Architecture", | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF2": { | |
| "uris": null, | |
| "text": "An Example of Label Ontology Hierarchy", | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "TABREF1": { | |
| "num": null, | |
| "html": null, | |
| "text": "", | |
| "type_str": "table", | |
| "content": "<table/>" | |
| }, | |
| "TABREF2": { | |
| "num": null, | |
| "html": null, | |
| "text": "", | |
| "type_str": "table", | |
| "content": "<table><tr><td>Model</td><td colspan=\"3\">Corpus Accuracy Mean Rank</td></tr><tr><td>word2vec-100</td><td>Base</td><td>0.76</td><td>1.43</td></tr><tr><td>word2vec-300</td><td>Base</td><td>0.77</td><td>1.41</td></tr><tr><td colspan=\"2\">c-word2vec-300 Custom</td><td>0.78</td><td>1.40</td></tr><tr><td>c-fasttext-300</td><td>Custom</td><td>0.82</td><td>1.33</td></tr></table>" | |
| }, | |
| "TABREF3": { | |
| "num": null, | |
| "html": null, | |
| "text": "Word Embeddings Evaluation", | |
| "type_str": "table", | |
| "content": "<table/>" | |
| }, | |
| "TABREF4": { | |
| "num": null, | |
| "html": null, | |
| "text": "", | |
| "type_str": "table", | |
| "content": "<table/>" | |
| }, | |
| "TABREF5": { | |
| "num": null, | |
| "html": null, | |
| "text": "Final System Submissions", | |
| "type_str": "table", | |
| "content": "<table/>" | |
| } | |
| } | |
| } | |
| } |