Although many computational problems can be approached using Deep Learning, in this position paper we argue that in the case of Information Retrieval tasks this is not mandatory and even detrimental whenever alternatives exist. Instead of learning (by training) how to solve the full problem, we suggest to split it into two sub-problems: a) producing structured data (speci cally knowledge graphs) out of the corpora, and b) providing usable tools (including natural language) to querying such structured data. Motivated by this two-step approach and its need of structured data, we introduce the Open Knowledge Graph (OKgraph) project, an initiative recently funded by Regione Autonoma della Sardegna aiming at providing insights on the rst part of the problem: a general way of generating knowledge graphs from text corpora, unsupervisedly.
Information Retrieval (IR) is about satisfying the information needs of users. From a very wide perspective, it means providing humans with a simple way to pose questions, for instance through natural language, and then computing the corresponding answer out of some given data, usually unstructured (texts, images, sounds, etc.). Depending on the complexity required to compute the answer, IR can be very challenging, ranging from simple keyword lookup to something close or even harder than passing the Turing test. Therefore, from one side we have an unstructured question (e.g., \how many countries in Europe?"), on the other side we have some mostly-unstructured big data to use in order to answer. How to compute answers from the data is the problem.
The most innovative way of dealing with virtually any problem, especially those where humans are someway involved such as in IR, is to resort to deep learning. In deep learning, what to do is induced from a vast amount of training data. In the case at hand, supposing the answer is not explicitly available in the data, the model should explicitly or implicitly learn, among the others, the concept of being a country, associate it to the appropriate entities, separate those in Europe from the others, and then count them.
This research path is disruptive because it generates useful computable functions no programmer is able to code. For instance, a function that given an image can list the objects contained within, or provide a textual description for it. But the greatness of deep learning comes at a cost, actually two drawbacks: on one hand, a) the generated function is somewhat a black box, that is, in most of the cases it does not provide any insight or knowledge about the problem; on the other hand b) it requires a huge amount of trained data, and for some problems such data simply does not exist yet.
We argue that for the speci c research objectives of IR, there is a di erent research path to approach the problem, that is, exploiting what we already understand well about the problem. By using the same example of counting the countries in Europe, we have that projections, ltering, and aggregates are wellstudied (for instance in the Database community) and every programmer can code this functions. In other words, we do not need to learn (using deep learning with its associated drawbacks) what we already know from decades of research in Databases. What we really need is to start from what databases need: structured data.
Knowledge Graphs are a machine representation of what Semantic Memories are for humans: general world knowledge that we have accumulated throughout our lives.
For instance, knowing that Rome is (an instance of) a settlement, and it is also the capital of Italy, which is a country. This kind of common sense knowledge is of paramount importance in a number of human tasks such as question answering, disambiguation, understanding, translation, learning, etc. The very same is true for machines, which are able to perform those tasks with high accuracy and recall whenever knowledge graphs about the given context are available.
Many successful crowdsourced projects focused on building curated knowledge that can be read by machines: Wikidata, Freebase, DBpedia (machine generated by the humanly-curated infoboxes of Wikipedia), Wordnet. Some other projects focused instead on (automatic) ontology alignment of speci c curated knowledge bases, such as BabelNet that produced a multilingual dictionary by aligning parts of Wordnet, Wikidata and other knowledge bases.
All these projects, and their produced knowledge (either \curated-only" or \augmented from curated data") have their merits and very successful applications, but they are also limited by two main weaknesses: 1. they cover only speci c (biased) areas of knowledge: very deep and narrow in a few cases (such as LinkedMDB, that knows about movie actors names but not about the number of their children) or wide but shallow (for instance, none of the previous projects are knowledgeable on, e.g., mobile device features, or recent news, or stock markets); 2. they are pretty much static, that is, once created they do not vary a lot; given the high cost of curating structured data, they tend to avoid time-changing topics (such as news or stock markets). Further, ontologies and automatic procedures (such as alignment) consists of ad-hoc, speci c purpose code that is expensive to maintain.
As a consequence, we do not have ontologies and structured data to
answer questions such as \what was the average rate on return of italian
market on December", or \what's new in last Android release", or \which deputy
was the rst to sign the divorce law in Italy". These limitations also a ect the
quality of existing knowledge bases. For instance, according to DBpedia Robbie
Williams is a \musical artist" while Lady Gaga is only a \person", therefore not
having information on her occupation nor musical genres. Our project SWiPE
(Searching WikiPedia by Example) [
Therefore, we want to address a currently under-investigated problem: learning machine-readable knowledge graphs from scratch, i.e., without necessarily relying on existing curated knowledge bootstrapping. Like a newborn can learn \from scratch" by listening other's people talk, an empty knowledge graph can be automatically feed with word information coming from natural language texts (such as generalist encyclopedias, news feeds, scienti c magazines), opening up a new world of applications which are currently not available or requiring too much e orts to be e ective.
This project is focused on investigating the fundamental relationship between unstructured data (as natural language text) and structured data (graphs representing knowledge), eventually leading to an autonomous way of inferring the latter from the former. The main outcome of our research will be a computer system, called OKgraph, that learns meaningful graph triples autonomously from scratch, that is, from non-annotated free text such as Wikipedia, WMT11 text data, UMBC WebBase and other available corpora, continuously updating a selfgenerated knowledge graph (that may resemble DBpedia, Wikidata or Freebase, plus statistics). The approach we are going to follow is based on the exploitation of linear regularities in word vector representations (vectors in a RN vector space with N indicatively ranging from 300 to 1000) obtained by state-of-the-art word embedding systems. In particular, we are interested in the analogies that such word vectors can represent. For instance, assuming vec(X) is the vector representing the word \X" and vicinity is given by cosine similarity, we have that vec(\Rome") - vec(\Italy") + vec(\France") is closer to vec(\Paris") than to any other word vector. Also, vec(\Germany") + vec(\capital") is close to vec(\Berlin"). This linearity relationship among vectors implies that not only words are semantically correlated through cosine similarity, but information on the speci c kind of correlation (concepts relationship such as "being capital of" or "married to") is hidden in the vector representation.
Also inspired by our successful SWiPE System, where structured
knowledge graphs (in particular entity-property-value relationships) are merged within
text/html data in order to be easily queried through a QBE-like interface [
While current mainstream research e orts are focused on either crowdsourcing (that is humanly-curated, such as Wikidata) or by writing speci c parsers for each attribute of interest (DBpedia), using word embeddings to generate knowledge graphs is a novel approach that is expected to be more scalable and more general than existing approaches, with possible disruptive outcomes for the scienti c research area.
Learning graphs is a process that will be conduced by using both very large corpora (such as English Wikipedia and books from Gutenberg Project) and medium to small corpora. For the latter, among the others, we want to investigate the use of the Sardinian Wikipedia project1, currently counting about 5000 entries. We expect many positive outcomes from the project, with possible large impact in the society. Having machine-learned open knowledge graphs allow to populate Wikidata and Wikipedia infoboxes could potentially have a positive impact on less-used languages such as Sardinian or other local languages. More generally, we remark that other than being useful as an interpretable instrument for humans (in contrast with other knowledge representations such as arti cial neural networks), knowledge graphs are also bene cial in many areas such as factoid and general question answering, disambiguation, search engines, etc., therefore addressing the need of structured data for IR tasks mentioned in the rst part of the paper. 1 https://sc.wikipedia.org/